Predicting Receipt of Social Security Administration Disability Benefits Using Biomarkers and Other Physiological Measures: Evidence from the Health and Retirement Study

Abstract

Objectives:

The objective of this study was to assess how well physiological measures, including biomarkers and genetic indicators, predict receipt of Social Security Administration (SSA) disability benefits among U.S. adults aged 51 to 65 years.

Method:

We used data from the 2006 to 2012 waves of the Health and Retirement Study (HRS), linked to SSA administrative data. Using logistic regression, we predicted benefit receipt (either Social Security Disability Insurance or Supplemental Security Income) using 19 distinct physiological markers, adjusting for age, sex, race, and select medication use. We then calculated the propensity (i.e., predicted probability) that each HRS respondent received benefits and assessed how well propensity score–based classifications could identify beneficiaries and nonbeneficiaries.

Results:

Thirteen percent of respondents received benefits. Using the propensity score cut point that maximized the sum of sensitivity and specificity, the model correctly predicted 75.9% of beneficiaries and 73.5% of nonbeneficiaries.

Discussion:

Physiological measures have moderate power to predict SSA disability benefit receipt.

INTRODUCTION

Since the 1990s, social and demographic surveys have increasingly incorporated physical measurements into their data collection. Examples of these surveys in the United States include the Health and Retirement Study (HRS), the National Longitudinal Study of Adolescent to Adult Health, and the National Health and Nutrition Examination Survey (NHANES). These surveys all collect anthropometric data such as height and weight. Some also collect blood or saliva samples to measure biomarkers—that is, substances in the body (for example, hormone levels) thought to indicate something about health or health processes—and some collect genetic information (for example, the presence of a gene variant linked to particular health outcomes). This collection of physiological data reflects an understanding that physiology and social circumstances are interrelated and contribute in combination to a person’s quality of life, disease onset, and even longevity. Nevertheless, it is unclear how well the current set of collected measures can predict certain health or economic outcomes, and whether the proliferation of measurement more generally has practical implications for public policy or service delivery.

This paper assesses the extent to which 19 distinct, objectively measured, physiological measures in the HRS, including select biomarkers and genetic data, can predict disability benefits among adults ages 51 to 65 from either of two programs administered by the Social Security Administration (SSA): (1) Social Security Disability Insurance (SSDI) or (2) Supplemental Security Income (SSI). These programs are designed to provide financial supports to working-age adults who are unable to participate in substantial gainful activity for a period of at least 12 months, or until death. The benefits provide critical income to most people who receive them, and many recipients live near or below the federal poverty line even with the benefits (Wright, Livermore, Hoffman, Grau, & Bardos, 2012). The benefits are also a large fiscal burden on federal and state governments. As of December 2015, 13.6 million working-age Americans received benefits from at least one of these programs (SSA, 2016a, Table 3; SSA, 2016b, Table V.F1). In fiscal year 2015, the federal SSDI and SSI expenditures for these working-age beneficiaries totaled $188 billion. Adding expenditures for Medicare, Medicaid, and SSI state supplements brings the estimated total public expenditures for working-age SSDI or SSI recipients in fiscal year 2015 to about $475 billion.¹

To the best of our knowledge, this is the first study to examine the extent to which biomarkers and genetic traits can predict receipt of SSA disability benefits. Understanding the relationships between these measures and benefit receipt could be useful to (1) gain greater insight into the disability experience and potential risk factors and (2) determine whether physiological measurement might have a role in future SSA program operations—such as in forecasting or helping to adjudicate disability applications.

We used HRS data linked to SSA administrative data to estimate how well physiological measures predict current SSA disability benefit receipt (that is, the extent to which measures in one period can explain variation in benefit receipt during that same period, as opposed to predicting future receipt), while holding constant other characteristics of the individual. Establishing the extent to which physiological measures can predict current benefit receipt, controlling for covariates, is a logical first step in a research agenda that could eventually consider use of such measures to predict future benefit eligibility or receipt. That is, conceptually, before one tests whether today’s physiological measures can predict future events, or before one considers how physiological measurement might improve SSA’s disability determination process, it is important to establish whether associations exist in the first place. If the measures do not provide information about current receipt, there is no reason to believe they would be informative about the future.

This study focuses on predictive power of the measures in aggregate—that is, overall, how well the physiological measures can distinguish disability benefit recipients from non-recipients. To assess this, we first estimated a logistic regression model predicting benefit receipt using the 19 physiological measures, adjusting for select covariates. Based on the model estimates, we then calculated for each HRS observation the propensity to receive benefits (that is, a predicted probability) given all physiological measure values. We classified observations as predicted to be receiving benefits or not at various cut points of the propensity scores. (That is, we classified observations with a score at or above the cut point as predicted to be receiving benefits, and those with a score below the cut point as not receiving benefits.) Finally, we evaluated the sensitivity and specificity of these various classifications—assessing how well the available physiological measures, in aggregate, could correctly identify SSA disability beneficiaries and non-beneficiaries.

BIOMARKERS AND OTHER PHYSIOLOGICAL MARKERS OF HEALTH

We use the term “physiological measures” to refer to several objectively measured traits. These include anthropometric measures (for example, height, weight, or waist circumference); measures of strength and function (for example, tests of balance or lung function); and biomarkers and select genetic indicators.

Biomarkers are measurable substances in the body that indicate something about health or health processes. Some biomarkers collected in population surveys will be familiar from routine clinical care—for example, cholesterol levels, which are used to assess cardiovascular risk, or glycated hemoglobin (often called hemoglobin A1c or HbA1c), which is used for diabetes diagnoses and monitoring. Other biomarkers are typically collected in clinical settings only among people with select medical conditions. For example, high levels of cystatin C, a protein filtered from the blood by the kidneys, can signify suboptimal kidney functioning. Some biomarkers have no known clinical applications. Telomere length, for example, is a measure of the length of the protective tips at the end of chromosomes; telomeres are believed to get shorter with each chromosome replication, so telomere length serves as a measure of cellular aging (HRS, 2013).

Biomarker measurement and genetic measurement are both relatively young fields. Several cancer researchers have assessed the extent to which biomarkers (for example, prostate-specific antigen levels) or genetic variation (for example, mutations in the tumor suppressor genes BRCA1 and BRCA2) might predict patient prognosis or even response to treatment (Ballman, 2015; King, Marks, & Mandell, 2003). For population health researchers, biomarkers and genetic markers have become a focus just in the last 15 years, with analyses of the relationships between these measures and mortality or major morbidity among general population samples. For example, high cystatin C has been linked to cardiovascular disease; high cystatin C is also associated with all-cause mortality risk (Crimmins et al., 2013). Similarly, telomere length has been linked with all-cause mortality risk among older adults, although the strength and consistency of this finding is uncertain (Mather, Jorm, Parslow, & Christensen, 2011). From population data with genetic indicators, there is evidence that variation in a single DNA building block (that is, a single nucleotide polymorphism) might in some cases predict major health outcomes at the population level. For example, Ewbank (2002, 2004) found that people with a particular risk allele, or gene variant, in the APOE gene have higher mortality and that this difference alone can explain 1% to 2% of the variation in life span among people older than 65 in six European countries (Ewbank, 2004). Research on biomarkers and genetics can thus complement population health findings about more established anthropometric measures such as body mass index (BMI), which has been linked to disease onset (Guh et al., 2009); mortality risk (Flegal, Kit, Oprana, & Graubard, 2013); and disability (Alley & Chang, 2007).

SSA DISABILITY BENEFITS

The SSA disability programs provide benefits to U.S. working-age adults who cannot engage in substantial gainful activity due to long-lasting, medically determinable, physical or mental impairments. The two programs use the same medical eligibility criteria, but the nonmedical eligibility criteria differ. SSDI is a social insurance program. Adults with qualifying disabilities are eligible if they and their employers have paid sufficient premiums (via payroll taxes) over a substantial and sufficiently recent period. Benefits paid are based on the recipient’s past earnings. SSI, in contrast, is a means-tested program; eligibility requires limited income and resources. The maximum benefit is low and income from other sources (including SSDI) reduces the SSI payment amount. People who receive SSI benefits due to disability automatically qualify for SSDI if they meet the SSDI work history requirement. Both SSDI and SSI beneficiaries may also receive benefits from workers’ compensation, private disability insurance, state programs, or the U.S. Department of Veterans Affairs. However, benefit coordination between these other sources and SSA varies.

Understanding the relationship between biomarkers and disability could be useful to policymakers and SSA program administrators for two reasons:

1. Forecasting future demand for SSDI or SSI.

The number of SSDI recipients has increased rapidly, roughly tripling between 1980 and 2015 (SSA, 2016a, Table 3). Although awards have declined from their peak (which followed the 2008 recession) and are likely to continue to decline as baby boomers age out of the working-age population, SSA actuaries predict that the SSDI Trust Fund will be depleted in 2023 (Social Security and Medicare Boards of Trustees, 2016). As Congress and SSA determine how best to fully fund SSDI in the future, accurate growth forecasts have never been more valuable. Forecasts could be improved with better measures of the physical and mental health status of the population and with a better understanding of how that status relates to benefit receipt. This idea is illustrated by Soneji and King (2012), who used estimates of the longevity effects of obesity and smoking to calculate the impact of those factors on the overall Social Security Trust Fund for old age, survivor, and disability benefits. Using similar methods, one could perhaps improve projections for the SSDI fund in particular by using projected physiological measure values (for example, projections of future cholesterol levels) and estimates of the associations between those measures and SSA disability benefit receipt.

2. Supporting disability determinations.

As physiological measurement continues to mature as a field, it is possible that SSA could add new objective measures to its disability determination process—that is, the process for deciding whether applicants qualify for benefits. At present, SSA uses a five-step sequential process to determine SSDI eligibility and a disability designation from SSI (see, for example, Wixon and Strand [2013]).

• First, the SSA adjudicator determines whether an applicant currently performs “substantial gainful activity”—defined in 2017 as earning at least $1,170 per month or $1,950 per month if blind (SSA, 2017a). Applicants are not eligible if they currently work that much.

• Second, the adjudicator determines whether a medical condition interferes with basic work functions and is likely to last at least 12 months or result in death.

• Third, the adjudicator compares the applicant’s health status against SSA’s Listing of Impairments, which lists the conditions severe enough to justify benefits without further review. SSA already uses some physiological measures in this stage of the determination process; a simple but important example is obesity that worsens the severity of other medical conditions (Schimmel Hyde, 2016). In addition, however, we hypothesize that physiological measures might be associated with benefit receipt because the measures reflect, or otherwise are correlated with, a benefit-qualifying condition—for example, cystatin C and chronic kidney disease.

• Fourth, the adjudicator assesses whether the applicant, given his or her “residual functional capacity,” can perform gainful work that he or she has performed previously. Physiological measures may predict residual functional capacity—that is, ability to perform sustained work—as adjudicators consider at this stage medical conditions that do not meet the severity criteria laid out in the Listing of Impairments, but that may be reflected in the measures. One example is hypertension (reflected by blood pressure) that limits ability to work but has not yet led to disability through its effects on other body systems: the heart, brain, kidneys, or eyes (SSA, 2017b).

• Fifth and finally, the adjudicator assesses whether the applicant could learn and perform a new job that would provide gainful employment, given residual functional capacity and “vocational factors”—namely, the applicant’s age, education, and work history.

In 2010, nearly 64% of applications were determined in the final two steps (Wixon & Strand, 2013), assessments that are time-consuming for the SSA adjudicator. In addition, the process is often perceived as subjective; many denials are ultimately appealed. Although SSA already uses some physiological measures, as noted, in the third and fourth steps of its determination process, SSA might be able to speed disability decisions and reduce the need for appeals by incorporating additional objective information into the process. That is, a strong link between physiological measures and benefit receipt could suggest opportunities for SSA to expand the list of explicit eligibility criteria (Step 3), while limiting the importance of the more time-consuming, potentially subjective assessment of residual functional capacity (Step 4).

Both applications of physiological measurement described here—that is, forecasting and aiding the determination process—would require substantial supporting evidence beyond the current study. Our more limited goal with this study is to provide policymakers, SSA program administrators, and others with foundational information on the use of physiological measures to predict disability benefit receipt. This information can help policymakers decide whether and how to incorporate expansion of physiological measurement into program operations in the future.

METHOD

We used restricted-access data from the HRS, linked to SSA administrative data, to estimate associations between physiological measures and receipt of disability benefits among adults ages 51 to 65.² Because our goal was to assess the extent to which physiological measures in aggregate can predict receipt of disability benefits, we simultaneously estimated the associations between benefit receipt and several distinct measures.

HRS sample

The HRS is a nationally representative U.S. longitudinal survey of noninstitutionalized adults ages 51 and older. The HRS started in 1990 and has interviewed its respondents either by telephone or in person roughly every two years since, periodically refreshing the sample with new, younger respondents to maintain representativeness of the target age range.

Starting in 2006, the HRS began augmenting its in-person data collection with physiological measurements. The HRS randomly selected half its eligible respondents for an in-person interview in 2006; the other half was selected in 2008. The HRS repeated in-person interviews for the 2006 group in 2010 and the 2008 group in 2012. In addition, the HRS refreshed its overall sample in 2010 and randomly selected half of the newly eligible respondents for in-person interviews in 2010 and the other half in 2012.

HRS data can be linked to restricted-access SSA administrative data, which contain information about respondents’ retirement, SSDI, and SSI benefits—provided that the respondent consented to SSA data linkage. If the respondent gave consent for this linkage at any time between 2006 and 2012, we can access his or her SSDI benefit data from at least 1962 through December 31, 2012,³ and SSI benefits data from January 1, 1974, through December 31, 2012. The HRS also collects self-reports of whether respondents are applying for SSA disability benefits but have not yet received a decision. We reviewed these data but did not include them in the analysis due to the low frequency of events (results not shown).

For this study, we limited the sample to people observed with (1) a nonzero survey weight in the HRS 2006, 2008, 2010, or 2012 survey waves; (2) at least one physiological measurement recorded during the wave; (3) a consent form sent to SSA to permit data linkage; and (4) an age of at least 51 years but less than full retirement age⁴ on the HRS interview date (because workers are eligible for full retirement benefits rather than disability benefits once they have reached full retirement age). Finally, we excluded from the sample any observations from the 2012 HRS wave with interview dates after December 31, 2012, because the SSA data linkage was complete only through this date.

Using this sample definition, a single person could have physiological measurements at more than one wave, but no more than twice—that is, in 2006 and 2010 or in 2008 and 2012. We updated the HRS sample weights (weighting for age, sex, and race) to account for linkage to the SSA administrative data (details available on request).

Physiological measures collected

Table 1 summarizes the 19 physiological measures used in this study. We included all physical markers and all biomarkers collected by the HRS, except for (1) body weight, which we used to create a measure of body mass index that we included instead, and (2) a timed walking test, because “inability to ambulate effectively” is included in SSA’s Listing of Impairments as a partial determinant of eligibility and thus its relationship with benefit receipt is not in question. We also included data on four gene variants that—based on genome-wide association scans (GWAS) among populations of mostly European ancestry—we expected might predict benefit receipt. Genetics information was available only for respondents who consented to genetic data collection in the 2006 or 2008 waves. However, for those observed at more than one wave, we can assume that status on the four gene variants did not change over time. Telomere length was available only in 2008. For the regression analysis described below, we standardized all continuous variables in Table 1 so that results would reflect the association between benefit receipt and a one-standard-deviation change in the physiological measure.

Table 1.

Summary of 19 physiological measures

Measure	Variable type	Units	Definition
Anthropometric measures
Height	Continuous	Inches	Distance from ground to point on a wall corresponding to the top of respondent’s head when standing against wall without shoes
Body mass index	Categorical	NA	Body mass index (BMI), calculated from height, described above, and weight, measured by scale up to 300 lbs or self-reported over 300 lbs; treated as a categorical variable with values of underweight (BMI <18.5), normal weight (≥18.5 and <25 [omitted from regression analyses as the reference category]), overweight (≥25 and <30), and obese (≥30)
Waist circumference	Continuous	Inches	Distance around waist, measured by tape measure at navel level
Measures of physical function
Grip strength	Continuous	Kilograms	Hand grip strength in respondent’s strongest hand, measured by spring-type dynamometer (highest value among two attempts for each hand)
Lung function	Continuous	Litres/minute	Peak expiratory flow rate—that is, maximum speed of expiration—measured by peak flow meter (highest value among three attempts)
Balance: Passed test	Binary	NA	Ability to stand with heel of one foot in front of and touching toes of the other foot for 60 seconds (30 seconds if age ≥65)
Pulse	Continuous	Beats/minute	Resting pulse (average of three readings), measured by blood pressure cuff
Systolic blood pressure	Continuous	Millimeters of mercury	Resting systolic blood pressure (average of three readings), measured by blood pressure cuff
Diastolic blood pressure	Continuous	Millimeters of mercury	Resting diastolic blood pressure (average of three readings), measured by blood pressure cuff
Biomarkers
Total cholesterol	Continuous	Milligrams/decilitre	Total cholesterol, measured from dried blood spot
HDL cholesterol	Continuous	Milligrams/decilitre	High-density lipoprotein cholesterol, measured from dried blood spot
HbA1c	Continuous	Percentage	Glycated hemoglobin, measured from dried blood spot
C-reactive protein (ln)	Continuous	Micrograms/milliliter (ln)	C-reactive protein, measured from dried blood spot, excluding values >10 (which likely indicate acute infection), and log-transformed to reduce skewness
Cystatin C	Continuous	Milligrams/litre	Cystatin C, measured from dried blood spot
Telomere length (ln)	Continuous	NA	Ratio of telomere to single copy gene (that is, a ratio proportional to the average telomere length in the cell), measured in saliva, excluding values >3.5 (which are difficult to interpret), and log transformed to reduce skewness
Genetic markers
APOE gene: Risk alleles present	Categorical	NA	Values of 0, 1, or 2: Number of copies of the E4 allele, as measured in saliva; presence has been linked to increased risk of cardiovascular disease and Alzheimer’s disease
COMT gene: Risk allele present	Binary	NA	Indicator for whether respondent is homozygous for the A allele, as measured in saliva; presence of this allele has been linked to depression risk
FTO gene: Risk allele present	Binary	NA	Indicator for any presence of the A allele, as measured in saliva; this allele has been implicated in diabetes
IL6 gene: Risk allele present	Binary	NA	Indicator for any presence of the C allele, as measured in saliva; this allele has been linked to diabetes and other metabolic disorders—at least among white Americans; because presence of this allele varies considerably by race/ethnicity, we interacted the IL6 variable with white race

Sources: Crimmins et al. (2008, 2013); Faul, Smith, and Zhao (2014a, 2014b); HRS (2013).

Note: The HRS standardized all biomarker values by assuming the distribution of biomarker values from the dried blood spot assays should be similar to the distribution observed in NHANES, adjusting for demographic characteristics. This standardization is intended to account for differences in biomarker values within the HRS across assays and across laboratories (Crimmins et al., 2013).

NA = not applicable.

Covariates

In our regression analyses, we controlled for seven additional variables: (1) age group (51 to 55, 56 to 60, or 61 to 65); (2) sex; (3) race (non-Hispanic white, non-Hispanic black, Hispanic, or non-Hispanic other); whether the respondent reported taking medications to manage (4) diabetes, (5) hypertension, or (6) lipids; and (7) time period (2006 to 2008 or 2010 to 2012). We controlled for medication use because it might alter the relationship between benefit receipt and physiological measures (including, HbA1c, blood pressure, or cholesterol levels). We treated 2006 to 2008 as a single time period and 2010 to 2012 as a single time period because the HRS combined data within each of these periods when cleaning and validating its physical and biomarker data.

Outcomes

We defined receipt of SSA disability benefits as a binary variable equal to 1 if, during the month of the HRS interview, either (1) the HRS respondent received SSDI benefits according to SSA administrative data or (2) the HRS respondent was younger than 65 and received SSI benefits according to SSA administrative data. Only people younger than 65 can qualify for SSI on the basis of disability.

Exploratory data analysis

Because of limited existing research linking physiological measures to receipt of disability benefits, we conducted three sets of exploratory data analysis to guide decisions about functional form for the regression models. First, we plotted age- and sex-standardized distributions of the measure values, comparing distributions for the populations receiving and not receiving disability benefits. These distributions provide information about the extent to which measure values vary by benefit status—for example, whether the measure means might be different for recipients and non-recipients, or whether the shape of the entire distribution might be different, or both. Second, we calculated correlations (using the Pearson correlation coefficient) among the 19 different physiological measures and the covariates. We used correlation information to judge whether multiple measures might reflect the same underlying trait. Third, we regressed benefit receipt on each of the physiological measures—first, with no other predictor variables in the model; then, in a separate model, controlling for all other physiological measures and the covariates. We plotted the residuals from these logistic regressions against each relevant predictor variable and analyzed the plots for evidence of nonlinear relationships between the predictors and benefit receipt. Evidence of nonlinear relationships could suggest a need for interactions or higher-order polynomial terms in the regression model.

Regression analyses and model selection

We conducted logistic regression, regressing disability benefit receipt on all physiological measures simultaneously, adjusting for the covariates. We assumed the relationship between the log odds of benefit receipt and each physiological measure was linear because, in the three sets of exploratory analysis described previously, we did not find evidence to the contrary (results not shown). We clustered standard errors at the person level to account for correlation between a single individual’s multiple (up to two) observations. We included a series of indicator variables to denote missing data on each measure so that we did not drop observations from the regressions due to partial missing data.

We considered three possible model specifications. The first included all physiological measures without interactions by sex or by time period. The second stratified by sex (that is, we fully interacted all variables with the indicator variable for female) and the third stratified by time period (2006 to 2008 or 2010 to 2012). We did this because relationships between physiological measures and disability could conceivably vary either by sex or by time period. For example, a woman who is 5’7” is relatively tall, whereas a man who is 5’7” is relatively short. Similarly, the relationships between physiological measures and benefits might vary by time period if, for example, program eligibility rules were applied differently over time or there were advances in medical treatment that made some conditions less disabling. We used Wald tests to compare between the stratified models and the unstratified model, and we selected the best-fitting model as our primary specification. We also used a Wald test to compare our primary specification to a model with only the covariates (not the physiological measures) to test the extent to which the physiological measures improved model fit.

Propensity for disability benefit receipt

We used the regression results to generate predicted probabilities of receiving disability benefits—that is, propensity scores—for each observation, based on physiological measures and covariate values. We used various cut point values of the propensity score to classify observations with scores above the cut point as those predicted to be receiving benefits and those with scores below the cut point predicted to be not receiving benefits. Then for each cut point chosen, we calculated the prediction’s sensitivity (that is, the proportion of positive predictions that are true positives) and specificity (the proportion of negative predictions that are true negatives).

We plotted receiver operating characteristic (ROC) curves to convey the information about sensitivity and specificity graphically. The ROC plot shows sensitivity on the y-axis and 1 – specificity on the x-axis, such that the ROC curve illustrates the trade-off between correctly identifying true positives and true negatives as we increase the cut point value used to classify observations. The area under the curve (AUC) for each ROC curve provides an index of the model’s predictive power. Bamber (1975) showed that the AUC is the probability that a predicted value for a randomly chosen observation receiving benefits is greater than the predicted value for a randomly chosen observation not receiving benefits. The AUC for a perfectly predictive model would be 1.0 (that is, with the ROC curve following the left and top boundaries of the plot area, with a vertical and then a horizontal line), while the AUC for random chance would be 0.5 (with the ROC curve following the diagonal, or 45-degree line, of the plot area).

Finally, we identified an optimal cut point as the point that maximized Youden’s Index, which is defined for each cut point as sensitivity + specificity – 1 (Youden, 1950). Thus, the optimal cut point was the point that maximized the sum of sensitivity and specificity.

Evaluating model prediction

The AUC, Youden’s Index, and Wald test results all provide information about a model’s predictive power, but the information from each differs. Youden’s Index is a measure of diagnostic accuracy at a given cut point, whereas the AUC is an index of diagnostic accuracy across the full range of possible cut points (that is, integrating the curve that relates sensitivity to 1 – specificity). Both generally improve as we add predictor variables to the model. In contrast, the Wald tests assess whether added predictors improve model fit beyond what is expected by chance—that is, whether the increased explanatory power justifies the added model complexity.

Robustness checks and additional analyses to explore policy implications

We conducted three additional analyses to explore the policy implications of this research for SSA.

First, we conducted the regression analysis using our primary specification with one exception: excluding the covariates for race (although we maintained the interaction between white race and the IL6 gene variant⁵). SSA does not consider race as part of its disability determination process, so policymakers might wish to know how well physiological measures predict benefits without controlling for race.

Second, SSA administrators are likely to be interested in physiological measures that an applicant cannot easily manipulate for disability determinations. We re-estimated the primary specification of the model excluding three measures that required respondent effort—grip strength, lung function, and the balance test—to assess how well physiological measures predict benefit receipt without these three.

Third, to assess whether physiological information could add explanatory value beyond some of SSA’s existing vocational factors, we re-estimated the model with the physiological measures and the covariates, but categorizing age as SSA does in its vocational factors (50 to 54, 55 to 59, and 60 and older), instead of using the age categories described previously, and we included a variable for education level (less than high school, high school or GED, more than high school). We were not able to include the third SSA vocational factor—work history and skills—because we could not easily standardize the information from the HRS data.⁶

Ethics oversight

This study was approved by the HRS Restricted Data Committee and the New England Independent Review Board.

RESULTS

Table 2 shows mean values of the outcome, physiological measures, and covariates from the 9,595 observations (7,426 distinct HRS participants) used in this study. Overall, from 2006 to 2012, 13% of the (weighted) sample received disability benefits. Exploratory data analysis suggested that many of the physiological measure values were correlated with each other (results available on request) and varied by disability benefit receipt status. Figure 1 shows four age- and sex-standardized examples of this variation by benefit receipt status from the 2010 to 2012 period.

Table 2.

The distribution of physiological measures and disability benefit outcomes among 9,595 observations in the Health and Retirement Study: 2006–2012

	Mean (weighted)	Mean (unweighted)	Minimum	Maximum	Number of missing values
Disability benefit receipt
Receiving benefits	0.13	0.13	0	1	0
Demographics
Age	58.68	58.76	51.00	65.92	0
Female sex	0.52	0.58	0	1	0
Non-Hispanic white	0.76	0.64	0	1	21
Non-Hispanic black	0.11	0.20	0	1	21
Hispanic	0.09	0.13	0	1	21
Other race/ethnicity	0.04	0.03	0	1	21
Anthropometric measures
Height	66.24	65.82	48	78	139
Underweight	0.01	0.01	0	1	461
Normal weight	0.19	0.19	0	1	461
Overweight	0.34	0.33	0	1	461
Obese	0.46	0.48	0	1	461
Waist circumference	40.32	40.38	20.5	78.0	244
Measures of physical function
Grip strength	36.43	35.01	5.5	100	310
Lung function	425.36	409.78	70	890	208
Balance: Passed test	0.79	0.77	0	1	81
Pulse	71.41	71.68	37	141	296
Systolic blood pressure	127.54	128.32	73.33	226.00	296
Diastolic blood pressure	81.31	81.73	47.33	145.33	297
Biomarkers
Total cholesterol	203.39	202.66	85.89	478.82	977
HDL cholesterol	54.65	54.74	12.52	145.81	1,385
HbA1c	5.80	5.88	3.62	17.26	943
C-reactive protein (ln)	0.50	0.54	−2.21	2.30	1,840
Cystatin C	0.99	1.00	0.05	8.56	1,127
Telomere length (ln)	0.30	0.31	−0.47	1.25	8,115
Genetic markers
APOE gene: Risk alleles present	0.30	0.30	0	2	4,085
COMT gene: Risk allele present	0.24	0.24	0	1	4,085
FTO gene: Risk allele present	0.63	0.62	0	1	4,085
IL6 gene: Risk allele present	0.59	0.55	0	1	4,085
Medications
Taking cholesterol medication	0.35	0.35	0	1	41
Taking diabetes medication	0.15	0.17	0	1	141
Taking hypertension medication	0.41	0.43	0	1	211

Note: Genetics information was available only for the 2006 and 2008 waves, but for the people observed at more than one wave we assumed that genetics did not change. Telomere length was available only in 2008. To generate the weighted means, we adjusted the HRS physical measure and biomarker weights for propensity to consent to SSA data linkage.

Figure 1.

Age- and sex-standardized distributions of physiological measures by disability benefit receipt status: Four examples from 2010–2012

In Wald tests, both the sex-stratified model and the time-stratified model fit somewhat better than the unstratified model (respectively, p = 0.024 [test statistic of 72.91 on 51 degrees of freedom] and p = 0.035 [test statistic of 67.26 on 48 degrees of freedom]). We selected the sex-stratified model as our primary specification because of its superior fit. Using a Wald test to compare this model to a model with the covariates only (also stratified by sex), we found substantial improvement in model fit when we included the physiological measures (p < 0.0001 [test statistic of 913.68 on 77 degrees of freedom]).

Table 3 shows the odds ratios associated with each physiological measure in the sex-stratified model. (Table 3 also shows 95% confidence intervals. However, with so many parameters estimated, we would expect some odds ratios to differ from 1 at a conventional 5% level of significance due to chance.) For women, the most predictive measures in this study—that is, the physiological measures with the most extreme odds ratios—were, in order, having two copies of the E4 allele of the APOE gene, the balance test, and lung function (respectively, odds ratios = 0.15 [95% confidence interval: 0.05 to 0.39]; 0.34 [0.28 to 0.41]; 0.51 [0.43 to 0.61]; we consider odds ratios less than 1 to be as extreme as their reciprocal, so that, for example, an odds ratio of 0.15 is as extreme as 1/0.15, or 6.67). The most predictive measures for men were the balance test, underweight BMI, and presence of the IL6 risk allele among nonwhites (respectively, odds ratios = 0.31 [95% confidence interval: 0.24 to 0.39]; 3.18 [1.09 to 9.30]; and 1.54 [0.75 to 3.12]). Almost all of these measures are binary or categorical. This likely reflects that, for the continuous variables, we estimated odds ratios associated with a one-standard-deviation change, which is smaller than a one-unit change for the binary variables (that is, the change from 0 to 1). Gelman and Hill (2007) recommend comparing coefficients from binary variables to those of a two-standard-deviation change in continuous variables. This would make the most predictive variables for women (transforming the results in Table 3) having two copies of the APOE E4 allele, lung function, and grip strength (odds ratios = 0.15 [0.05 to 0.39]; 0.26 [0.18 to 0.37]; and 0.30 [0.19 to 0.46]). For men the most predictive variables would be the balance test, underweight BMI, and cystatin C (odds ratios = 0.31 [0.24 to 0.39]; 3.18 [1.09 to 9.30]; and 1.69 [1.28 to 2.22]).

Table 3.

Odds ratios, by sex, of the association between physiological measures and receipt of Social Security Administration disability benefits, controlling for demographic factors and select medications

	Females		Males
	Odds ratio	95% CI	Odds ratio	95% CI
Demographics
Age (ref: 51–55)
56–60	1.29	(1.03, 1.61)	0.91	(0.70, 1.19)
61–66	1.05	(0.80, 1.37)	0.68	(0.49, 0.94)
Race (ref: Non-Hispanic white)
Non-Hispanic black	1.15	(0.84, 1.59)	1.25	(0.85, 1.82)
Hispanic	0.71	(0.48, 1.05)	0.75	(0.48, 1.18)
Other race/ethnicity	0.67	(0.35, 1.28)	1.37	(0.76, 2.47)
Anthropometric measures
Height	1.14	(0.97, 1.34)	0.99	(0.82, 1.20)
BMI category (ref: Normal weight)
Underweight	1.35	(0.35, 5.18)	3.18	(1.09, 9.30)
Overweight	0.71	(0.51, 0.99)	0.71	(0.49, 1.01)
Obese	0.92	(0.64, 1.33)	0.70	(0.45, 1.08)
Waist circumference	1.35	(1.18, 1.54)	1.10	(0.93, 1.32)
Measures of physical function
Grip strength	0.54	(0.44, 0.68)	0.73	(0.62, 0.85)
Lung function	0.51	(0.43, 0.61)	0.77	(0.68, 0.88)
Balance: Passed test	0.34	(0.28, 0.41)	0.31	(0.24, 0.39)
Pulse	1.01	(0.90, 1.13)	1.06	(0.95, 1.18)
Blood pressure
Systolic	0.85	(0.71, 1.02)	0.98	(0.81, 1.20)
Diastolic	1.09	(0.92, 1.30)	0.94	(0.78, 1.15)
Biomarkers
Cholesterol
Total cholesterol	0.95	(0.85, 1.06)	0.92	(0.81, 1.05)
HDL cholesterol	0.91	(0.81, 1.02)	0.91	(0.78, 1.05)
HbA1c	0.89	(0.79, 1.01)	0.92	(0.82, 1.04)
C-reactive protein (ln)	0.95	(0.84, 1.08)	1.25	(1.08, 1.45)
Cystatin C	1.22	(1.12, 1.33)	1.30	(1.13, 1.49)
Telomere length (ln)	0.86	(0.69, 1.07)	0.98	(0.77, 1.26)
Genetic markers
APOE gene: Risk allele present
One E4 allele (ref: 0)	0.95	(0.68, 1.32)	0.99	(0.67, 1.45)
Two E4 alleles (ref: 0)	0.15	(0.05, 0.39)	0.70	(0.20, 2.38)
COMT gene: Risk allele present	1.10	(0.78, 1.55)	0.83	(0.55, 1.24)
FTO gene: Risk allele present	1.22	(0.91, 1.63)	1.10	(0.78, 1.57)
IL6 gene: Risk allele present
Among whites	0.52	(0.33, 0.81)	0.85	(0.49, 1.46)
Among nonwhites	1.22	(0.71, 2.10)	1.54	(0.75, 3.12)
Medications
Taking cholesterol medication	1.61	(1.30, 1.99)	1.60	(1.25, 2.05)
Taking diabetes medication	1.14	(0.86, 1.53)	1.30	(0.94, 1.79)
Taking hypertension medication	1.08	(0.86, 1.36)	1.37	(1.06, 1.76)
Time period
Period: 2010–2012 (ref: 2006–2008)	1.01	(0.83, 1.24)	1.12	(0.86, 1.46)
Model constant
Model constant	0.15	(0.08, 0.28)	0.58	(0.27, 1.22)

Note: For all continuous variables, we show results for a one-standard-deviation change in the variable. We present 95% confidence intervals for each estimate. However, with so many parameters estimated, we would expect some to differ from 1 at a conventional 5% level of significance due to chance.

BMI = body mass index; CI = confidence interval.

Table 4 shows the cut point value that maximized Youden’s Index (that is, maximized the sum of sensitivity and specificity) under our primary model specification. Specifically, if we used a cut point value of 0.120 to predict benefit receipt (that is, predicting that observations with a propensity score value below 0.120 did not receive benefits while those with values above 0.120 did receive benefits), our classification scheme would have a sensitivity of 75.90% and specificity of 73.50%—meaning that we would correctly identify 75.90% of people receiving benefits and correctly identify 73.50% of people not receiving benefits. Sensitivity and specificity would remain comparable if we allowed a different cut point for men than for women. Figure 2 shows the relationship between sensitivity and specificity graphically in an ROC curve. The figure demonstrates that the primary specification of the model, including physiological measures, has substantially higher AUC (0.803) than the expectation with no predictors—represented by the 45-degree line (AUC = 0.5)—or the model with covariates only (AUC = 0.667).

Table 4.

Using propensity scores to identify Social Security Administration disability beneficiaries

Model	Optimal cut point for identifying beneficiaries (maximizing sensitivity + specificity)	Sensitivity (proportion of beneficiaries correctly identified) using the optimal cut point	Specificity (proportion of non-beneficiaries correctly identified) using the optimal cut point	Youden’s Index (sensitivity + specificity – 1) using the optimal cut point	Area under the curve
Primary specification (sex-stratified)	0.120	75.90%	73.50%	0.494	0.803
Men only	0.120	75.97%	73.42%	0.494	0.799
Women only	0.124	75.03%	74.50%	0.495	0.808
Without race variables	0.123	74.57%	74.05%	0.486	0.801
Without grip strength, lung function, and balance test	0.131	66.09%	69.91%	0.360	0.741
With covariates only	0.116	66.10%	60.79%	0.269	0.667

Figure 2.

Sensitivity and specificity of physiological measures for explaining receipt of Social Security Administration disability benefits

Note: Points on the ROC curve represent deciles of the propensity score distribution. The 45-degree line represents the expected curve when predicting at random (that is, predicting by chance alone).

In addition to results from the primary model specification, Table 4 shows optimal propensity score cut points and their associated sensitivity and specificity for two other analyses with policy implications for SSA. The model appears comparable, with a similar AUC, when we remove variables for race (AUC = 0.801). However, the model has a much lower AUC (0.741) when we remove variables for three physiological measures that are easy to manipulate: grip strength, lung function, and performance on a balance test. This suggests that together these three variables provide considerable information to predict benefit receipt—moving the AUC from 0.741 to 0.803, whereas all the other physiological measures together move it from 0.667 (the value in a model with covariates only) to only 0.741.

Finally, when we redefined the model covariates to include SSA vocational factors, we found that adding physiological measures gave a far superior model fit over the vocational factors and other covariates alone (p < 0.0001 [test statistic of 825.82 on 77 degrees of freedom]).⁷ This means that the model explains benefit receipt much better with physiological measures than with only the available vocational factors and other covariates.

DISCUSSION

This study tests the extent to which 19 distinct physiological measures can predict current receipt of SSA disability benefits among adults ages 51 to 65. Better understanding of the relationship between physiological measures and benefit receipt could provide greater insight into the disability experience, or have program implications for SSA. To the best of our knowledge, this is the first study to examine relationships between SSA disability benefit receipt and either biomarkers or genetic indicators.

We found that, collectively, the 19 physiological measures added substantial predictive power beyond the model covariates, which included (among other things) age, sex, and race—improving the AUC from 0.667 to 0.803. Using the propensity score cut point that maximized Youden’s Index, our model correctly predicted 75.90% of SSA disability beneficiaries in the linked HRS-SSA sample and 73.50% of non-beneficiaries in the sample. Nevertheless, this means the model still incorrectly classified roughly a quarter of all observations. We thus consider the physiological measures to have moderate predictive power overall.

The findings we present are exploratory. We have no way to assess causality between the measure values and benefit receipt and we do not claim any causal link. Furthermore, this study assesses predictive power of the physiological measures taken collectively—estimating the associations of several variables simultaneously—so we interpret odds ratios on each individual predictor with caution. For example, we estimated the relationship between obesity and benefit receipt (Table 3) holding constant at least one other measure of adiposity (waist circumference), use of three medications that are linked to obesity (to manage hypertension, lipid levels, and diabetes), and several factors plausibly in the causal pathway between obesity and health outcomes (systolic and diastolic blood pressure, total cholesterol, pulse rate, and HbA1c). In this context, interpreting the odds ratios on obesity or related (collinear) measures is not straightforward. In addition, interpretation for the genetic variables is complicated because we selected gene variants to include in our model based on GWASs of people with mostly European ancestry. It is possible that our model does not fully account for interactions between the genetic variables and other characteristics that would be needed to give the odds ratios a meaningful substantive interpretation.

Our study results suggest that physiological measurement is relevant to SSA program operations. Taken together, physiological measures do provide useful information to predict disability benefit receipt. It is also likely that one could improve benefit receipt prediction further by using richer data than was available here (although, to date, the HRS remains one of the best available sources of physiological data for a general population). Other studies show that additional biomarkers—including, for example, hormones such as cortisol and dehydroepiandrosterone or proteins such as insulin-like growth factor 1—can also predict health outcomes (Baylis et al., 2013; Goldman et al., 2006). Future research might also add measures that better reflect mental health status, as opposed to only physical health status.

As noted previously, we see two possible policy applications of this work for SSA—although both would require substantial research beyond the current study before they are practical. First, it might be possible to incorporate population-level information about physiological measures into forecasts of demand for SSDI and SSI benefits (for example, using expectations about future cholesterol levels and the estimated association between cholesterol and benefit receipt to help forecast future benefit demand). Second, it might be possible to add new physiological criteria to SSA’s disability determination process, in addition to the existing medical criteria and vocational factors. This second application might speed the disability determination process, but would depend, in particular, on research assessing the extent to which benefit receipt is predicted by physiological measures that are not easily manipulated by program applicants.