Kenneth Manton, Ph.D., Duke University
Korbin Liu, Sc.D., Urban Institute
KENNETH MANTON: We want to give you a sense of how the various data elements fit together when you are referring to the NLTCS.
For example, in 1982 of the ICS, there are 1,926 caregivers, both continuing and a small sample of people who discontinued care.
Then in 1984, we have quite a different group of surveys, the community-based survey, the next- of-kin survey (for people who were deceased in the 2 year interim period), and the institutional questionnaire (for people who were institutionalized on the date of which the survey was delivered).
The projected 1988 survey would not do the next-of-kin survey, but would do the community and institutionalized surveys. A four year follow-up would be a long period over which to rely upon next-of-kin recall on health services. The other component is the Part A data, which is the acute hospitalization use, and home health use.
Currently the files are linked from 1980 through the end of 1985 with some trail-off information through three quarters of 1986.
The public use tapes that will be generated will be composed actually of two sets of tape files: a rectangularized form of the survey data and the longitudinal file structure. This is integrated on the sample person so that you will have one very large record with all the information tied to that from the various record types on the public use tape that the Bureau of the Census released. The second set of tapes will be the individual Part A bills with a survey respondent identifier so that there will be multiple bills for each individual. You will be able to link the survey response information on the other tape file, with the detailed Part A bill information for that entire period 1980 through the end of 1985 with some information through 1986.
QUESTION: Will the information from the 1982 survey be included in the longitudinal file?
KENNETH MANTON: The full longitudinal sample, of roughly 25,000 individuals that were in the 1984 sample, will be in the file.
A total of 36,000 people drawn off the HIMF formed the sample frame for 1982. The 25,000 people who formed the longitudinal core will be together on the file both with 1982 and 1984 information in a rectangularized format. On the same person records, you will have the 1982 and 1984 information.
The work on the documentation and the files is essentially done, so it is a matter of clearance and then integration into the National Technical Information Service (NTIS) system.
QUESTION: Can you talk about some of the types of questions in the different surveys?
KENNETH MANTON: The documentation package has a series of appendices, background documentation on the survey, both the interviewer instructions for 1982 and 1984, and the community instrument. That tells you a lot about certain questions.
You would have the interview survey manuals, and then there would be a code book for the survey files themselves for the rectangularized file, with the information on them.
The community survey instrument asks about both chronic and current medical events. There are 29 conditions up front and broken down into things that occurred in the past 12 months versus chronic conditions that were experienced over a longer period of time.
There is a very extensive set of questions on ADL, IADL, and what I call IADL-II, functional limitation measures which have to do with physical functioning like holding packages or climbing stairs.
With each of those questions, there are a number of detailed questions that go along with what type of personal care or informal care, who might be delivering certain types of services to deal with those functional impairments, and the type of equipment that might be used. There can be up to 25 questions associated with particular ADL or IADL measures.
There is a lot of information on income and assets. There are some things on housing characteristics. There are questions on behavior. There are questions on health service use, visiting physicians, and other types of health service professionals.
There are various types of general questions on program entitlement in terms of Medicaid use.
KORBIN LIU: The NLTCS was designed, in part, to be a backdrop for the National Long Term Care Channeling Demonstration data. You will find fairly similar Items between domains of information between the NLTCS and Channeling.
There was a major concern about who provided what level care, what types of care, and what types of providers, in terms of whether they were family members or were formal paid providers, and in terms of costs.
For example, if somebody receives assistance in bathing, the questions are: who provided that care; was it a member of the family; was it an outsider; if it was an outsider, was it paid care. There are some dollar amounts associated with out-of-pocket payments.
On the income information, they ask sources of income, such as Social Security, Veterans Administration (VA), dividends and interest income, and so forth. They ask this for not only the sampled person but other members of the family living in the household.
In 1982, they basically had eight questions about home ownership and the market value of the home. There was no information in 1982 on liquid assets. On the other hand, there were questions about the income derived from liquid assets, so one might be able to estimate the liquid assets.
In 1984, for the community survey, the survey designers added a few more asset questions, such as stocks, bonds, CD's and so forth. As time went on, the amount of information on types of financial status improved.
In the 1982 community survey, they asked about institutional history, such as: Did you spend any time In a nursing home? When was the first time? When was the second time?
For each nursing home episode, they would also ask, how many days, weeks, or months did you spend there? For those people in the community in 1982, you have information on their prior nursing home use.
Part of that 1984 sample was the 1982 sample. They again asked in 1984: Over the last 2 years, did you spend any time in a nursing home? How many times did you enter a nursing home? How many days did you stay in a nursing home?
In some preliminary tabulations, we estimated that about 15 percent of community disabled population in 1982 went into a nursing home at some point in time in those 2 years. By 1984, 7 percent were still in nursing homes.
For these sample members, you had the community characteristics in 1982, and then you had the information about what kind of nursing home history they had over the 2 years. By 1984, you had the information on what their payment source was at admission, and their payment source at the current time.
A particular focus was the Medicaid transition. For private pay patients, there were specific questions and an institutional questionnaire about payment source at admission. There is some information there on spend-down phenomena.
Most of you are aware that the nursing home length of stay distribution tends to be bimodal, from the standpoint of having short and long stay patients. I think we are able to see that kind of phenomenon between 1982 and 1984 with this survey.
The other strength of this particular data base right now is that we have also got the Medicare bills attached to them, and the Medicare bills include not only the hospital but also the Medicare SNF, and the home health bills.
QUESTION: When will the public use tape be available.
KENNETH MANTON: It will be available through the NTIS. Herb Silverman from the Health Care Financing Administration (HCFA) might want to say something.
HERB SILVERMAN: Because of the sensitivity about the release of information, the government has put a lot of safeguards on the public availability of survey data, so there are a lot of little hurdles that have to be jumped before we can release it.
We have our Freedom of Information Officer who will have to review the documentation on the nature of the information, and make sure that the provisions for public notice have been met.
Every data file that is potentially available to the public is recorded in the Federal Register, so we have to make sure those requirements are met.
The most conservative statement on availability is before the end of 1987.
KENNETH MANTON: I think when Korbin was going through the instruments, we covered a little bit the community instrument, the institutional instrument, and a next-of-kin instrument for the deceased.
You also have the information on the Part A bill files. The bill file records obviously are contingent upon the Freedom of Information Officer clearance. They would contain an edited version, i.e., eliminating redundant bills and going through the standard editing procedures.
In the ICS for 1,926 individuals, they asked caregivers questions that reflected on the psychological motivation and the acceptance of the burden of care.
There are actually four or five survey instruments that were used. The community-based for the chronically disabled is the most detailed.
We linked them all to the person-based record so you did not have to crosswalk among four or five record types.
Regarding the disabled, 65+ noninstitutionalized, those who were given the detailed instrument in 1982, what happened to them In 1984? How did they sort out? A total of 409 became non- disabled; 4,114 ended up in the disabled component; 487 were institutionalized; and 1,383 died over the 2 year interval.
You have four donor states, and four receptor states telling you where people shifted back and forth between 1982 and 1984. You have to be careful about this since what is in 1984 Is not a proper subset of what is in 1982.
They did not go back to all 25,000 individuals who were non-disabled in 1982; they pooled a 47 percent sample of those because of budget constraints and did telephone screens of that 47 percent sub-sample. Then they brought in the 5,000 individuals who were 63-64 in 1982.
By the time you get to 1984 in one sense, you have looked at 36,000 plus 5,000 individuals who have been somehow touched upon in the two survey dates, but they have sub-sampled the group that was non-disabled in 1982.
For the institutional block, there was not an instrument in 1982, whereas there was in 1984.
The 1,992 is based on two components. In one, 1,708 people who on the April 1 freeze date for the sample who were in institutions, and there are about 280 people who became institutionalized by the time the instrument was delivered.
QUESTION: A person who was non-disabled in 1982, and non-disabled in 1984 would not be interviewed?
KENNETH MANTON: That is right.
QUESTION: Was the 1982 sample a cross-section of the whole population over age 65, or was there some kind of stratified sampling going on there where you took a smaller sample of the non-disabled population?
KENNETH MANTON: The total sample frame was the Medicare eligible population.
It is the vast majority of those 65+, taken out of the HIMF. From that they pulled out 55,000 individuals, because they did not know what their yield rate was going to be for the disabled. They had blocks of cases, which they would screen to see, in a given block, how many were disabled. They roughly wanted a target number of 6,000.
They worked backwards from the 55,000 and at 36,000 they identified the 6,393. The design effect in terms of this survey is relatively mild because it was not strongly clustered. The Census statisticians agreed that the sample design effects were less major for the 1984 survey.
One approach to the variance adjustment procedure which is normally done by the survey statisticians says, because of clustering effects and other things I have got a higher variance. When I go to a spatial cluster of cases, I do not have independent responses, so I do not have as many degrees of freedom as I might think. I need to adjust and increase my variance for that. They have various different models for randomizing effects to increase their variance estimate.
Another approach is to say that clusters in these other effects are important. If I were looking for local area effects, and I had a clustering in my sample design, I might be worried about a confounding between the spatial clusters and the sample design in my local area effects if I was looking for local market conditions.
I might want to go with the model-based approach where I build in the stratification and cluster effects as part of my model for analyzing that data. There is a larger discussion between sample statisticians and other people over this finite population approach versus model-based approach.
QUESTION: I did not hear mentioned the availability of insurance coverage for long term care, probably because in 1982, and even in 1984, such was not widely available.
Was there a question in 1984, and is one intended in 1988, for discriminating between utilization patterns on insured versus non-insured individuals?
KORBIN LIU: There was not a specific question on whether you had long term care insurance. In that section where they were asking about helpers, in that whole series of questions, they did ask, is it paid for by Medicare, by Medicaid, or other? That is probably as close as we can get with the 1982.
The designers for the survey were aware of a small prevalence of private long term care insurance, and I think it is one of those compromises that we had to make in terms of keeping the questionnaire within bounds, but clearly it seems like any future surveys are going to have a bigger interview section asking about long term care insurance.
We know some of the financial information can be improved, probably unlike the assets, and there is a scientific advisory group that is supposed to make recommended changes in content.
One person you can conceivably talk to is Bob Clark from the Office of the Assistant Secretary for Planning and Evaluation (ASPE), about your concerns, because he has been reviewing some of the documentation. You might talk about what would be useful to put in there in terms of long term care insurance, because it is obviously a topic that has to be covered. It has probably evolved to the point where some specialized instrumentation for that is appropriate.
QUESTION: You have got some users. Will NTIS be doing the support? I would also like to know if you have any comments on the difference in the prevalence estimates between the national health interview survey (NHIS) and the NLTCS.
KENNETH MANTON: In terms of further formal support, that has not been specifically addressed. If there were to be an 1988 survey, and currently that funding decision is being made. Obviously, there will be people working in the field and there would be support available in an informal sense.
HERBERT SILVERMAN: Making the public use tape available is just the beginning of the process, and it is a very difficult situation for agencies in the cutback mode to provide this kind of support. There is no in-house plan to provide an ongoing source of consultation formally. We are putting a great deal of store in the completeness and detail of the background documentation to make it as helpful as possible to understanding the data base.
QUESTION: We have just begun working on the 1982 survey, so I have some direct experience with the tape. I have some issues about the comprehensiveness of the documentation and some problems that we have found.
Actually, one problem that is still bothering us is just the very basic coming up with numbers that match the numbers that have been printed so far about the ADL's and IADL's, because my understanding of the documentation is that there were different sources for this information. On the data tape, there are a number of items where the source came from the control card information, from the initial screening. Then there is information from the detailed interview, with the sampled person themselves, and these numbers do not really match. It has been a problem for us to decide which is the appropriate identification of the ADL's and IADL's to go with, and I really do not know what resource you used to come up with your numbers for the number of people with ADL deficiencies.
KENNETH MANTON: On the screen there were 6,393 individuals identified. Of those 6,393, 6,088 actually had the detailed interview given. Of those 6,088, there were about 5,600 that completed question 13. Question 13 queries the person on what medical condition caused the disability or functional limitation.
I know that when we were starting to work with it, if you used the 6,393 in the weighted estimate, that comes out with 5.07 million weighted cases that have either a chronic ADL or IADL, and if you used the 5,600 number, it was about 4.65, so, to some degree, it depends on what questions you pick. The 5.07 million which comes from the screen or control card information is the most comprehensive. Regarding the difference between 6,393 and 6,088, we have done some analysis of that dropout group, and we have around 50 percent mortality; and it is a group that is very, very ill, and is probably responding for health reasons, to a large extent.
The difference between the 6,088 and the 5,600 is a more definitional question. For example, we have been looking at some data on the social health maintenance organization (HMO) evaluation and if you use nurses as interviewers, you get somewhat different numbers and different levels of screen.
Depending upon what criterion you use you can get somewhat different numbers, but to some degree that is a research question. We were doing some work on a GAO analysis, and they asked the question about unmet need for ADL types of problems, and IADL problems, and depending upon how you define that, you could get vastly different numbers.
QUESTION: I am not sure I understood when you spoke earlier about over-sampling of the age group 85 and older. Did you talk about it as it applies to the 1982 survey, to 1984 survey, or only to the future survey? Is that reflected in the weight that you assigned?
KENNETH MANTON: Both 1982 and 1984 had over-samples; in 1988 there is planned an oversample. The weights will reflect that.
QUESTION: For the community survey in the NLTCS, why are the estimates nearly twice as high as they are in the NHIS, and 1979 home care supplement?
KENNETH MANTON: What we are talking about is specifically home care supplement. If you look at those numbers, you find that they are very close when you look at the 85+, while the numbers are different at the younger ages.
As I recall, they might have used a different set of IADL's and ADL's. That could throw the distribution off.
We, for example, did not use incontinence as an ADL, but used walking around inside. Incontinence is a fairly severe ADL. One would expect a younger cohort to have less experience with that. On the other hand, walking around inside would be a problem that you would find more with the younger age group, relative to incontinence, and so by using walking around inside, we might create a larger prevalence because of that particular ADL.
Cathleen Yordi might comment on that because she was working on the social HMO data. If you go to a core group of more severely impaired, and you have got a nurse interviewer, you get higher agreement than if it is more of a judgment.
There is an article by Joan Huntley on a series of people in the mailbag that compared not only the 1979 Home Care Supplement but also talked about the NLTCS and also the three Established Populations for Epidemiological Studies of the Elderly (EPESE) at that time. The estimates for disability levels across the three sites in the EPESE studies vary quite broadly to include numbers that are above the NLTCS, and one of the sites at least has numbers that are very low.
KORBIN LIU: In another session, Jennifer Madans is probably talking a little about the next-of- kin versus the NNHS regarding the ADL's. In a briefing that I heard, you basically would find a lot of people with no need for assistance in bathing and dressing, when they were reported by the next-of-kin, but, 2 days later, you would find from the nursing home these same people were depending on bathing and dressing assistance. There are a lot of false-positives and false- negatives, and there are many artifacts with the function.
KENNETH MANTON: I would be more concerned about false-negatives losing people outside the sample than a certain small marginal false-positive rate, because there, you are obviously depending upon the definition you used. If you are thinking of an insurance product, you might be taking people and allowing them at a relatively low level of functional ability to still qualify. You might be using various types of cut-points and you want to look at the implications of different cut-points to see what population that is marketable with reasonable risk would be. The thing that would kill you is if you lost a lot of people who potentially could have qualified and did not have them in.
We are also talking about, for example, the case of the NNHS. Logically what we have here is two dimensions: one dealing with population coverage, or coverage for a certain sub-population; and the other dealing with time. The fact is we have three nationally representative longitudinal surveys: the NNHS 1985-1987 follow-up; the NLTCS 1982-1984, with potential 1988 follow-up; and the LSOA sample to the NHIS.
Obviously, there is a marginal difference in time, but the difference is probably more in population coverage than instrumentation. The population coverage in terms of the SOA, for example, is not stratified, even though they have a larger group, a little over 16,000 people, 55+; and about 11,000 people over the age of 65. They have 876, I think, in 1974 that are over age 85, whereas I think in 1982, the numbers are more like 2,400-2,500. Obviously, there is the age stratification fact, which gives you more precision on the 85+ group, which is going to have the long term care service needs.
Also, the full 6,393 that you have seen in the NLTCS has passed a telephone screen, using the criterion of a 90 day plus chronic ADL or IADL disability.
You are targeting in on the chronically disabled community-based population, and have greater precision. I do not know exactly what the numbers are for the full SOA for 1984, but if the proportions held up, you could expect about 1,500 total community-based chronically disabled people in the SOA or LSOA, so you have got roughly four times the sample size, plus the over- sample of the 85+.
There is quite a bit of a difference in sample coverage. You are going to get a large yield on the non-chronically disabled in the SOA.
They do overlap to some degree, but some overlap is comforting in the sense of coordinating results.
The NNHS is a different beast with roughly 11,000 people in the 1985 round. I think it is 5,000 in the current resident and 6,000 in the discharge sample. They have two sample components, one of people who are residents in an institution at the time of the survey, and another where they find out about discharges in the previous 12 months. This is not an admission cohort. It is the flip side when people come out rather than when they come in, and if there are very different cohort sizes, there are analytic models to try and take discharges and turn them into admissions by certain types of calculations. The sample sizes in the NNHS are focused on the nursing home population; it is much larger numbers and has specialized instrumentation, which is the other dimension of differences, the instruments and their focuses.
QUESTION: In the surveys of people who may be disabled but not institutionalized, do you capture data to indicate the percent of elderly who are dependent on high technology home health equipment and their payment source for it, such as ventilators and renal dialysis?
KORBIN LIU: I am not sure that it has got that level of detail.
QUESTION: Do you first have to have a ADL limitation before the subsequent questions are asked about what source of assistance you then have, either mechanical or in terms of personnel?
KORBIN LIU: They go through each of the ADL's, so in the questions, there is a fairly elaborate skip pattern. There is no sense in asking follow-up questions if that is not a problem in the first place.
QUESTION: You could be, if I understand it correctly, totally dependent on total parenteral nutrition (TPN), and not have an adl. I understand that there are people who are fully functional who are on TPN. TPN is a fairly high tech way of obtaining nutrition into the body. If you get that you are supposed to be fully functional.
KORBIN LIU: In the way that those ADL's were met, I do not recall a specific question on TPN. They were, for the most part, geared toward personal assistance, rather than high tech or mechanical.
QUESTION: I am really impressed by the creativity and forethought that went into putting this remarkably complex survey together. I have two questions: ONE, I wondered if we could just have a little brief descriptive history of how this got started; and TWO, could you tell us a little bit about what might be in the works right now in terms of research being done with this data set.
KENNETH MANTON: I guess in terms of past history, Herb and Korbin might want to say something. I can talk a little bit about more current usages.
HERBERT SILVERMAN: I am not the final repository of the history. There were a lot of discussions in the early 1980's about the long term care population and about how well were their needs being met. There were several interagency committees meeting at that time. I believe there might be members of some of those committees present here.
Several people got together from ASPE and HCFA back in 1980/1981, and developed the initial 1982 survey, and then HCFA took the initiative in picking up the 1984 rerun. One of the moving spirits in that was a fellow named Jack Sharp, who has since retired. I was involved in helping expand the scope of the 1984 survey. One of the shortcomings we realized from the 1982 survey was that we really did not get a full cross-section or look at the aged, impaired population. We did not include the institutionalized; we wanted to know more about that, so we deliberately went into the 1984 survey wanting to get both a longitudinal dimension on how changes took place across time, and a cross-sectional dimension, what was the prevalence of the impairments in the aged population.
We consulted extensively with Census, and the staff at HCFA, in terms of generating the sampling frame.
Basically, there was a combination of a broad policy perspective, and then the nitty-gritty technical work that went into its development.
KORBIN LIU: I really do not have too much to add. I was there too in HCFA's research office.
One of the major reasons was to write a backdrop for Channeling. This was about 1978-1981, when ASPE, HCFA, and the Administration on Aging (AOA) were going to invest in this enormous demonstration project on community-based care. Prior to that point we had funded a number of other community-based care programs like TRIAGE and ON LOK. At the same time HCFA, ASPE and AOA realized that the issue of generalizability would be raised, and it would be important at that point to start a nationally representative tivestudy or survey at the same time we were starting Channeling.
The incentive for the 1984 survey, I think, was in large part that we had this very rich cross-sectional data source. It would be such a waste not to use that information to estimate the risk of entering nursing homes. You would have a follow-up which dealt with this 1984 sample, which was clearly the cohort of people who were in the 1982 sample. We wanted to see what happened to them in that intervening period, based on the characteristics they had in 1982.
KENNETH MANTON: We looked at some of the impacts from 1982 to 1984 in terms of changes of service use, the Medicare service use, and the change in hospitalization versus home health use. There were some very dramatic changes from 1982 to 1984, which might be associated with reimbursement change impact. One of the strong motivations for 1988 is that a lot more of the changes will have worked themselves through the system.
We know from 1982 to 1984 there was a decrease in the rate of hospitalization and a shortening of length of stay. Where did those people go? Those types of questions could be approached.
We had requests for information from both the NLTCS and the 1977 NNHS to help in the design of some insurance products. We provided a series of tables for use of retirement communities and provided information for people from Travelers and Marriott.
KORBIN LIU: I might just add that there are a number of other groups that are starting to look at that data. The National Center for Health Services Research (NCHSR) has a research program that Robyn Stone and Peter Kemper are involved in. Based on a recent conversation with Peter, the issues they are looking at are risk of entering nursing homes, the patterns of informal care, what services are received in the community, and ultimately various packages of community care, both formal and informal that are received by disabled elderly persons in the community.
Sandy Newman at the Urban Institute is conducting a study on the risk of entering nursing homes. I think that Beth Soldo and Doug Wolfe at Georgetown are also using that data set to look at informal care and various service combinations.
QUESTION: I know that the conference is emphasizing accessing the data tapes, but prior to that it would be important to look at a psychometric package on each of those three instruments. I am wondering if such a package is available through NTIS. The package would consist of the instrument, the coding manuals, the training manual, and especially the validity and reliability studies on the instrument prior to fielding.
HERBERT SILVERMAN: It is intended that that will be part of the documentation accompanying the tape.
QUESTION: If that is the documentation, do we have to request the tapes to get the documentation, to decide whether we want to use the tape?
HERBERT SILVERMAN: There is a limited supply of questionnaires available.
QUESTION: It is not just questionnaires. The validity and reliability studies, the coding manuals, the training manual--all of that?
HERBERT SILVERMAN: The real hope is in packaging it all together as an accompaniment to the data tapes rather than as a piecemeal activity. Just write to me and make the request. We will try to comply with it.
QUESTION: I have done some work with the ICS, which is a subset of the 1982 NLTCS, and there are some discrepancies in both the number of caregivers and the number of impaired elders who are receiving care from those caregivers. It is beginning to become a problem in that some of these estimates are being used for policy purposes; and I have been asked a number of times which estimates are correct.
Basically these surveys were done by two different entities, the NLTCS was conducted by Census, and the ICS was conducted by the National Opinion Research Center (NORC). I just wanted to throw it out as an issue, because there are discrepancies in the national estimates. maybe you could speak to that in terms of some of the possible reasons, some weighting problems, or some other methodological issues.
KENNETH MANTON: Beth Soldo asked me some specific questions or numbers about that. Census did the core community-based survey. NORC did the caregiver; they picked up individuals supposedly from the sample.
QUESTION: I think that the sample clearly came from the control card. The control cards were sent to NORC, so clearly in defining the ADL's and the IADL's. You should use the control card in at least trying to get comparability in estimates. Even after doing so, there are some discrepancies. I think, if nothing else, people ought to know that.
I have one other question, and that is more a data cleaning question. I have started using the 1984 data, and have begun constructing information on the caregivers, and there are discrepancies in using the different records.
For example, if you use the household record, and then you look at the helper record, and you match up the person numbers, sometimes you will get somebody on the Household record who is a spouse; they will turn up on the helper record as a child. That is just as an example.
You have got two different relationship types with the same person number. The question I have is if, in fact, you have set up a file which is hooked into the sample person, and from what I understand that you could merge all the files so one would not have to go to each one of these separate files, how did you handle the discrepancies? I am just curious because we do not know whether to call the person a spouse or a child. Have you handled them?
KENNETH MANTON: Obviously, you go as thoroughly as you can to perform the contingency checks. A complete set of contingency checks across all variables is never possible.
The first Census public use tape we got missed six million elderly people when you added up the sample weights. When we went through that, they went back through all their software and it turned out that they left truncated a nine-digit field to eight digits. They lost probably six million people when they weighted that up.
We had to convince them that we were not tabulating things wrong. That took 2-3 weeks, then we got them to go through their software. Then Herb and I had a number of discussions, concerning the variables used in calculating the weights.
The age variable that I would recommend using was the one that was finally used to generate the final sample weights. How that variable was arrived at was to look at the multiple age variable and do contingency checks to see what the greatest consistency was.
It is in no way the same thing going back to an independent data source and validating. There have been efforts to generate consistency on the basic variables, and that was gone fairly intensively.
We had another question in terms of an 1984. There is one person who showed up in prison, and we wondered if that was a valid code as an institutional code.
I cannot guarantee that every variable is absolutely squeaky clean, but if you did not have the demographic variables and the sample weights right, you were nowhere.
A lot of effort was spent on those, and then the basic demographics, the marital status and other things. There are a lot of contingency checks.
That does not catch the single case that is off, but it tells you if there are general inconsistencies or patterns. The bill file information, on which we have done a lot of cross-checking seems in aggregate to make sense.
We had straightened out the 1984 sample weights, and then we just happened to go back and check the 1982 cross-sectional final weights, and found that they did not add up to 5.07 million any more, but to 4.98 million.
Then Herb got a new tape in the mail, and the numbers did add up to 5.07 million.
There is no way of catching everyone, but there has been a fair amount of consistency checking.
KORBIN LIU: On your first point about the discrepancies between the NORC ICS and the NLTCS, it might be helpful for the group to give some sense of how big the discrepancy is.
KENNETH MANTON: Is the discrepancy thousands, tens of thousands, or hundreds of thousands?
KORBIN LIU: The bottom line, is it so big that people should not even try to use it?
QUESTION: I think they should try. I am just saying that there is a difference. The other problem is in trying to look at the number of caregivers in the NLTCS, and the number of caregivers who are estimated in the ICS. You have got some differences in that the ICS differentiates between primary and secondary caregivers, whereas in the NLTCS, there really is not that differentiation. They are very different instruments, but one is contingent on the other.
KENNETH MANTON: The ICS is contingent upon the NLTCS?
QUESTION: The ICS is much more detailed if you are interested in caregiver information, but, again, you are not getting the exact same population that you would looking at in the NLTCS. The ICS also is just cross-sectional. It was only for the 1982.
KENNETH MANTON: Yes, and you would get some numbers and differences. Remember, I was talking about the institutional component. If you talk about the freeze date of April 1, you are talking about 1,708 people. They also had 284 people approximately who were indicated to become institutionalized between the freeze date and the time in which they attempted to deliver the questionnaire. You have got a series, because you have got a longitudinal sample. You have got different survey dates. Unless the ICS was contemporaneous with the delivery of the household interview, you would expect some systematic measurement differences.
QUESTION: In the documentation there are several good write-ups on the ineligibility problems, the non-response rates, and things like that. There are these discrepancies, some of which, at least to date, I have not been able to reconcile completely.
KENNETH MANTON: If they have to do with survey date, you may not directly have the Information to deal with but you can only be aware of it.
To some degree, when you look at a survey, you have always got stochastic error. Then you have certain measurement problems that will be systematic in our measurement error. Some of those will be associated with sample design artifacts like date of survey delivery.
There is a certain amount of slippage that can go on because you have got different instruments delivered at different dates. The reason why the 1,982 versus 1,708, the 284 difference shows you what can go on in 3-4 months with respect to a phenomenon as volatile as institutionalization.
Institutionalized status is relevant as a non-response, a variable for 1982 only. In 1984, we had institutionalization.
Then you have got "moved outside the country," or "moved within country beyond the sampling limits" where they were going to pursue people in a particular area. The people are essentially ruled out of the sample due to either being deceased or institutionalized in 1982, or for geographic mobility.
Next is the standard types of non-response that one typically thinks of when one thinks about a survey non-response. This is both for 1982 and 1984. I believe it was roughly in 20 percent of the cases that in order to try to track down the screener information and be certain of that Census actually went and did a household screen. They simply just did not do a telephone screen. If there were questions, they did a sizable number of personal visits to screen the individual.
There are essentially two, if you will, denominators in terms of the non-response rates in 1982 and 1984. One is going to be for the screener, which in 1982, for example, is going to be roughly the 36,000 individuals. Then you have the non-response at the level of when the subgroup or the chronically disabled group is identified when they tried to do the household interview.
At each of those levels the response rates are just about 96 percent. For example, 6,393 dropped down to 6,088. Our analyses of that subgroup, the difference between 6,393 and 6,088 was almost 50 percent mortality for the 1982 non-respondents.
They asked the reason there was the need for a proxy respondent. Once that was fixed, based on that screener interview, then the person received that diagnosis of senility or cognitive impairment, and then the proxy carried through the rest of the interview.
In addition to that, there is, in the instrument itself the SPMSQ, so that for those individuals who are not too impaired to respond, you can get some measure of cognitive performance.
QUESTION: Who was used as proxies in the institutional component?
KORBIN LIU: In the institutional questionnaire, the sampled person was asked the SPMSO questions and then somebody on the staff of a nursing home, a nurse, aide, whoever would be asked the questions on ADL, and ultimately the institution would be providing all other information including payment source information.
KENNETH MANTON: The extent of proxy responses especially for the extremely elderly population and highly functionally disabled is always a question in survey design. I think that is one of the reasons why in the SOA they decided not to over-sample. On average, we are getting a 96 percent response rate, overall, and that is with an over-sample of the 85+. I think the Census field operation staff did quite a good job on that.
QUESTION: Were there sample design effects for the 1984 survey?
KENNETH MANTON: The sample design effect for the 1984 survey is smaller, but Census did do analyses of sampling variances and produced various simple types of variance adjustment procedures with some look-up tables that you can use for, at least, first order analyses of precision of various types of estimates.
There are two different approaches when you are trying to deal with the sample design effect. One is more model-based; i.e., trying to explain the design effects within the analysis. The other is saying, it is going to be hard for me to get the true model of the phenomena. I have got ways of conservatively adjusting my variance estimates for what might be the perceived level of contamination or correlation and stay within a cluster.
One says I am going to search for the substantive model that underlies the design effects, and directly analyze it as part of my analysis. The other says, I do not ever know when I have the correct model, but I have ways of adjusting my variance upward, and will deal with that to give me an upper bound estimate.
One of the results that was interesting, referring back to an earlier question, in terms of analyses that have been done is that if you look at the population weighted number of people with ADL's 5 and 6, and compare that to the number of people who have that profound level of impairment in institutions, there apparently are more people in the community with that high level of impairment than are residing in institutions.
It is also very interesting that if you look at the relationship of economic characteristics and resources to ADL status for people in the community, it is curvilinear. What that means is that when you go one to two to three to four, there is a drop-off in status and resources. When you go to five to six, those individuals tend to have fairly good economic status; they tend to own their own home; they tend to have a spouse available.
In other words for a person to be retained within the community, at that profound level of disability, they have to have three or four factors going for them to keep them out of institutions.
Not only is it a larger group than you would find in an institution, but it is also a very specialized group.
QUESTION: What was the length of time that someone had to be disabled to be in the survey?
KENNETH MANTON: Remember it is 90 days, so you might think of somebody with hip fracture who could undergo rehabilitation. There are situations where you could conceive of rehabilitation, where there would be a medium-term disability, and obviously one of the challenges is to subdivide this table and see if that is associated with particular types of diagnoses.
If you adjust for the 37 percent who died, you are talking about 35-36 percent who got better among survivors; and what I think you are seeing is people of five and six ADL's are probably bimodal at least as a population. People who are that severely impaired because of some underlying acute type of condition, if you want to consider hip fracture, are different than others, such as an Alzheimer's type where there would be progressive, monotonic course.
You will find that there is a big difference between the 65-74 and the 85+ in terms of the probability of getting better or worse. There are about 3,200 deaths over this 2 year period, and over the 6 years total, you would expect a little over 10,000.
For any individual who is in the survey, you have the 5-6 years worth of Medicare service use, and you can aggregate that in the service use in different windows.
You would look at the full 5 years, or you might take a 12 month window that centers around the mean survey date, and ask for plus or minus 6 months from the survey date what was going on.
We took calendar year 1982 and calendar year 1984 and not using data from the survey but simply using data on the sample components, which include the identification of people, whether they were disabled or not, or in institutions or not, and looked at what was implied by those sample components in their use of hospitals.
Think of 100 percent of all people for all hospital episodes. We have abstracted out the total number, as in a life table calculation. That is like a rate X, and then we show how many people leave in the first day? How many leave in the second day? How many leave in the third day? We do this for all hospital episodes for 1982, as it is for the 1982 calendar year, and as it is for the 1984 calendar year.
What you can see are fairly major differences, in terms of hospital length of stay, in terms of the sample component.
The non-completed curve is that group that is the difference between 6,393 and 6,088 that I said had the high mortality rate. They also tend to use a lot of hospital services. They tend to be the longest stayers.
You can take that community disabled group, I mean the people who screened in and answered the questionnaire, the 6,393 people, and see how these service use measures varied as a function of disability level.
You could get similar types of life table calculations for hospitalization. You can not only do hospitalization with the Part A files, but you can also do home health services. You could take all home health service use for 1982 and 1984, and see what changes there were between 1982 and 1984 in terms of mean number of visits or mean number of days this person spent in a given home health episode.
You can do that for SNF's, as well. For the Medicare data, you can use life table methods to look at the episodes. I use the life table type procedures, because I want to be able to adjust for a certain type of censoring events. Once a person dies in the hospital then he is no longer eligible to accumulate length of stay. I want to make an adjustment for the fact that I now have missing data, because the person died.
The basic point here is that you not only have the survey data, but the associated continuous time use of the major types of Medicare service files, and there are ways of using the survey data in conjunction with those Medicare files to tell us a lot about chronic functional health status impact and acute service use. It allows us to examine questions like who is not becoming institutionalized or who is not using hospital services for 1982 or 1984.
From some of these results and from some things that we saw in the California MSSP, people with certain chronic disease diagnoses, like diabetes, seem to have a drop in hospitalization.
There seem to be certain types of chronic diseases where there were short hospitalizations for people without informal caregiver help. They were going to the hospital, rather than remaining in the community.
KORBIN LIU: One large component of the 1984 sample was the 1982 sample of noninstitutionalized disabled elderly people. Between 1982 and 1984, you do have an incidence of nursing home admission; and what we found was that approximately 15 percent of the 6,000 people in the 1982 sample had spent some time in a nursing home between 1982 and 1984.
The 1984 NLTCS also had an institutional component which described various characteristics of people who were in nursing homes in 1984. Some of the people who were in that 1984 institutional component were people in the original 1982 community sample, but not all of them. Not all of the people in the 1982 sample who went into nursing homes were found in the 1984 sample.
The reason is that some of the people returned to the community, or they were deceased, but you subsequently have the information to look at both long and short stay nursing home patients with the 1984 information and combined with the 1982 information.
QUESTION: On a deceased questionnaire, who were actually interviewed in regards to next-of- kin? Was it actually the next-of-kin or the person who assumed responsibility for the person?
KORBIN LIU: It was not necessarily next-of-kin. Essentially they were trying to find the most knowledgeable person about the sample person.
QUESTION: How did you determine that a particular person was senile?
KORBIN LIU: The question was asked to a proxy, and the proxy was responding. If the sample person were responding, that question was not asked.
KENNETH MANTON: When they identified the proxy respondent, they had a two-phased screen. They first tried to do a telephone screen, but in about 20 percent of the cases, they had to make a personal visit. At that time when they had to use the proxy respondent, they asked the reasons or determined the reasons we were using a proxy respondent.
If it was a profound cognitive impairment that prevented the person from responding, and that was on the list of reasons for proxy respondent, then that is where the diagnosis of "senility" came in. If they actually talked to the individual, then the short portable was given and for those individuals who were able to respond directly without full proxy responses you had the distribution, but it was a truncated distribution in the short portable.
QUESTION: You said that this was a backdrop to Channeling, were there any case management related questions in the NLTCS?
KORBIN LIU: It was primarily focused toward specific services that were being received for specific ADL's and IADL's.
QUESTION: Did you ask any questions about drug use? If so, did you ask for specifics, or ask them to get out the drugs that they were taking, either prescription or over-the-counter?
KORBIN LIU: There were questions in the instrument about medical services, including prescription drugs. The questions were oriented, not toward the quantity, but toward the expense. There was an attempt to find out how prescription drugs expenses were covered. There were other questions about whether, if there were no costs, then was it because of professional privileges, or was it covered because you were part of an HMO plan, or so forth.
QUESTION: What was your degree of confidence in the responses that they were not gaming their response to protect a home health care service, or to describe the client as more dependent than they might, in fact, really be to ensure that home care services might continue?
KORBIN LIU: Is your question that the sample member under this national survey would have been afraid of losing eligibility.
QUESTION: There might be a difference between getting 8 hours of attendance at home, or 2 hours, depending upon the degree of disability. I am just wondering how did you validate what was described was what, in fact, was going on?
KORBIN LIU: I am not as familiar with the confidentiality and clearance procedure. This is a survey that was conducted by Census, and so I think the lead-in would be the letters and the screeners, essentially just indicating this is a national survey to collect information for research purposes and so forth.
KENNETH MANTON: We think of the 1982 and 1984 surveys which have been conducted, and the proposed 1988 repeat, for each of the three dates there are a different survey components or instruments, delivered to a different survey population.
In 1982, we have primarily the community-based survey for the disabled elderly, people with a chronic impairment of 90 days or who anticipated that it would be for 90 days.
In the NORC ICS, they went back to a set of individuals for a sub-sample of the total community population. They talked to the caregivers about the burden of delivering care, both for a sample of individuals who were continuing to deliver care and for a group of individuals who had stopped delivering care.
The 1984 survey was more complex. There was a total of about 10,000 interviews conducted, in addition to the community-based interview, which is very similar between 1982 and 1984. You had the institutional survey and then you had the deceased questionnaire. In both dates, you also determined the status of this population.
That turned out to be useful for some aggregate level analyses. You have the detailed community-based survey for people who screen in who have chronic disability, but you can also then cut the pie apart and have the non-chronically disabled apart from the total institutional population. This allowed us to look at some basic aggregate level changes for the entire population and put these people in contact.
In addition to the survey instruments in 1988, it would involve a repeat of the community and the institutional surveys.
In addition to the survey instruments delivered at particular dates along this time line, you also have those survey records linked into Part A service file. This file includes the exact dates, reimbursement amounts, and number of visits for home health, for SNF's, and for hospitalization, for roughly the full survey plan, a 10 year period.
Currently, the data is available, and linked from 1980-1985 complete with some additional data in 1986.
There have been even some discussions of the possibility of Part B linkage, but currently the anticipated plan or projected plan is that both Part A bill files and the individual records with a survey respondent ID, i.e., a non-meaningful ID would be prepared and available as one file. Then a rectangularized version of the 25,000 individuals with the 1982 survey and the 1984 survey responses would be available in another file, with a common survey respondent identifier which allows you to link. You could have multiple bills associated with a particular survey record in a rectangularized format. Instead of doing what Census did with their public release tape, which has four or five distinct record types, and which had a separate component for the 4,916 people who were in the aging sample, we put them in one person-based file with all record types linked.
That basically takes the data processing away from one of writing some PL-1, Fortran code or whatever into various merged types of files. It is inefficient data storage wise, but it simplifies some of the programming tasks.
One thing to recognize in the design is that the 1984 sample is not a simple subset of the 1982 sample in that you have the 4,916 people over age 63-64 in 1982 who are aged in. A second feature is that there was only a 47.4 percent sample of the non-disabled people who screened out in 1982, because of budgetary constraints; also people who were disabled and received the detailed interview in 1982 were reinterviewed automatically in 1984. You get detailed characteristics on people who are chronically disabled at one date but who improved.
In addition to that, there were the various components, the aging sample in the non-disabled sub- sample were telephone screened and had the personal visit screens that determined who became newly disabled.
You get the full flow composition of that population. As you can see, there is an institutional group identified in 1982 of 1,992 individuals. In terms of screening, they found that the people that were institutionalized as of April 1, 1982, which was the freeze date of the sample, there were, I believe, 1,708, and there were about another 284 who became institutionalized between the freeze date and the date of survey.
The NLTCS is one, the SOA/LSOA conducted by the National Center for Health Statistics (NCHS) is a second, and the 1985 NNHS with the 1987 follow-up is the third survey. One can see that the one axis is time, and that they are roughly contemporaneous, or cover reasonably comparable experience, but they differ very much in terms of the portion of the population covered by the survey and in terms of an instrumentation.
The procedure was to draw 55,000 individuals and then to take reduction sets until they achieved a total sample size of 6,000 cases. They did not know exactly how many people were going to pass their criterion that they were using to quality for the detailed community survey. They took an upper bound of 55,000, and, as they went through the reduction sets, when they hit 36,000 they reached their targeted projected goal of 6,000 community surveys. The SOA/LSOA is more of a community-based sample with no such screening procedures.
If the yield rate that was achieved between the 36,000 and the 6,000 cases were to be applied to either the 16,000 individuals over 55 or to the 11,000 individuals more property over 65 you would expect about 1,500 people impaired in the SOA/LSOA, despite the large sample size.
In addition to that, in the SOA/LSOA, they did not do an age stratification, as I understand, so they only had 876 people over the age 85, despite the large base sample size, as opposed to on the order of 2,300 persons over age 85, in both 1982 and 1984.
The 1986 LSOA, the longitudinal follow-up, is about 5,500 persons, and had a complex sample selection procedure where they did pick up all people over age 80, and then did pick up certain groups between 70-79.
All minorities were picked up, and spouses of people over age 80 were picked up as well as the 50 percent sample of the residual, which ended up with a total of 5,500 individuals.
Again without the disability screen, and as a general population sample. The NLTCS has much more precision on the disabled community dwelling population and on the 85+, but you are missing a component in terms of the non-chronically disabled individuals who are back in the population.
There are also major differences in terms of instrumentation. Because of the target focus and the NLTCS on the disabled population, the instrumentation on ADL, IADL, personal services, and equipment used is much more detailed than in the SOA/LSOA.
The NNHS is an institutionalized population, about 11,000 individuals total, 6,000 in discharge, and 5,000 in current resident, although I tend to flip those numbers.
The instrumentation is very different focusing on an institutionalized population and looking at much more detailed aspects of the institutional experience than is available in the NLTCS.
The NNHS in the discharge sample will tell you something about an admission cohort, but you have to do some backward mathematics and analyses to make assumptions about the size of the entry cohorts to be able to know what the admission experience was that could have corresponded to the observed discharge experience.
The point is that there are complementarities among the three different surveys; there is also some common instrumentation that can allow you to relate the results of various surveys to one another. For all three surveys, there are plans to link to mortality data and to Medicare service use data, so you will have detailed endpoint to compare the experience of the various sub- populations and all the instruments that contain some degree of approximation of functional status and health status measures.
KORBIN LIU: One of the questions that came up is which data base is preferable for looking at specific questions like risk of entering nursing homes and length of stay. In line with what Ken has saying, the two data bases that start out with community population are the logical ones.
In both cases we are going to have fairly short nursing home use experience, so if one wants to look at length of stay, then the obvious source is the NNHS.
One of the problems with the 1977 NNHS was that it included multiple segments of nursing home stays for given individuals. If you try to compile those lengths of stay, they turned out to be a shortened pattern than one would expect. In 1985 they corrected that problem.
Effectively, if one wants to answer the question, what is the risk, duration and cost of nursing home care, you would have to use multiple data sources.
KENNETH MANTON: Type C is a survey eligibility type of screen, where you have people who drop out, because of death before April 1, or death between April 1 and the survey date. Institutionalization was a reason for screening out in the 1982 survey only.
Obviously, in 1984, there is an institutional sample, and they were brought directly into it. The numbers in 1982 were 1,992 people. In 1984, people dropped out because they were not in the geographic range of the survey operation.
Then the next type of non-response is the type A and again these are frequencies rather than rates just to show the various types. The response rates in the survey were remarkably high, about 96 percent so that the non-response problem was rather minimal.
The largest proportion are sample persons who answered. Things were reasonably stable between 1982 and 1984. Proxy responses go up proportionately as disability level goes up. The proportion of proxy failures in the previous table is fairly small.
Obviously, the field operation group at Census was able to do a fairly effective job here in the proxy identification methodology that they were using.
QUESTION: Were the people who were most disabled in 1982 most likely to be non-respondent in 1984?
KENNETH MANTON: The proxy response rate increases. You have got two types of responses, the person or the proxy. For the persons who are already disabled you can find an effect of proxy respondent? If one were to rely only on sample person responses, you would find, obviously, a huge gap.
QUESTION: There is a couple of hundred non-answers, right? If they were all from those with five or six ADL's, then that means a lot of the people that had five or six ADL's in 1982 did not answer.
KENNETH MANTON: It was a slight increase, but I do not have those exact numbers at my fingertips.
First of all, you are talking about 4 percent and 4.4 percent total non-response, and then you are subdividing the 4 percent non-response into certain other types of non-response, so you are talking in terms of maybe dealing with a total of 1-2 percent. Then there are people at the various levels, the one, two, three, and four. The numbers of the cases were certainly small, and it is not that all the nonresponses occurred at ADL five and six.
QUESTION: Can you talk a little bit about what the 1984 sample represents, in particular, since everyone who is impaired in 1982 was followed up in 1984, and you do not have screeners on everyone in 1984? How do you figure out what the 1984 population represents? How do you do the weighting for that?
KENNETH MANTON: You have individuals who are given the instrument in 1984, who show up non-impaired. Indeed, to compare service use between 1982 and 1984, there are certain additional adjustment factors needed to get to the comparable population. In that particular case, It is to look at the detailed responses and find individuals who came up non-impaired in terms of the detailed instrument.
One could not just take the sample component and deal with that. One has to go back and look at the ADL and IADL responses. There are certain check-questions between 1982 and 1984 that you can go to as control variables to help you subset that. You do have to be careful to decompose it that way.
Another instance that you have to be careful of is in terms of the institutional sample, because if you remember there are 1,992 in 1982 who were not interviewed because they were institutionalized. That could have been at two dates. To get our institutional comparable sample, we had to walk out those people who were institutionalized as of April 1, as opposed to additional people who were institutionalized when you attempted to interview.
The non-respondents in terms of the difference between people who screened in during 1982 was 6,393 versus 6,088, but that sub-sample was a high Medicare service use group, and had 50 percent mortality.
The comparable identified subgroup of non-responders in 1984, which is not defined exactly the same way, had about 25 percent mortality, was a little less health impaired, but still probably was not responding to a large degree because of the nature of the medical problem.
There are some fairly subtle differences that you do have to be careful about in terms of making sure that you are defining the same population in 1982 and 1984, because 1982 and 1984 are not precise enough.
QUESTION: Let me just ask two follow-up questions, if you use the 1984 survey tape in a cross-sectional sense, without worrying about the longitudinal part, what is it supposed to represent?
KENNETH MANTON: You have both final cross-sectional and final longitudinal weights.
The weights themselves are obviously affected by several factors, not just the aging group. You have got mortality over the interim, and non-response of certain types, but also you have got the fact that the people who were drawn from the 36,000 cases who were non-disabled in 1982, you only had a 47 percent sample.
There are several additional factors in terms of calculating the sample weights, but the longitudinal weights are adjusted appropriately to deal with those various factors, and the cross- sectional are derived to give you the appropriate cross-sectional distribution.
QUESTION: Where could one go for more information on some of the technical details, like the difference between these dates and the fact that some of these adjustments had to be made?
KENNETH MANTON: For the documentation, from the NTIS release. I have a short document which describes the system file structures, the reasons for putting them in that form, and some processing characteristics.
Then we redid the documentation, which is about 300 pages, describing all the different types and variables. We put that into an integral sample person-based format. In addition, we also included the 1982 and 1984 survey instructions for the interview manuals. You have to dig through the interviewer manuals in order to pick out the subtlety.
You have system file documentation, the documentation on the variables, the interviewer manuals, and then copies of all the instruments.
Then there was a question asked earlier in terms of any studies of validation. We have looked at some things in terms of proxy respondents in the response categories by the different subgroups, but that is validating against an independent experience.
There is always this measurement question. Can I have one definition of my long term care population? It seems to me that varies depending on the policy and research purpose. I do not think you can have one simple definition.
I think with respect here to the instrumentation, you have got a fairly broad screen with some small proportion of false-positives. That makes me feel comfortable. The false-negatives is what would kill you in terms of trying to represent the population, if you were losing individuals who should have been in who were dropped out.
To a certain point if the documentation is there, you are left with some hard work to go back and track down things, but you have to be careful not only with instrumentation but longitudinal aspects of the survey.
If the weights are calculated correctly, they will absorb a number of these adjustment factors, so the simple answer is the longitudinal weight should capture a lot of this. If you are not a trusting person, you have to look at a lot of details.
When we went back and forth with Census, we spent the last 4-5 months working on longitudinal weights. In the first set we lost six million people, because they had a nine-digit field that was truncated one from the left, so they lost one decimal point.
The next thing we went through was the variance calculations to adjust for sample design effects. Certainly the survey statisticians have one approach which involves some model of randomization which inflates the variance usually to adjust for cluster effects, or adjust the variance estimates for the fact of not having a simple sample design having a stratification and clustering structure built into it.
The NLTCS is relatively simple. There are strata, for example, on age and eligibility status, and a couple of other variables. The clustering effects are relatively minor and small. Census did calculate a first level of approximation at least for 1982, and they redid some of the material for 1984.
This is one way they have done it. For two sample components which have different f-factors which are based weight differences, one being black and those are Medicaid eligible, and then all others. You have a simple regression function where you plug in the population weighted value of people in that subgroup; (a) you simply put in the population value, and (b) you put in the square for that term.
If you are trying to do certain types of things like forecasting or if you are dealing with local area phenomena, like market phenomena, there is an alternative approach, which is based on what they call super-population models, where you attempt to view these sample effect as either, (a) irrelevant, in which case you ignore ft, or, (b) if relevant, to include it in your modeling effort and in your substantive analysis.
The notion of looking at the longitudinal 1982/1984 component, and at the information from the survey instrument in terms of functional status is shown on a table based on changes from 1982. The 1982 survey status is down the side, and the 1984 survey status is across the top, with the additional factor that sample members could have died over the 2 year period. Even for people with five and six ADL's, it looks like there is some improvement about 25 percent or a little less.
If you take out the non-survival group, the 36 percent who died, or the ADL five and six, which is roughly similar to the institutional group, you are talking about almost 35 percent survivors, who improved in terms of functional status. Some of the preliminary analyses suggest that for people at that impairment level, there may be two types of processes. One, they really have an acute problem like hip fracture which causes the 90 day impairment but they are rehabilitative versus something like an Alzheimer's condition, or a stroke, where you would figure the capacity of rehabilitation is less and the damage might be more permanent.
You are finding rehabilitation potential at the younger ages, at the higher disability levels, which is consistent with what one would expect.
We talked about the fact that these are linked to the Medicare service use files.
Those represent the continuous service use over a 5-6 year period; and the question is how do you relate those to the information on the survey data. One way to do this is to utilize various types of simple life table models as a way of adjusting for various types of sensory and competing risk effects.
One could think that if one were already in the SNF that certain types of hospitalization would not occur. In that sense the amount of time spent in the SNF would reduce the exposure time for hospitalization.
Likewise, death is also clearly a censoring event. If somebody is very ill, and then dies early in the year after measurement, then their exposure time for service use is greatly truncated. By recombining the exposure within a given interval, using life table methods, you can adjust for the exposure time differences.
Remember I talked about the non-completers for the high mortality rate. They also tend to be people staying in the hospital the longest time. In other words, the proportion staying in the hospital each length of stay tends to remain higher. They are a very high service use group.
People who are not chronically disabled are staying in the hospital shorter times than people with chronic disability.
My basic point is that one can think not only of analyzing changes in the survey characteristics but of linking those survey characteristics to the continuous services and conducting combined analyses to get some rather rich types of analyses on Medicare service use of a functional status for nationally representative population.
QUESTION: Did you keep track of ADL people that had the same or different impairments?
KENNETH MANTON: In the sample record, you will have all the survey responses, so you will have the ADL's on both dates, and you can see changes, not only in level but in type of ADL.
KORBIN LIU: We used a simple summated Katz ADL score. The reason we tried it this way was in part, because a lot of nursing home studies had looked at the Katz hierarchy and how it correlated with an ADL sum. For example, Pete Shaughnessy did that at the University of Colorado, with multiple samples of nursing home patients, and found a very high correlation between the two.
KENNETH MANTON: We wanted to look at certain types of intervention scenarios, in terms of health status changes. For that type of projection work, it was useful to have a summary score indicating disability level.
QUESTION: Could you give us any traps or problems with the data that you have come upon that might not be apparent to somebody just starting out?
KENNETH MANTON: We mentioned the date at which the survey is sample frozen or the interview date. You have got to be a bit careful about that because that can have a major impact in terms of interpreting certain types of changes.
There are probably lots of questions in terms of instrumentation. There are 29 medical conditions and they break down in 17 and 12, roughly, in terms of events that occurred within the last 12 months. There was a question 13 where they used a certain re-code list, where there were questions about what conditions caused disability. Some of the re-code lists would have made It difficult, for example, to get hip fractures out from other types of fractures.
The senile diagnosis that is based on proxy response is manifestly cognitive impairment. Other causes are the Alzheimer's related disorder question versus just general cognitive impairment, plus social factors and medication which can cause degradation of cognitive impairment. Trying to disentangle that complexity gets to be a knotty problem.
You have got behavioral problems in there. For certain types of measures that are softer you have to look at multiple indicators. The instrument is rich enough that you do have some consistency checks. For example, out-of-pocket payment at times was a questionable variable, but it seemed to make sense both within the survey, and also in terms of other ancillary data.
We tried to do some things that would determine measures of unmet need. Depending upon how that was defined, we could have broadly different numbers. The incontinence question could mean you increase your numbers by a factor of five or six. Was incontinence an unmet need, in the sense of what we could do to meet that problem? Clearly they had a problem, and there was discomfort in whatever involved. Maybe for urinary incontinence someone could do something but maybe not for bowel. There are questions of definition, if you are talking about new technology or surgical procedures that you could cure the bowel incontinence, so that might be relevant.
I think that, in general, for the purposes to which that was being put, it was decided not to work with the incontinence. First they gave very large numbers, that are a little implausible and the other numbers seemed more comfortable in terms of the unmet need.
KORBIN LIU: Just a couple of more examples. We are working on the out-of-pocket payment questions: did you pay; was this paid care; did you pay any of it? Answer yes: how much did you pay. Blank: did you pay any of it. Blank: then there was a number at the next line. It is the kind of problem that you frequently find in survey responses.
Similarly, with income some parts do not add up to the whole. There were questions on specific income sources and amounts, and there were summary questions. You are going to find variations, if you add up all the individual sources, and then the summary number.
Robyn Stone has been working on the ICS. The NLTCS was conducted by Census, and the ICS was conducted by NORC. Basically, they had samples from the NLTCS sample, and went to identify all caregivers of that particular sample.
When Robyn weighted up the number of caregivers and the number of sample persons from the ICS, she found a fairly large discrepancy between the number of disabled elderly people, with ADL limitations indicated by the weighted NORC survey and the numbers that we found with the NLTCS.
If you are going to be working with both files, and you had notions about merging them, it is definitely worth while to first talk to Robyn and others who have begun working in that file, and to look very carefully at the variations that currently exist from her analysis.
QUESTION: Have you looked at the usefulness of the information and diagnosis in the Medicare files versus the diagnostic information from the survey?
KENNETH MANTON: First of all, diagnosis is primarily on the hospital bills, though there are a certain number of SNF and home health bills where you get some diagnostic information. That information starts really to phase-in during October 1983 with respect to Prospective Payment System (PPS).
Before that, the information on diagnosis is generally spotty. Remembering the way the PPS came in, I believe it was the fiscal date of the hospital that caused it to trigger in during the 12 month period between October 1983 and October 1984. If you look at the distribution of completeness of diagnostic reporting in terms of even the hospital files, it got better during the year.
One can look at the diagnostic information from the bill files, but remember that is for an acute episode, and what you are tracking here are two types of diagnostic information. There are 29 medical condition variables, a certain set of events within the last 12 months, and another set are chronic conditions and then question 13, which is the medical condition which caused the disability with that particular coding.
QUESTION: Is the 1984 community survey similar to the 1982 survey?
KENNETH MANTON: There are just a very few small changes.
QUESTION: Individual caregivers could be identified and longitudinally followed in the same way as sample members?
KENNETH MANTON: My impression is that sample persons can be followed longitudinally and if they have a spouse in both times, then you could probably make an inference.
In terms of having like a common ID number where everything could be exactly matched up, I do not think that is the case. I have not looked at that specific question, but I do not think so.
QUESTION: Why was not there a 1986 survey?
KENNETH MANTON: It would have been desirable to have a 1986 survey and a 1988 survey. When 1982 was done questions were raised about the advisability of doing 1984. That was allowed to happen. In the context of the 1986 round, several other large data collection efforts like the National Medical Expenditure Survey (NMES) were underway. There was a question of budgetary constraints. The 1988 survey is actually being handled through a very different mechanism now under a regular grants program at the National Institute on Aging (NIA.)
You have got the 6 year window. What you lose is the short run changes in the functional dependency.
If you think of something like hip fracture and rehabilitation, I have a feeling that those things which can turn around probably will turn around in 2-4 months. The difference between a 2 and a 4 year loss, in terms of that information, might not be as great as what you would pick up if you did a 3 month reinterview to look at the short term changes.
You are going to miss some intermediate shifts. The recall information will give you part of that, but it will not be like having a continuous time line on service use.
QUESTION: How could state agencies use the national data base to estimate the prevalence of disability in a given state?
KENNETH MANTON: That is the reverse question of how to combine it with demonstration data for localized sites. If I have got Channeling and two experimental interventions, and if that affected service outcomes in certain ways, how do I know what the implications are at the national level, given those intervention effects?
What is needed is joint information and measurements between the demonstration data set and the national data set. If that information is good enough in terms of functional health outcome measures, then in effect you are taking the response that occurred for an individual with a given set of characteristics in the demonstration, and reweighting it to the national distribution from the survey data.
The question now is how can you back down to a state level. If you were willing to assume that the transitions or the service use characteristics for individuals of a certain type in the national survey were going to be similar to what was occurring within a given state, then you could take the responses for individuals with set characteristics in the national survey and reweight back to the state distribution.
There are various strategies whereby you might be able to identify a set of states with similar Medicaid programs. Then you might have more confidence to use combined information from five or six states that are the most similar.
QUESTION: I talked to Peter Kemper and he told me, in fact, that the Maine site really stood out, compared to the other nine channeling sites.
KORBIN LIU: It also gets back to the question that Tom Wan asked about area variations. Just anecdotally, there was a Senate Finance Committee Hearing on Medicare. Senator Long had asked Larry Bartlett, basically, "is there an access problem for Medicare SNF patients?" Larry Bartlett said, "Sir, I think there is because you do not have any Medicare SNF's in the state of Louisiana."
QUESTION: What if we just wanted to look at the 1984 NLTCS for the prevalence of disability and we wanted to somehow apply those estimates to a certain state, just cross-sectionally. Obviously, you would have to adjust somehow for the age and sex composition of a particular state. Beyond that, what would you suggest?
KENNETH MANTON: I would need additional factors to control for. Age, sex, and marital status would be the obvious ones.
Certainly, marital status would be a general one for remaining within the population of who remains within the community or who becomes institutionalized. There might be other factors that you might want to adjust for.
The adjustment could be done in a more or less sophisticated fashion. I mean, the first level of adjustment is like an indirect age, sex, and demographic characteristic standardization. They could go to a higher level of control with some simple regression model. You could get more sophisticated than that, depending upon how detailed you wanted to make things match up.
KORBIN LIU: You might also look at hospital discharge records and the kinds of cases you see. If there is a short term disability associated with stroke and hip fracture, by looking at the hospital records, that may also give you some additional information to estimate the disability levels in the types of disability.
QUESTION: Are the weights for the 1982 survey okay?
KENNETH MANTON: One of the problems is the 1982 survey cross-sectional weights on the new file did not match to the NTIS file. That was corrected, so now they correspond and match. The 1982 survey public use file from NTIS, as far as I know, has no problems. The weighting on that should be correct, so you should be able to use that and get your national population weighted numbers.
One can never guarantee every single variable in that file, but there should not be out-of-range values. When you look at the different age variables, there can be inconsistencies and you have got to come up with a single age variable that is the best compromise in order to produce your weights. The weights, I believe, are clean.
QUESTION: I am actually looking at issues and options in financing long term care. I would like to know some of the new applications that you are working on.
KORBIN LIU: In addition to the pre/post-PPS study which we talked about before, we were about to study the patterns of nursing home use and the effects on financial status of the disabled, and possibly the effects on spouses.
Between the 1982 and 1984 surveys about 900 people went into nursing homes. By the way, if you want to look at who goes in the nursing homes, there is a lot of information. You have got the sample person who goes into a nursing home and then there is a question at the end of the questionnaire that said, "For the sample people who were still in the community in 1984, did your spouse go to a nursing home?" You actually in one sense have doubled the size of the nursing home use population possibly.
Because of all the information that is available on income and assets, you can look at changes in income and assets over time.
QUESTION: What kind of asset information was collected?
KORBIN LIU: It is not as good as we like. The 1982 survey, basically, had a question about house ownership and given house ownership, what do you think the market value of the house is? There is no question at all on liquid assets in 1982. On car ownership, you do not know whether it is a Cadillac, Mercedes or my old Ford Stationwagon. I am not sure that helps you in terms of asset information.
There were questions on source of income, and one of them was, "Did you get income from savings and so forth?" You had to make that heroic assumption that every piece of asset is used as an income generator, and that may not be a safe assumption to make.
In the 1984 community survey, they enhanced the income information. There were questions on stocks, bonds, CD's and so forth, and that was an add-on for the 1984 survey, which I think is an improvement. Generally, I think there are better surveys for income and assets.
QUESTION: Beth Soldo did an analysis on the 1980 NHIS supplement, to look at the receipt of formal care. Have you done anything like that on either the 1982 or 1984 surveys?
KORBIN LIU: I think Beth is still doing that. She is working with Doug Wolf with the 1982 NLTCS on family living arrangements, formal care, informal care and so forth.
QUESTION: Have you looked at home health issues?
KENNETH MANTON: I worked with Tony Hausner in the Office of Demonstrations at HCFA on some issues with respect to home health service use and substitution. We did a 1982 study to look at case mix measures using the home health service users off the 1982 survey file. We were doing some replications looking to see how those case mix measures replicate for 1984.
QUESTION: Have you written up these analyses?
KENNETH MANTON: For the 1982 survey there is a paper that should be appearing in Health Care Financing Review. We have gone past the galley proof stage, I think, and it should be coming out shortly.
QUESTION: Long term care is not the same as long term care for the elderly, for children, or for handicapped people. Are you planning to expand your survey to other types of long term care?
KENNETH MANTON: Not the NLTCS. I have looked at some of the other long term care issues with other data bases.
For example, for the Supplemental Security Income (SSI) program, there was a file that took 12 monthly entry cohorts and looked at Medicaid Type D living arrangements, which are the nursing home use. It was in an administrative record file, so it did not have anything like a rich set of co-variates. They drew a sample out and then followed them for 36 months, but they took the entry cohort for each month during 1982. They got around 10,000 persons a month. They were followed up for 36 months to track people in and out of facilities. You could look at things in terms of reasons for qualifying, change in service use, and mortality. You obviously did not have anything like the rich set of variables that you do in this survey.
QUESTION: How appropriate do you feel these data bases may be for constructing claims costs and loss distribution estimates over a wide range of ages?
KORBIN LIU: It is very hard to do and obviously it is not a direct, insured environment we were talking about. What we were really talking about is some level of need and how that need, after being mediated through a number of other factors, translated into some kind of behavioral response to nursing homes.
I think you have got some of the best national data on the subject. I find it very hard, whether it is an insured or non-insured environment, for somebody who is heavily disabled with incontinence, with a caregiver that is been working constantly for 6 months, not be at high risk of going in nursing homes under any circumstances.
I think that there is some very basic information on risk and length of stay from the three or four data sets, but I was thinking about the NLTCS, the LSOA for predicting risk, and the NNHS for predicting length of stay. Between the two sets of information, risk and length of stay, you can develop a fairly good total liability or total utilization for a cohort, which I think is what you need for designing policies and costing amount.
There are other data bases around, too. There are specific ones for states. As you probably are aware, Connecticut is developing that nursing home inventory and their inventory's been translated into patient level files for length of stay.
If your policies are oriented toward home care as well as nursing home care, then the basic disability/morbidity relationships are going to be extremely important. If somebody is disabled because of epidemiology, then they are going to be requiring some kind of assistance.
There is a lot of information here on the basic biological underpinnings of long term care. That, in itself, should be extremely valuable. You will see variations because of the other factors.
KENNETH MANTON: Associated with that with people by their various risk factors, you will also be able to look at income and assets, although, we talked about assets being somewhat problematic.
If you are talking about a private insurance product, you have marketability, function of ability to pay, and how to cost the thing out. You do have economic ability to pay. You do have a basic need measures. You have the current service utilization for the acute services, at least, and other types of reports of long term care usage, home health service usage, etc. There is a lot of information there.
What you will not know are certain specific behavioral factors in terms of how willing people will be to purchase a given product, that is, what the actual experience under a given type of insured product would be.
I think any time you are starting a new market area, you are going to lack that experience. The best one can do is use the type of data that is represented here.
You can also look at things there in terms of various types of health insurance coverage, at least on a superficial level and coverage under various federal programs so you can see what coverage patterns are like that might also be relevant to filling gaps in service delivery.
KORBIN LIU: I would like to make an unsolicited point. On this question about predicting nursing home admission, it seems to me that we were assuming that a nursing home admission is a nursing home admission is a nursing home admission, and they are not.
People use nursing homes for different reasons. I think Ken has been pointing that out that you have people who go to a nursing home for post-acute convalescent care. You can have people who go to nursing homes for long term care. People who go to nursing homes for terminal care.
My guess is that if you want to try to predict nursing home use, it is important to decide which types of nursing home use you are predicting, because if you are just predicting nursing home use, it is all going to be scrambled. The variables that you are using to predict nursing home use could all wash out because the dependent variable is different.
One thing is to first identify what that nursing home use is and then begin to establish different functions for predicting different types of nursing home use.
QUESTION: How might you be able to differentiate different types of reasons for nursing home entry?
KORBIN LIU: Diagnosis is certainly very important, I think. Basically looking at the 1977 NNHS per se, a chronic brain syndrome had a mean length of stay of about 750 days. Cancer had a mean length of stay of about 39. Diagnosis could be critical to predict long versus short.
You can look at those lengths of stay. About 15 percent of the people in 1982 spent some time in nursing homes between 1982 and 1984. Only half of them were still in nursing homes in 1984. We clearly had some short stay people in there between those 2 years. We can take a look at them, see what kind of diagnoses they had, see what the discharge status was and so forth, and then you get into identifying what kind of a stay it was.