[This Transcript is Unedited]




February 27, 2009

Hubert Humphrey Building
200 Independence Avenue SW
Washington, DC

Proceedings by:
CASET Associates, Ltd.
Fairfax, Virginia 22030
(703) 352-0091

Committee Members:


P R O C E E D I N G S (8:35 a.m.)

Agenda Item: Call to Order and Welcome

DR. STEINWACHS: I am Don Steinwachs, a member of the National Committee and co-chair of the Populations Subcommittee. I'd like to welcome you to this hearing on modeling health insurance data. This is the second in a series of hearings that we have had. The first one looked at the adequacy of data and information in being able to assess access and patterns of insurance coverage and uninsurance in the nation, as we look forward to health reform initiatives.

This hearing is being broadcast live on the Web, so I ask you to speak into microphones. Apparently they will become accustomed to you as you become accustomed to them; it is a process.

Before we start the hearing and introduce our invited guests, and we are very happy to have you with us today, I would like to ask that members of the committee introduce themselves, and then also we will go around the room and ask everyone here to introduce themselves, and then we will start the hearing.

I am Don Steinwachs, Johns Hopkins University, a member of the committee and of the subcommittee.

DR. SCANLON: Bill Scanlon, Health Policy R&D, member of the committee and member of the subcommittee.

DR. HORNBROOK: Mark Hornbrook, Kaiser Permanente, member of the committee and member of the subcommittee.

DR. BREEN: Nancy Breen, National Cancer Institute, member of the subcommittee.

MR. HITCHCOCK: I am Dale Hitchcock from ASPE here at HHS, in the data policy shop. I am a staff person to the subcommittee.

MS. BULLHEIMER: Linda Bullheimer at the National Center for Health Statistics. I am an advisor to ASPE on this project.

MR. GARRETT: Bowen Garrett. I am a senior research associate at the Urban Institute and Health Policy Center.

MS. MC GLYNN: Beth McGlynn. I am the Associate Director of RAND Health.

MR. BAUMGARDNER: Jim Baumgardner, Deputy Assistant Director of Health and Human Resources from the Congressional Budget Office.

DR. SUAREZ: I am Walter Suarez. I am a member of the committee and a member of this subcommittee.

DR. GREEN: Larry Green, University of Colorado, member of the committee and member of the subcommittee, no conflicts.

MS. JACKSON: Debbie Jackson, National Center for Health Statistics, committee staff.

DR. GIROSI: Federico Girosi, RAND. I am leading the modeling team of the COMPARE Project.

DR. BANTHIN: Jessica Banthin, Director of the Division of Modeling and Simulation in the Center for Financing, Access and Cost Trends at AHRQ.

MS. NIVASSAN: Sherovasi Nivassan, Health Disparities Research Coordinator, Division of Cancer Control and Population Sciences, NCI.

DR. SHORTRIDGE: Emily Shortridge, NORC at the University of Chicago. I am a research scientist.

MR. SHEILS: I'm John Sheils with the Lewin Group.

MR. HOTT: Randy Hott, the Lewin Group.

MR. MUSCO: John Musco with the Office of Health Policy at ASPE.

MR. ARNETT: Ross Arnett with Econometrica.

MR. GOREITCHEV: Dimitri Goreitchev with Econometrica.

MR. SACARADE: Alex Sacarade, Econometrica.

MS. CARROLL: Monique Carroll, GE Health Care.

MR. O'HARA: Brad O'Hara from Census, small area estimates of health insurance.

MS. BICKFORD: Carol Bickford, American Nurses Association.

MR. ROSHOWSKI: Jim Roshowski, Center for Studying Health System Change.

MR. MERMAN: Gordon Merman, Office of Management and Budget.

MS. JAMISON: Missy Jamison, National Center for Health Statistics, CDC, staff.

MS. KANAAN: Susan Kanaan, writer for the committee.

MS. SASHMONY: Nina Sashmony, ASPE.

MR. POISELL: John Poisell, CMS.

DR. STEINWACHS: Thank you very much, and welcome again. Before turning this over to Bill Scanlon to introduce the first panel, I want to express a special thanks to Dale Hitchcock and to Roshita Dorsey, who made this possible by organizing and inviting all of you, and encouraging you to come, who are on the panel and very much appreciated. Also, Debbie Jackson and Cynthia Sydney, who made this possible too. So thank you very much.

DR. SCANLON: Let me extend my thanks to all those who made this possible, as well as to all of you who agreed to come and talk with us.

As Don indicated, this is the second hearing that we are having with respect to data for health reform. These two hearings have been done at the request of ASPE, but they are continuing a focus of this subcommittee on the issue of data availability for analysis that can help inform through policy choices with respect to health and health care.

It goes back at least for the entire time that I have been on the full committee. There have been two themes for those hearings. One is the issue of access to data that exists. That is a real question not only for researchers and analysts outside of government, but as it turns out, an issue within government. So we have done a letter to the Secretary saying that we at least need to think about how we can streamline this process within government.

The IOM has raised in its report earlier this month other sets of issues which may apply more outside of government, but those are things that we need to be thinking about too, how do we rationalize the access to data that exists.

The other theme is the issue of data that don't exist but should, or we would hope that they would. That is something that we need to also think about some more seriously. Yesterday the full committee approved a letter to the Secretary indicating that there is a real need for investing more resources in data collection. There is an opportunity in the newest bill, the Stimulus Act, to make some of these kinds of investments, but they only should be perceived as a first step. Data collection is not a one time activity; it is something that needs to be established on an ongoing basis, and there needs to be enough adequacy for adequate investments so that we have the richness of data to be able to understand the issues that we are facing.

In particular, it is totally inadequate to know what is happening on a national level. We know too much about the variation in the country in terms of that it exists, and we know too little about what explains that variation and how we might change it in a positive way. So thinking about how to improve data on a sub-national basis is something that is also important to us.

The last hearing was a focus of people who are data producers and some of the data users from an analytic perspective. Today we are moving to what I would think of as the heavy lifters, people who are trying to model various aspects of the health care sector, which is the challenging activity in terms of drawing upon multiple data sources, and trying to both make them work better or harder, to learn more from the interactions, but it also increases the demands or the requirements for the data. So it is something that we need to know about in terms of making recommendations about future data activities within the Department.

We recognize that while we are doing these hearings in the context of anticipation for health reform, that a lot of things may happen with health reform before additional data can be collected.

DR. GRUBER: Hello?


DR. GRUBER: Hello? John Gruber calling in. I can't really hear someone on the line.

DR. SCANLON: John, this is Bill Scanlon. We were just starting.

DR. GRUBER: I can't really hear you.

(Remarks off the record.)

DR. SCANLON: John, thank you for joining us. We were just getting started. We were just introducing the hearing and emphasizing that what we are interested in here is the issue of data, access to data as well as production of data that is adequate to be able to do the kinds of analyses that we need to underpin health reform.

As we know, the health reform process is in some respects already underway with respect to proposals. We are anticipating that this is going to be a process, not an event, that there will be opportunities to weigh in with additional information, additional analyses over time. So there is a question in terms of what it is that can be done to try and facilitate and enhance this process.

At the last hearing, we heard strongly about this whole issue, since most of our users of data were from government, about the problems and the need to expedite access to data within government. This is a key issue.

Today we are interested in expanding this in terms of the area of modeling, to know about some of the hurdles that you face as well as how you have overcome them, the capacities that you have developed, and then things that you could suggest for us to think about in terms of recommendations that the Department may do in terms of making the capacity to do modeling stronger for the future.

We are going to turn now to the first panel. We are very happy that you agreed to participate. Beth McGlynn from RAND is going to be our first presenter. If you would like to start, Beth, you have the floor.

Agenda Item: Panel 1: Modeling Health Insurance Data, Practice Versus Perspective

MS. MC GLYNN: I want to start by thanking the committee for the opportunity to talk to you about what we have learned about in particular the federal data sets that are available for use in microsimulation modeling. I want to in particular acknowledge Federico Girosi, who is the real brains behind the operation. He will answer all your hard questions as we go along.

This is the outline of what we are going to talk about . I am going to give you a brief overview of the model that we have developed. That is primarily to set in context comments we make about the data sets that we have tried to use and some of the challenges that we have encountered.

We are going to talk about our experiences with some of those data sets, and then talk about dreams we have in our off hours about data sets, if they only existed, what could we possibly do, and then end with some priorities for where we hope you might recommend that the government go.

COMPARE is an initiative. It is an acronym that stands for Comprehensive Assessment of Reform Efforts, that was begun at RAND about three years ago. It has two goals, to provide a factual foundation for a national dialogue on health reform options, and then to develop a set of tools that would help with the development of legislation and policy options, both at the federal and state level, as well as those things that might be under consideration by large private companies.

We look at the effects of policy changes on multiple dimensions of system performance, including cost, quality, access. Then we have another category which is operational feasibility, how hard is it to get from where we are today to the future state of the world that is envisioned.

The dimensions are not so well highlighted, so I will read them to you. The microsimulation model itself produces results in the areas of health, national health expenditures, that is both in the aggregate and distributionally, consumer financial risk, which looks at the impact on different kinds of households, and the area of coverage. All of these are both aggregate and distributional results.

COMPARE is an agent based microsimulation model. The agents are shown in green here. Let me in particular highlight what is in the box. We have developed a way of looking at not only the person level, but some other units that are important for policy making purposes, so families, health insurance eligibility units and tax units. As you know, those come together in different ways as you are working. The arrows are some of the messages that go back and forth between the different agents.

One of the things that is a little bit different about our model is that we estimate premiums and insurance status endogenously. So that is one of the features of this model. It gives you an overview.

We use a variety of data sources. The first three that are listed, the SIPP, the MEPS-HC and the HRET/Kaiser are the core data sets. So the base population is built out of the SIPP. That is linked to the MEPS-HC for the purposes of coming up with health care expenditures, health status and utilization patterns, and then the HRET/Kaiser data are used to describe employer characteristics and employer health benefits.

You can see that we use a number of other data sets. Those are used in a variety of ways, including to make sure that distributions look like national distributions. In many cases we have to take the MEPS for instance and inflate it to look like national health expenditure accounts. These are issues that are familiar to everybody, but we do use quite a variety of data sources.

The agents that are in the microsimulation model, their behavior is modeled in a number of different ways. Our starting point was a series of linear progressions that describe the behavior of individuals and HIE use.

We have more recently moved into modeling behavior using utility maximization methods and by the end of next month pretty much everything will be converted to the utility maximization framework. Then the behavior affirms is principally modeled using cost-benefit analysis.

The way the model works is, we are looking at a fairly short term before-after. The results of the model describe the new state of the world at equilibrium following the policy change. We estimate that that is about a two to three year process. We use 2007 right now as the base year on which changes are estimated.

We don't do much in terms of macro facts. The employment variables are static. We have been talking about how we can reconfigure the current state of the world to reflect the economic downturn, and I think we have a method in mind for doing that.

Any time you have a big shop like that, you have to have some kind of timely data, and you have to have a base data set that you are using, and ability to have access to the variables that help you understand things like, are the unemployment effects located in a particular sector, and do we have something in the data set that lets us know that there are differential effects by sector, for example.

We don't in our model have people changing jobs or being laid off in the course of the simulation.

The analysis right now is performed at a national level. We can and have a method in place for doing state level analysis. That would be done by reweighting the data to make the synthetic universe look like the state that we were looking at. But to do really good state level analysis requires some additional state specific data, and we will talk in some of the limitations about some of those issues.

To date we have modeled the following policy options, an individual mandate with the National Health Insurance Exchange and subsidies, an employer mandate, Medicaid SCHIP expansion, refundable tax credit, more like the Bush tax credit than what McCain was proposing in the campaign, and then a Medicare buy-in modeled after the recent Baucus proposal.

I should say, each of these were originally modeled as independents or stand-alone policy options. Most recently what we have implemented is the combo package proposed through Baucus. So that was the first time we put a number of these options together. It gives you a sense of the kinds of things we have been doing.

We are also in the process of developing a provider module. The focus of that will be on the behavioral responses of physicians and hospitals to changes in payment policy and what we characterize as health services delivery interventions. Those are things that are a bit more on the ground, like disease management, the medical home, public reporting, pay for performance, those kinds of changes.

Initially the module will operate independently. Over the long term we will be able, because of the modeling architecture that we have built, the microsimulation model, to connect that and have it interact with the model structure that I showed you before.

One of COMPARE's main commitments is to advancing policy analysis through transparency. The results, and had we had more time, I would have bored you with it through the website, but in your copious free time you can take a look at the results, which are publicly available on the website, randcompare.org.

Also on the website you can find a white paper that describes the microsimulation methodology. This is what we refer to internally as the short form. It is only 40 pages long. Federico has a blockbuster 400-page version coming soon to a theater near you. It will have quite a bit more detail.

There is also an aspect of the website that lets users interact with basically the modeling results which has a data set underneath and people can change parameters and assumptions, and look at how that affects the results of these different policy options. So I think of it as fantasy policy maker.

Here are some of the strengths and limitations of the data sets that we are using. I'm not going to spend a lot of time on the strengths, because I think those are pretty well known. Let me just highlight some of the limitations that we have encountered with the SIPP.

A big issue there is that it is not very timely. It was last fielded in 2004. We are using the 2001 with a 2002 panel update. It doesn't within that data set have reliable information on utilization, and it doesn't have any premium data.

The MEPS-HC, which is a terrific data set, is a little bit problematic because it is smaller than the SIPP and the CPS, which is difficult when you get to much more subunit analysis. There is not unrestricted access to state codes, so this has caused some challenges when there are important regulatory differences state to state, and even with some of the Medicaid structure. So for modeling some of that, one has to get a bit creative.

I will talk in a minute about why we use the Kaiser HRET as the main source of information on benefit design, but it is quite small, and it is limited to aggregate information on employees. So you don't have any individual level information that helps you understand some things about firm behavior.

Let me talk a little bit about data sets we had some challenges with. The first is the MEPS HC-IC, which is a data set that is available through restricted access. It is the kind of thing we use for modeling the choices that are made by employees.

The problem we had with it is, it is fairly small. The data are quite noisy and not particularly nationally representative. After spending quite a bit of time working with this data set, we ended up using a simpler analysis than we had anticipated, which meant that we couldn't really take full advantage of the information about health benefits. So we like a lot of other people assign people to an average type plan, and you can't answer questions about what if we did this, that or the other thing to mandates, for example. So that kind of fine level of analysis.

We had pretty timely access to the data through AHRQ, but it is still inconvenient to have to go out to AHRQ to use it. You can't import data directly into the simulation model. You can only export regression coefficients, so it is a fairly in efficient way to interact with the data. Then the final plea for equipment upgrade. Our programmers complain that the screens were small, the computers were outdated, and they therefore could only spend limited amounts of time there before they were going home with migraine headaches. So maybe in the stimulus package we can buy AHRQ some new computers, I don't know.

Then let me turn to the MEPS IC. Again, a restricted access data set. It is a very large and unique data set, but it takes a long time to get access to it. You have to put a proposal in.

Interestingly enough and certainly not helpful, you can't get approval to do analyses that somebody else is already doing, so there is actually no opportunity to have groups run head to head with each other and learn from each other by replicating analyses that are being done.

You have to propose doing something that benefits the Census. It is not always clear that the things that would be useful in terms of health policy analysis are things that you can -- although we are creative -- arguably benefit the Census.

Again, this is inconvenient access. You have to go to a Census office. You not only can't import data into the simulation. They are very picky. In fact, Federico recommends for those of you who have not reading the page that describes access to the MEPS IC as something that tells you the difficulties. They are not even really happy about you taking tables away. They would prefer that you just take regression coefficients away, and that is problematic for its utility.

I think it took us about a year to get access to the data. We have just gotten access after we have put our initial round of results out, so now we are trying to figure out how we can make use of that.

Then to fantasy data sets we wish existed. It would be great to have a very large cross sectional population survey that gave us an ability to understand and model people's preferences with respect to choice of insurance. We think this can be fielded about every five years. It could include scenarios and vignettes. It is the kind of thing that would have helped us with the utility maximization and some of the other work that we have done.

Another data set that would be quite helpful is what we call the birth to death HRS, which gives us an ability to understand transitions over the life course, and would be particularly helpful for some of the health estimates that we are producing out of the microsim, which we base on the HRS and then we do lots of creative things to age it down to younger ages. But we could certainly improve a lot of those health outcomes estimates that we are producing right now with a much more detailed data set at younger ages.

We would love to have a large linked employee-employer data set that has information about the choices that people face. This is something that we would put together all in one place, things that we cludged together from multiple data sources, that would allow us to better understand firm behavior. Right now you have just aggregate data on the employees and firms, and you have a hard time looking at individual behavior within firms. So this kind of a data set would help with that.

Then the final thing that I think a lot of people are struggling with as we look at policies like the National Health Insurance Exchange is understanding something about the non-group market today, which is small but growing. There isn't any decent data out there, so I think all of us who are operating this base are probably creatively making things up.

Then finally, as we look at the provider module, it would be helpful to have a physician longitudinal data set that would enable us to model behavioral responses to different kinds of incentives. It would be helpful to have that mapped together with some understanding of physician characteristics and be able to observe changes over time that are linked to those kinds of characteristics, and that would allow you to link other kinds of data sets in like Medicare data.

I worry a lot that policy options are being modeled today strictly on the Medicare population. We don't have much of an understanding of what goes on outside of the Medicare population. From some of the analyses we have done at the provider level, I can assure you that it is probably quite different behavior within the commercial market than it is in the Medicare market.

So I will end with these three priorities. We really do support the letter yesterday, which is improving access to data sets that already exist, and in many cases to help with the documentation around using some of those data sets. We would like to have data available in a more timely manner. We would like the data that becomes publicly accessible to be no more than a couple of years old.

We hope that there is a priority on developing new data sets that allow us to look at changes to the health care system beyond just coverage expansions. I think there is a big focus on that. That is where a lot of the microsimulation modeling has been, whether or not you believe that we are about to get universal coverage in the next 90 days or whatever. I think that will only then begin the next round of things we need to look at. We believe microsimulation modeling can be helpful for that, but the data sets for doing those kinds of analysis are much more limited.


DR. SCANLON: Thanks very much, Beth. If it works for everyone, I think what we are going to do is have everyone present, and then we will have a joint discussion for the first panel.

So Jim Baumgardner from CBO, you're on next.

MR. BAUMGARDNER: Today I am going to primarily just talk about CBO's central model for examining proposals to expand health insurance coverage.

We call it HISim, the health insurance simulation model. We have been developing the model since early in the decade and have used it for looking at some kinds of proposals since around 2004.

Essentially it is aimed at looking at proposals aimed at the non-elderly population and proposals that are aimed to affect the amount of health insurance coverage. Obviously one of the effects we are interested in also is effects on the federal budget, since that is our primary responsibility, although looking at coverage and premiums and also looking at some distributional effects is also something we are interested in doing.

Recently CBO released two big volumes on health in December of 2008. One of those is called Budget Options. About a dozen of the proposals in there used this health insurance simulation model to develop ten year federal cost estimates as well as effects on coverage.

There is a December 2007 paper that we have that is on our website, CBO.gov, that gives more details on what I am going to talk about today in terms of the model itself.

It is a microsimulation model. SIPP is the platform for the model. Underlying everything are the underlying observations from SIPP. When I get to the next slide you will see how we have used a lot of other data sets and imputed onto SIPP people from other data sets.

We allow for the multiple coverages that some people have that you see in SIPP. We convert SIPP people into being part of health insurance units. In some cases you will get a family within SIPP that has more generations than you could plausibly have in one health insurance unit. Like, you might have a child who is in their upper 20s, let's say, who couldn't really be on the parents' policy. We may have grandparents present. So we do divide these into what seem to be coherent health insurance units.

It is what we call an elasticity based model. We start with a baseline with people in the coverage category, and then there will be some policy that comes in and in essence changes the effective premium that is available to that unit. Then people respond to those changes in premiums. So their coverage in the next period depends on where they start from, and then depends on the change in prices they see. We use elasticities from the literature to develop the behavioral responses. Many of the papers are referenced in that 2007 document that I mentioned before.

We also put each worker within SIPP as a synthetic firm created around them. So we have both firms making decisions to offer coverage or not, and then we have individuals, if they are offered coverage from the firm, they make a decision whether to take it or not. Then they also have an option to go to a non-group market, and we are developing other options besides the non-group market.

This is a rough schematic of how the model is set up and the changes that go on. In the upper left are the main data sets that are involved. So again it is a SIPP platform, but we use Bureau of Labor Statistics, some firm data they have to help us with our wage distribution in these synthetic firms.

CBO projections are relevant. Obviously we are dealing with a major federal program such as Medicaid and the SCHIP program, and we need to have people assigned in our model, so that the numbers that we say are in Medicaid in our model have to match what CBO says is in the baseline. So there are those kinds of calibrations that are done. We use data from MEPS. That helps us with things like health status spending, the contribution of employers, MEPS IC helps us with that.

We use the MBER tax simulation model to get marginal tax rates. Of course most people have employer sponsored insurance, so most people's insurance is excluded from taxes. That has an implication for what effectively the price of insurance is after you take those tax effects into account, and that depends on the marginal tax rate that a person is in. So we do need that kind of tax information.

Then the national health expenditure accounts are also used to calibrate up, so that we want to match total premium numbers that are in the national health expenditures.

In your base case, you have a person who is maybe in a family, who maybe are not, who has a firm that is offering insurance. We have their family income. Then a proposal may come along with could for instance be a voucher or some other subsidy or change in the tax treatment of insurance or a regulatory change that is going to affect the premiums people face, or you may ultimately have mandates on either individuals or employers to have coverage or to offer it. That is going to change the after tax premiums that a person faces in the employer sponsored market, and it will also change the alternative premium outside of the employer based market. Then as I mentioned before, both firms and individuals are responding then in this model.

Then ultimately we get changes in coverage, changes in the source of coverage. We can do distributional analyses by poverty level, and ultimately we want to be looking at budgetary effects. Federal budget effects is of course a focus of ours.

I will talk about four kinds of things we have used this model for which are proposals or potential proposals which we have in this health options volume that we released in December. Then I will end by saying where things are going down the road.

One such proposal we looked at was maybe what we would call in the category of a regulatory change. One of the things that we do in building up premiums in the model, if there is employer based insurance involved, the person is assigned to a firm, and then we have a synthetic cohort of fellow workers, and then we calculate based on information on health status and conditions and hospital utilization that we can pick up in SIPP. We get the expected health costs of each person in the firm, and then along with an administrative load we calculate a premium for everyone in that synthetic firm.

Now, for people who are in a small firm we look at the state regulations on insurance in the state the person is assigned to and take into account the rate compression regulations in that state. We also do this in the non-group market for people. A small set of states, a very small set, have full community rating. A somewhat bigger set of states have a modified community rating, and then many states in the non-group market don't have a rating restriction like that. But we model that to then get the estimated premium that the person would face in the non-group market.

This slides gives you an idea for non-group premiums in 2009 by reported health status of people. We have broken out the range of premiums that they face. This is the national look. Some people in a community rate state are just going to see a much narrower, or one premium.

So anyhow, that is what we do in non-group.

This first option that we looked at was a regulatory change that would allow people to buy a non-group policy from any insurer, not necessarily one that was licensed in the state the person lives in. What this does is, it undoes community rating, in the sense that a young healthy person in New York who is facing a high premium relative to their expected cost because of the community rating in New York, could buy a policy from an insurer that is licensed in Arizona and could probably find a lower premium there.

So you get a situation where some high expected cost people in what I will call more regulated states ultimately would face a higher premium as the lower cost people in those states unwind and go to other insurers outside of the state where they could buy a less expensive policy. So you crank all this through.

Also there is an issue of mandates to cover in vitro fertilization and an assorted list of things. That is also something one could avoid the rules in one state if this were the policy.

So in the end we calculate about a half a million reduction in the uninsured, but you do get this differential by health risk with some higher health risk people in more regulated states seeing higher premiums. The biggest premium reductions are for the people in those same states who had lower expected health costs.

We also get some firms dropping insurance, in this case because as more employees find a cheaper option in the non-group market, there is some impetus for some firms at the margin to stop offering. That is part of the whole calculation there.

A different option we looked at was a subsidy to purchase non-group coverage. Roughly a subsidy was allowed of up to $1500 for a self only policy, up to $3,000 for a family policy. That ends up being about a 40 or 50 percent subsidy for the average uninsured person.

We set up this policy so that it would phase out starting at 200 percent of the federal policy level, the subsidy would phase down to zero at 250 percent of the poverty level. At most, 70 percent of the premium could be covered by the subsidy.

So what we got from the model then, about four million people ended up using the voucher. About 2.2 million of those would have been otherwise uninsured people. We calculated about an $8.6 billion budget cost in 2014. That is a one year budget cost. So that is another kind of policy we looked at.

Then a third proposal we looked at would extend the tax deduction for all non-group premiums. The currently self employed people can take a deduction, but other people in the non-group market cannot, and this would extend the tax deduction for anyone buying a non-group policy. In that case, we estimated that we would get about 700,000 newly insured people as a 2014 number, as well as a $6.3 billion budget cost.

If you do division and look at how many newly insured people do you get per budget dollar of expenditure, this policy doesn't do as well by that metric as that previous one, the vouchering to low income people. Part of what goes on is targeting, targeting, targeting. To just have a tax deduction, obviously people at higher income are more likely to pay taxes if the deduction is worth more to them, but that is where you have fewer of the uninsured people. So that is what is going on there.

Yet another option we looked at was to look at this notion of federally subsidized reinsurance for high cost claims. So the policy here was when a person's claims exceeded $50,000, the federal government would pay 75 percent of the rest of their claim costs above that.

I think we looked at two options in our volume. I am only going to talk about the one that was aimed at smaller firms, those with fewer than 100 employees in the non-group market, but I believe we also had an everybody policy we put into the book.

We concluded though that this would be about a 14 percent subsidy, obviously higher for individuals that had higher than expected health costs and for firms that had a group of workers and covered family members that had higher expected costs.

We used the MEPS, but we also used -- and here is a data issue. As we talked to actuaries about trying to model something like this, I think the consensus view was that within MEPS it is the right tail of the spending distributions then compared to what you would actually see in claims data. So we used some Society of Actuaries data to fill that tail in some more and to do this estimate.

So ultimately we estimated that the $2.1 million reduction in the number of uninsured has a very large budget cost of $32 billion as the 2014 number, so that is just an annual number. In essence you are buying a very large base out in this proposal. Obviously under current law lots of people to have health insurance, and that health insurance is covering all the claims. Here you are taking 14 percent and have the government pay that.

Then finally, we also started to experiment with an employer pay or play type of proposal. The proposal we looked at had a $500 fine per employee for employers if they did not offer health coverage. This only was to larger firms, firms with more than 50 employees. There was a requirement that they cover at least 50 percent of the premium, those larger firms.

We estimated to begin with that there were about ten million uninsured people, independents who are in firms that are large enough to be affected by this type of pay or play system. In the end, we calculated that there would be about a three or four percent reduction of uninsurance within that ten million group, so pretty small actually, is the reduction of the number of uninsured under that policy. The proposal does make some money for the federal government though, and that is mainly firms paying the fine.

Of course, with a larger -- $500 is pretty small. As you get up to what the premium cost would be, you would expect much larger participation by firms.

So going forward, we of course are going to be in the heart of big proposals to change the health insurance system. The other kinds of options, some sort of subsidy proposal, some sort of subsidies probably targeted toward low income, lower income populations, and other offerings possibly besides non-group that is available today, those are probably the directions things are going.

Federico is here, but the brains behind our operation are working very hard at a location south of here right now.

I can't add much. I thought Beth did a nice job of covering some of the data issues. I really can't top that. We certainly would like everything. We would like a huge sample of people and their employers and their fellow workers and details on the insurance coverage all in one big sample, along with the people's health conditions and what their health spending had been. That is the top-down approach to what we would like.

Thank you.

DR. SCANLON: Thanks very much, Jim. John, are you set? We will give him a second. John?

DR. GRUBER: Hi, sorry, I got dropped off just as Jim was finishing up. Are you ready for me now?

DR. SCANLON: We are ready for you to go. We are trying to find your slides, and then we will be set.

DR. GRUBER: Tell me when you are ready.

DR. SCANLON: Okay, we're set.

DR. GRUBER: I have a less detailed presentation than what has been given so far, but I think a lot of the things we have been discussing are pretty similar. I'm glad you guys convened this meeting. I'm sorry I can't be there. I think it is very exciting that so many people are working this space and thinking hard about putting numbers on questions which often are very hard to put numbers on. I think the difficult issue for ASPE is going to be that even when very smart, well meaning people get together and try to put the same number of something, they are often going to differ. The difficulty for you is how to work through that and figure out which numbers to rely on, so I think it is useful to get a sense of what everybody is up to, so it is a useful meeting. It is useful for me too to hear it.

I am going to talk about what I have been working on with my microsimulation model. If you go to the second slide, there is a highly detailed schematic of how the model works that I had to have my eighth grader do, because I don't know how to do this sort of thing.

The way the model works is not so different from a lot of other microsimulation models. I have got two inputs, the data sources and the policy parameters. This enormous black box which is 15,000 lines is data code, which basically converts everything to price responses and makes behavioral assumptions on how individuals and firms will respond. Then there is an output, which is population flows and dollars. So I will talk a bit through how this actually works in practice.

The next slide talks about the data. I thought Beth's presentation was excellent in terms of laying out the alternative data sources. Unlike Beth and CBO, I used the CPS instead of the SIPP. I think there are pros and cons there. One pro is, the CPS is a little bit more recent. There is a March 2005 CPS which can be matched to February 2005 to get both coverage and employer offering. So that is the data I used. On the other hand, the SIPP has the problem that it doesn't have dynamics, which is obviously important for modeling things like state COBER subsidies. So that is a downside of the CPS.

The other upside of the CPS of course is that the samples are larger, and for reasonably sized states you can do state specific estimates with reasonable confidence. So that is another pro of the CPS, but I think that is something that we need to discuss.

My model recalibrates the later CPS by age and income cells, recalibrating insurance coverage in particular to where we are in the most recent CPS. Going back to what Beth was saying, often if I do state work, states will often have their own surveys. In Northern California this is true, in Connecticut this is true. States have their own surveys, where they will have certain targets states think are right for number of uninsured and such, so in those cases I will recalibrate to those numbers, because on the state basis it is useful to be using the same numbers the policy makers are using.

Employer premiums comes from the METS IC, from their public use data. They have tables by state on premiums by firm size, so that is what I use there. I also got untabulated data a few years ago from the METS IC folks on the distribution of premiums within firm size, so what the tenth, 15th, et cetera percentiles were. So when I assign firms a premium, I essentially assign them a draw from the distribution for their firm size, and then renormalize by state.

Then for non-group premiums, what I do is take the MEPS micro data on people and use that to get underlying health costs. I basically take people who have employer insurance, say what do they spend in a year based on their age and reported health status, and then I take that age reported health status, I assign it to an underlying health care cost, and then I assign a load factor to turn it into a non-group premium.

The tax rates are imputed using everyone's favorite tool now, TaxSim, which is pretty straightforward and that is the way you do tax stuff right now. Right now the model is projected to 2009.

On to the next slide. What is in the black box is the key to the way I think about things, let's take everything and turn them into prices. So take any proposal you want, or any market change and ask, how is that changing the prices that face different people as you consider different avenues of insurance.

Here I should have had another nice diagram which I didn't, but my model is divided into four insurance groups, employer insured, non-group insured, publicly insured, which is Medicaid because I am doing non-elderly, and the uninsured. Then I have different ways they can get insurance through non-group, through ESI or through public.

For any given policy shock, how does that policy shock affect the prices each different group faces for each different avenue. So you can think of a four by three matrix, four different insurance groups considering three different avenues to buy insurance. We are saying, when there is a price shock, how does that affect the prices that face each different group in each different avenue.

So for example, if you give me a $1,000 credit towards non-group insurance, it costs $10,000; that is a ten percent subsidy. Then one place I differ from some other models is, I don't try to internally estimate as part of the model the key elasticities. I think that these elasticities are very hard to estimate in cross sectional contexts. I think you need more precise empirical strategies, like drawing on natural experiments and other things, to estimate the key elasticity. So I draw the literature to estimate the best elasticities to use for each response rather than internally to the model, estimating the actual elasticities, which I think you can get misleading answers. So I draw on and use the best estimates from health economics and more credible empirical strategies to try to figure out what the right elasticities are to use.

The next slide gives an example of some of the various avenues one needs to think about. Let's say for example the government announced, as Bush had in his budget for a number of years, we are going to have a credit towards non-group insurance. One of the questions one has to think through is, why is this 15,000 lines of code and not ten lines of code. The answer is, because with every given change there are a lot of different questions one needs to think about.

For example, to what extent would those who already have non-group insurance take up this new entitlement? We know that people do not always take up even money that is sitting on the table. With the earned income tax credit, which is free money for people, take-up is about 85 percent. So we know that take-up is less than complete, even among those who don't have to change their behavior much to take it. So that is the first question, how do you model the take-up among those who already have non-group insurance.

The second question is, to what extent will the uninsured buy that subsidized insurance, and how will that vary by their income and by their health status. So I have a series of take-up equations which are going to say, when we subsidize this non-group insurance, to what extent is that going to cause the uninsured to take it.

Those take-up equations depend on a number of factors, the uninsured's health, their income, whether they are offered ESI. For example, we know from work that Jessica has done and I have done and other people have done, Linda Blumberg and Len Nichols, there is a whole series of articles which show that people who are offered ESI but turn it down are not very price sensitive.

So what I have in my model is the notion that if you were already offered employer sponsored insurance, you are uninsured, and you turn it down, you will respond less to a non-group than if you weren't offered, because you have already revealed that you are not that interested in having health insurance. So that is the kind of thing that is in those behavioral equations.

Then you have the question, to what extent will those who have employer insurance decide to move to this new option. For most employer insurance, even with $1,000 or $2,000 credit, it will still be cheaper to stay on ESI than to go to the non-group market. But for others who are charged a lot for their employer insurance and who are healthy so non-group is cheap, they might want to move over. So that is another reaction you have to model.

Then probably most importantly is how do you model the employers to all this, how will employers respond by dropping insurance or raising employee contributions, or how they will respond by altering the generosity of what they provide, which in turn feeds back to government budgets because of the tax exclusion costs of employer provided health insurance.

This is just an incredibly brief overview of what is in this huge black box, how do we convert these policies into price changes and how those price changes change a move to behavior.

The next slide talks about something I feel pretty strongly about, which is doing a fully integrated policy analysis. What we are seeing more and more, you are seeing it with everything Baucus and Kennedy and Obama are talking about, is not something that Bush talked about, let's do a non-group credit and nothing else.

These are very integrated policy analyses, with ten or 12 different moving pieces, from non-group reforms to subsidized care to subsidized low income pools to small business credits to pay or play tax on firms.

What is very important when you are doing things in price terms is to consider these all at once. I think you can get in trouble if you stack these and say, let's first do this one, then this other one, so let's first do the effect of pay or play and then let's do the effect of low income subsidies.

It turns out you can get a lot of what we call path dependence, and very different answers, depending on which order you stack them in. It is very important to do these things all at once, so you are not path dependent based on what you are doing first or second.

The other thing that I think is important that I built in my model is not to have -- to try to minimize as possible knife-edged distinctions. Once the credit gets to be X big, then firms will pay attention. I think that may in fact be right, but it is incredibly sensitive to an unknown, which is, we don't really know what X is.

I will often get asked by policy makers, how big does the credit have to be before people will respond. The answer is always, that is not a well defined question. As the credit gets bigger, people respond more. I think it is very important in my view, the way I do my modeling, to not have a knife-edged cutoff, but rather have it be continuous, and reactions continuously increasing in size.

The next slide talks about modeling firm behavior, similar to what a lot of people are doing here. The model is based on individual data, but so much of what we draw in a firm reaction. So that creates synthetic firms, where I have data from the BLS for a given wage category. So for a given worker who earns say $15,000 what is the distribution of earnings of their coworkers. I will have it separately by firm size and locations in the country.

So I can say, given that you are in a firm of ten to 25 employees in the West, given that you earn $10,000 to $15,000, I know the probability distribution of the earnings of your coworkers. I know what share of your coworkers are likely to earn $5,000 to $10,000, $10,000 to $15,000, $15,000 to $20,000, all the way up to $100 plus.

So what I then do is go to the CPS and pluck out synthetic coworkers for you to match that earnings distribution. So I have built up the synthetic firm to match what I know is the underlying correlation of wages across firms. Then once I have a synthetic firm I can say, now let's ask how the firm reacts based on the characteristics of workers in that firm. So that is what goes on with the important part of modeling firm behavior.

The last slide, I want to talk about a couple of important caveats that I think apply to all the models we are talking about today. We have talked some so far about data as well, but I think we need to step back and think about the general issues.

The main thing you have to remember is that it is a garbage in-garbage out process. We are just relying on a lot of assumptions here, we all are, and the better the assumptions, the better the answer. I think that is why it is critical to not worry just about the structure of what we do, but where we get the elasticities that drive all the behavioral responses, and using the best available evidence, using groups like this to discuss what the best available evidence is on behavioral responses. Then when that is not available, figuring out how you can tie to that evidence as closely as possible. So that is the first caveat.

The second caveat which I think is true once again, it is certainly true for my model and I would argue for most, is the effort is going to vary depending on the magnitude of the change. I feel very confident that if you tell me you are going to put in a $1,000 non-group credit, I can tell you with reasonable precision what that is going to do. If you tell me you are going to do a new pool where everybody gets subsidized for up to five times poverty, financed by taking away the employer tax exclusion, I am on much thinner ice.

So once we move to grander changes of the type that we are talking about now, I think we have to recognize that the confidence intervals go up, I presume for all of our models, certainly for mine. The confidence intervals are going to go up, and I think we need to recognize that uncertainty. Folks like CBR aren't allowed to do that; they have got to give one answer, but I think we need to be internally honest, because we are going to be getting less precise answers as the policies deviate further and further from where we are today.

Then that leads to my last point. As much as we can do in groups like this to be transparent about modeling the process, I call my model a black box, but in fact I have now put together a spreadsheet which lays out every single one of the assumptions that is in the model, in a document which describes it all. I think we all need to be as transparent as possible in what is going into the models, so that ultimately folks like ASPE and CBO and others who need to use these to make policy can understand why we are getting different answers and what is going on.

So that is all I have.

DR. SCANLON: Thank you very much. Hopefully you can stay around for the discussion?

DR. GRUBER: Yes. I might hop off, but I see you have some more presentations, and you are regrouping at noon for the discussion?

DR. SCANLON: No, actually probably about ten minutes or so. Bo Garrett is going to present, and then we will have a discussion.

DR. GRUBER: Okay, good, I'll just hang on then.

DR. SCANLON: Thanks. Bo, you're on.

MR. GARRETT: Great, thank you. Good morning. Let me thank the committee for inviting me today to talk about the health insurance policy simulation model that I am creating with a team of my colleagues at the Urban Institute. I am really here to talk today on behalf of the whole team.

For people who may have come in before I did, I brought handouts with me, so there are handouts on the table. I am going to talk about our model. HIPSM is a detailed microsimulation model of individuals, families or HIUs as they technically are in the model, and employers making coverage decisions within a health insurance market. You are going to be hearing a lot of variations on a theme now, at this point in the morning, having heard about three other models.

This one builds on the Health Policy Center's experience with the health insurance reform simulation model, or HIRSM. That was used to model reforms in Massachusetts in building the road map to coverage document, Blumberg, Holahan and Weil.

This model takes a lot that we learned from that, but it is designed to be faster and easier to tailor to new policy specifications and state specific analyses. This was jointly developed with the Tax Policy Center.

To give you an overview of my talk, I am going to cover HIPSM's capabilities, the model structure, current applications, and I'll just make a few comments about data needs.

HIPSM is designed to cover a very wide range of policies and to be expandable to ones that we haven't yet put in. That, just to name off a few, includes Medicaid and SCHIP eligibility expansions, individual and small group market reforms, reinsurance, income related subsidies, purchasing pools, individual employer mandates. Basically it is all the components that you need to model something like a Massachusetts style expansion, and then there are others as well.

The output, just to go straight to the end, the output of the microsimulation model is a bunch of detailed tables of the estimated effects of the reforms, including insurance coverage status of people and baseline and reform. We have the four categories, uninsured, Medicaid, public, non-group and ESI. The costs of the reforms to different groups, government employers and individuals, changes in ESI premiums from sponsorship, a range of things, cut by interesting factors like income, age, health status, firm size where applicable. These tables, we have a specific set, but they can be modified and extended for specific needs to look at very narrow populations that one wants to.

I like the word cludge you used for bringing all the different data sets together. I love that word. We have had to use a lot of that for our database.

We need a data set, an individual level data set that resembles the health insurance situation in the United States in 2009, or maybe some other year. What we start with as our core file is the 2005 CPS, annual demographic file. That is matched with the February file contingent work supplement to get information on employer offer and some other variables. It is also linked to the MEPS HC to bring in individual level expenses and other health status variables, the survey of income, which has a lot of information on tax related income, and tax variables from the Tax Policy Center's tax model, which would include marginal tax rates.

Given the goals of this talk, one thing to point out which is very clear is that we have to bring all these variables together from different data sets. We do our best to statistically match them in ways that preserve the relationships among variables. But surely there are things that are lost. If I knew a particular factor was lost, I would try to make sure the data reflected that in the end, but you can't look at everything. Ideally one would have this all on one big file, one big survey.

Like some of the other models, workers are organized into synthetic firms, defining what the population of firms looks like out there. We have used information from both the MEPS IC and Statistics of U.S. Business. They both have characteristics that the other ones don't, and together they fill out a picture of the distribution of firms.

Then these data need to be reweighted and adjusted to match various benchmarks that we have for coverage, income, health care expenditures. Some of the benchmarks come from work by Selden and Sing that does reconciliation of the MEPS and the national health expenditure data, because we want the aggregate amount of health care expenses in the model to be nationally representative.

We have an aging module that gets us from the '05 survey related to '04 data for coverage and incomes, and bring those to 2009. When we do that, we can use a variety of -- first we will get it to '07 data, because we know what the '07 CPS is now, and then going beyond that we will make some additional assumptions to get to '09 and even further, as needed.

The premiums in HIPSM are endogenous as they were in HIRSM. They are built up from the risk pools in their underlying health care costs. We apply typical rating rules for a typical state; we don't do each and every state's rules at this point, but we take typical rating rules in the individual and group markets and then apply those to the national population within each risk pool, whether it is an ESI risk pool or a non-group rating.

This results in expected -- we can compute expected costs conditional on the rating rules, multiply that by a loading factor for administrative load, and then we get our premium estimates that would vary by firm size and the different characteristics of synthetic firms and the different non-group rating groups. We benchmark these with the ESI at least to values in the MEPs IC and the Kaiser HRET employer benefits survey. In the latter case, if you look at the tail data you can get a distribution of what the premiums out there look like by firm size, for example. It helps us target those better.

Talking about the structure of the model, the behavioral effects in the model, it uses a utility based approach. When we started this a few years ago, we thought that would be very helpful in helping us -- we intended for it to help us get better estimates of the effects of reforms that are well outside our historical experience, rather than an elasticity base model which is more locally relevant.

It also provides a general framework. In the Gruber model, all these potential effects were converted into prices; we leave ours in their natural form, because we specify this utility approach, so everything does get eventually expressed into changes in relative utilities of the different available options.

In the model, individuals choose the option that provides them the highest utility as we determine it to be that firms offer. If there is total willingness to pay, it sees the total cost of offering.

To get at things like people who look like they want to be in Medicaid but they are not, you have to deal with these latent factors. We describe total utility as a specified utility function which I will talk about, plus a latent utility error term. That is because we need to assume that existing coverage is optimal at baseline. We find you in this particular category with these particular characteristics. For the model to work, that must have been what was best for you at the time, at least in some sense.

The key to HIPSM's mechanisms then are imputed error terms that do two things. They have to insure that the baseline is optimal, and they have to yield premium elasticity and take-up rates consistent with the targets. This is where all the behavioral stuff is going on. This is our black box. This is where we have to do calibrations.

What we do here, we have got the underlying utilities. We have got error terms. Locally, for say a 20 percent change in the ESI premiums, we want firms to quit offering if premiums were to rise by rates predicted by the best elasticities from the literature. So we are targeting firm level take-up, firm operated individual ESI take-up rates, non-group elasticities, Medicaid take-up rates by whether you are uninsured, by whether you have ESI, which would yield the assumptions we make for CREDA.

One of the things that the utility framework also does is, it automatically will generate some heterogeneity in effects. We may have targeted an overall aggregate elasticity for non-group, but then people who have higher health expenses may be more likely to move because any change for them may be magnified by the fact that it is all in a higher amount of dollars given their spending levels.

The utility functions themselves are developed based on economic theory. I believe the ones used in the RAND model are quite similar. It depends on the expected out of pocket health care expenses, variants of out of pocket health care expenses, the value of health care that you consume. That is going to differ, depending on if you go from being uninsured to insured; you will have an increase in your health care consumption, and that is going to add some value. Out of pocket premiums, tax incentives, and we also have an affordability component which is your expected out of pocket expenses of that option over your income.

Let me walk you through how the flow goes in HIPSM. Reforms change the available options to you. It may change rating rules, it may change relative prices of ESI versus non-group coverage. This would also have potentially a change in workers' willingness to pay for ESI within the synthetic firms. Firms will then react to these altered worker preferences.

Given what firms do, individuals or families will choose the new best available coverage option given the firm's decisions. The premiums need to adjust to reflect the characteristics of the new risk pools. Then the model will iterate until coverage changes across iterations and is stable.

To walk you through a particular example of Medicaid SCHIP. Let's say public coverage becomes available for new eligibles. I should say that the current eligibility model in HIPSM is a detailed Medicaid eligibility simulation model that we start with, but then we can specify an expansion as covering everybody, like all children at 300 percent of poverty and adults at 150 percent of poverty. This may be particularly attractive for new eligibles that had higher out of pocket costs. So this reduces the demand for ESI, because now you have got this new option that is free or very low cost. Fewer firms will offer ESI, so that will cause some people who may have had ESI and aren't eligible for Medicaid to lose it.

Overall, Medicaid SCHIP coverage should increase, uninsured and other types of coverage would likely decrease. Government costs and total health care spending rises. Private premiums will adjust, then we go through a round of second order effects.

The model up to this point that I have described is mostly about coverage changes. We can bring in alternative assumptions to the model results that take into account various supply side considerations and others.

For example, we were looking at a particular proposal. It may have proposals to invest in various cost containment strategies. You look at the literature and consult experts and come to a conclusion about what kind of effect those would have. Those can be then included in the model's output.

Another example would be supply constraints that might occur if you have universal coverage. The model may say costs are going to go up by this much because it is initially assuming that there is going to be supply to meet demand. That may not be true. If not, we can take that into account.

Our baseline assumptions would have consensus public program participation rates from the literature. A proposal might include the fact that there is going to be a huge outreach effort. Let's say the best estimate of what that would do from synthesizing the literature would be a ten percent rise; we can include that in.

Also, because we have got this aging module, we can have various scenarios for future wage, employment, health care costs and insurance premium growth.

Let me talk about some of the studies, well, I guess this is all of them that we have underway. There are national studies and state specific studies. The national studies, we are doing a variety of Medicaid SCHIP expansions for Kaiser, where we are specifying different levels of expansion as a percentage of FPL for children, parents and non-parents separately. We are doing this with and without enhanced outreach efforts.

We also presented some preliminary analyses at the AEA meetings of a set of reforms at the national level that would resemble the types of reforms that would be contemplated. These build on each other in sequence in the way that we present them. They are not sequential in the way they are applied, as John Gruber was speaking of. They are integrated. These happen together when we are specifying them as a group. So we looked at a Medicaid SCHIP expansion only if you reformed the non-group and small group markets, like applying age rating and giving low income premium subsidies, and what would happen if you had the two of those. Then what would happen if on top of value added employer pay or play mandate, et cetera.

We are also looking at what would happen if there is no reform. If there is no reform, we can make various explicit assumptions about what the growth in health care costs would be, if there is any growth in premiums over that, what the growth in wages would be in the next five years by different quintile of wages perhaps, and include that change in health care costs relative to changes in the model which would reduce the amount of coverage and increase the number of uninsured, but also have the demographic growth by age and gender cells, and predict changes in employment and patterns like that.

We are also doing some state level analyses. I should talk about some of the data issues particular to that in a little bit, but we are modeling a large array of expansions for New York State. We are looking at a wide range of things likely to be considered by a state or the federal government. We are doing Massachusetts style expansions, looking at a single payor example. Within the combined public-private approaches, we are doing everything from little tweaks to much more comprehensive things.

We are also doing something like, if there is no health care reform, for the Colorado Health Care Foundation and Medicaid SCHIP expansions and some other things in Colorado.

I think Beth did a great job covering the data needs. I am just going to make a few points. I'll second the need for employer linked data. We would like to have a large data set linking the two at the most superficial level for better construction of synthetic firms, or it would be really synthetic at that point.

Also, for the testing of core hygienics about how firms make decisions given how workers sort into firms and how they arrive at those decisions to offer or not.

There is no national data set that has all the variables we need. There is certainly no state level data set that has all the variables we need. Sample size becomes a problem. A state specific database that is commingling all the characteristics we need would be extremely useful.

I should say that we do the best we can. I think we do a reasonable job of compensating for the data we don't have. We will take several years of data in a state or a region that we think can represent a state and then reweight that data to look like particular data in a particular year, and bring in various pieces of state information to try to take what we can know about the characteristics of the state and make sure our database reflects that, before we proceed to simulations. But obviously this would be much better if we had real data to do those.

Thank you very much.

DR. SCANLON: Thank you, everyone. I think you gave us a lot to think about. I'll open this up for questions.

DR. STEINWACHS: Two of the modeling efforts use elasticities alone, if I understand correctly. You are moving to utilities in RAND.

Could you say a little bit about two things I was interested in? One is how much difference you see between using elasticities alone and using utilities. Then, I wasn't very clear about what data you use or what research findings you use to get at the utilities themselves.

MR. GARRETT: In our case, we are probably looking at the same elasticities that John Gruber would be looking at and John Gruber in many cases estimated, and making sure that for a 20 percent change in non-group premiums, that we are going to get in that particular case an effect similar to what a CBO study says, in terms of what is going to be the percent change in the utility if there is a 20 percent change in the non-group coverage.

So locally with that just one thing going on, we seek to be consistent with what we determine to be the fast read of the literature. But then if you had some really big change, if you had two changes at once, we wouldn't just be applying a particular loss in sequence; they would be interacting in the model according to peoples' utilities, so a different elasticity would come out than the one you targeted to because the situation is different.

If you had a really big expansion, the intention of having this utility approach is that you won't get a constant utility as you go out or a constant function going out. It will depend on who is being affected and what they value in the model. We hope that that way we can get better answers to some of the big what-if questions.

DR. STEINWACHS: So when you estimate a utility for a subgroup of the population, how are you arriving at that estimate?

MR. GARRETT: That is a very good question, how do you start with utilities to begin with. There, there were several papers in the literature that we looked at, what would go in them. Some of the people at AHRQ did some papers. Sherry Reid did a paper, Colleen Herring did some papers.

They would specify, here is what a utility function for different coverage types might look like. There might be very particular reasonable coefficients of risk aversion that one would have various assumptions. We would probably use an assumption close to one Sherry Reid used in a model, about the value of additional health care. So we will draw on those. We can experiment with them, too. We will start with those. To the extent that that doesn't yield what we need in terms of cost, this is what the error times yield.

DR. GIROSI: I think Bo and I read the same papers.

DR. STEINWACHS: That sounds good. Evidence based approach, I like evidence based approaches.

DR. GREEN: I've sure got a flood of questions, but I think I will save some of them until the next session, too. Before I ask a couple here, as a recovering physician, I'd like to thank you for the therapy session. Ever since the publication of the Chasm work, you realize you go to school for a quarter of a century, and then it appears you don't exactly know how to answer the patient's question. It has been very reassuring that we are all in good company.

Do you guys do autopsies?

DR. STEINWACHS: On the health insurance system.

DR. GREEN: Once you have done some projections and then somebody makes a decision, do you then go back and see how close your model was to what actually happened? If so, do you have anything to say about any of these approaches, which one is demonstrably more likely to hit the target?

MR. BAUMGARDNER: We go back on things if they actually happen. With the non-elderly market, I can't think of examples recently, but when we look at Medicare and Medicaid projections at the end of the year, we have to then go ahead with a baseline forecast for the next ten years, and are looking at what happened in the last few years, how that turned out compared to what we had expected at the time.

So there is that kind of low reactor that goes on. If we had more time, I think we would do more of it, but we do some within the limits of the time we have got to do it.

MS. MC GLYNN: We are looking at going back and looking at the Massachusetts experience. I think that is the main big one we have, and look at how well we would have predicted or replicated what happened. It does give you an opportunity to learn what was unexpected when things get put into the real world. We have a way to incorporate that back in.

We had to first do the step of putting the multiple options together, which we have just done, because we previously had just looked at options one at a time. Since Massachusetts was a multi option policy change, we weren't quite there yet. We are now pretty much at the point where we can -- once we finalize the utility maximization, we can go back and try to replicate that. But we think that is a very important way of continuing to refine the capabilities of the model. The big challenge as Jim said is that we don't have that much real world observation to draw from, but it is important.

MR. GARRETT: One thing about the Massachusetts expansion that would certainly provide feedback to me is, before the individual mandate was in place, a lot of people responded.

It might be more people than responded to a $500 penalty that hadn't even occurred or more, but the penalty hadn't even occurred. If it does occur, now that it has, is it commensurate with any old thousand dollar penalty of not doing something. If there is an overwhelming response, you might think people maybe are not being penalized much more than they would getting the equivalent amount.

So you say, what to do with this fact. When we do our individual mandate, we might just do it to assume straight out, rather than what is the response to this financial penalty, for the individual mandate we might say let's assume it is 100 percent effective or 90 percent effective. Then how much would it cost, and then you can do sensitivity analysis around that and create some different scenarios.

DR. GREEN: I listened to all of your presentations, the title of the hearing here, modeling insurance data including coverage, access, utilization, quality and cost. But I was struck with the lack of consideration of results to people, like emotional status, to cartoon it in this model. There would be an expectation that more people would get better or fewer people would get better or more people get worse or fewer people get worse, even measuring something related to morbidity, comorbidity or functional status.

My question is, is that because it is out of scope and it was not what you were asked to do? Or is it that you use it as a modifying or confounding variable in the analysis at some point that I missed? Or is it a dependent variable that is not available? Why do these types of analyses not drive it home to the fact that all of this eventually needs to track to whether or not people benefit?

MS. MC GLYNN: We actually do have health outcomes in our model. I greased by that fairly quickly. I hoped to have Internet access where I could have shown you those results.

We produce estimates that are primarily right now life expectancy estimates. We hope over time to make those more detailed.

The challenge there frankly is the data limitations. When I talked about the dream data set of the birth to death HRS, right now a lot of what we have done is to build on work we have done previously called the future elderly model, where we looked at changes in Medicaid policy and the health effects on the population. We were able to draw that down with HRS with some comfort to the 50 and older population, but then to look at health effects for those under 50 population, which fortunately have fewer deaths you have to worry about. It is a bit more challenging.

But we actually do produce health effects, and that is something we are going to continue to try to build on.

DR. GREEN: Can you help us understand those data limitations? Those are really important to us. We are interested in data limitations.

MR. GARRETT: In a particular application, I think it would be reasonable in many cases to say -- our model at this point mostly is about the coverage changes. That would be very reasonable if you were writing this up to say, these are the coverage changes that we observed. Here is what we know about how these will affect people's life span, morbidity, et cetera, and talk about those. If one had some marginal effects or something from the literature that one wanted to add to that section, you could.

Of course it is a point that is very important to discuss. We would want to highlight those in applications. It is not critical I think to getting to the coverage estimates, but then once you have got the coverage estimates you can talk about what further implications those would have.

DR. GIROSI: If you want to get into the health part of it, you do need to follow people for a long time. The SIPP and the MEPS follow people, yes, but for a relatively short amount of time. So if you want to see people developing diabetes or developing a heart condition, you need to follow them for a number of years. That is why the health and time survey is a wonderful survey which follows people for up to 12 years biannually. It is a great source of information of how people transition from one health state to another. You can also correlate that with the insurance they get, the access to treatment they get. That is why we were able to do some of that for the 50 and older, basically. But the HRS does not do anything for age less than 50, 45, depending on which ways you get it. So that is one of the data limitations.

DR. GREEN: I take it from those responses that -- you mentioned this, Beth, in your presentation too; is the key data limitation insufficient longitudinal data about particular individuals?

DR. GIROSI: For doing things in health. You really need to follow people for a long time.

MS. MC GLYNN: I guess I would also argue that one of the challenges I saw when we were looking at the health results is, there isn't a lot of good detailed linkage about coverage, what that means for access, what that means about whether people are getting the care they need, which I know from my own work is not -- coverage is not a guarantee; there is quite a bit of gap.

The thing that many of these things lack is any kind of real data about the continuation of receipt. So we make lots of assumptions based on pretty narrow bands of observation about, if I observe a patient who has been observed with hypertension today and is getting hypertension meds today, we know what that is going to look like in a year, five years, ten years.

So we made for chronic disease a lot of assumptions about the maintenance of effort over a long period of time, which is probably the real key to the ultimate health outcomes. We really have very little good national data. We have lots of cross sectional data that we use to make those kinds of estimates, but I think they are fairly heroic with respect to where we are really at. But we are certainly trying to do the best we c with the data.

But I agree, if you are really interested in looking at health policy from the perspective, as your committee, I presume does, of what the impact is on the health of the population, which is ultimately what we ought to care about, then I think we are sadly lacking data sets that help us with answering those kinds of questions.

DR. SUAREZ: One of the interests I have is minority health and health disparities. One of the things that I didn't see and I wanted to explore that in these microsimulation models and approaches, is to what extent -- and there is some literature out there about the difference in decision making by minority groups of health insurance -- to what extent the models can allow you to look into the effect of policies on the decision and the insurance coverage and other things on minority health groups, specific minority groups, then look at the effect of those decisions on the outcomes of utilization, quality on minority groups.

Is that something that your models can look into?

MR. GARRETT: I'll just start. We have not looked at that much with the model yet, but those data are there in the baseline file, and one could tabulate the effects generated by the model by race and ethnicity.

We haven't made specific tailoring to the utility function specifications according to that. If there were some evidence to bear on that, we could.

The people would be found in different types of jobs with different incomes, and to a large degree that ought to be reflected, and would have different health statuses. To a certain degree that ought to be reflected in the final results already. But that is an interesting question.

MS. MC GLYNN: We also haven't done that much with it. We do have race data. Often when we get into the analyses you end up with white/non-white being about as discrete as you can get, even if you start with more detailed data than that on the starting sample.

One of the things is a good capture in terms of all of the different places people are getting care, how well we really understand. So we do know for a lot of minority populations there are different avenues of seeking care. We don't think those are particularly well described in the data sets that we are using. That would be helpful to have more data on that.

From some very small studies we have done, I suspect some of those suggest some of those providers are delivering better care than what people are getting in the middle. So the Medicaid population for instance may be less well served than people who are going to free clinics, from some work we have done.

But it is something we certainly are interested in. We just haven't gotten very far with that.

MR. LAND: You mentioned the various problems with the data sets that are available. They are not very timely, they don't have the longitudinal approach, they don't have the data items you need and so forth, even down to the fact that you don't have very good computers that you can access.

I know the National Center goes hat in hand to other federal agencies to solve some of their financial problems. Does AHRQ and Census Bureau and any other data sets that you are using go to you all hat in hand and say, we could do better if you could contribute to the cause, even to buying another computer or to going to all the think tanks and research companies and saying, how about kicking in some money? Does that happen? Is it possible? Would you participate?

We all know that the federal statistical agencies are challenged by resources. I am just wondering if they have looked into other ways of helping them out that would help you all out?

MS. MC GLYNN: As far as I know, the main solicitation that we get frankly is to be asked to participate in efforts to lobby on the Hill for better funding for these agencies. I did not use the word lobby. Strike that from the record, please.

But anyway, as far as I know, we have never been asked to do anything like that, other than try to be a voice of support for the broader needs. I recently got asked this by a committee staff on the Hill, about what the broader research data needs were in a particular area.

So we get asked those questions, and we certainly try to lend support for these kinds of efforts. For instance, on this longer HRS type longitudinal data, we think there was something like that on the books, and then it got cancelled. We certainly tried to speak in favor of that. We have not been asked directly for contributions, old computers that we could offer.

DR. SELDEN: I think if there is any sort of issue like migraine related with the screens that we have, I don't know if that has been raised. But I would be happy to talk to you about that, and see if we can --

DR. BANTHIN: AHRQ is not a -- we cannot collect funds from people we serve. We are prevented by law from doing that. But we can look into better computer screens. We did buy a new server, a new high speed server recently for the data center.

DR. HORNBROOK: I was wondering whether in today's world, with the emphasis on loss of jobs and job loss and employers' decisions about where to make jobs, in foreign countries or in the States, how important is it for us to start thinking about modeling the employers' behavior under mandates versus under totally being free of any employment related health insurance that goes into a different kind of direct access through Medicare or Medicaid or some new system?

Beth indicated that the COMPARE model doesn't talk much about the decisions about jobs and loss of jobs. Do you have a sense of how important it would be to focus on the employer behavior in these models?

MR. GARRETT: Our model also doesn't endogenously have people losing their jobs. We take the distribution of individuals as we get it from the CPS. We will weight it differently to get to '09, according to what happened or what we believe happened, and farther out as well. But that is not having individuals in the model lose their job or not.

All the models have people able to lose their health insurance coverage should some other alternative become available, or lower cost alternative become available.

MS. MC GLYNN: Colleagues of ours, Nera Sute and Danny Goldman, have recently done an analysis which I think will be published in the next three years in health services research, whatever the time lag is these days, that looks at the effects on employment and productivity for industries related to the proportion of employees covered by health insurance. It does an analysis over two decades and does show an effect and does some comparisons to like industries in Canada to try to get some sense of what it would look like in a different policy framework.

So we have been talking to them about how we can take those kinds of results and import them back into our model over time. We are just at the early stages of talking about that. But it does look like that might be -- that is a way in general that we can bring things into the model to try to better understand some of those more macro effects.

At the outset we tried to limit what we thought was possible. The general feeling was that a lot of the macro stuff was harder to get your hands around in the context of the microsim. But we are continuing to look at how we can do that, particularly from where we started to today, the importance of that as a policy context has certainly increased dramatically. Now we are re-upping those efforts, plus we have gotten a lot of work under our belt so we can entertain those possibilities. So we will see how that goes.

DR. HORNBROOK: One of the things with national health insurance would be an auto rescue package.

MR. BAUMGARDNER: What we do, and I think I have to separate what is in the health insurance model I talked about today from things that are done to put in the HL settings.

We typically are dealing with a ten year budget window, so there is some baseline coverage by coverage type that we are starting with every year, and then you let the model run. How do we set that baseline insurance number, the number of people without insurance every year? We do go through a separate modeling exercise. Part of it is going on right now, where we do look at changes in the unemployment rate within CBO's macroeconomic forecasts. So some of the issues you are referring to, it starts off with CBO's macro forecast of the unemployment rate, and then feeds into what we think the baseline uninsurance will be, along with Census projections on changes in the population, and at some level ethnicity also feeds in as well as our projections based on the national health expenditures on premium growth. Both evidence in the literature and some stuff we have done in house to put together the population by demographics, and premium growth and unemployment rate changes feed in to get baseline uninsurance amount.

Most large scale proposals, by the time they could possibly be implemented, at least within CBO's macroeconomic forecast these tough times we are in are over by then, so you are back to a more normal rate of unemployment by the time of implementation.

MR. J. SCANLON: Could I ask, how would you approach modeling some of these relatively newer concepts at a conceptual level that were mentioned in the various campaigns currently in the health policy community? Such things as a national or regional health insurance connector or exchange, and then coupled with that some sort of a public plan option. Is it possible to -- have you done any thinking about what assumptions would you make, what would it look like, how would you model those parameters?

MS. MC GLYNN: We have a national insurance exchange as part of what we have already modeled.

DR. GIROSI: This is something which is reasonable to do. It is within the scope of the simulation. Things in the pipe within the range of all the models that have been presented, I believe.

MR. J. SCANLON: You would just need specifications?

DR. GIROSI: Yes, specification work needs to be done, because that is the problem; sometimes we don't know exactly what does the plan look like. But that is certainly within the scope.

MS. MC GLYNN: Our basic approach to that is, generally when we talk to policy makers, getting detailed specifications is challenging, is that we basically have a checklist of what we need to know in order to model this option. We begin filling that in with what we have observed in other policy proposals or what people we talk to or what the literature suggests might be reasonable ways to think about it.

We do stress testing the limits of a design so that we think about things that should push the model to do either disastrous or wonderful magical things to try to understand how it behaves under different assumptions, and then have available -- and this is what we have done in all these policy options -- a fair amount of flexibility to mess with the parameters, I'm sure there is a better word than that, do permutations of the combinations of the parameters in order to come up with a sense of how that option would behave under lots of different design scenarios. So that is the basic approach.

Then a lot of times, that starts a productive discussion about whether these things make sense once you have seen these results. We have had this experience, you show policy makers this is what it looks like when it is put together the way you were thinking, and wow, that isn't really what I was hoping, what if we either add this element or jury rig with that element. That is all stuff that we can do.

In fact, the design of what we put together was intended to allow people to engage in that sort of what-if exercise and get a better sense. I think in general what we know is that people's priors are not always right. Not that the modeling, to Jonathan's earlier point, is perfect, particularly the more out of nature we get with the modeling, but it begins helping have a discussion that is certainly better than making stuff up, which is the more normal approach.

DR. SCANLON: Nancy, you get the last question before the break.

DR. BREEN: This is more about the wish list. If we were to obtain additional data that would help with the modeling and have more evidence to work with to maybe make fewer assumptions.

I am wondering, I think all of you used either the SIPP or the CPS as your starting match file. I know that a lot of states are starting to collect data, and a lot of them are telephone surveys, and they don't have much identifying information associated with them. I wondered, is that a limitation.

So my question is, what identifying information do you need in order to be able to do the matches you need in order to model? Then a follow-up question to those is, if you were able to model at a more local level, would that allow you to bring more of these macro variables in and look at health disparities in a more meaningful way than you are able to at the national level, just by virtue of having better information and knowing what the political parameters are.

I was struck, Bo, when you said that you had a typical state insurance, and I thought, what could that possibly be?

So anyway, I just wanted to know about the information you need on the base file for matching, and then if you had more local data, would that allow for better macro modeling and modeling of health disparity outcomes.

MR. GARRETT: In terms of matching, it is an interesting question. When a state has a smaller survey of the residents, they may not have all the covariates you want on there. The CPS would have the variables you need. You could either learn what you can learn about coverage rates as the state sees it from that file and bring that information over, and make sure that what you use is related to reflect those assumptions. Or if you had enough variables you could do a separate matching, if you had a really detailed state level survey.

In terms of identifiers, given that we are not matching the same individual across these surveys, we are matching on the various covariates of age, health status, gender, a laundry list.

DR. BREEN: All you need is a rich set of covariates, then. You don't need any --

MR. GARRETT: When used as a rich set of covariates, ones that have real data. The more covariates you have, the better. So for example if we had had state available from the MEPS public use file that we used, we would have matched on state when we matched the CPS as a factor.

But in terms of a typical state, in our case the typical state would have, for the purpose of non-group payment, to answer your question, would have age rating. It would have rating on health status, and it would not be community rated. That is the assumption we make with the local state.

MS. BULLHEIMER: Could I just follow up? Do you match on race and ethnicity, to Walter's point? Is it possible to do that?

MR. GARRETT: In the MEPS? We do the match, yes. The short answer is no.

PARTICIPANT: Could you repeat the question?

MR. GARRETT: The question is, when we match observations from the CPS core file that we start with, to the MEPS, do we match on race and ethnicity. The answer is, we take race and ethnicity into account in the match. What we do is, rather than have hundreds or thousands of cells for various characteristics, we start with a few set of broad cells.

MS. TUREK: There are very different sized surveys, so how often is a MEPS person matched to the CPS?

MR. GARRETT: They are not the exact same people, and some people get reused. So a person who has a particular age and health status in the CPS is going to get some person with the same in the -- we define some very broad cells on the basis of gender, age and income. Within each of those cells, we have a very detailed regression of health expenditures that we estimate on MEPS data.

From that, one gets an expected value, an expected expense. That is like a propensity score, but it is ultimately more of a continuous measure. Then we match -- within these broad cells, we match people who have similar expected costs. Then arriving at those expected costs, we use race and ethnicity as factors, as well as employment and a host of other things.

We do a fairly similar thing. Basically we started with a wish list of covariates along which we want to match, age and health status and whether the person works, the region where they live, race, ethnicity, gender.

So we start from a longer wish list and then we divide the data in cells. But sometimes you can't find someone in the same sets, especially because the MEPS are kind of small. We are going to get two years of the MEPS.

So sometimes you can match along, take the race and ethnicity into account if you are in a bigger cell, if you start from a bigger cell. But then if you are looking at the non-group market, people aged 25 to 30, you are not going to be able to match on race and ethnicity, because you don't get enough people there. There is not too much information there.

MR. GARRETT: I should mention this algorithm that smartly computes cells based on a bunch of characteristics that will come out using some kind of artificial intelligence to come up with that. We haven't put that to use yet, but I want to give them credit for assisting and making me known about it. We expended it once, it was fairly comparable with what we had already, and we didn't pursue putting it in. But it is on the agenda.

DR. BANTHIN: Tom will talk about it when we are up next.

DR. SCANLON: Let me thank you all. John, thank you for hanging in with us on the phone.

DR. GRUBER: I'm sorry I couldn't participate more actively, but thanks for including me.

DR. SCANLON: That's okay, I understand. The phone is not the ideal situation, but we really appreciate your contribution. Thanks again.

DR. GRUBER: You bet.

DR. SCANLON: Debbie is going to warn everyone about the new security. We are going to take close to a 15-minute break. We will come back at five of 11 for the second panel. Thanks.

(Brief recess.)

Agenda Item: Panel 2: Modeling Health Insurance Data, Practice and Perspectives, Continued

DR. SCANLON: We had the dilemma in the first panel that we could have gone on a lot longer in terms of questions. Now we are going to add to that dilemma, I'm sure, through what we are going to hear.

Jessica, do you want to start?

DR. BANTHIN: Yes. Thank you very much. I will be talking with Tom Selden. We both work within the Division of Modeling and Simulation. We are a small group of six economists who work within the Center for Finance Access and Cost Trends at AHRQ.

We wear many hats here, unlike some of the other presenters. We are involved with collecting data, putting out public use files. My group in the Division of Modeling and Simulation, we are very much involved in developing augmented files and tools to help other modelers and simulators use the MEPS data. Then we also develop and maintain our own simulation models.

Because we build models ourselves, we are very much aware of many of the gaps, so we will talk about that in our presentation. We even have NISBE statistical tools that will help us a lot in filling some of these gaps, and these are available to others as well.

We are going to divide the presentation into two parts. We will first talk about the household data and our household based models and tools, and then we will talk about the MEPS insurance component, which is the employer survey, and some of our work there. Again, it is public use files, data products. We also do a lot of basic research to estimate some of the key parameters that are inputs to many of the models around this room.

The MEPS household survey is a one stop data source for many of the key elements in microsimulation models. At this point we have about 13,000 households, 35,000 persons. It is the non-institutionalized population from 1996. It is ongoing. We follow people for two years. We have monthly insurance coverage. We have many insurance coverage variables as the committee knows, I'm sure. We also have sections on employment offers, take-up access to care, expenditures, utilization.

What is missing here is, we have detailed measures of health status, health conditions. These are self reported data. They are valuable. Disability status, work loss days, things like that. Some of the things that are missing though that modelers, economists, would like to have. Although we have non-group premiums and employee contributions to premiums, we are missing the employer portion of the premium. That is really critical when you want to look at tax exclusions and tax subsidies. We also don't have premiums for those people who have turned down coverage. Economists like to know that because they want to know what someone has given up, what the price was when they gave it up.

Furthermore, except for one year, 1996, where we did collect policy booklets and we abstracted detailed information on the benefits, we have not repeated that data collection, so we are missing information on the generosity of the coverage or the adequacy of the coverage that people do have.

This is an issue of growing importance. With rising health care costs, employers are looking at all sorts of ways to cut back on the generosity of that coverage in ways that people don't notice or don't mind. We have very different policies out there in the market. There are catastrophic policies that cover the high end and nothing much up front, but there are a growing number of front end policies. This is like what Walmart provides, policies that cover some preventive care, regular care, but you are not necessarily going to be protected in the event of a catastrophic illness.

We do a lot of things to make the MEPS better to use for modeling and simulation purposes. We have detailed income data which we then -- we use the MBR TaxSim and we compute federal income tax simulations.

Very importantly, we align our data. We periodically reconcile the MEPS expenditure estimates with the national health accounts. We have done that twice, the last time for 2002. We create a file where we have augmented the MEPS and bumped it up so that it benchmarks to the NATA. We then use that file to project forward to 2016. Those files are available for downloading by request on our website. We are improving them. They will soon be available for downloading.

We also have developed imputed employer contributions from the MEPS IC. Those are not available publicly yet, but we have them. In addition to benchmarking our expenditures to NHEA, Tom has done some additional work to add in provider tax subsidies spending that is not tied to specific patient events, which is out of scope for MEPS, but in scope for a modeler who is trying to benchmark to national health accounts data for this population.

We have immigration citizenship status that is imputed based on some collected data, and then we impute documentation status, but only through '05. We have fully imputed jobs variables which are very useful. What is not here is, we have a monthly expenditure file that some simulators like. It is being used currently by people at Treasury. We have allocated the annual expenditures in the MEPS to the 12 months of the year.

Let me just go back and talk about the importance of reconciling our expenditure estimates to the NHEA. This is critical. I was involved in the 1993 efforts with the health care tax reform way back then, and I remember some of the very early estimates where there were three different groups providing preliminary estimates of what it would cost to reform the health care system. There were three wildly different results. We realized that we had all started at three very different places. We hadn't aligned our benchmarks.

Simulation modelers have come a long way since then, so we undergo this very time consuming detailed reconciliation of our expenditures to the national health accounts. We work very closely with our colleagues at CMS to take things out of the national health accounts as precisely as possible. These results are published in a paper.

Since then we have updated this. Tom has updated it. We have a working paper that is downloaded, Selden and Sing. You can download it from the Web, in which we have done additional work to make up for example for the sim tail in the MEPS. It seems likely we are not sampling really high expenditure cases. These very sick people may be attriting from the survey, so we have done a little statistical correction to make the tail look more like what we have seen in some commercial claims data.

Then taxes. We have a detailed income and assets section in the MEPS. This is an important part of our simulations, to look at the tax effects of any of these reforms.

Then I just want to mention, these are a selection of some of the papers. I included a handout with the full citations of selected papers, where we have done some work that estimates some of the -- we can make use of the MEPS because it is such a rich database to estimate some of these key elasticities and parameters that are key inputs into our models and other people's models.

Before I turn it over to Tom, we also look at burdens of health care. MEPS is particularly well suited for looking at financial out of pocket burdens when we have good measures of income and we have good measures of out of pocket spending. This is one of the strengths of the MEPS, and it is a key output that policy makers really do care about, who is going to be affected by these reforms, who wins, who loses. By modeling the effects of reforms within the MEPS, we can get very accurate estimates of changes in burden across subgroups, including minorities, although many of the key differences are along education, income and source of insurance sources. But this output measure is very important to our model.

DR. SELDEN: As Jessica mentioned, we are taking a two-pronged approach. We are trying to develop and enhance the data products to facilitate these modeling efforts that we have been hearing about. We also undertake our own modeling efforts.

One of the key areas in which we have been providing estimates for quite a while is through our KIDSIM model, where we have a set of spreadsheets with program parameters. I think it is up to about 120,000 cells now of details about all the different programs in all the different states to try to think about -- we merged this into MEPS to try to think about which kids are eligible for which programs for health insurance. We developed then estimates of take-up and we also look at the number of eligible uninsured children.

There are some differences among the different models. The CBO came down on our side on that. One of the advantages that we have is that we start off a lot closer in terms of the number of kids who are on Medicaid and public coverage in general. So it is a real advantage when you go try to figure out who is eligible and uninsured.

We use this to track progress of these programs over time, but also to think about alternative scenarios for expanding eligibility or reducing it. We have looked at the impacts that might have in terms of crowd-out and program take-up. We have also estimated the cost that expansions have. It is not just the budgetary costs, but there are also offsetting savings. If a child is crowded out, then you don't have to have the act subsidy for ESI. If a child would have been uninsured but at risk of falling down into a Medicaid spend-down, those costs are saved but they end up in the SCHIP budget. So there are offsetting costs there that we have tried to take a look at. We are currently updating those estimates. We are at 2007.

We have also expanded KIDSIM to include program eligibility for non-elderly adults, so that is a major enhancement of providing some estimates to CRS, by using this in the model.

One of the key aspects here is that we do have detailed information on employment and disability status as well as insurance and expenditure by source of payment. So we can triangulate to get at a very hard to model aspect of this, which is eligibility through disability.

We have in the past had a much more integrated model where we would look at specific health reform scenarios. We are not quite there. We are more looking at ways that we can provide data products and basic research that assists other programs, but we are also -- as we build more and more, we are getting closer and closer to having something like this in the future.

The big shift here in the talk is when we start talking about the MEPS IC. People have mentioned it before. You have got a basic problem. You have a household survey. You see a worker, and the worker's spouse potentially, but you don't see all the other coworkers at that firm or even at that establishment. If you have an establishment survey you see all the workers at the firm, but you don't even have any of their spouses.

You really have to have both pieces together. But it is very difficult to do that. You can start off with a household survey and go talk to people and their employers and find out about the employers, but then your sample response rate is still way down.

So we have developed a methodology for bringing things together, the HC and the IC. The IC is a really nice data source. It has got a very large sample. It has got a very good response rate. Right now I think it is primarily being used for modeling microsimulation efforts. It is primarily being used to some sample means maybe by firm size that people are using to calibrate their model. But there is so much more that can be done with that model add we are trying to do that.

A few years ago, an effort was made. That has lapsed, I won't go into that too much, but we are trying to resuscitate that by hopefully gaining access to the IC data again so that we can populate these establishments with MEPS HC workers.

One of the tools we use to do this, and it was really developed for other things, but we have adapted this tool a bit, and it is available to everybody, is this SASMAC we call HD call-up. It uses a lot of artificial intelligence to think about the problem of how you match two data sources together when you have thin cells, you have a very long wish list of dimensions on which you would like to match up, and you have different sorts of parameters that you want in terms of the minimum cell size and the rules of collapsing cells. You can specify this stuff as a user and have this cranked through and do your matching for you. That is available for anyone to use.

We use that here. We also use some raking plus stratification methods that try to do the very best job we can on aligning the two data sets on about 20 different variables. What we end up with, we think, is a pretty sophisticated model that then can at a minimum be used to produce the sorts of benchmarking estimates that people really feel are most necessary.

So if you are dissatisfied with the MEPS IC benchmarks that you are currently getting and you said, what I would really like is to look at people in this wage range who don't take up coverage but who are eligible. I want to know what their premiums are, what was the lowest price option that they faced that they turned down, and what does that distribution look like.

As soon as we have these two data sets merged together, we can provide that benchmark that you are interested in. So if you have wish lists that you would like to give us, we would love to hear them.

DR. BANTHIN: So we are in progress. We have many things in the works here, but these are our main models, tools and plans.

DR. SCANLON: Thank you very much. John?

MR. SHEILS: Good morning. It is a pleasure to be here. These meetings where modelers get together and talk are actually good therapy for all of us.

We have been using our model in various forms for about 22 years now. The first time we used it was with the bipartisan Commission on Universal Coverage, 1988. That was the Pepper Commission, if some people remember that. It was a great opportunity to do analyses of a few different alternative plans. A lot of the plans we are talking about now are the same ones.

We just finished a study, we did a study last fall of comparing the impact of President Obama's proposal and the impact of Senator McCain's proposal. In that we looked at the public plan and what happens under different parameters, different design issues.

With McCain, it was particularly challenging, because he took a very different direction. It had a logic to it that you needed to explore if you were going to do an objective job. You had to follow that logic and give it the greatest credibility you can. But it was a plan where in effect you would have an increase in medical underwriting. In states where they say you can't turn down somebody just because they have got a health condition, you can't charge them more just because they have a health condition, in his plan that would go away, and you would have an enormous high risk pool and all kinds of segmentation of risk and people shifting around.

You can't answer that question, what its effects will be, unless you model those flows. If you take the sick people out of the system, CREDAs go down for the people you c help, so you get some takeout there just because of the way you have broken it out.

So as I say, you have to think about those things pretty carefully.

We just finished something with the Commonwealth Fund, where we modeled ten different Congressional plans. We have some detailed tables in addition to that report, if anybody is interested in it.

We found that the key here is to get a baseline database that depicts the characteristics of the population, individuals and employers. We found that the best way to do that is through using the MEPS data. The MEPS data is interesting because it will give you a lot of things that can be relevant. It gives you month to month coverage data, it gives you a little bit of information on utilization of health services over the course of a year, you know when people have changed jobs.

There is also information on health conditions. That was very useful when we had to start figuring out how to model underwriting. We could identify those people who were going to be excluded. We could take them to the pool that they are creating. And because we also have information on their costs, we can look at what the pool's costs are going to be. We can also look at what happens to what remains in the individual market, people who are basically just healthy. That brought down the premium quite a bit, which is what you see in states with large high risk pools. The individual coverage actually turns out to be pretty low cost because they don't cover anybody who is sick.

So all of that information has just been so helpful to us over time. It allows us to generate month by month income data, too, which is very useful for modeling public assistance programs.

The model is set up to model Medicaid and SCHIP expansions, premium subsidy programs such as those proposed by McCain through a tax credit or those proposed by others through vouchers. We model the pay or play employer model. There are at least three different variations on pay or play that we feel have very different outcomes.

So every one of these proposals we look at, there is a great deal of review of the assumptions for appropriateness of what we are modeling. So it is important to remember that. There is individual coverage and so on. There are individual mandates that have some weird effects on employers, too.

Insurance market regulations. We have a model that will simulate -- and you heard a similar one described earlier -- will simulate the premiums for a group based upon the characteristics of their population, and the rules that are permitted in their state. There is age rating, for example; we will be able to model that explicitly with an employer database. We also model that for the individual market as well.

The databases we use here in HBSM start with the MEPS. That is the basic household data file. We take the data for four years and get a large sample. We also use CPS updates on coverage and population characteristics. This tends to come out more recently so you want to stay current as possible. Optics are more important than its impact on the actual results, but it is important. We also use SIPP data in many places to develop some behavioral parameters, participation rates in Medicaid and SCHIP, for example, is an example of what we will look at.

The model is controlled to an array of program data. We rely very heavily on data from the Office of the Actuary at CMS. There is once in history where we felt they were wrong, and thank God, in the end it turns out we were a little closer. You may remember that. We were doing the President's analysis. But we do control to that carefully.

We also go through this process to try to reconcile what is in the file with what is in the accounts, keeping close tabs on what those numbers are, because it will make a difference.

CBO projections are used where we can. CBO is not final because it is infallible, but it is infallible because it is final. So we go with their conclusions wherever possible.

We have a synthetic firm portion of the model. You know how certain poisonous frogs are a bright color, to warn people not to go near them? That is why we use the word synthetic, so you understand we are going into something that we have to concoct from the available data out there.

We decided early on that it is very important to have a file that gives you information on the employer, and then the characteristics and health utilization features of everyone who is working in that firm plus their dependents. To do that right, we feel you have to have a database that accounts for a great many things, such as part time/full time status, wage level certainly. There are a couple of other items such as how many people in the firm have coverage, how many are eligible but turn it down, how many are ineligible. So we go through a rather elaborate process of doing that, which leads us on a strange trail here.

Our basic data on health plans is the 2006 Kaiser HREP data, which others are using. But we match that with data from the 1997 RWJ survey of employers. That is important, because it gives us a list of information on the characteristics of the workers in the firm.

So for example, you could have a firm where most of the workers are low wage females who work part time. When we match together and construct that, those are the kinds of people who will be attached to that firm. If you have a plan that treats part time workers differently, it will enable you to take that into account, because for some individual firms that will have a profound effect. At least, that is what we find with our numbers.

When we model proposals dealing with medical underwriting, with this match we have the information on not just firm, but we have the information on their health expenditures. That health expenditure information and their condition data can be used to model health status rating where it is required in the small compartment.

So we spend a lot of time doing that. We match between -- the RWJF data is quite old. Its health plan data is probably just short of useless because it is so old, but we match the plans on the basis of just ranking of these health plans by actuarial value. So the firms with good benefits get assigned to the HEET firms that have good medical and so on.

I think I have already explained this. We use the METS data because it gives us a database that has a great deal of detail in it. We don't have to worry so much. If you don't have information on health conditions you have to impute it to your file. You always have to worry about, did I impute it the right way, did what I do make sense. We think it is more likely that it will make sense if it comes from a database that contains its own unique joint distributions.

Another thing about the MEPS that we like is that it is conducive to a month to month simulation methodology. This might be a distinction between ours and some other peoples' plans. We set the data up such that we have data for 12 months. For every individual we have their coverage in each of the 12 months, thanks to the MEPS survey where they collect things.

We know about their employment. We can take their income and say their earnings must have been distributed over these weeks where they worked and the unemployment insurance while they were employed and so on. We are able to construct a month by month accounting of their income and their employment information. You can line that up with their coverage information, and that coverage information is lined up with date of service for the services that they use. So it is pretty powerful.

It is an enormous amount of housekeeping. It is no fun for the programmers certainly, but we feel we can be a little bit more comfortable with it in that way.

The month by month simulation feature is very important as well from a policy perspective. Earlier I think John was mentioning this. A lot of these proposals that are coming up have many moving parts. Some affect employer coverage and some affect Medicaid or the coverage for non-group.

If you are a person, and many people in the file are like this, where you have somebody who is in for six months in Medicaid, they get a job, they perhaps pick up coverage for an employer for the next six months or maybe work for an employer and don't take coverage, either way, it is interesting and it is important, because those policies will affect what happens during those two periods differently. Understanding how people are distributed across those snippets of time will make a difference, part your eligibility, part your costs, are important.

As I say, we merged the data for 2005.

Since you all are interested in data, I thought I would do something interesting here and present some data. I am the only one who did this at this point, so I hope I didn't miss the mark too much.

This is a basic breakdown of some of the issues we face with just the simple question of how many people are uninsured. Here are the data files we worked with.

With the MEPS 2005-2006, we get the first column of numbers, of 50 million people uninsured at a point in time, 37 million uninsured a full 12 months, 68 million sometime over the year, and any time in 24 months we can do that as well for MEPS and get about 83 million people.

Now, some of the other databases are a little more difficult to work with. The CPS for example, if you take the questions literally, and the questions are asked literally, what it tells you is the number of people who didn't have coverage from any source over 12 months, over a year. It is an all year number.

The problem is there is a great deal of under reporting of Medicaid coverage in the CPS. If we like CPS because it gives us some state level data, this is an important thing to hang onto. So we correct for under reporting, and we do this by this month to month simulation approach, distributed income across the months of work, earnings across the months of work, weeks of work, months, and the various income sources. We construct month by month income levels. It is very much like what they do in the TRIM model.

From there, we go one further step. We take the coverage from employment and spread it over weeks of employment, of course. So you get a synthetic representation of what the distribution of coverage was for each individual in the file for each of 12 months. We correct for under reporting controlled to the average number of people who are supposed to be on that program, who were ever on the program.

The third column shows us what we wind up with. We get the number of people uninsured all year. We get 34 million, which reflects the under count in the CPS of Medicaid. The CPS gives us 47 million at a point in time, 62 million at any time in the year.

The difference between that and MEPS is interesting, because I think the MEPS under reports coverage a little bit for Medicaid, and we do make an adjustment for that. So once you get through doing adjustments, you get some different perspectives.

We took the SIPP data and used that. We really got stuck on this for awhile. These are the numbers for the 2004-2006 period, the 2004 panel. We have done some reweighting to try and replicate the figures. We get a very different set of answers. We only get 30 million uninsured all year, about the same number of uninsured at any point in time, but only 30 million all year, only 58 million any time in the year, and only 74 million any time over 24 months. Either the SIPP is right or these guys are wrong.

This chart shows what happens in the 2001 panel of SIPP and the 2004 panel of SIPP. Same thing; average monthly of uninsured in 2001 is 43 million, but in 2004 we are getting 45 million. That makes sense; the number of uninsured we know did increase over that period. And we get an increase in the number of people who are uninsured all year; it goes from 24 million to 30 million. That is rather a big jump, but it is in the right direction.

But then sometime in the year we are getting less now in 2004 than in 2001. We are running the data through the state programs saying this and that, so it is a red flag for us.

It also shows the number of people uninsured sometime in 24 months, but down from 79 to 74 million people, and that is in the wrong direction, we believe. There is a great deal of variability in the estimates it gives you on a month to month basis.

We are concerned about the 24 month business, because we do see the MEPS going in the opposite direction. The number of uninsured for 24 months has gone up steadily; MEPS shows that it is going down.

I want to go back to this one thing here. Going back to the CPS table, a lot of us know, many people have said, that it is not an estimate of the number of people uninsured all year. It is really an estimate of the number of people who are uninsured at that point in time of the survey.

A lot of people feel that fairly strongly. I think that is wrong, but we went through this process of saying we are going to figure out what it really means and correct for under reporting and try to do everything right. What we wound up with was an estimate of the number of people who are uninsured at a point in time, 47 million people.

Well, people have been saying that CPS says there are 46.5 million, almost the same number who are uninsured at a point in time. Do you remember Emily Latella from Saturday Night Live, Gilda Radnor? She would get all upset about this issue, somebody out there, it is not a problem. She turns and goes, oh, never mind. That is basically where I am going to end.

DR. SCANLON: Thank you.

MR. SHEILS: Actually there are a couple of suggestions of things we would like. It would be great to have the state identifiers for MEPS. It is very unsporting that you have them and we don't.

I would like to see the SIPP data longitudinally stabilized. It is nerve wracking. Program cost data by eligibility group is important for getting some of our costs so we can get a little better data from the Medicaid program in particular. It would be interesting if at some point we would get micro data on employers from MEPS. I think you have already heard about that.

MS. TUREK: This is a joint presentation. I am the government project officer, and Linda is the project officer at the Urban Institute. We did it slightly different, because we were told 15 minutes. So we figured we would give you a fast introduction to TRIM, talk about what we do in the health area, and then if there is time we can give you a lot more detail.

TRIM is a modeling system. We model pretty much all of the major federal tax and transfer programs. We apply rules at the individual record level, basically the program rules, to see who is eligible, and then we talk about who participates.

We do this because one of the big questions in earlier years was, this person gets this many dollars from Medicaid, that many dollars from welfare; how much money does this person or family get in total? And because the program rules vary from program to program, the only way that you could get it was through this kind of system where you modeled the individual and then could add up the family or household level.

We primarily use the CPS, but there are a lot of databases. We statistical match in the HIS and pretty much all the other databases everybody else uses.

CPS is in the process now of adding point in time health insurance questions based on a bunch of research funded by Census, ASPE and SHAGAC, which matched the surveys to Medicaid statistic information system records. That would be a good benefit.

TRIM has been around almost as long as I have. I have been involved with it since 1976 either as the project officer or when I was the division director managing the project officer.

Urban Institute is our current contractor. We have it set up so that for each TRIM product we have a technical representative in house that does the review. So the health models are reviewed by Tom Musco and Don Cox because they know what the programs are about.

It is a public use model. It is a client server system that you can get to over the Web. We give you a complete detailed description of how everything that is done in the model for every program. We give you our databases annually, I think going back to '93, and you can get them back to 1980 if we use the CVs. We give you the program rules year by year and you can actually, if you can figure out how to do it, go out and run your own simulations on current parameters.

We had once wanted to fix it so that you could change the program parameters. They had somebody who developed a straight language, but unfortunately we ran out of money. To get out there and model, we don't give Urban any money to help you, so you have to be pretty good at guessing what to do. Budgets are tight

The programs we model are health, Medicaid SCHIP and employer sponsored insurance, the cash income programs, non-cash benefits and the taxes, and medical tax credits would be included there.

What its main purposes are, are to identify people who are eligible for means tested programs. As people have pointed out, the CPS under reports many of them, and so we run people one by one through the eligibility rules from the programs. We do this state by state. We can do all of our estimates state by state, and we come up with a list of eligibles. Then we impute participation. Then once we get a baseline that looks a lot like the world is today, where we have each program matching what the administrative data says or close thereto -- there is never quite an exact match because the administrative data will include people we don't. Medicaid would have the people in long term care institutions. But we use the baseline and then we do what-ifs, what if we changed this, how does it change the world from today.

We also go for quality review. Right now the Urban Institute has -- outside of our TRIM team we have technical experts who review all deliverables. I think you are one of them. We also had outside reviews of the model. John Treika reviewed it a couple of years ago. They looked at the Medicaid and SCHIP modeling, and we did in the late '80s pay the National Academies of Science an enormous amount of money to do a complete and entire review of the model. A lot of what they said we could afford has been implemented.

I am going to turn it over to Linda, who will tell you about some of the work we have done with the health models.

MS. GIANNARELLI: Thank you. I am just going to talk about the aspects of TRIM that relate to today's topic, so don't worry, I'm not going to start talking about the SSI model.

I want to talk about three different pieces of TRIM. The first, what is our Medicaid and SCHIP modeling, the second, our employer sponsored insurance modeling. This wasn't specifically on our list of bullet points, but since matching has come up quite a bit in today's discussion, I also want to mention our statistical match with the NHIS data.

First, our Medicaid and SCHIP modeling. As Joan mentioned, we simulate eligibility and we simulate enrollment and we can do what-if modeling. In terms of simulating eligibility, where all the other models were looking person by person, household by household, person by person for everybody in the CPS data. Like the Lewin model, we do this on a month by month basis. We also feel that modeling eligibility month by month is very important because we do find a lot of people who appear to be eligible for part of the year and not the entire year.

We use the same kinds of things that the Lewin model does as far as how do we go about allocating annual income across the months of the year. I think sometimes people think, all you have in the CPS is annual income, so you are really making that up, but there is a great deal more in the CPS than maybe someone would initially think of. You know for instance how many weeks people have worked. You know how many spells of unemployment they have had. You know the number of weeks of unemployment.

So there actually is quite a bit to go on. It is not just making it up, although to continue with the frog analogy, it is a somewhat brightly colored frog, but not a very brightly colored frog.

We do the eligibility modeling month by month, and we capture all the detailed state specific rules for each type of eligibility, as you see on the list.

These are a few of our eligibility results, just to give you a sense. Obviously we can cut this a lot of more detailed ways, but we like the Lewin model can look at eligibility in the average month of the year or we can look at eligibility on an annual basis, is someone ever eligible for Medicaid or SCHIP in any month of the calendar year.

This chart is looking at eligibility by main reason for eligibility, by user group. This particular slide is dividing it by mandatory due to SSI, other mandatory SCHIP and some other optional categories, and medically needy, and finding a total of 59 million people eligible for Medicaid or SCHIP in at least one month of the year, and these estimates are for 2003.

After we do eligibility we also model enrollment. Again picking up on the Lewin presentation, we are trying to correct for the under reporting that is in the CPS data, not only for Medicaid and SCHIP, but for all the social programs. In fact, for Medicaid and SCHIP it is somewhat less bad than it is for TANF and some other programs.

So basically we are taking the people who are eligible for Medicaid and who report that they are enrolled in Medicaid or SCHIP, but then we are taking a subset of the eligible non-reporters to augment that caseload in order to come acceptably close to size and characteristics of the actual caseload. As far as that actual caseload, we are getting that from the EMSIS data. We are not using the EMSIS data unadjusted; we are starting from the EMSIS data but we are subtracting the institutionalized who wouldn't be in the CPS data, adjusting for cross state duplication, making some other adjustments in order to arrive at our final targets. For SCHIP we are using targets from the SEDS data.

That gives us a baseline that can be used for a variety of different purposes. One can compare eligibles to enrollment to estimate take-up rates. One can examine the characteristics of eligible individuals, decompose changes from the prior year. So for example, to what extent does it look like changes in eligibility are due to changes in economic circumstances versus changes in program rules. One can do estimates of uninsured people and eligible uninsured people. One interesting use is to be able to do alternative poverty analyses, for instance the National Academies of Sciences poverty proposals have different thresholds depending on peoples' insurance status, either uninsured, publicly insured or privately insured. So by correcting for the under reporting of Medicaid and SCHIP it allows one to implement those NAS poverty definitions a little bit more accurately than using the unadjusted CPS data.

Some examples of Medicaid and SCHIP alternative policy simulations. One can change one or more specific eligibility rules. One can play a what-if game, for instance, what if all states have the eligibility rules of a particular state. One can play other what-if games of applying rules from a prior year. You can leave eligibility unchanged and modify take-up. Playing off of TRIM's comprehensive nature, one can look at cross program interactions. You can say what if there were a particular change in SSI, for instance something like eliminating the 209B rules so that everyone in SSI was automatically eligible, what would that do to Medicaid enrollment. So you can look at those kinds of interactions.

In employer sponsored insurance, what the model does is impute premiums. The Kaiser HRET data was mentioned earlier, and that is also what we are doing for matching in premiums. We are matching in both the employer and the employee premiums. That is what we are doing for the private. You can see on the slide the data sources for the federal employees. So we are doing that match in order to bring into our data the premiums as well as the plan type of a particular person's imputed coverage.

There is also an element of that model that imputes whether or not a person has been offered health insurance, whether they are in a firm that offers to some people but just not them individually, or whether they personally received an offer but they chose not to take it. So that imputation is available for potential alternative simulations.

The ESI model can be used for a couple of different kinds of uses. You can analyze insurance status combined with whether or not insurance was offered or accepted. You can do some what-if analyses, what if more individuals accepted the offers. The imputed premiums are available for modeling tax credits as part of TRIM's federal income tax modeling.

I don't have a slide for the NHIS match, but we do a statistical match with the NHIS data every year. We bring in the ADL and IDL information and we also bring in reasons for limitations. So for people who have activity limitations we are bringing in the variables that indicate why they are having those activity limitations. We also bring in through the NHIS match the private non-group premiums. We used to have a different way of doing that, but now that we are doing that NHIS match, we are getting those premiums from the NHIS data. That information again is available for use in alternative simulations.

Interestingly, the CPS pretty soon is going to start having some limitation data directly in the CPS, so that aspect of the match will no longer be necessary because we will actually have real data on that in the CPS. So that will be great.

Some capabilities in TRIM that were used for 1990's health reform analysis. We had in the 1990s some elements maintained in TRIM that allowed us to look at changes in employer provision of ESI, look at the employee decision between public and private coverage, and look at a broader range of health spending. Those elements of the model have not been kept up to date.

Some ongoing and planned work. Something that picks up on a couple of the comments made by other presenters is the modeling of eligibility and the problems with under reporting in various surveys.

One thing we are doing right now under the TRIM contract is, we are looking at that matched data file that matched to the CPS with the EMSIS data. We are basically looking at people who we simulate to be eligible for Medicaid and/or enrolled in Medicaid, are we finding them in the EMSIS data, and conversely for because

who are enrolled in Medicaid according to the EMSIS data, do we find them to be eligible. We are analyzing those results right now, and I think that is going to be extremely helpful in terms of improving both our eligibility and our enrollment modeling.

We are finishing up our 2006 baselines right now. There is a task planned on the agenda for fiscal 2010 to bring some work by Rich Johnson on long term care insurance, to bring that work into the TRIM framework.

Joan mentioned a lot of the ASPE uses of TRIM, so I won't go into this too much. There have been a lot of ASPE uses of TRIM in the health area. The second bullet point there, looking at the effects of a non-group health insurance task credit, and lots of different kinds of baseline analyses. Looking at eligibility by AGI level and by immigrant status. Immigrant status is one thing that is imputed into the CPS data as part of our pre-processing. At various points we have done some very detailed tables of eligibility cut different ways that has then been provided to folks doing detailed modeling at Treasury so that they could feed those pieces of information into their modeling.

Lots of different non-ASPE uses of TRIM. CBO has used TRIM, has used our longitudinal series of eligibility and enrollment flags from TRIM. CRS has used TRIM to look at the impact of Medicaid and SCHIP restrictions on benefits to non-citizens. I will let you scan the other bullet points.

I think some of the strengths of TRIM as far as health, its ability to do detailed state specific eligibility modeling, the ability to look at cross program interactions. One that I didn't think to put on this slide is the ability to model a package of changes, not just to one program, but to look at a package of changes and to do distributional analyses. For instance, if there is a legislative package that is including changes not only in let's say Medicaid eligibility but also taxes, also food stamps, to be able to model all of that in one internally consistent framework, and then to look at distributional impacts and winners and losers and other kinds of impacts of the entire package of changes at the same time.

MS. TUREK: And you can easily look at minorities, because we have that data on the CPS.

MS. GIANNARELLI: Right, you can cut results by whatever variable is available in the CPS.

As Joan mentioned, the model is available online. Trim.urban.org is the website. Registered data users can extract variables, so you can extract eligibility flags once they have been made public use and put on the server. If you request simulation access, then you can run simulations.

There is a not so pretty picture of Medicaid documentation available online so you can read the documentation of the various modules online. This is just a picture of looking at a particular household through the micro data viewer online.

DR. TUREK: This is basically what we can do. This shows you the steps that TRIM goes through when it is doing the simulation. I think the ability to look at immigrants is very important because there will be programs that do not relate to illegal's or will depend on when the immigrant arrived and their status. We incorporated that capability quite awhile ago. We used them to do a lot of analysis of immigrant programs.

Although the last charts are not very pretty, all you have got to do is go to Trim.urban.org and you will get these charts. It is very comprehensive documentation that explains in great detail how each model is developed. We believe that models used for policy analysis need to be transparent. So if somebody else wants to reproduce your results, they have the information they need to do it.

We were fortunate when we went from a mainframe computer to a client server system that my boss was a systems programmer. So TRIM was built not only by the analyst, but by computer professionals who established it to standards in the industry for computer programs. It is a C++ model, and it is updated daily pretty much. But I like it. It is kind of fun. We are always adding something new to it.

DR. SCANLON: Thank you very much. Let's open it up for some questions before we have the committee discussion.

DR. SUAREZ: This is a question for Art and maybe the other members of the panel can join in as well.

We heard this morning different models built on utility based approaches, model built on elasticities based approaches, models that use CPS or the infrastructure of data and some that use SIPP. I was wondering if there has been an evaluation of the various models in a way that shows where the different outputs are and why those differences are.

An example. This is an evaluation of the various concepts using a common policy. So we have at one end a policy situation and you drive the policy through the various models and come up with all these results. Where are the differences in those results, and why are those differences happening? So that at the end of the day maybe we get a better understanding of where the research issues are.

DR. TUREK: The National Academies study did look across models. I don't really know of anything since then that has taken that kind of a comprehensive look.

DR. BANTHIN: Linda Bellheimer can speak to this because she was the project officer in an extensive comparison funded by Robert Wood Johnson.

DR. BELLHEIMER: Yes, we actually did this with Dr. Jerry Glied at Columbia.

DR. TUREK: When?

DR. BELLHEIMER: It was about five years ago. We took three of the best known national models, all of which you have heard from today, three of the seven that you have heard about today were involved in this.

We started with a meeting in which we brought all of the microsimulation modelers into a room, and we spent a whole day going through assumptions that underlie models. Sherry did a wonderful job of getting agreement on, these are the parameter assumptions on which there is agreement. These are the assumptions on which there is not agreement. This is a gray area in the middle where people are going to adapt depending upon the issue.

So we had that as the starting point. Then we designed a fairly plain vanilla health reform with some public coverage expansion, some tax subsidies, and we asked three different groups to model it, and to document where their assumptions did not follow the standard set of assumptions on which we had agreed to, where they were in the gray areas or where they firmly had their own assumptions.

It was a very interesting exercise. The three models did not come out with the same sine on the impact on coverage. When we peeled back the layers of the onion to find out why, there were two things that drove this, and you have heard a lot about both of them today. One was the assumption about the employer's response. The behavioral response of employers was critical in the assumptions. That was where there had been a lot of difference in the meeting that we had talked about, about where people were.

The other was about premiums in the non-group market. I think there was a fundamental difference between models in which premiums were established endogenously within the model and models in which premiums were imported from market studies.

We subsequently had another meeting in which because of the interest in the premiums in the non-group market, we brought a whole lot of members of the insurance industry in to talk about this. They were very perplexed by the notion of endogenous determination of premiums and said, this is not the way we think at all.

I think it was a very interesting meeting. We talked all day. I think probably by the end of the day there was a sense, over time, yes, what you are saying is happening in your models would probably work its way out in the marketplace, but if you are looking for 2009, 2010, 2011, that this would be the response that you would see in the marketplace, probably not.

But I think what we learned from this was that there is an enormous amount to be learned from having multiple modeling groups modeling. We learned more about the things that you really need to be sensitive to in health reform by looking at what grows the differences than we would any other way.

So I think thinking about transparency and for the committee to be thinking about how you get multiple modelers to look at a proposal and then share the reason for the differences is probably going to tell us more about health reform than any one model looking at it alone.

DR. SUAREZ: It sounds like there is an opportunity for repeating that experience, but maybe even with a larger group of models.

DR. BELLHEIMER: Yes, I think it would be an excellent thing to do.

DR. SUAREZ: Thank you.

MR. SHEILS: Once you have done an analysis of the plan, it is good to try and reconcile them with the other guy's numbers if you can. We have done that in the past and have learned a lot. I think the other guys learned something too, but we learned an awful lot about it.

We do have this joke at work, though. If somebody else gets a modeling result that looks like ours, we declare that they must be making the same mistakes.

DR. GREEN: I would like to explore with any of the panelists two quick areas pretty quickly here. Before I do, an unanticipated consequence of the hearing this morning for me as a member of the committee is to be reminded how unnecessarily complex the system is. The take-home message for me is, the models would be a lot better.

The first area has to do with something that I believe Linda brought up. It has to do with where politics and policy often mix, and that is in geography. You mentioned geospatial data; it is my impression that people working to take full advantage of these investments in public data and national data sources spend a great deal of money geocoding information that is in them. I am wondering about your opinions about the sufficiency or the strategy that we should be contemplating about these data sources arriving geocoded before distribution versus, it depends what we are going to use them for, and everybody just needs to plan on that being a cost of business.

DR. SELDEN: We have files available for people to use in house that go to quite narrow detail on where people are located. But the problem is that we cannot even release a file to you here to use outside the data center that has state on it, because the sample sizes are so small. We are talking about medical conditions and specific procedures that are done and total amounts of money spent, and the timing of the event. It is very easy to use that information with state identifiers to violate confidentiality. So we just have real limitations there.

DR. BANTHIN: Let me just add, we do geocode the data in MEPS. You can access it through the data center. You have to submit a proposal. If you bring in other data like location of hospitals, we will match it and then give you a file back. Sometimes you have to use it within the center if you are going to continue to use the identifiers, but if not you can take it away.

DR. GREEN: That set up my second area. I wanted to ask you your opinions about privacy considerations that restricted the use of the data to answer urgent policy issues.

MR. SHEILS: I think it would be a disaster if people somehow got the idea, I mean the public got the idea that their material won't be treated confidentially. Reporting in all these surveys could disintegrate, sort of like our economy disintegrated. That I think is the worst fear.

But I have got to say, I would love to see the state code on MEPS. It would be interesting to see if there is some way to -- I guess you could do that by eliminating some of the other variables, but it would have to be for all of the files.

Maybe there is some way to prioritize on what types of geographic disclosure is preferable. Urban-rural is important to a lot of researchers, but state code might be important to another group of researchers. You can't release the files both ways. You have to pick one or the other, or else you defeat the purpose of doing it.

So it would be interesting if there is some forum for figuring out what the prioritization should be of those things for research purposes.

DR. SCANLON: Let me follow up on this with reference to the IOM report that came out at the beginning of the month. Even though it is not dealing with survey data per se, I think it has applicability to the issue of survey data.

This was about HIPAA and privacy rules and how that impacts on research. They came up with three sets of recommendations, two of which in some respects are what are going to be the governmental rules. The third one though referred to researchers. It said, the researchers need to think about how they can increase the security for data that they are using, and in some respects how they can make the public confident that they have increased the security.

I guess I find this interesting, because it is saying that there is going to be an obligation here on both sides. If we are thinking about changing the rules by which data that are potentially identifiable would be available, what is going to be the quid pro quo.

The reason this becomes particularly important is because we have had hearings before for the full committee and for the work group where the premise that all data are going to become identifiable at some point in time, that there is not going to be any core that you can't match to some other data set and through that identify an individual. So the public use files are becoming leaner and leaner under the current rules.

So the idea would be, if we have to relax those rules, how do we do it with confidence. So both panels, the question is, what can we say in terms of response to this kind of recommendation that is going to convince the public that we can think about changing rules for disclosure?

DR. TUREK: I think if you are talking about the kind of policy analysis that is being done in response to program changes, you also have to think about timeliness. If it takes you six months to a year to get the data through some kind of long research agreement, by the time you get it you won't need it anymore. So I think timeliness is critical.

DR. BANTHIN: It seems to me that one ought to be able to come up with some kind of certification standards, and that there ought to be -- we are required internally to go through fairly strong steps to protect data. We actually collect a lot of primary data, so we pay a lot of attention to security of those data, because we are also a defense contractor. It is hyper ridiculous actually, but it is very secure, and there are very strong standards there. There are people in charge of that. There are plans in place. Those get audited and checked. There are limitations on who can have their hands on what data. There are significant consequences to employees who violate any aspect of that.

So I would wonder whether it would be possible to create some certification program and allow people to become certified because they have a program in place that allows them to handle those data in a secure manner, and somehow cut through some of the processes that are in place today that are aimed at protecting privacy, but that get in the way of nimbleness or utility of the data for policy research, particularly policy research that is trying to be responsive as things are under discussion. So I offer that as one possibility.

MR. SHEILS: One thought too is that a lot of these services are done year after year after year. We could I suppose do a regional identifier for a couple of years, and then when you get the new samples of people, this time you use state and don't put out the rural.

The researchers will always have access to a data file that has a geocode. It may be two or three years old, but at least you will have some degree of access.

DR. SUAREZ: Just a quick follow-up on that. I am a member of the Privacy and Security Subcommittee of the NCVHS. Thank you for that recommendation. I think that is an excellent concept, the development of a certification program.

But for purposes of the discussion, all the databases that I have heard of, I'm not sure they fall under HIPAA, in the following sense. I don't think AHRQ is a covered entity and MEPS is not a HIPAA covered data source. The reason why MEPS cannot release state level data comes from other sources, not HIPAA.

So there are issues beyond HIPAA with respect to the release of identifiable information on public databases or databases held by government agencies.

DR. BANTHIN: We are sensitive to this issue, because we do as AHRQ employees have access to all these variables, and we don't like to be called as having unfair advantage, unsporting. So we do try to make efforts to release stuff. But it is hard. We have not -- our agency or center has not devoted creative energy to thinking about this. We are limited in our funds, so we have this data center, we now have agreements with the Census so that people don't have to come to Rockville. They can go anywhere Census has a data center and use our data.

The MEPS IC, the employer data, that is the most difficult to get of all. It has everything to do with the IRS. They approve all those proposals. We don't even have access to these data, and AHRQ has funded the collection of the data. We will wait seven to eight months, and I am crossing my fingers, to get our proposal approved. IRS reviews those. They don't want anybody to think that your tax records are being used irresponsibly.

DR. SCANLON: Walter, you were absolutely right, in terms of -- this is not HIPAA data. One of the things about thinking about research and privacy is that we need to think about across the board what are all the rules, how do we harmonize them in a way that we are both protective of the privacy interests that we care about as well as promoting the research. It is thinking very broadly in this process.

DR. BREEN: Another thing that Tom referred to briefly but I think we need to be mindful of is that a lot of these data sets are just too small. The sample size is just too small to be able to look at regional or particularly state or even large metropolitan areas. So I think we really need to think about investing more resources in these data sets.

I just want to point out the MEPS, which is very efficient, it is done on top of the National Health Interview Survey, but it is constrained by the sample size of the National Health Interview Survey. The resources used for that survey have diminished over the last ten, 15 years. So that needs to be corrected as well.

Then I think there could be more centralization among the data collection as well. You have got a number of surveys, some of which are redundant, and those could be explored more and the resources could be used better. But I think we need more resources also on some of these key surveys which are used to build for modeling purposes.

DR. SCANLON: We did try to address that yesterday with a letter to that effect.

MR. GARRETT: I believe there are people like in the LAHD program who have tried to take confidential data with the right access to that and blur it, but preserve as many characteristics as you can. I don't know how good this can be done or not, but there might be a blurred real data set that is far superior to matched data sets.

DR. BELLHEIMER: We are blurring data. In our linked databases where we link our surveys through administrative data, clearly once you have got a survey that is linked to Medicare or to Social Security, the chances of identification become much higher. There is a great demand for public use files from these data.

We have started what we call data perturbation, which is an evolving science, which involves perturbing the data very slightly and maintaining the internal structure of the variance/covariance matrices.

We have put out our first perturbed data files. There is a very recent article in the American Journal of Epidemiology about the methods that are being used in this. Some researchers feel very comfortable, and we have on the website a table which shows you the variance/covariance from the perturbed data to see how close they are. Some are comfortable using those, others uncomfortable and want to use the original data.

So what we suggest is that they use the public use files to go on the website, design their analyses, and then send their final programs to us and we will run them on the original data so you feel more comfortable. So they get a tool quickly that they can use, and they can develop their programs, and then they can get a fast turnaround on the actual data if they don't want to use the perturbed data.

But I predict that over time, researchers will start feeling more comfortable just using the blurred data and not wanting necessarily to get the original data.

DR. GREEN: So building off of that, the committee members will correct this, but I think we folks on the committee are struggling to contend with our hopes that the population and individuals in the population reap benefits with these investments of public money into developing data sets and organizing them so that huge policy decisions that affect everybody can be well informed.

With that hope on one side of our discussion, then we have these fears on the other side that we will do harm, and that in the process of doing this, we subject individuals and the population to risk for harm. This is something that will not go away. We can title the hearing whatever we want to, and the issue is resources, it seems to me.

So I would welcome the panelists' suggestions as to what are the next steps. Is there an urgent need to systematically undergird progress on the nation's health statistics system if we only could reframe it, get the right people in the room for a day. But again, maybe that is overstating it, but I feel like we are struggling to think fresh and anew about this in ways that really get the benefit. Meanwhile, we have precious few examples of where harm has been done.

PARTICIPANT: By researchers.

DR. GREEN: By researchers, right, by investigators. So if you have got ideas, they would be well received.

DR. TUREK: I wanted to say something about synthetic data. It is very important that the covariates that relate to the programs you are interested in don't change much. The one synthetic data that I saw greatly changed the covariate on the disabled who are a major group getting federal programs.

So I think I am going to be looking, because I wouldn't have used synthetic data. I would have gotten enormously different results. So the question is, can you do it in such a way that the data that people use to estimate the impact on public programs doesn't change; which variables do you select to look at.

DR. STEINWACHS: I have a question, but there was an oversight. Dale pointed out to me that we need to give Linda Bellheimer credit for helping identify all of you and getting you here to this panel. Also, Roshita Dorsey was not in the room when I gave her thanks for doing this, too. Dale keeps telling me he just fronts for the operation, but he is a very good front person, I think.

I would like to go back to the discussion earlier about the health outcomes, because ultimately if you are asking the nation to invest a lot of money in health insurance, and that is always what comes back in the messages on this, is the so-what question. Some of the so-whats are, we cut administrative cost to other things, but the other is the quality and outcome equation.

The point was made before that to do it well, we need some longitudinal data sets to be able to observe that pattern of insurance, uninsurance over time, and recognizing that that is not going to be here for anything in the near term other than what you have taken advantage of already.

There is a lot of cross sectional data that talks particularly around associations like potentially preventable admissions or ambulatory care sensitive admissions. You could go sometimes and link the failure to get certain services to certain expected outcomes. For example, for most common chronic diseases, you could make some estimates. I guess that is what the RAND model does in terms of mortality. So I was interested in trying to hear from you, are there some things that with the currently available data you see could be done to build out another piece of these models that tries to answer the question of, if we were fully insured versus how we are currently, what would be expected over time in terms of improved health, productivity, longevity, other measures that we don't provide now?

PARTICIPANT: What kind of insurance are you talking about, Don?

DR. STEINWACHS: Offline Beth and I were talking about how insurance gets characterized. She was pointing out, it is the actuarial value of the insurance, and is that a good proxy for coverage. But a lot of what the public sees is, you are insured or uninsured.

A lot of the research when you get down to it looks at, are you insured, are you uninsured, and some of those results are rather striking when you say what happens to use of services and what happens to some chronic disease outcomes.

What I was looking for was your insights for what could be done to build out in directions that would help answer people's questions, saying what are we going to get for it in terms of better health and a more productive nation.

DR. SELDEN: I don't have any answers to those questions, but I think you could actually expand the question in one other dimension. The outcomes are not just health outcomes and productivity outcomes, but also financial outcomes in terms of what financial stresses households are facing, not just whether they are insured or not, but what happens to the frequency with which families have ten or 20 percent out of pocket burdens as a percentage of their income.

So it is something that is relatively low-hanging fruit for most of these models. It would be great to see result tabulated. That is an easier question to answer.

DR. STEINWACHS: It ties also to the bankruptcy. You hear every so often about people going into bankruptcy.

DR. SELDEN: Exactly.

DR. BANTHIN: I just want to add one thing. We know that even in countries with universal coverage, there still remain disparities in terms of chronic conditions, use of preventive services, use of the health care system, that are generally along socioeconomic lines.

MEPS has been used. Thomas applied some World Bank measures of equity in delivery and use of health care using that data. I am looking at the role of assets, exogenous changes in wealth, controlling for insurance and their effects on use of preventive care.

DR. STEINWACHS: So Jessica, could you give us some news about what is going to happen to us if we lost some of our wealth recently?

DR. BANTHIN: I have preliminary results that say if you have an increase in wealth and you go get that colonoscopy and those dental checkups, the implication is, if you lose all that wealth, you may postpone your colonoscopy and your dental checkups.

MR. SHEILS: There are lots of confounding factors. For example, a lot of the uninsured, they just don't want to buy any insurance because they are healthy, their kids are healthy, and they have other needs, very serious needs. Did I really say you shouldn't get the car fixed so you can have insurance, because you have got to get to work.

So there are a number of factors in play there, but to the extent that the uninsured are somehow self selecting into that situation, it is quite possible that they are healthy and they are not going to be a better health outcome.

It is unfortunate, but there are little bits of work you can do. I know the IOM did a study on the uninsured, trying to establish the case. They were unable to demonstrate any conclusive productivity. They couldn't find the surveys, they couldn't find the data to say productivity goes up.

We took the NHIS and we fooled around with it one day, and we got an estimate of what the increase in worker productivity is, or rather the reduction in the number of work lost days. It was something like .4 work lost days. It wasn't that hard to do that. But you may not be terribly pleased with the output. I'm not sure that productivity will be affected, or even our international competitors would be affected.

DR. STEINWACHS: It won't be the first time I haven't been pleased by some of the outputs. The other point that I think you are making is that it is not just receiving services, it is quality too. Some of the things that we are concerned about are conditions that are untreated, undiagnosed and therefore you don't get a benefit.

MR. SHEILS: I just want to say, we almost always show that the uninsured are spending more, not less, as a consequence of reform, because a lot of them are now required to get them.

MS. MC GLYNN: We find the same thing. I guess it is important to ask if that is the right question. Honestly, to use the favored buzz word in the nation today, one might construct a comparative effectiveness frame and ask if you are interested in advancing the health of the population, what are the tools at your disposal, and where frankly does insurance fall within that.

If you look at just in terms of contributors to mortality, we know the health system is a relatively small contributor compared to things like behaviors that aren't clearly influenced by insurance or not. So we might be better off investing in education, investing in putting recess back into schools, investing in teaching our kids nutrition.

I think there are a lot of those things that insurance is coming in and trying to pick up the mess that we already created. So I think these things are important to look at. There is a fairly strong critical religious element that doesn't want to hear that, that thinks they know the answer. I think that poses some challenges when either you do a document and you come up with not finding a big effect, or you suggest that insurance isn't the only mechanism.

So I think those things ought to be increasingly going hand in hand with models that look at just the financial aspects. When I talk about quality of care stuff, I really do think that what we have seen is that -- and this is true not only in this country but in other countries -- just getting people coverage and assets does not guarantee that once you walk in the door of the doctor's office, you are going to get what you need.

That frankly is the bigger determinant of health outcomes. There are certainly a lot of factors that are tightly linked to the factors that we know are true for some subsets of the uninsured, that may make them either at higher risk or higher benefit from also having coverage. But we have to look more comprehensively at putting a system together that delivers the care that is needed.

So just fixing the problem of uninsurance isn't going to fix all these problems as well. Not that it is unimportant, but it is not the only thing.

DR. TUREK: The Daschle book, many of the examples there were of people who had been employed with private insurance who ended up with something like cancer couldn't work and lost their insurance, and ended up going bankrupt.

One of the questions here is what happens to insurance if you get really, really sick.

MR. HITCHCOCK: I was just going to note that we have been here for over four hours, and I haven't heard anything about the concept of uninsured and then the concept of employers choosing to raise deductibles or copayments and the effect that that might have on enrollment. I don't really know if we need to discuss it, but I would just like to note that I was surprised that those concepts weren't more part of our discussion today.

DR. SCANLON: That may be a question for all of our modelers, to what extent are the characteristics of policies built into the models. I think we heard a little bit of that.

The other thing I didn't hear much of is the whole issue of insurer behavior. I think there is this potential assumption that health insurance is going to involve some rigid set of rules and all insurance is going to conform to that. Well, health insurance could involve modification of current rules where there is latitude, and what insurers do may be an important part, particularly if it is not across the country, but we have some state options in there, what insurers do may then affect the outcome.

So maybe talking a little bit about the insurance offerings and the insurers' responses.

MR. SHEILS: That is a really good point. Selection effects are a difficult thing to model. What actuaries will do is say, we just assume you are getting 20 percent sicker people, because that is what we have seen elsewhere. That is valid from an insurance perspective.

I'm not sure how well any of the models follow that. We model the selection that comes from the obvious; we try to keep it here for you rather than over here. We can model that.

An example. There was this guy down in Florida who was trying to enroll people in his Medicare HMO. He did it by having street dances. The theory was, if they are healthy enough to go dance in the street, they are a better risk. How do you legislate that option?

So we are at this point are a little bit -- maybe other people here have got this mastered, but we are a little bit stumped on how we are actually going to model it. But at the micro level the way we do it, this kind of selection without resorting to an actuarial mechanism.

MS. MC GLYNN: I don't think we have much data on insured behavior, because we don't largely observe who gets denied, who applies. A lot of that is just not in the data sets in any way that helps you model insurer behavior. So with big scale changes that is a little bit out of the model right now.

MR. GARRETT: What is in the model? At least with the utility based framework, we have a standardized ESI type of package and we have a standardized non-group type of package, and those are very different from each other in terms of their coinsurance that we convert it to, and deductibles.

So to the extent that you have different types of coverage, it is not always the same thing, we attempt to reflect that. When there is a reform, it is not what is cheapest in our models, it is what is the value for you, given various cost sharing arrangements and how it is going to affect the out of pocket spending.

DR. SELDEN: Just quickly on details of what coverage people hold, it has been quite a long time since a natural forum for collecting that information and presenting nationally representative information about that, and how it interplays with the burdens that people experience and whatnot, would be in MEPS. It has been a long time since we have collected and abstracted those. I'm not sure what it would cost. But it is fairly low-hanging fruit.

Another piece of low-hanging fruit might be to ask questions regarding medical debt in the MEPS.

DR. SUAREZ: I don't know how this plays into the models that I didn't hear today was these so-called non-traditional health coverage approaches like HSAs and MSAs. Is that something that is part of the consideration in those models? Health service accounts.

MR. SHEILS: We do model it. I can't tell you we do a great job of it, but we do model it. We model the savings effects. One of the things that you find is that, they talk about a high deductible plan, like $1,000. If you break you ankle, you will spend $1,000 in 20 minutes.

It turns out that about 84 percent of health spending occurs above $1,000 deductible. If you take everybody who spends more than $1,000, add up the amount in excess of $1,000, 84 percent. Once you have walked into an emergency room, your insurers are gone. That is too harsh, but it is an aspect. It is not going to have quite the effect that one might hope for.

MR. J. SCANLON: In all of your models, I think this is true, when you are estimating, you take a baseline case in your current policy and you run it into the future, and then you apply the potential policy changes.

Can you estimate the dynamics as well? Movement from an employer sponsored plan to another plan, movement into Medicaid or SCHIP or something like that? You can show that dynamic among and in between those enrollments as well?

MR. SHEILS: Right. A public plan is an opportunity to cover your workers at a lower price. It models that. There is also the crowd-out issue. If your workers become eligible for all these benefits, with the firm file, we have got all the characteristics on these people, so we will figure out what their premiums would be and if they are eligible. Then you can compare that to what the premiums would be if the employer provides it.

You find a lot of cases where it is just cheaper to buy it. You have identified a group that has a high risk of crowding out.

MR. J. SCANLON: Can I ask John the same question I asked the previous panel? You have simulated -- I know you did work for Commonwealth where you tried to simulate a health insurance exchange and a public plan option. What sort of assumptions did you make about a public plan?

MR. SHEILS: We did it for the Presidential comparisons. When Hilary Clinton and when Senator Edwards talked about a public plan, they were very clear it would be modeled on Medicare. Senator Obama was very careful not to say that. We had no idea.

I had to cover my butt, so I just took a rate that was halfway between private and Medicare, which was higher than any payment rate under the federal programs today except FCHPP. We used that, and on the basis of that we were able to identify those groups under the proposal with all of the other factors considered now, it is better to go into the public plan.

The decision to move through the plan, not everybody is going to -- I can save ten cents, I am going over here. We had plan change elasticities that we used to model that, but that is an extrapolation with probably very different data.

DR. BANTHIN: You can add inertia factors. We did that in our '99 paper. Economic theory predicts the long run equilibrium. That can take ten years to happen, and politicians and policy makers want to know what is going to happen next year and the year after. So you can work in some of these effects that slow down changes.

Agenda Item: Wrap Up, Final Comments

DR. SCANLON: We are getting close to one o'clock, and people probably have some commitments. According to our agenda, we are about 50-some minutes past the time when the committee was going to resolve these issues that we are struggling with. I think we are fortunate that we are 50 minutes past that, so now we can put off resolving it to another day.

I want to thank everybody, and then open it up for last comments from the committee members. I think this has been a valuable day in reinforcing a number of things and adding a lot of new information to what we need to think about.

Our challenge, Larry portrayed it very accurately in terms of balancing various perspectives. The challenge is to think about how to do that, how to communicate that. We don't make policy, we present information to the people that do make policy; how to present the information on what seems to us like a reasonable balance, how that balance does not involve unnecessary risk and yet provides significant benefits, because that is what they make decisions upon when they are thinking about policy.

So I'll open it up to questions and comments.

DR. HORNBROOK: I just want to remind us that the data lag here is really bad. We look at Oregon just for example. Our economy is going down the tube, school districts are cutting back, the state is cutting back. We watch the enrollment figures for health plans. They are losing enrollment. Those people aren't leaving the state, they are going on the uninsured roles. So there is some very, very recent changes that are very marked now in certain parts of the country, especially where the economic downturn has been very serious. It is affecting families, and of course the mortgage crisis is affecting them. So we are going to see people coming in on Medicaid. The Medicaid programs are going to have their expenditures go up, but you are going to see the state budgets can't afford it. It is going to be a strange simulation that is way beyond some of the dynamics in the models we have now.

DR. GREEN: This has been a week like no other in my personal life. I came in Monday, and we started a pre-hearing before the NCVHS, and we are now having a post hearing after the NCVHS meeting, and they have all had different names.

DR. STEINWACHS: Just to give you variety.

DR. GREEN: Well, actually I think it is to give us the illusion of variety.

It has been a very convincing week. We actually have a fairly desperate emerging need. Mark crystallized it again for us. The way we go about doing statistics and analytical work with the intention of influencing the directions we take is insufficient. We don't seem to have any dissent about that.

We are getting glimpses of the new world in the information age where data and information flow differently, connect differently, are available in different weights and different scale and different velocity. I am beginning to think that, as is often the case in policy, we may not have it framed right. I am looking forward to continue to work with the committee about us trying to get our minds around how this needs to be framed in order that we can get to a well ordered solution.

MS. MC GLYNN: I just want to make a comment on that. I really do feel like this is a fundamental issue in health care that for decades now, we are trying to do stuff with an unbelievably inadequate data infrastructure. Until we bite the bullet and start with what would a functioning health industry look like that had data like almost any other industry you can look at to make real time decisions.

So forget about research and simulation and modeling; what about just making decisions day after day by individuals, by physicians, by insurance companies, by firms, et cetera? What would that information infrastructure look like? In some ways, that needs to be the focus. I say this at some hazard of having my health research union card pulled, but almost more of that is important than fixing the research survey database, or at least it ought to also be on the table.

I think that until you get to that point, the ability to respond to nuanced ideas about changes in how we pay for health care, nuanced ideas about comparative effectiveness information available, none of that can get translated into really different behaviors on the top of the information infrastructure that we have today.

So I think you are absolutely on target in terms of saying, are you asking the right question, and what would it look like if you came at it from a different direction. We are starting over today, and we are going to build a system where every day in real time I have information on how many people died, how many people have disease X, how many people got treatment Y, how much did I pay for that, and that it looks like models that we can see in retail, so that you can have response in real time, solving today's problems, not problems that we now well understood five years ago.

Sorry, one of my soapboxes.

DR. SUAREZ: I was actually doing a little bit of modeling here with the wish list that you all provided us. I think that is a great place. I think it all builds into the comments that have been made.

I was looking at the various suggestions, and I want to point out a few of them very briefly here. We need to increase the size of population surveys, and we need to have better longitudinal data sets, like longitudinal coverage data, creating a long term longitudinal survey.

We talked a lot about state specific data and trying to add that level of analysis and trying to drill down into better understanding of the demographic at that sub-national level. We also talked about linking employer and employee data sets. We talked about having a better employer micro data set, and even getting some better non-group market data. RAND has been developing a provider data set, a longitudinal larger perspective data set.

Then I think we heard a few other major themes, like the privacy issue and the concerns around what privacy will continue to do, and might even do more as we get into the next iteration of privacy protections, and this concept of creating a certification program.

We talked about the need to look at subsets of population like minority health and health disparities and the effect of policies on minority population choices for health coverage.

But the most fascinating thing was your comment about, five years ago doing these analyses of the most prominent microsimulation models, and trying to perhaps go back and do on in a larger scale, and perhaps use that as a way to look at your question, Larry, are we looking at this in the right way, or should we be thinking of taking a different approach altogether.

So I think it has been a rich morning. I really want to thank you also for coming and giving us all this wonderful information.

DR. STEINWACHS: I just want to thank all of you. This has been to me very insightful and useful. You did discourage me away from my health outcome model. It was part of my comparative effectiveness approach. Beth has just taken it right on and said, forget it.

I think this has helped, really. You have heard the questions from the committee, to think about what are the data needs and how do we begin to meet those. This really ties into a vision of the Nationwide Health Information Network and health information infrastructure, of which statistics is a piece, but not the whole. So we keep talking about it and we keep having this vision, but at the same time feel that we are moving all too slowly to connect those pieces together, and we need to. So again, I thank you very, very much.

I thank all the people that helped make this possible, Roshita and Dale.

(Whereupon, the meeting was adjourned at 1:00 p.m.)