| HOME | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|---|
| ||||||||||||||||||||||||||||||||
RESEARCH ARTICLE |
a Polisher Research Institute, Madlyn and Leonard Abramson Center for Jewish Life (formerly Philadelphia Geriatric Center), North Wales, Pennsylvania
b Population Council, New York, New York
Vicki A. Freedman, Polisher Research Institute, Abramson Center for Jewish Life, 1425 Horsham Road, North Wales, PA 19454 E-mail: vfreedman{at}abramsoncenter.org.
Decision Editor: Fredric D. Wolinsky, PhD
 |
Abstract |
|---|
 TOP Abstract MethodsResultsDiscussionReferences |
|---|
Methods. Using the 1993 Asset and Health Dynamics of the Oldest Old study (n = 7,443) and 1998 Health and Retirement Survey (HRS; n = 7,624) the proportion of persons ages 70 and older with severe cognitive impairment was calculated under various assumptions about item nonresponse, differential loss to follow-up, and the size and composition of the nursing home population. Impairment was measured for self-respondents using a modified version of the Telephone Interview Cognitive Screen; for proxy respondents ratings of memory and judgment were used. Chi-square tests were adjusted to account for complex survey designs.
Results. Ignoring loss to follow-up, excluding nursing home residents, and assigning a low score to those refusing subscales yielded a statistically significant decline in severe cognitive impairment from 5.8% in 1993 to 3.8% in 1998, or an average annual decline of 6.9%. When cumulative effects of survey design issues were considered and design effects incorporated into statistical tests, statistically significant declines persisted, albeit at a reduced average annual rate, ranging from 2.5% to 6.9% per year.
Discussion. Previously reported improvements in severe cognitive impairment appear to be robust to a variety of specifications. Replication with future waves of the HRS and other data is warranted.
IN a recent article (Freedman, Aykan, and Martin 2001
) we showed declines in the percentage of Americans aged 70 and older with severe cognitive impairment from 6.1% in 1993 to 3.6% in 1998, or an average annual decline of 8% per year. We tested and rejected the hypothesis that changes in the demographic and socioeconomic composition of the older population explained the improvement. Our analysis drew on the first wave of the Asset and Health Dynamics of the Oldest Old (AHEAD93) and the fourth wave of the Health and Retirement Study (HRS98), two national surveys of older Americans.
In that article, we (Freedman et al. 2001) speculated about a number of potential causes for the improvements we observed, including the dramatic increases in the education levels of more recent cohorts of elderly adults; earlier detection and treatment of dementia and related comorbid conditions such as depression, strokes, and circulatory and cardiovascular diseases; and improved behavioral and lifestyle factors such as dietary habits, vitamin supplementation, and exercise. We also raised numerous survey design issues that could potentially confound these results and, when possible, attempted to quantify those effects. We discussed, for example, the change in preferred mode of interviewing for the youngest respondents, potential learning effects among HRS98 respondents, differences in proxy recruitment, item refusals, potential differential loss to follow-up in the HRS98 by cognitive status, and potential shifts in the institutionalized population outside the scope of the surveys. For three of these issues—mode changes, learning effects, and proxy recruitment—we argued there would likely be small, if any, effects. For the remaining three issues, we presented or referred to sensitivity analyses demonstrating that the declines in severe cognitive impairment persisted even under reasonable to conservative assumptions.
A weakness of that analysis is that the authors (Freedman et al. 2001) raised and dismissed each of the three survey design issues separately, rather than focusing on their cumulative effects. It is possible, for example, that no single issue negates the findings, but that the combined effects could offset improvements completely. Moreover, probably the most troubling survey design issue—potential differential loss to follow-up—could not be fully investigated with the data available at the time the article was written. Consequently we made assumptions about the extent of differential loss to follow-up by cognitive status on the basis of the first and second waves of AHEAD. An additional weakness was that we did not have access to sampling information for the HRS98 so we instead made assumptions about the likely size of the design effects on the reported test statistics.
As reported in a note at the end of that article, revised weights and additional information on loss to follow-up became available just prior to publication, making it possible to more fully explore differential loss to follow-up. In that note we reported that the new data yielded declines in severe cognitive impairment from 5.8% in 1993 to 3.8% in 1998 and that preliminary sensitivity analyses suggested these findings were robust to a variety of assumptions. We did not, however, fully develop our exploration of the cumulative effects of survey design issues on estimates of aggregate changes in cognitive impairment. Since that time information on sampling units also has been made available, so the implications of the survey's complex design for statistical tests can now be addressed more accurately.
A growing number of articles have considered issues of missing data in the context of survey design on point estimates and statistical tests. In their work on missing data, for example, Little and Rubin 1987 have highlighted the important distinction between data that are "missing at random" or "not missing at random" and have discussed the implications for point estimation and appropriate variance calculations. In this particular application, older persons refusing to answer cognition questions, lost to follow-up, and in nursing homes that lay outside the scope of the survey represent three different sources of uncertainty, each of which likely involves data that are not missing at random. For example, Herzog and Wallace 1997 have shown those refusing to answer certain subscales scored worse on other cognitive items. Similarly, as reported in our earlier work (Freedman et al. 2001), those respondents lost to follow-up by 1998 were more likely to have severe cognitive impairment at baseline. Moreover, older persons in institutions undoubtedly have higher rates of severe cognitive impairment, with published estimates reaching over 10 times those reported for the community-based older population (Rovner and Katz 1993; Magaziner et al. 2000).
The purpose of this article was to investigate whether our previously reported results (Freedman et al. 2001) are robust to assumptions made about missing data. The effects of three potentially confounding sources—item nonresponse, differential loss to follow-up, and the exclusion of the nursing home population—are considered simultaneously. Statistical tests that take into account the complex sample designs of the AHEAD and HRS are also presented.
|
Methods |
|---|
|
TOPAbstractMethodsResultsDiscussionReferences |
|---|
hrswww/. In the initial AHEAD and HRS waves, and in the cohorts added to the HRS98, households were selected on the basis of a multistage area probability sample plan, with oversampling of Blacks, Hispanics, and Florida residents. Sampling weights provided by the University of Michigan compensate for the differential probabilities of selection and are poststratified to the Current Population Survey by age, sex, race, ethnicity, and marital-status groups to account for differential nonresponse and loss to follow-up. The most recent version of the weights (available in the Tracker 2.0 file and used with all estimates presented here) involved identical procedures in constructing the AHEAD93 and HRS98 weights.
We used the public release version of the AHEAD93 and the preliminary release of the HRS98. Both surveys were funded by the National Institute on Aging and conducted by the University of Michigan. The HRS98 data have not been cleaned and may contain errors that will be corrected in the final public release version of the data set.
Measures
Self-respondents were asked to complete a variety of items, which combine into four subscales: immediate recall of 10 words, delayed recall of 10 words, subtracting 7 from 100 five consecutive times, and answering several other questions (e.g., naming the president and vice president, counting backward from 20 by 1). Respondents answering the survey themselves were considered severely cognitively impaired if they scored 8 or less out of a possible 35. Those unable to complete the survey had a close family member or friend report about their memory and judgment. We classified those reported as having poor memory and poor judgment as severely impaired.
Scenarios
We varied our assumptions about three different sources of missing data, presenting the cumulative effect of each subsequent source on estimates of aggregate changes in the percentage of older Americans with severe cognitive impairment. Rather than model the missing data or adopt complex imputation strategies, which we could not do for the nursing home population and which seemed unwarranted given the small amount of missing data, our overall strategy was to impute refused information in two different ways and to adopt a range of reasonable to conservative assumptions about the percentage with severe cognitive impairment among those lost to follow-up by 1998 and nursing home residents in both years.
First, we varied the treatment of refusals by self-respondents. In both years, a modest proportion of self-respondents (11.7% in 1993 and 10.0% in 1998) refused to answer one or more of three subscales—immediate word recall, delayed word recall, and subtracting 7 from 100 several times. The majority of these self-respondents had scores above the threshold of 8 and thus did not require imputations, but 3.2% of respondents in 1993 and 1.9% in 1998 refused and had scores of 8 or less. Here we present two scenarios, both of which assume data are not missing at random. In the baseline scenario we assigned a low score based on the algorithm proposed by Herzog and Wallace 1997(and used by Freedman et al. 2001) to those refusing a given subscale. In the second scenario, we assigned a zero for those refusing a given subscale.
We then added the effects of differential loss to follow-up under several assumptions, again assuming in each case that the data are not missing at random. Three scenarios are presented, each assuming progressively more severe cognitive impairment in 1998 among the members of the AHEAD93 cohort lost to follow-up (6.27% of the original 1993 cohort). First, the rate of severe cognitive impairment in 1998 among those lost to follow-up by 1998 was assumed to be the same as observed for this group in 1993—4.25% or 5.30% depending on the treatment of refusals. Second, we allowed incidence and recovery rates to be the same as those we observed for respondents appearing in both 1993 and 1998, yielding impairment rates for those lost to follow-up of 5.74% and 6.55%. Third, we assumed that incidence was twice the rate of that observed for respondents in both 1993 and 1998 and that there was no recovery, yielding impairment rates for those lost to follow-up of 12.39% and 14.22%.
Note the approach we adopted in assessing loss to follow-up is conservative for two reasons. First, our adjustments essentially ignore the fact that the sampling weights used in the analysis already adjust for loss to follow-up on the basis of age, sex, marital status, and race/ethnicity. Such weights should already account for at least some of the differential loss to follow-up by cognitive status. Second, we did not take into account the very small amount of loss to follow-up among spouses of original HRS cohort members who would have been age 70 or older in 1998. The rate of severe cognitive impairment among this group is likely to be lower than average, given that the overwhelming majority were in their early 70s and all were married.
Finally, we added the effects of potential shifts in the prevalence of severe cognitive impairment among the institutionalized population. The AHEAD93 excluded respondents living in nursing homes. To ensure comparability between the 1993 and 1998 surveys, Freedman and colleagues 2001 excluded from the HRS98 data anyone who had entered a nursing home. Investigating the sensitivity of results to the exclusion of the nursing home population required further assumptions about the size of the nursing home population and about the percentage with severe cognitive impairment within that population. From our reading of the literature based on nursing home surveys from 1985 and 1995 (Dey 1997; Hing, Sekscenski, and Strahan 1989), we found that the nursing home population has declined as a percentage of the older population. Using simple linear interpolation and linear projecting through 1998, we calculated 5.58% in 1993 and 5.27% in 1998 as our estimates of the proportion of the 70-and-older population in a nursing home. The percentage with severe cognitive impairment was more difficult to pinpoint, and may have increased because of an increase in special care units or declined because of the increase in post-acute care being provided in nursing homes. We have seen published estimates of severe cognitive impairment in the 60%–70% range for the mid-1990s based on local samples of nursing homes (Rovner and Katz 1993; Magaziner et al. 2000). We present two scenarios, first assuming that 60% of nursing home residents have severe cognitive impairment in both years, and then assuming a conservative increase from 55% to 65% nationally over this period.
During this same period, the size of the assisted living population—although still relatively small (see, e.g., Hawes, Rose, and Phillips 1999)—undoubtedly also grew. The lines are often blurred between community, assisted living, and nursing home. The AHEAD93 included respondents living in small assisted living facilities but potentially excluded those respondents in facilities that qualified as or were part of larger institutions. The HRS98 included respondents who entered into assisted living facilities (of any size). Assuming older persons in assisted living were more likely to be cognitively impaired than were those living in the community, the exclusion of some assisted living residents in 1993 would bias results against finding an improvement in cognitive impairment. We therefore omitted consideration of this issue from the following analysis.
Statistical Tests
We calculated chi-square statistics for the difference between 1993 and 1998 in the percentage of Americans aged 70 and older with severe cognitive impairment. The chi-square statistics are presented two ways: first without design effect adjustments and second with design effect adjustments that compensate for clustered sampling, weights, and the overlap in cases between the samples. We calculated the design effects in two steps. First, we calculated the ratio of the Pearson chi-square statistics under complex design and simple random-sample assumptions. We obtained the complex design chi-square from SUDAAN (which compensates for weighting and hierarchical sampling; Research Triangle Institute 1992) and the simple random sample chi-square from SAS (SAS Institute 1999). This produced a design effect of 1.42. Second, on the basis of the formula proposed by Kish 1965(pp. 457–458) for variances of differences between two overlapping samples, we adjusted this design effect to compensate for the 4,812 cases appearing in both samples. The final design effect, which compensates for weighting, hierarchical sampling, and the overlap of the samples, was 1.19.
|
Results |
|---|
|
TOPAbstractMethodsResultsDiscussionReferences |
|---|
|
How are the aggregate changes for all age groups combined affected by alternative assumptions about missing information? In the baseline scenario, we observed an average annual percentage decline of 6.9% (see Rows 1 & 2 of Table 2 ). When those refusing to answer were instead assigned no points for the refused subscales, the decline was larger in percentage points terms (from 6.67% in 1993 to 4.38% in 1998), but was still an average annual decline of 6.9%.
|
Adding in the nursing home population further attenuated results (see Rows 9–20 of Table 2 ), in part because of the increase in the population base, but declines remained substantial and statistically significant. When the composition of the nursing home population with respect to severe cognitive impairment was assumed to be stable over time, declines were in the 3.6% to 4.7% per year range. As shown in the bottom panel, when we assumed the nursing home population had become more cognitively impaired over time, declines were in the 2.5% to 3.7% range. Under the worst-case differential loss to follow-up scenario that we considered and the more conservative item-nonresponse scenario, the percentage of severely cognitively impaired older persons including nursing home residents declined by 2.5% per year.
|
Discussion |
|---|
|
TOPAbstractMethodsResultsDiscussionReferences |
|---|
). This analysis does not—and, indeed, cannot—take into account all survey design issues that could be influencing rates of severe cognitive impairment. Even seemingly minor changes in fieldwork protocols could potentially influence the types of survey measures analyzed here. Four such issues that have come to our attention are worth mentioning briefly.
First, the AHEAD93 included a single list of words for all respondents to recall immediately and after a short delay, whereas the HRS98 used four lists, one of which was randomly assigned to each respondent. This modification was intended to make it more difficult for spouses to learn by overhearing one another; hence we would expect its introduction to make the recall tests more difficult in 1998. Moreover, in our earlier work (Freedman et al. 2001), we observed improvements on all subscales, not just the immediate and delayed recall, so it is unlikely this change is confounding the results presented here.
Second, it could be that selection into the proxy group has somehow changed over time in relation to cognitive functioning of respondents. However, as we reported in 2001, the proportion of proxies was lower in 1998 than in 1993. Such a change is consistent with (although not definitive evidence of) improvements in cognitive functioning.
Third, because a substantial number of persons in the HRS98 had been exposed to the cognitive tests in earlier waves, learning effects are of some concern. That is, exposure to prior tests could potentially influence the scores of those with relatively good cognitive functioning, if they practice during the interim or in anticipation of the follow-up interview. However, it is difficult to believe that learning effects are substantial enough to influence measures of the type of severe cognitive impairment reported here. Can someone who is unable to count backward from 20 or unable to name an object used to cut paper remember that they had difficulty with specific items 2 years prior and then plan and practice those items to improve their performance at a subsequent interview? We think this is unlikely except in the rarest of circumstances.
Finally, a change in the preferred mode of interview—from telephone to face-to-face—among the youngest respondents raises concerns about whether respondents might be able to cheat on the phone and thereby perform better. It is possible, for example, for respondents on the telephone who have relatively good cognitive functioning to jot down words or use pencil and paper to subtract. However, it is again difficult to believe that this type of mode effect could influence the severe cognitive impairment measured here. Can someone with severe cognitive deficits plan and execute the steps necessary to cheat while participating in an interview? Again, we remain skeptical about the size of such mode effects for these measures. Moreover, experimental evidence from the second and third waves of AHEAD suggests mode effects are minimal for cognitive performance (Herzog and Rodgers 1999).
Irrespective of the robustness demonstrated here, caution in interpreting these results is still warranted. The data are considered preliminary, and although no major changes are anticipated, it is possible they will not match final public release files. In addition, we have not formally incorporated into the statistical tests the extent of uncertainty we have about severe cognitive impairment among refusals, those lost to follow-up, and nursing home residents. Multiple imputation strategies discussed by Rubin 1987 and others would be necessary to further refine the chi-square test statistics presented here. However, given the relatively small amount of missing data and the wide range of plausible and conservative assumptions we have investigated, we believe it is unlikely that a multiple imputation strategy would substantially change the conclusions of our results.
More importantly, because the data contain measures for only two time points, they are merely suggestive that improvements may be occurring over this time period. Certainly, replication of the results presented here with future waves of the HRS and other data sources is in order before one can draw firm conclusions that there is a trend toward improvements in cognitive functioning among older Americans.
|
Acknowledgments |
|---|
Received for publication March 2, 2001. Accepted for publication June 22, 2001.
|
References |
|---|
|
TOPAbstractMethodsResultsDiscussionReferences |
|---|
This article has been cited by other articles:
 |
|
 |
|
  M. G. Parker and M. Thorslund Health Trends in the Elderly Population: Getting Better and Getting Worse Gerontologist, April 1, 2007; 47(2): 150 - 158. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 V. A. Freedman and L. G. Martin Commentary on "Trends in Scores on Tests of Cognitive Ability in the Elderly U.S. Population, 1993-2000": Beyond Inconsistent Results: Finding the Truth About Trends in Late-Life Cognitive Functioning J. Gerontol. B. Psychol. Sci. Soc. Sci., November 1, 2003; 58(6): S347 - 348. [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||
| HOME | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|---|
