Assessment of Mortality Disparities by Wealth Relative to Other Measures of Socioeconomic Status Among US Adults

Key Points Question How does the disparity in mortality by wealth compare with disparities by other socioeconomic status (SES) measures and by smoking history? Findings In this US cohort study involving 6320 participants, disparities in mortality associated with wealth were statistically significant and larger than the disparities associated with education, occupation, income, or childhood SES; at high levels, additional wealth was not associated with lower mortality. However, the wealth disparity was much smaller than the smoking differential, which was evident across all ages 25 years and older. Meaning These findings suggest that mortality disparities by wealth are larger than those associated with other indicators of SES but that the smoking differential is far greater.

This supplemental material has been provided by the authors to give readers additional information about their work.

eAppendix 1. Mortality Follow-up
Vital status was based on: 1) searches of the National Death Index (NDI); 2) Wave 3 tracing (conducting between May 2013 and 2017), and 3) longitudinal sample maintenance. 1 The most recent mortality file for MIDUS includes deaths that occurred as late as March 2018. However, the Wave 3 fieldwork began in May 2013. Thus, we suspect mortality follow-up is incomplete beyond May 31, 2013 (e.g., there is no way to know whether someone who was interviewed in May 2013 was still alive at some later date unless they rechecked vital status). The last NDI search that included all MIDUS survivors covered the period through December 31, 2009. MIDUS conducted a more recent search of the NDI through the end of 2016, but they included only 713 respondents (out of 5,806 MIDUS respondents thought to be still alive). For this search, they selected MIDUS respondents for whom they had sufficient identifiers for matching and who they suspected might be deceased (e.g., no recent contact, invalid addresses, etc.).0 F * An examination of the mortality rates by sex, age, and calendar year provides further evidence suggesting that mortality follow-up is incomplete beyond 2013. We estimated age-specific mortality using a Cox model with age the time metric and controlling for sex and calendar year dummies (as a time-varying covariate). Compared with 1995, the age-specific mortality rates did not differ significantly for any year between 1996 and 2014. However, mortality rates began to decline precipitously in 2015 (HR=0.70*, p<.05, relative to 1995) and continued to decline in every subsequent year (HR=0.47, p<0.001, in 2016; HR=0.25, p<0.001, in 2017; and HR=0.017, p<0.001, in 2018). Mortality decline of this magnitude is particularly suspicious given that, among the US national population, life expectancy actually decreased (i.e., overall mortality increased) for three consecutive years between 2014 and 2017. There is no reason to expect the mortality rates among the MIDUS cohort to decline so dramatically after 2014. We believe that most (if not all) of the apparent mortality decline is a statistical artifact resulting from incomplete mortality follow-up. Thus, we have restricted our analysis to mortality through May 2013.
Finegood et al. 2 modeled mortality through October, 2018,1 F † but we suspect that their estimates are attenuated because of incomplete mortality follow-up during the last five years of that period. As a sensitivity check, we refit Model 1 from Table 1 for mortality through 12/31/2016. The HRs for wealth (e.g., HR=0.51, 95% CI 0.35-0.74 for those with $1M+ in wealth relative to those in debt) are weaker (i.e., closer to 1.0) than those shown in Table 1 (e.g., HR=0.41, 95% CI 0.24-0.69 for $1M+ vs. those in debt), but they follow a similar pattern. * It is unclear how "recent contact" was defined. The documentation states that they excluded respondents who had been recently contacted via participation in MIDUS projects or who were identified as alive during Wave 3 refielding in 2017. 1 However, the refielding included only those respondents who did not complete the self-administered questionnaire (SAQ) or cognitive component in 2015. Among the original MIDUS cohort, 3294 respondents were re-interviewed for Wave 3 between May 2013 and June 2014 (only 1 respondent was interviewed in May, 1527 were interviewed in June-July 2013; 94% were interviewed by 12/31/2013). We do not know whether they rechecked vital status at a later date for those individuals or whether they were presumed to be still alive. Information regarding the date when the SAQ was completed is not provided on the public-use file.

Education
Educational attainment was measured in terms of degree completion ("What is the highest grade of school or year of college you completed?") using the following 12 response categories: 1=No school/some grade school (1-6); 2=Eighth grade/junior high school (7-8); 3=Some high school (9-12 no diploma/no GED); 4=GED; 5=Graduated from high school; 6=1 to 2 years of college, no degree yet; 7=3 or more years of college, no degree yet; 8=Graduated from a two-year college or vocational school, or associate's degree; 9=Graduated from a four-or five-year college, or bachelor's degree; 10=Some graduate school; 11=Master's degree; and 12=Ph.D., Ed.D., MD, DDS, LLB, LLD, JD, or other professional degree.

Occupational Socioeconomic Index (SEI)
The occupational SEI was created by Hauser and Warren 3 based on the three-digit 1980 census occupational codes. Scores range from 7.1 (shoe machine operator) to 80.5 (physician).

Household income
Income from each source (i.e., wages/salary, social security, government assistance, and all other sources such as pensions, investments, child support, or alimony) was reported in categories, which we recoded to the mid-point of the range within each category. Income from each source was top-coded at $200,000 (except government assistance, which was top-coded at $50,000). We recoded top-coded values to the harmonic mean of a Pareto distribution. As suggested by von Hippel et al., 4 we compute the harmonic mean of a Pareto distribution with equal to the maximum of one or [ln where is the number of cases in the top category; −1 is the number of cases in the penultimate category; is the lower bound of the top category; and −1 is the lower bound of the penultimate category. Restricting α to a minimum of one ensures that the value of the top category is no greater than twice the lower bound of that category. We then summed across all sources to compute total income. We were unable to make an equivalence adjustment based on household size and composition because MIDUS did not collect that information at Wave 1.

Assets
The self-administered questionnaire included the following questions, which we used to determine the current net wealth of the respondent and his/her spouse (if applicable).
List of Categories shown on previous page (above J8): For J15, we recoded the responses to the mid-point of the range within each category, with those who reported themselves to be in debt on J14 coded as a negative number. The top category ($1,000,000+) was recoded to the harmonic mean of a Pareto distribution as described above for income.

Inflation adjustment for income and assets
For interviews conducted in 1996, we converted income and assets to 1995 dollars using the Consumer Price Index (CPI) provided by the Bureau of Labor Statistics (https://data.bls.gov/cgibin/cpicalc.pl). The median month among interviews completed in 1995 was April (which is treated as the reference), and the median month for those conducted in 1996 was July. Thus, the multiplier was 0.97 for data collected in 1996.

Adult SES
The composite measure of overall SES was based on six variables measured at baseline: educational attainment and occupational SEI of the respondent (and, if applicable, his/her spouse or partner), annual household income, and current net assets of the respondent and spouse/partner combined. A similar measure has been used in prior studies. [5][6][7][8][9] Because income and assets were positively skewed, we applied a square root transform to those two items. [If assets were negative, we used the untransformed values (i.e., everyone with net $0 assets or debt was assigned to zero on the transformed variable).] We standardized the six items and computed the average across relevant items (e.g., six items if married/partnered and both respondent and spouse/partner have ever been employed; three items if not married/partnered and respondent has never been employed; Cronbach's α=0.77 at baseline).

Childhood SES
The index for childhood SES was based on baseline measures of Mother's and Father's educational attainment and occupational SEI 3 as well as perceived financial status (i.e., respondent's rating their family's financial status relative to others, coded on a 7-point scale ranging from "a lot worse off" to "a lot better off"). We computed the average across the five standardized items (Cronbach's α=0.78).

eAppendix 3. Multiple Imputation Procedures
Among the 6,320 respondents included in the analysis, the predictors with the highest percentage of missing data were assets (11%), household income (10%), waist circumference (7%), and father's health status when the respondent was age 16 (5%). We used the "ice" command to perform multiple imputation. For the multiple imputation process, we used information for all the analysis variables as well as whether the respondent ever drank alcohol regularly and wage/salary income for the respondent (and spouse/partner, if applicable).
For continuous variables, departures from normality may result in implausible imputations when using the default draw method. Assets, household income, and number of hospitalizations had a skewed distribution. Prior to imputation, we applied a square root transformation to the income variables, but that was not possible for assets because it has negative values. To ensure that imputed values were within the range of observed values, we used prediction matching for all those variables and several others for which imputation generated out of range values (i.e., age, education, occupational SEI score, and physical limitations).
For imputation of ordinal variables (e.g., self-assessed health status), we used an ordered logit model for imputation. We performed five imputations and then used the "mim" prefix command to re-estimate the model for each imputation and combine the five sets of estimates using Rubin's rules for more information about the SES-related measures. We initially categorized wealth into the 10 categories shown above, which retains the first two response categories (i.e., in debt, net $0) and all the response categories above $100,000 from the original question (see p. 4). However, we collapsed the remaining 28 categories into two groups ($1-49,999 and $50,000-99,999). For the models presented in Table 2, we used the simplified 4-category version of wealth (i.e., In debt/Net $0; $1-49,999; $50,000-299,999; $300,000+), and for comparability, the other SES measures were categorized to have a similar distribution-or as close as possible given the level of measurement. (%) f 429 (6.8) IQR = Interquartile range a Includes respondents who identified as Asian or Pacific Islander; multiracial; Native American or Aleutian Islander/Eskimo; and other. b See eAppendix 2 for more information about the construction of this variable. c The vast majority (89%) of the sample had no hospitalizations in the past 12 months. d Physical limitations counts the number of the following physical tasks for which the respondent reports at least "a little" health limitation: 1) Lifting or carrying groceries; 2) climbing several flights of stairs; 3) bending, kneeling, or stooping; 4) walking more than a mile; 5) walking several blocks; 6) walking one block; 7) vigorous activity (e.g., running, lifting heavy objects); and 8) moderate activity (e.g., bowling, vacuuming). e Alcohol abuse is based on four items from the Michigan Alcohol Screening Test (MAST), 11 which has been used in many prior studies. 5,7,12 f Drug abuse is based on the Drug Dependence scale of the Composite International Diagnostic Interview Short Form (CIDI-SF) 13 and has also been used in prior studies. 5,7 MIDUS asked about the same types of drugs as the CIDI-SF (i.e., sedatives, tranquilizers, amphetamines, prescription painkillers, inhalants, marijuana/hashish, cocaine/crack/free base, hallucinogens, heroin, prescription anti-depressants), but the MIDUS question referred only to non-medical use (i.e., "on your own"-that is, "without a doctor's prescription, in larger amounts that prescribed, or for a long period than prescribed") whereas the CIDI-SF screener includes any use of those same drugs.

eFigure 2. Probability of Surviving From Age 65 to 85 by SES Measures and Smoking History, Demographic-Adjusted
Note: The predicted survival curves are based on a model of age-specific mortality above age 65 that controls for sex and race; those variables are fixed at the mean for the sample.