Ethnic/race differences in the attrition of older American survey respondents: implications for health-related research.
Objective. To compare models of attrition across race/ethnic groups of aging populations and discuss implications for health-related research.

Data Sources. The Health and Retirement Study (1992-2008).

Study Design. A competing risks model was estimated using a multinomial logit model when respondents faced competing types of risks, such as dying, being lost from the study, and nonresponse in some years for different groups of elderly. Key explanatory variables were foreign birth, health insurance, and health status. Principal Findings. Variables describing foreign birth, health insurance, and health status differed in their prediction of attrition across ethnic groups of aging populations.

Conclusions. Differences in the predictors of attrition across ethnic groups of elderly could potentially lead to biased estimates in health-related research using longitudinal data sources.

Key Words. Attrition, race, ethnicity, health, insurance, health status

Aged (Surveys)
Aged (Research)
Health insurance (Research)
Risk assessment (Research)
Gerontology (Research)
Zhivan, Natalia A.
Ang, Alfonso
Amaro, Hortensia
Vega, William A.
Markides, Kyriakos S.
Pub Date:
Name: Health Services Research Publisher: Health Research and Educational Trust Audience: Trade Format: Magazine/Journal Subject: Business; Health care industry Copyright: COPYRIGHT 2012 Health Research and Educational Trust ISSN: 0017-9124
Date: Feb, 2012 Source Volume: 47 Source Issue: 1
Event Code: 310 Science & research
Product Code: 6322000 Medical Care Insurance; 6320000 Accident & Health Insurance; 9912200 Venture Analysis NAICS Code: 524114 Direct Health and Medical Insurance Carriers; 5241 Insurance Carriers SIC Code: 6321 Accident and health insurance
Accession Number:
Full Text:
A large literature is devoted to the investigation of race/ethnic health disparities relying on data analysis. While the minority population is growing (U.S. Census Bureau 2008; Arias 2010), quite often past and current data have small samples of Mexican Americans, Latinos including Mexican Americans, blacks, and others, especially when stratified by basic demographic characteristics, such as age and gender (Dunlop et al. 2002). Representativeness of the participating respondents of the population of interest and nonrandom loss from the study are one of the biggest concerns leading to bias estimates and nongeneralizability of some results (Hausman and Wise 1979; Diehr and Johnson 2005). While representativeness of the sample is difficult to test due to lack of information about nonparticipants, attrition has been analyzed and addressed in some studies on health disparities. However, none of the studies have compared models of attrition across different ethnic and racial groups of aging populations. Using the Health and Retirement Study (HRS) 1992-2008, this study compares models of attrition for Latinos including Mexican American, Mexican American separately, black, and white elderly Americans. The main hypothesis is whether variables describing foreign birth, health insurance, and health status are associated with attrition and whether these associations differ across racial/ethnic groups of elderly.

This study falls into the strand of literature about determinants of attrition among elderly individuals (Mihelic and Crimmins 1997; Van Beijsterveldt et al. 2002; Chatfield et al. 2005; Kapteyn et al. 2006; Tyas et al. 2006; Stimpson and Ray 2010). A recent literature review of attrition among elderly found that age, cognitive impairment, and poor health were associated with high attrition rates (Chatfield et al. 2005). However, none of the studies provided a comparison of determinants of attrition across different races and ethnicities. Such attrition is essential to the assessment of race/ethnic health disparities. Kapteyn et al. (2006) analyzed attrition in the HRS and found that being born outside of the United States, being Hispanic, being in poor/fair health, or having the onset of a health condition was associated with a higher probability of attrition than remaining in the HRS, depending on the type of model estimated. However, their analysis did not address the question of whether Hispanics and blacks with certain characteristics, for example, being in poor health, were more likely to attrite compared to similar whites. Rather than estimating a separate model for each race/ethnicity or interacting race/ ethnicity with other covariates, the authors estimated a multinomial logit model for the pooled sample of whites, blacks, and Hispanics while introducing binary variables for each race/ethnicity category. This approach imposes a "common-effect" assumption when effects of other covariates on attrition are the same across different groups of elderly, which is relaxed in our study (Jha et al. 2007). Documenting differences in the attrition by race/ethnicity is also relevant to the literature on the "Hispanic paradox," a concept that describes a mortality advantage among Hispanics compared to similar whites and African Americans (Markides and Coreil 1986). Among other explanations, some researchers have argued that this phenomenon is a result of the selective return migration to the country of origin by older people in poor health, the so-called salmon bias hypothesis (Markides and Eschbach 2005). However, we cannot test this hypothesis due to lack of data on migration.

A competing risks model was estimated using a multinomial logit model when respondents of the study faced competing types of risks, such as dying, being lost from the study, and nonresponding in some years (Kapteyn et al. 2006). That is, when respondents entered the HRS in 1992, some of the participants were still present in 2008, the latest year available. However, some of the respondents died, others were in and out of the study during this period, while the rest of the respondents were lost from the study. Only one event/failure, such as death, nonresponse, or loss, can take place exclusively to the others with some probability, which defines a "competing risks" situation. The main assumption is the independence of risks of each event/failure type conditional on explanatory variables. The key explanatory variables are foreign birth, health insurance, and health status.

Given the substantial financial and human resources dedicated to the issue of racial/ethnic health disparities, this study may assist in developing future studies on health disparities in aging populations. Moreover, differential attrition by race/ethnicity in panel surveys has significant negative implications for health research in general due to growing ethnical diversity of the U.S. population.


Data Source and Sample

This study analyzed data from the HRS. The HRS is the most current, comprehensive, longitudinal study of aging (Juster and Suzman 1995). The HRS is a multistage clustered area probability sampled study of 51- to 61-year-old individuals that started in 1992 by collecting information about 12,600 people born between 1931 and 1941 and their spouses. These data were supplemented with individuals born before 1923 as part of the Study of Asset and Health Dynamics among the Oldest Old (AHEAD) in 1993 and combined with younger cohorts in 1998 and 2004. All individuals were re-interviewed every 2 years with a response rate of more than 80 percent across all waves Ouster and Suzman 1995; Institute for Social Research 2010). The HRS contains detailed information on demographics, health status, and health care utilization that is uniform across all ethnic/racial groups and perfectly suited for the analysis of race/ethnic differential in attrition.

To maximize the Latino sample, our analysis includes all individuals from the HRS who were observed at age 51 or older resulting in a sample of 1,492 Mexican Americans, 2,487 Latinos including Mexican Americans, 4,367 blacks, and 21,845 whites. Table 1 reports the distribution of the birth cohort in the sample. Samples of whites, Mexican Americans, Latinos, and blacks have 1.5, 2.5, 2.2, and 2.4 percent of observations with missing values for one or more variables respectively and were excluded from the analysis.


The outcome variable is the vital status of each HRS respondent in 2006. The status of each respondent can fall in four exclusive categories. We labeled "always in" those respondents that responded in each wave up to 2006; "died" those respondents who were reported dead between the first wave they were interviewed and 2006; "ever out" those respondents who did not respond in some years; and "attritor" those who entered the study at some point but were continuously absent from the study up until 2008. Since 2008 data are already available, we used the status in 2008 to determine whether nonresponding individuals in 2006 or earlier were attritors or if they were back in the study in 2008.

The set of explanatory variables included continuous variables such as age, age squared, and binary variables for being female, married, for not completing high school, completing high school, having some college or higher, being foreign born, being born before 1931, and being born after 1941. We also controlled for lack of health insurance and being on Medicaid. Health status was described by the number of self-reported health conditions (high blood pressure, diabetes, cancer, heart problems, stroke, lung problems, and arthritis) and number of activities of daily living (ADLs) (dressing, walking, bathing, eating, getting in and out of bed, and using the toilet). Since severity of health conditions may vary significantly across individuals and ethnicities, we controlled for self-reported health with a binary variable for reporting poor or fair health status. Controls for financial well-being were measured by the log of household income in the HRS.

Analytic Approach

We estimated a multinomial logit model of the status of each respondent by 2008 as a function of baseline characteristics--characteristics when observed for the first time--of those who are 51 and older across all groups of elderly combined and later by race/ethnicity to determine whether predictive effects of covariates vary across race/ethnicity. We also tested our model to insure that it satisfies the Independence of Irrelevant Alternatives assumption using the Hausman test (Hausman 1978; Long and Freese 2006). Since the sample of Mexican Americans in the HRS includes only 168 attritors total across seven waves, we did not estimate wave-by-wave determinants of transition across race/ethnicity due to significant loss of power.

Table 1 reports summary statistics and sample sizes. Table 2 reports relative-risk ratios and 95 percent confidence intervals by race/ethnicity, while Table 3 reports predicted probabilities for selected variables based on race/ ethnic-specific models.


Table 1 provides summary statistics of the key variables, sample size, and percent of people who died, ever were out, or were lost by race and ethnicity from the HRS between 1992 and 2006. Compared to whites and blacks, Latinos were younger, with fewer years of schooling, more likely to be foreign born, and less likely to have health insurance. Blacks and Latinos were more likely to be on Medicaid compared to whites. Blacks had a greater number of health conditions, while the number of ADLs was most prevalent among blacks and especially Mexican Americans despite their younger age. According to Table 1, 22 percent of whites, 24 percent of blacks, and only 17 percent of Mexican Americans died by 2006. Surprisingly, Mexican Americans had the lowest percent of respondents being lost from the study.

Table 2 presents estimates for the multinomial logit model predicting the relative-risk ratios for being dead, ever out, or being an attritor by 2006 compared to being "always in" as a function of baseline characteristics. The first column reports the result for the combined sample of white, black, Latino, and other race elderly populations. Consistent with the well-documented "Hispanic paradox" literature, Hispanics faced a smaller probability of dying compared to whites. Foreign born were also less likely to be attrited due to death. As we would expect, lack of health insurance, greater number of health conditions, and ADLs were associated with greater probability of dying, while higher income protected from greater mortality. Blacks, Latinos, those who were foreign born, and those without health insurance were more likely to nonrespond in some waves; that is, to be "ever out." Surprisingly, estimates of the combined sample shows that Mexican Americans were less likely to attrite from the study. On the other hand, foreign born were more likely to attrite. As it was previously documented, those in poor health were more likely to attrite than to remain in the study, although respondents having an additional health condition or reporting being in fair/poor health were more likely to die than to attrite.

Columns 2 through 5 report estimates of a similar model by race/ethnicity. Foreign-born Latinos had a smaller probability of dying, while lack of health insurance was associated with greater probability of dying for whites and blacks but not for Mexican Americans and Latinos in general. While the number of health conditions and self-reported poor/fair health was associated with higher probability of dying across all groups, the number of ADLs was significant only for whites and African Americans. The number of health conditions and ADLs was a significant predictor of being an attritor across all groups of aging populations as well. Surprisingly, being foreign born was not significant for Mexican Americans, while it was significant for others. However, Mexican Americans with no health insurance had a greater probability of dropping out of the study.

Interestingly, separate models demonstrated that chances of attrition associated with health conditions were greater for Mexican Americans than for white counterparts. That is, according to estimates, Mexican Americans having an additional health condition had almost similar chances of dying and dropping out of the study. However, white elderly having an additional health condition had greater chances of dying than being lost from the study. Different formatting (boldface, italics, underlining) indicates whether relative-risk ratios for becoming an attritor were statistically different from the relative-risk ratios of dying. For whites and African Americans these ratios are significantly different, while for Mexican Americans and Latinos in general these ratios are statistically the same.

To facilitate interpretation of the results, Table 3 presents predicted probabilities for selected variables based on race/ethnic-specific model estimates that were applied to the whole sample while setting a variable of interest to a particular value. Thus, a foreign-born Latino had a 19.3 percent probability of being an attritor and 15.5 percent probability of being dead, while for a non-Latino white the corresponding probabilities were 15.9 and 20.4 percent. Across all groups of elderly, foreign birth was associated with greater probability of attrition than death. Lack of health insurance was associated with about 25 percent probability of death for whites and blacks, while for Mexican Americans and Latinos these probabilities were smaller. However, a Mexican American without health insurance had a 19 percent probability of becoming an attritor; for similar non-Latino whites this probability was only 11.6 percent. Medicaid status was associated with lower probability of attrition for all groups of elderly. Having two health conditions led to significantly greater probability of death relative to probability of attrition for non-Latino whites. For Latinos and Mexican Americans these probabilities statistically were the same.


This study compares models of attrition across different racial/ethnic groups to demonstrate some implications of growing diversity of the U.S. population for health-related research and assist in planning of future studies of health disparities in aging populations. Variables describing immigration status, insurance, and health status differed in their prediction of attrition across minority groups of elderly that may lead to biased estimates of the relationships under investigation. Our findings suggest that work examining racial/ethnic disparities in mortality, health decline, and disability onset using the HRS was likely to underestimate health disparities due to higher relative attrition of Latinos in poor health measured by the number of health conditions and ADLs (Sudano and Baker 2006; Dunlop et al. 2007). Sudano and Baker (2006) found that lack of health insurance was responsible for small fraction of disparities in health outcomes that could be explained by the lack of attrition adjustment in their work. Similarly, our findings suggest that work debating about the relationship between lack of health insurance and mortality is likely to be sensitive to attrition process by health insurance status, particularly for Mexican Americans (McWilliams et al. 2004; Kronick 2009). McWilliams et al. (2004) found that while the lack of health insurance was associated with significantly higher mortality rate among whites, there was no mortality differential for Hispanics referring to Hispanic paradox as an explanation. Ethnic differences in the attrition processes by health status and health insurance could explain their finding as well. Due to growing ethnic diversity of the U.S. population, researchers would be advised to pay a closer attention to potential ethnic/ racial attrition differentials in health-related research.

This study has several limitations. The sample of Latinos in the HRS was still small, forcing us to combine several cohorts to conduct analysis. Due to the small sample of Latinos, we were unable to test interactions of covariates and ethnic/racial binary variables to test statistical significance of differences in the effects of covariates across races/ethnicities. The small sample size problem may also have been responsible for some of the nonsignificant results for Latinos. As a share of minority population becomes more prevalent, findings of our study can be validated in the future. Larger representation of the minority population in surveys will allow additional testing of our model.

To summarize, there are important differences in the attrition process across different groups of aging minorities. Our study indentified respondent characteristics that may improve attrition adjustment in health-related research and guide future effort to retain minority population in panel surveys.


Joint Acknowledgment~Disclosure Statement: Natalia A Zhivan and Alfonso Ang are supported by the Network for Multicultural Research on Health and Healthcare, Department of Family Medicine, David Geffen School of Medicine, UCLA, funded by the Robert Wood Johnson Foundation. No conflict of interest exists. Each author has made substantive intellectual contributions to the study.

Disclosures: None.

Disclaimers: None.


Arias, E. 2010. "United States Life Tables by Hispanic Origin." Vital and Health Statistics 2 (152): 1-33.

Chatfield, M.D., C.E. Brayne, and EE. Matthews. 2005. "A Systematic Literature Review of Attrition between Waves in Longitudinal Studies in the Elderly Shows a Consistent Pattern of Dropout between Differing Studies." Journal of Clinical Epidemiology58 (1): 13 9.

Diehr, P., and L.L. Johnson. 2005. "Accounting for Missing Data in End-of-life Research." Journal of Palliative Medicine 8 (suppl 1): S50-7.

Dunlop, D.D., L.M. Mahneim, J. Song, and R.W. Chang. 2002. "Gender and Ethnic/ Racial Disparities in Health Care Utilization among Older Adults." Journal of Gerontology: Social Sciences 57B (3): S221-33.

Dunlop, D.D., J. Song, L.M. Mahneim, M.L. Daviglus, and R.W. Chang. 2007. "Racial/Ethnic Differences in the Development of Disability among Older Adults." American Journal of Public Health 97 (12): 2209-15.

Hausman, J. 1978. "Specification Tests in Econometrics." Econometrica 46 (6): 1251-71.

Hausman, J., and D.A. Wise. 1979. "Attrition Bias in Experimental and Panel Data: The Gary Income Maintenance Experiment." Econometrica 47 (2): 455-73.

Institute for Social Research. 2010. University of Michigan [accessed on March 29, 2010]. Available at

Jha, A.K., D.O. Staiger, F.L. Lucas, and A. Chandra. 2007. "Do Race-Specific Models Explain Disparities in Treatments after Acute Myocardial Infarction?" American Heart Journal 153 (5): 785-91.

Juster, F.T., and R. Suzman. 1995. "An Overview or the Health and Retirement Study." Journal of Human Resources 30: S7-56.

Kapteyn, A., P.-C. Michaud, J.P. Smith, and A. van Soest. 2006. "Effects of Attrition and Non-response in the Health and Retirement Study." RAND Working Paper 407. Santa Monica, CA: RAND Corporation.

Kronick, R. 2009. "Health Insurance Coverage and Mortality Revisited." Health Services Research 44 (4): 1211-31.

Long, J.S., and J. Freese. 2006. Regression Models for Categorical Dependent Variables Using Stata. College Station, TX: Stata Press, 2d Edition.

Markides, K.S., and J. Coreil. 1986. "The Health of Southwestern Hispanics: An Epidemiologic Paradox." Public Health Reports 101: 253-65.

Markides, K.S., and K. Eschbach. 2005. "Aging, Migration, and Mortality: Current Status of Research on the Hispanic Paradox." Journals of Gerontology: Psychological and Social Sciences 60B: 68-72.

McWilliams, J.M., A.M. Zaslavsky, E. Meara, and J.Z. Ayanian. 2004. "Health Insurance Coverage and Mortality among the Near-Elderly." Health Affairs 23 (4): 223-33.

Mihelic, A.H., and E.M. Crimmins. 1997. "Loss to Follow-up in a Sample of Americans 70 Years of Age and Older: The LSOA 1984-1990." Journal of Gerontology: Series B, Psychological Sciences and Social Sciences 52B: S37-48.

Stimpson, J.P., and L.A. Ray. 2010. "Attrition of Older Mexican American Survey Respondents." Journal of Immigrant and Minority Health 12 (3): 414-7.

Sudano, J.J., and D.W. Baker. 2006. "Explaining US Racial/Ethnic Disparities in Health Declines and Mortality in Later Middle Age: The Roles of Socioeconomic Status, Health Behaviors, and Health Insurance." Social Sciences &Medicine 62:909 22.

Tyas, S.L., R.B. Tate, K. Wooldrage, J. Manfreda, and L.A. Strain. 2006. "Estimating the Incidence of Dementia: The Impact of Adjusting for Subject Attrition Using Health Care Utilization Data." Annals of Epidemiology 16 (6): 477-84.

U.S. Census Bureau. 2008. "An Older and More Diverse Nation by Midcentury" [accessed on March 21, 2011]. Available at: releases/archives/population/cb08-123.html.

Van Beijsterveldt, C.E., M.P. van Boxtel, H. Bosma, P.J. Houx, E Buntinx, and J. Jolles. 2002. "Predictors of Attrition in a Longitudinal Cognitive Aging Study: The Maastricht Aging Study (MAAS)." Journal of Clinical Epidemiology 55 (3): 216-23.


Additional supporting information may be found in the online version of this article:

Appendix SA1: Author Matrix.

Please note: Wiley-Blackwell is not responsible for the content or functionality of any supporting materials supplied by the authors. Any queries (other than missing material) should be directed to the corresponding author for the article.

[Corrections made after online publication 9/23/2011: Some highlighting that was in the submitted version of Table 2 of the article did not appear in the originally published version. These highlights are now included in the forms of bold, italic, and underlined text.] Address correspondence to Natalia A. Zhivan, Ph.D., Department of Global Health Systems and Development, School of Public Health and Tropical Medicine, Tulane University, 1440 Canal Street, Suite 1900, New Orleans, LA 70112; e-mail: Alfonso Ang, Ph.D., is with the Department of General Internal Medicine, University of California, Los Angeles, CA. Hortensia Amaro, Ph.D., is with the Institute on Urban Health Research, Bouve College of Health Sciences, Northeastern University, Boston, MA. William A. Vega, Ph.D., is with the School of Social Work, University of Southern California, Los Angeles, CA. Kyriakos S. Markides, Ph.D., is with the Department of Preventive Medicine and Community Health, University of Texas Medical Branch, Galveston, TX.

DOI: 10.1111/j.1475-6773.2011.01322.x
Table 1: Sample Size and Means for All the Variables by Race and
Ethnic Groups When the First Time Observed, Health and Retirement
Study (HRS) 1992-2004

Variable                        Whites         Blacks

Age                                64             62
Female (%)                         55             60
Married (%)                        72             50
Born outside the U.S. (%)           5              5
Less than high school (%)          24             47
Some college plus (%)              41             26
No insurance (%)                    7             15
Medicaid (%)                        3             15
Household income (U.S. $)      69,077         39,859
Household wealth (U.S.$)       329,79.5       91,819
No. of health conditions            1.14           1.43
No. of ADLs                         0.23           0.42
Birth cohort (%)
  Born before 1931                 44             34
  Born 1931-1941                   34             41
  Born 1941-19,54                  22             25
Final status (%)
  Always in                        56             49
  Died                             22             24
  Ever out                          8             14
  Attritors                        13             13
No. of observations            21,845          4,367

Variable                       Mex. Anger.    Latinos, All

Age                                 60             60
Female (%)                          54             56
Married (%)                         72             67
Born outside the U.S. (%)           44             56
Less than high school (%)           73             65
Some college plus (%)               13             18
No insurance (%)                    32             28
Medicaid (%)                        16             17
Household income (U.S. $)       42,471         41,751
Household wealth (U.S.$)       114,910        116,693
No. of health conditions             1.03           1.04
No. of ADLs                          0.45           0.42
Birth cohort (%)
  Born before 1931                  28             30
  Born 1931-1941                    39             39
  Born 1941-19,54                   33             31
Final status (%)
  Always in                         57             55
  Died                              17             15
  Ever out                          15             16
  Attritors                         11             14
No. of observations              1,492          2,487

Notes. About 70 percent of respondents entered the HRS in Waves 1 and
2 across all ethnic and racial groups, the rest of the respondents
were allowed to enter between Waves 3 and 7 to maxi mize the size of
the minority population. Samples of whites, Mexican Americans,
Latinos, and blacks have 1.5, 2..5, 2.2, and 2.4 percent of
observations with missing values for one or more vari ables,
respectively, and were excluded from the further analysis. ADL,
activities of daily living.

Table 2: Multinomial Lop-it Model, Base Outcome "Always in," Health
and Retirement Study (HRS) All 51+

                                     All                   Whites
Variable                           RRR/CI                  RRR/CI


Black                           1.076                        --
                               [0.972,1.1901                 --
Mexican American                0.801 **                     --
                               [0.666,0.964]                 --
Latino, other                   0.700 ***                    --
                               [0.550,0.889]                 --
Born outside the U.S.           0.855 **                0.932
                               [0.741,0.9861           [0.771,1.126]
No insurance                    1.270 ***               1.296 ***
                               [1.105,1.459]           [1.084,1.550]
Medicaid                        1.078                   1.053
                               [0.936,1.242]           [0.850,1.304]
Log of household                0.932 ***               0.898 ***
  income                       [0.906,0.959]           [0.864,0.932]
No. of health                   1.313 ***               1.301 ***
  conditions                   [1.271,1.357]           [1.253,1.352]
Number of ADLs                  1.139 ***               1.157 ***
                               [1.092,1.187]           [1.096,1.221]
Poor/fair self-                 1.825 ***               1.972 ***
  reported health              [1.676,1.988]           [1.782,2.181]

Ever out

Black                           1.748 ***                    --
                               [1.564,1.953]                 --
Mexican American                1.382 ***                    --
                               [1.150,1.661]                 --
Latino, other                   1.647 ***                    --
                               [1.328,2.042]                 --
Born outside the U.S.           1.325 ***               1.395 ***
                               [1.145,1.533]           [1.119,1.738]
No insurance                   1.281 ***                1.324 ***
                               [1.125,1.457]           [1.108,1.583]
Medicaid                        0.900                   1.287
                               [0.741,1.092]           [0.948,1.749]
Log of household                0.961 **                0.952 **
  income                       [0.932,0.991]           [0.907,0.998]
No. of health                   0.991                   0.978
  conditions                   [0.948,1.035]           [0.926,1.033]
Number of ADLs                  0.976                   0.983
                               [0.915,1.041]           [0.897,1.076]
Poor/fair self-                 1.117 *                 1.225 ***
  reported health              [0.997,1.250]           [1.056,1.421]


Black                           0.996                        --
                               [0.891,1.1141                 --
Mexican American                0.731 ***                    --
                               [0.601,0.891]                 --
Latino, other                   1.119#                       --
                               [0.902,1.388]#                --
Born outside the U.S.           1.401# ***              1.257# **
                               [1.221,1.608]#          [1.036,1.524]#
No insurance                    1.004#                  0.913#
                               [0.875,1.153]#          [0.760,1.097]#
Medicaid                        0.758# ***              0.938
                               [0.639,0.898]#          [0.732,1.201]
Log of household                0.967~ **               0.948@ **
  income                       [0.938,0.998]~          (0.909,0.988]@
No. of health                   1.188# ***              1.178# ***
  conditions                   [1.145,1.233]#          [1.128,1.229]#
Number of ADLs                  1.139# ***              1.144 ***
                               [1.087,1.194]           [1.077,1.214]#
Poor/fair self-                 1.312 ***               1.459# ***
  reported health              [1.191,1.446]           [1.299,1.638]
Pseudo R2                       0.1413                  0.1433
No. of Obs.                28,523                  21,286

                                   Blacks                Mex. Amer.
Variable                           RRR/CI                  RRR/CI


Black                                --                      --
                                     --                      --
Mexican American                     --                      --
                                     --                      --
Latino, other                        --                      --
                                     --                      --
Born outside the U.S.           0.806                   0.647 **
                               [0.508,1.278]           [0.449,0.931]
No insurance                    1.485 ***               1.095
                               [1.113,1.982]           [0.685,1.751]
Medicaid                        1.460 ***               0.829
                               [1.135,1.877]           [0.520,1.3201
Log of household                0.976                   0.974
  income                       [0.915,1.040]           [0.879,1.079]
No. of health                   1.304 ***               1.422 ***
  conditions                   [1.201,1.415]           [1.208,1.673]
Number of ADLs                  1.163 ***               1.149 *
                               [1.066,1.269]           [0.982,1.345]
Poor/fair self-                 1.497 ***               1.548 **
  reported health              [1.219,1.839]           [1.055,2.271]

Ever out

Black                                --                      --
                                     --                      --
Mexican American                     --                      --
                                     --                      --
Latino, other                        --                      --
                                     --                      --
Born outside the U.S.           1.518 **                1.242
                               [1.041,2.214]           [0.878,1.756]
No insurance                    1.431 ***               1.148
                               [1.091,1.878]           [0.787,1.675]
Medicaid                        1.100                   0.429 ***
                               [0.788,1.536]           [0.233,0.791]
Log of household                1.038                   0.918 **
  income                       [0.964,1.118]           [0.855,0.986]
No. of health                   0.963                   1.198 **
  conditions                   [0.872,1.062]           [1.007,1.427]
Number of ADLs                  0.98.5                  0.936
                               [0.869,1.117]           [0.763,1.148]
Poor/fair self-                 0.841                   1.423 *
  reported health              [0.661,1.071]           [0.988,2.049]


Black                                --                      --
                                     --                      --
Mexican American                     --                      --
                                     --                      --
Latino, other                        --                      --
                                     --                      --
Born outside the U.S.           2.129# ***              1.306#
                               [1.434,3.160]#          [0.881,1.935]#
No insurance                    1.079~                  1.688 **
                               [0.769,1.514]~          [1.065,2.674]
Medicaid                        0.833#                  0.765
                               [0.600,1.157]#          [0.431,1.3.57]
Log of household                1.043                   0.923 *
  income                       [0.957,1.136]           [0.849,1.002]
No. of health                   1.179 ***               1.449 ***
  conditions                   [1.069,1.301]           [1.205,1.742]
Number of ADLs                  1.174 ***               1.216 **
                               [1.060,1.301]#          [1.018,1.452]#
Poor/fair self-                 1.064#                  0.820#
  reported health              [0.832,1.360]           [0.534,1.257]
Pseudo R2                       0.1357                  0.1.553
No. of Obs.                 4,225                   1,427

                                Latinos, All
Variable                           RRR/CI


Black                                --
Mexican American                     --
Latino, other                        --
Born outside the U.S.           0.764 *
No insurance                    0.976
Medicaid                        0.971
Log of household                1.001
  income                       [0.923,1.085]
No. of health                   1.408 ***
  conditions                   [1.235,1.604]
Number of ADLs                  1.079
Poor/fair self-                 1.483 **
  reported health              [1.095,2.009]

Ever out

Black                                --
Mexican American                     --
Latino, other                        --
Born outside the U.S.           1.237 *
No insurance                    1.023
Medicaid                        0.567 ***
Log of household                0.947 *
  income                       [0.896,1.001]
No. of health                   1.092
  conditions                   [0.955,1.250]
Number of ADLs                  0.974
Poor/fair self-                 1.188
  reported health              [0.895,1.576]


Black                                --
Mexican American                     --
Latino, other                        --
Born outside the U.S.           1.610# ***
No insurance                    1.291
Medicaid                        0.559 ***
Log of household                0.998
  income                       [0.934,1.065]
No. of health                   1.304 ***
  conditions                   [1.140,1.493]
Number of ADLs                  1.152 **
Poor/fair self-                 0.854#
  reported health              [0.630,1.158]
Pseudo R2                       0.1313
No. of Obs.                 2,388

Note: Estimates are not reported for female, age, age squared,
married, educational attainment variables, cohort variables, and
variable for "other" race. Formatting indicates whether relative-risk
ratios for being an attritor are statistically significantly different
from the relative-risk ratio of dying: boldface: statistically
significant at 0.01 level; italic: statistically significant at 0.05
level; underlining: statistically significant at 0.1 level. Hausman
test does not reject the IIA for the samples of Mexican Americans,
Latinos including Mexican Americans, blacks, and for the sample of
white and pooled sample for the "Dead" and "Ever out" categories and
rejects hypothesis for the "Attritors" category for the pooled and
white samples.

*** Statistically significant at 0.01 level;

** Statistically significant at 0.05 level;

* Statistically significant at 0.1 level.

ADL, activities of daily living.


Boldface: statistically
significant at 0.01 level is indicated with #.

Italic: statistically significant at 0.05
level is indicated with @.

Underlining: statistically significant at 0.1 level
is indicated with ~.

Table 3: Predicted Probabilities for Selected Variables Based on Race/
Ethnic-Specific Multinomial Logit Models

Race/Ethnicity                     Always in (%)    Died (%)

Born outside the U.S.
  White non-Latino                     52.4           20.4
  African Americans                    43.1           15.6
  Mexican Americans                    52.1           13.6
  Latinos, all                         48.6           15.5
No insurance
  White non-Latino                     52.2           25.7
  African Americans                    44.0           25.4
  Mexican Americans                    48.3           17.9
  Latinos, all                         49.7           18.9
  White non-Latino                     53.9           22.8
  African Americans                    47.4           27.0
  Mexican Americans                    61.0           19.3
  Latinos, all                         58.4           22.8
Having two health conditions
  White non-Latino                     52.6           24.6
  African Americans                    47.6           24.0
  Mexican Americans                    48.3           20.8
  Latinos, all                         47.9           22.5
Poor/fair self-reported health
  White non-Latino                     48.2           28.0
  African Americans                    48.0           26.0
  Mexican Americans                    49.6           22.4
  Latinos, all                         49.2           23.9

Race/Ethnicity                     Ever Out (%)     Attritors (%)

Born outside the U.S.
  White non-Latino                      11.3         15.9 ***
  African Americans                     18.8         22.5 ***
  Mexican Americans                     17.0         17.3 ***
  Latinos, all                          16.6         19.3 ***
No insurance
  White non-Latino                      10.5         11.6 ***
  African Americans                     18.1         12.5 *
  Mexican Americans                     14.9         19.0
  Latinos, all                          14.8         16.6
  White non-Latino                      10.6         12.7
  African Americans                     15.1         10.5 ***
  Mexican Americans                      7.6         12.2
  Latinos, all                          10.0          8.8 **
Having two health conditions
  White non-Latino                       8.3         14.6 ***
  African Americans                     13.8         14.5
  Mexican Americans                     15.0         15.9
  Latinos, all                          14.6         15.0
Poor/fair self-reported health
  White non-Latino                       8.9         14.9 ***
  African Americans                     12.7         13.3 ***
  Mexican Americans                     17.0         10.9 ***
  Latinos, all                          15.8         11.2 ***

Notes. Statistical significance indicates whether particular variable
is a significant predictor of being an attritor compared to being dead
and is based on Table 2. To obtain predicted probabilities, estimates
from race/ethnic specific multinomial logit models were sequentially
applied to the entire sample of 28,.523 individuals while the variable
of interest was set to one (for born outside the U.S., no insurance,
Medicaid, and poor/fair health) or two (no. of health conditions).

*** statistically significant at 0.01 level;

** statistically significant at 0.05 level;

* statistically significant at 0.1 level.

ADL, activities of daily living.
Gale Copyright:
Copyright 2012 Gale, Cengage Learning. All rights reserved.