Contributions of Socioeconomic, Demographic, and Behavioral Risk Factors to All-Cause Mortality Disparities by Psychological Distress in the United States: A Blinder-Oaxaca Decomposition Analysis of Longitudinal Data

2.1. Data

The data for this study were from the 1997-2014 National Health Interview Survey (NHIS) prospectively linked to the 1997-2015 National Death Index (NDI).28 As a nationally representative, annual cross-sectional household interview survey, NHIS provides demographic, socioeconomic, and health characteristics of the civilian, non-institutionalized population in the United States. The National Center for Health Statistics (NCHS) developed public-use versions of NHIS prospectively linked with death certificate records from the NDI. For this study, we used the latest available 1997-2015 public-use linked mortality file containing 19 years of mortality follow-up data from the date of survey participation through December 31, 2015.29

2.2. Sample

The study sample consisted of adults aged 18 and older from 1997 to 2014 NHIS sample adult files. Since the main decomposition analysis allows only dichotomous group variable, we restricted the study sample to adults with no PD (K6=0) and adults with SPD (K6=13+) as defined below. The final pooled eligible sample size after excluding missing values was 263,825; the listwise deletion accounted for 1.61% of the total sample size of 268,145. For missing values (accounting for more than 1% of the total observations), for poverty status (18.74%), body mass index (BMI) (3.54%), and alcohol consumption (1.66%), we created missing categories so as to not lose a substantial number of observations for the analysis.

2.3. Outcome Measurement (Dependent Variable)

We analyzed crude death rates for all-causes combined. The mortality outcome was dichotomized, with 1 being dead and 0 equaling alive.

2.4. Psychological Distress Classification

Psychological distress (PD) was measured by the Kessler 6 (K6) nonspecific distress scale7 of six symptoms. Respondents were asked: “during the past 30 days, how often did you feel, 1) so sad that nothing could cheer you up, 2) nervous, 3) restless or fidgety, 4) hopeless, 5) that everything was an effort, and 6) worthless.” We created a 5-level categorical variable with scores of 0, 1-2, 3-5, 6-12, and 13-24 from the K6 scale.8 Since the Blinder-Oaxaca technique allows only a dichotomous group comparison, we only used individuals with no PD (K=0) as a reference group and individuals with SPD (K=13+) as another group. We did sensitivity analysis by changing the reference group from adults with no PD (K6=0) to adults with no SPD (K6=0-12).

2.5. Contributors (Independent Variables)

Based on the previous literature, we selected the following contributors for model estimation: 1) sociodemographic characteristics including age, sex, race/ethnicity, nativity/immigrant status, education, marital status, poverty status, housing tenure, and region of residence, and 2) behavioral characteristics including BMI, smoking status, and alcohol consumption.5,8,9,12,19–21

2.6. Analytic Approach

The Blinder-Oaxaca decomposition analysis was conducted to quantify the contributions of individual characteristics in explaining the observed mortality differences between adults with no PD and those with SPD.30,31 While regression analysis reveals the association between mortality and its risk factors, the decomposition analysis allows to estimate the association between the mortality disparities and their contributing causes or risk factors, possibly identifying factors for reducing health inequalities.31

In the given equation, Y̅ _SPD and Y̅ _{no PD} no represent means of the rates of deaths for adults with SPD and those with no PD, and X̅ _SPD and X̅ _{no PD} are vectors of group-specific means of the individual-level characteristics studied. β̂ * is the vector of coefficients from the pooled linear probability models that reflects the average outcome difference associated with each variable. β̂ _SPD is the adjusted difference in outcomes between adults with no PD and those with SPD that remains unexplained by covariates. The “explained” portion of the decomposition reflects the change in the mean of mortality rates for adults with SPD, compared to those with no PD, due to the differences in the individual-level characteristics. The percentage of the outcome difference explained by each covariate can be calculated by dividing the explained difference by the total difference:

We estimated two models of Blinder-Oaxaca decomposition analysis: model 1 included only socio-demographic characteristics; model 2 included both socio-demographic and health behavioral characteristics as contributors.

Collinearity between model covariates was assessed using variance inflation factors, and all values across models were within an acceptable range (≤1.61). Complex survey design procedures, svy in Stata, were used, to account for clustering, multiple stages of selection, and disproportionate sampling.32 To correct the bias from the ineligible adults for linkage to the NDI due to insufficient identifying data, we used eligibility-adjusted weights developed by NCHS, instead of the standard sample weight.33 The sample weights were adjusted by dividing by the number of pooling years. All the analyses were conducted by Stata 17 and the Blinder-Oaxaca model was fitted using the Oaxaca procedure.32

3.1. Mortality and Individual Characteristics by No Psychological Distress and Serious Psychological Distress

Overall, there was a significant difference in weighted unadjusted death rates between adults with no PD and those with SPD (Table 1). Adults with SPD had higher all-cause mortality, compared to those with no PD. Differences between adults with no PD and those with SPD were observed in individual characteristics including socio-demographic characteristics and health behaviors, known to typically affect health and mortality (Table 1). Adults with SPD were more likely to be female, non-Hispanic Blacks, Hispanics, American Indians/Alaska Natives, or non-Hispanic others, US-born, with less than high school education, not currently married, below poverty level, renters, Midwest or South residents, with higher BMI, current smokers, and former drinkers (P<0.001).

3.2. Blinder-Oaxaca Decomposition

Blinder-Oaxaca decomposition analysis results (Table 2) showed that 38.86% of the difference in all-cause mortality between adults with no PD and those with SPD (-6.23 percentage points, 95% CI = -6.92, -5.54) was explained by the sum total of the sociodemographic characteristics (Model 1). Similarly, 46.86% of the difference in all-cause mortality between adults with no PD and those with SPD was explained by socio-demographic characteristics and health behaviors combined (Model 2).

The sum total explained by measured variables includes counterbalancing factors, some of which contribute to and others diminish the disparity by SPD. The lower educational level for adults with SPD accounted for 1.51 percentage points of mortality difference by PD, or 24.27% of the total 6.23 percentage point mortality disparity (Model1). Thus, holding all other sociodemographic variables constant, higher educational attainment for adults with SPD would eliminate 24% of the all-cause mortality disparity between adults with no PD and those with SPD. Other sociodemographic factors made a smaller contribution to the disparity, including lower-income (14.24%), higher proportions of renters (5.59%), and higher proportions of non-currently-married individuals (4.45%) among adults with SPD. Some other factors that were positively associated with all-cause mortality mitigated the disparity, including higher proportions of females (-7.65%) and younger individuals (-3.29%). These are the factors that favored adults with SPD and would serve to increase all-cause mortality disparities between adults with no PD and those with SPD if equalized.

Similarly, the difference in educational attainment (-1.32 percentage points or 21.13%) was the highest contributor to all-cause mortality disparity between adults with no PD and those with SPD in Model2, which additionally examined the contribution of behavioral risk factors. The second highest contribution to all-cause mortality disparity was due to higher rates of current smoking among adults with SPD. The contributions of education, poverty status, marital status, and housing tenure slightly decreased, indicating that health behaviors such as smoking and alcohol consumption partly accounted for the impact of socio-demographic factors.

Decomposition analysis using a Fairlie non-linear approach34 as a sensitivity analysis yielded similar results, although the magnitude of the contributions varied (Table 3). There is a possibility of overestimation of the impact of the difference in poverty or education between adults with no PD and those with SPD on all-cause mortality difference.34 An additional sensitivity analysis was conducted to assess variations in the contributions of individual characteristics by using 2, 5, and 10 years of mortality follow-up. The longer the follow-up years, the larger the component explained by measured variables (Table 4). This might be because the longer follow-up allows for a greater manifestation of the lagged effect of smoking, BMI, drinking, and socioeconomic disadvantage on mortality. A third sensitivity analysis was conducted to examine variation by a reference group. We changed the reference group from adults with no PD (K6=0) to adults with no SPD (K6=0-12) (Table 5). The portion of explained difference by each determinant, statistical significance, and the direction of contribution were generally robust regardless of whether adults with no PD or adults with no SPD were used as a reference group, except for age and region of residence.

4.1. Limitations

This study has some limitations First, our study used an NHIS sample eligible for linkage to the NDI, which is, on average, 94.8% of the original NHIS survey participants.33 Excluding records ineligible for linkage may lead to biased mortality estimates. To address potential bias, we used the eligibility-adjusted sampling weight to account for the NHIS-NDI mismatches.33 Second, while our Blinder-Oaxaca decomposition included a variety of risk factors of all-cause mortality, there could be other potential factors of mortality, confounding existing risk factors. Third, in the NHIS-NDI database, all the risk factors were available at baseline as of the survey date, while SES, behavioral-risk factors, and PD are expected to vary over the long follow-up. It is worthwhile to examine the extent of changes in SES or behavioral factors from childhood, working-age, to older ages contributing to the disparities in mortality by SPD status, using the life course approach36 with longitudinal data.