The Health and Retirement Study A Longitudinal Study of Health, Retirement, and Aging Sponsored by the National Institute on Aging

Home Intro/Guide Documentation Data Publications Search

Welcome Co-Inv Data Flow Study Design How to Use Site Map Contact Help Desk What's Available Participants Links

Sample Sizes and Response Rates (2002 and beyond)

Two of the five samples interviewed to date by HRS, and a majority of a third sample, came from a screening of 69,337 housing units conducted in 1992. That sample of housing units was generated using a multi-stage, clustered area probability frame. Of those housing units, 14% (9,419) were determined to be non-sample (unoccupied, or non-households). In all but 214 of the 59,918 identified households, the eligibility of the household members for inclusion in the HRS, AHEAD, or WB samples was determined, for a screening response rate of 99.6%.

The original HRS sample consists of individuals born between 1931 and 1941, inclusive. This sample came from the household screening described above. At the baseline data collection for the HRS sample in 1992, a total of 15,497 individuals were eligible for interviews. This total included persons identified in the household screening, plus their spouses or partners regardless of year of birth. Of those identified in this way, interviews were obtained with 12,654 respondents (7,704 households), for an overall response rate of 81.4%.

The second sample was generated for what began as a separate study: Asset and Health Dynamics among the Oldest Old (AHEAD). This sample consists of individuals born in 1923 or before. Those born between 1914 and 1923, and about half of those born in 1913 or before, were identified through the 1992 household screening operation. The other half of those born in 1913 or before were identified using the Medicare enrollment files maintained by the Health Care Financing Administration (HCFA, since renamed the Centers for Medicare, Medicaid Services, or CMS). For the AHEAD sample, interviews were obtained with 8,222 respondents (6,046 different households), a response rate of 80.4%.

The HRS and AHEAD studies were merged, with a single interview schedule in 1998. At the same time the third and fourth samples were added. The War Baby (WB) sample consists of those born between 1942 and 1947, inclusive, and was obtained from the same 1992 household screening. The Children of the Depression Age (CODA) sample consists of those born between 1924 and 1930 (the 'missing' birth cohorts between the HRS and AHEAD samples). These individuals were identified from the Medicare enrollment file. Since many members of these birth cohorts were already part of the study because they were current or former spouses and partners of those in the HRS and AHEAD cohorts, the new samples excluded those individuals with spouses or partners who were born in 1923 or before, or between 1931 and 1947. The baseline response rates for the CODA and WB samples in 1998 were 72.5% and 70%, respectively

In 2004 a new sample cohort of individuals born in 1948-53 (age 51-56 in 2004) was introduced, which carries forward the steady state aspect of HRS., The Early Baby Boomer (EBB) sample was obtained through the screening of 38,385 households. Eligibility was determined in 91.3% of the screened households, and a total of 4,420 individuals in 2,755 households were found to be eligible. Interviews were completed with 3,340 individuals in 2,159 household for individual and household interview response rates of 75.6% and 78.4%, respectively. Factoring in the screening response rate yields overall baseline response rates of 69.0% for individuals and 71.6% for households.

The ground rules for following baseline respondents and their spouses in subsequent interview waves were as follows:

1. Persons interviewed in the baseline data collection (i.e., in 1992 for the HRS sample, 1993-94 for the AHEAD sample, 1998 for the CODA and WB samples, and 2004 for the EBB sample) will be included in the reinterview sample, regardless of their year of birth and regardless of any change in marital status. The only exceptions are: a) those who had died and for whom exit interviews were obtained in a previous wave; and b) those who have been permanently removed from the sample at their insistence; there were 1593 such cases (out of a total sample of about 32,000) at the end of the 2004 data collection period.
2. Spouses and partners reported by respondents at any previous wave will be included in the reinterview sample, new spouses and partners reported at the time of a reinterview will be added to the sample, and interviews will be sought with them. This includes spouses and partners who declined to be interviewed at the baseline and/or follow-up data collections.
3. For panel members who are found to have died, exit interviews will be sought with a proxy informant who is knowledgeable about the family and financial situation of the deceased. If the deceased is survived by a widow or widower, the exit interview will be sought with the former spouse.
4. If an exit interview has been conducted at a previous wave but the information obtained in that interview was incomplete, generally because the estate had not been settled, a short post-exit interview will be sought to fill in the missing information.

2. Response rates and numbers of interviews

The response rate takes into account exit interviews (conducted for deceased sample members) as well as core interviews. The base on which each response rate is calculated is limited to those in households in which at least one baseline interview was conducted. It excludes deceased respondents for whom an exit interview was obtained in an earlier wave. Information about the number of interviews and the number eligible at each wave is provided in Table 1a, below.

The experience in obtaining the cooperation of sample members in follow-up waves is summarized in Table 1b. The response rates range from a low of 84% to a high of 93%. Across samples, there was a downward trend from 89% to 87% between 1994 and 1998, stability at just below 87% from 1998 to 2002, and then a percentage point increase to 88% in 2004.

The overall response rate at any follow-up wave is a mixture of the response of three major types of individuals: those who participated in the prior wave (referred to as re-interview cases), those who were eligible to participate in the prior wave but did not (referred to as re-contact cases), and new spouses who become eligible for the first time. A complete overview of response rates thus involves the baseline response rates, plus three wave-specific response rates for each follow-up wave. Details on these rates are given in Tables 2, 3 and 4, and described briefly here.

The reinterview and recontact response rates appear to show more stability than the baseline response rates discussed earlier. For the HRS sample, reinterview response rates track in the low to mid-90% range, with a slightly upward trend from the first reinterview (91.8% in 1994) to the sixth follow-up wave (95.1% in 2004). For the AHEAD sample, the reinterview response rates are slightly higher than for those for the HRS sample, and like the HRS sample show a slight upward trend from 93.6% in 1995/96 to 95.9% in 2004. The CODA sample reinterview rates track closely those for the AHEAD sample across the first three follow-up waves, while those for the WB sample are somewhat lower than those of the AHEAD or CODA samples, and somewhat higher than those for the HRS sample.

Recontact response rates are much lower than the reinterview response rates. The recontact rate at the first follow-up wave of the HRS sample was extremely low (8%), but this reflected a decision not to put any effort into trying to recruit the baseline non-respondents (i.e., the non-responding spouses and partners of baseline respondents) out of concern that this could jeopardize the cooperation of their spouses. At the second follow-up, that concern was overlooked and the recontact rate increased sharply, to 38%. There was a rather sharp downward trend across successive waves from 1995/96 to 2000: from 38% down to 24% for the HRS sample, and from 55% down to 31% for the AHEAD sample, followed by an improvement in 2002 for both of these samples. In 2002 we increased our respondent payments and placed special effort in the field on interviewing recontact cases. Preliminary data from 2004 indicate a decline from 2002 to 2004 in the recontact rate for the HRS, CODA and WB samples (Table 3). This reflects in part a policy decision to limit the effort to convert resistant cases because of concerns about losing them permanently, and in part the fact that the recontact cases of 2004 (noninterviews from the more successful 2002 wave) were a particularly resistant group.

Response rates for added spouses and partners are erratic, probably reflecting the rather small number of cases encountered at each wave, but the overall response rate across samples and waves is 66%, considerably lower than the baseline response rates obtained for the HRS and AHEAD samples, and somewhat lower than those for the baseline response rates of the CODA and WB samples (Table 4).

Disposition of Cases across Waves. To date, seven waves of data have been collected from the HRS sample; six waves from the AHEAD sample; four from the CODA and WB samples; and one wave from the EBB sample. In this section we examine the patterns of cooperation of members of the four older samples across follow-up waves.

A total of 13,565 individuals are in the HRS sample: 12,899 in the original sample (95.1%), and the remaining 666 added as new spouses and partners since the baseline interviews in 1992. Over two-thirds (68.9%) of this sample have complete interview histories from their initial entry through 2004 or, for those who died prior to 2004, through an exit interview. The remaining 31.1% have missed at least one interview: an average of 3.0 interviews. The total AHEAD sample is 8,494 individuals, 98.4% of whom were in the original sample. More than four out of five (82.1%) have been interviewed (self or proxy, core or exit) at every wave they were eligible. The remaining (17.9%) missed an average of 2.5 waves. The WB and CODA samples have shorter histories in the study and thus fewer opportunities to have refused to be interviewed. Complete interview histories have been provided by 84.0% of the 2,426 individuals in the CODA sample, and the remaining 16.0% have missed an average of 1.9 interviews. For the 2,745 individuals in the WB sample, all requested interviews have been obtained from 78.8%, and the remaining cases missed an average of 1.7 interviews.

3. Mortality Ascertainment

Tracking. The HRS typically learns of the death of a respondent when an interviewer attempts to reach the respondent for an interview during the main data collection period. The respondent's spouse or another close family member or friend is asked to provide a final interview on behalf of the respondent (the Exit interview). In some instances, HRS staff are notified of a respondent’s death by a family member between waves. Such notification is not considered definitive. Interviewers are alerted to this in the Respondent Profile and the main data collection period is used to confirm the deceased status of the respondent and pursue an Exit interview with a spouse or another close family member or friend. There are also situations in which the respondent is in tracking because we were unable to locate him or her during a previous wave of data collection. Tracking resources for confirming the deceased status of respondents and locating an Exit proxy reporter include the Social Security Death Index (SSDI), which is accessed through the SSDI website and also through a paid subscription to Insight databases (available only to members of a national tracking team which conducts advanced tracking of respondents).

Exit Interviews. HRS has conducted Exit interviews for deceased respondents in every wave starting with the 1995 wave of AHEAD. The Exit interview was developed to find out about the status and activities of the respondent from the time of the last interview until his or her death and to find out about the circumstances of the death. HRS attempts to complete an Exit interview for all deceased respondents, with two exceptions. New spouses for whom a baseline interview was never obtained and respondents who had requested to be removed from the sample prior to their death are excluded from the Exit interview.

The Exit interview is administered to someone knowledgeable about a respondent who died since the last wave, preferably a surviving spouse or close family member and is typically obtained during the field period when the death is first reported. Thus, for most respondents it occurs less than two years following the respondent’s death. If it is not possible to obtain an Exit interview on the first attempt, the family member or proxy is recontacted during the next data collection period. If an Exit interview has been conducted at a previous wave but the information obtained in that interview was incomplete, generally because the estate had not been settled, a short "post-exit" interview is sought to fill in the missing information.

The response rates for the Exit interview have been very high throughout the study. The overall response rate for Exit interviews conducted through 2002 is 93 percent. There appears to be little if any systematic bias in Exit response rates for key demographic groups. Response rates are slightly lower for those born between 1931 and 1947 (the years corresponding with the HRS and WB cohorts) than for those born prior to 1931 (88-89% vs. 94%, respectively), and for non-Hispanic Blacks (90.3%) compared to non-Hispanic Whites (93.4%). By far the strongest predictor of Exit response is interview status in the prior wave. An Exit interview was completed for 95% of respondents who completed a Core interview in the prior wave versus 69% for those who did not.

National Death Index. The HRS conducted linkages to the NDI following the 2000 and 2002 waves, and will conduct another following the 2004 wave. The public-release tracker file is updated with information about which respondents were submitted to NDI, the match status, and the month and year of death for verified matches. Cause-of-death data from NDI+ are available as restricted data only. The NDI data for a given calendar year are available and complete roughly sixteen months after the end of the year. Thus, for the 2000 wave we submitted a finder file in mid-2002 and for the 2002 wave in mid-2004. We submit in the finder file every person whose death has been reported or discovered during field operations for that wave, plus everyone who was not verified as alive by giving an interview in that wave, plus anyone whose death was reported previously but who was not successfully linked in prior NDI searches.

As of the 2002 wave, 5,644 deaths had been reported and submitted to NDI at least once. Of these, 5,366 or 95.1% were successfully matched. By far the most important determinant is whether or not we know the Social Security number of the decedent; the absence of that information lowers the match rate by nearly 11 percentage points. SSN is missing for about 14% of cases; were it available, our match rate would be about 96.5%. The rate for African-American decedents is about 1.5 percentage points lower and that for Hispanic-Americans about 6 percentage points lower than non-Hispanic whites and others. Return migration likely accounts for at least some of the Hispanic differential. The match rate is not systematically related to cohort of birth or to whether or not an exit interview was obtained with a proxy for the decedent.

There were 1,086 HRS participants whose vital status was not determined during the 2002 wave. Of these, 70, or 8.2% were matched to the NDI. Based on the very high match rate of known decedents mentioned above, we believe the number of deaths among those not successfully matched must be very small. There remains the question of whether the NDI matches are accurate given that we have no other verification of death. To address this type of error (false positive matches to NDI) we conducted an experiment after the 2000 wave by submitting 1,000 names of persons who gave live interviews during the 2000 year. Of these, the only successful matches were for persons who had died during the calendar year 2000 but after they had given their interview. Thus, the false positive rate is essentially zero.

Taking these results together, of the combined total of 5,714 deaths, our tracking procedures recovered 98.8% of them, and the NDI recovered 95.1%. Thus, users can be fairly confident of using the HRS tracking data in mortality studies even before the NDI confirmation is complete. Limiting analysis to deaths confirmed by NDI will understate mortality.

Validation of Mortality Rates. The number of deaths found at each wave among the original sample members in each of the four older samples are shown in Table 5. Cumulatively, 16% of the original HRS sample had been found to be deceased by the 2004 wave, based on information from the interviewers and from the National Death Index. Over half (55%) of the AHEAD sample died in the decade between their baseline in 1993/94 and 2004. For the CODA and WB samples, the cumulative proportions who died between their baseline in 1998 and 2004 were 17.9% and 3.5%, respectively.

Table 6 compares the observed cumulative mortality in each of the cohorts comprising the HRS to expected mortality based on national life tables, by sex and race. The HRS data are weighted to reflect differentials in the probability of selection and baseline response rates. The matching to life tables (produced by the National Center for Health Statistics) was done using successive annual life tables by single year of age for the years 1992 through 2002. The only statistically significant differences seen in this table are for the AHEAD sample, for which the observed mortality is lower than the expected mortality for all four subgroups, and significantly so for white males and females. This may be attributable to underrepresentation at the baseline of individuals who are in poor health, in particular due to the exclusion of nursing home residents from the AHEAD sample, but this hypothesis has not been tested. The very high four-year mortality rate in vital statistics for black males in the War Baby cohort is not fully matched in our data. For the remaining samples, however, the correspondence between the observed and expected mortality is reassuring.

4. Number of cases by birth cohort

Each of the five samples is defined in terms of a specific range of birth cohorts, but because interviews are sought with the spouses and partners of selected individuals, the actual range of birth cohorts observed in each sample is considerably broader. If those birth years fall within the range of another sample, then these spouses and partners form part of the sample of their own cohort and are given an appropriate sampling weight. If, however, a spouse or partner’s year of birth lies outside the range of any sample, then he or she is given a zero sampling weight until their cohort is fully represented. Thus, for example, the spouses and partners of HRS sample members born in 1924-30 or in 1942-47 are assigned zero weights in 1992 through 1996, but are given non-zero weights starting in 1998 when their birth cohorts became fully represented with the addition of the CODA and WB samples.

Those in the original HRS birth cohorts (1931-41) constituted 78% of the respondents in 1992, while the remaining 22% were the spouses and partners of those individuals, born in earlier or later years. The number in the HRS birth cohorts dropped by about 900 for the 1993/94 data collection, and as a proportion of all respondents they dropped to just 45% because of the introduction of the AHEAD sample (those born prior to 1924 along with their spouses and partners). In 1998, the CODA (1924-30) and WB (1942-47) samples were added, though at a lower sampling rate than for the HRS sample, and the overall sample, when properly weighted, was representative of the U.S. population born prior to 1948. Table 7 provides details about the distribution of birth years for respondents at each data collection.

5. Self and proxy interviews

Although our goal has been to conduct interviews with the sampled individuals themselves whenever possible, we do permit a proxy interview to be conducted when an individual is unable to do so because of physical or cognitive limitations, and also occasionally when the individual is unwilling to take the time to be interviewed but consents to having someone else (almost always their spouse) be interviewed as their proxy. A substantial number of proxy interviews are in fact done when there is clearly no alternative, following the death of the sampled individual. Most proxy interviews are done because the interviewer finds that the sampled person clearly cannot complete the interview, or is given that message by a gatekeeper. A small number, starting with AHEAD 1995, have been triggered by the low score that a respondent obtains on a test of their cognitive abilities. When that occurs, the interviewer is encouraged (via an interviewer prompt) to either terminate the interview and start again with a proxy informant, or to continue the interview with the help of a care provider, if possible.

The proportion of proxy interviews is higher among older respondents. For example, in the 2002 wave less than 9% of those born in 1931 or later had proxy interviews, compared to 19% of those born before 1924. The proportion of proxy interviews has increased across years, from 5% in 1992 to 11% in 2002, before falling back to 9% in 2004. This rising pattern generally holds within birth cohorts as well, and likely reflects in part the increasing age of the sample members. The decline in proxy interviews in 2004 likely reflects the higher proportion of interviews done face-to-face rather than by telephone. The proportion of interviews done by proxy informants is shown in Table 8 for each birth cohort and for each wave of data collection.)