Association Between New-Onset Medicaid Home Care and Family Caregivers’ Health

Key Points Question How does family caregivers’ health change when the person they care for begins to receive Medicaid home care services? Findings In this longitudinal cohort study, family caregivers’ self-rated mental health improved significantly after their family member began receiving Medicaid home care services. Their self-rated physical health did not change. Meaning The findings of this longitudinal cohort study suggest that Medicaid home care programs may have spillover benefits, affecting not only their direct recipients but also those recipients’ family caregivers.

= 0 + + + 1 * ℎ * _ + 2 + i indexes individuals and t indexes time in rounds. is a vector of round indicator variables (fixed effects), is a vector of person indicator variables (fixed effects), is a vector of time-varying confounders, and is the outcome.
* ℎ is the difference-in-difference estimator: it is zero in rounds in which no one in the household receives Medicaid home care and one in the first round in which someone in the household receives Medicaid home care.
_ indicates whether an individual is a non-disabled adult (likely caregiver), disabled adult, or Medicaid home care recipient.
Individuals who were never exposed to Medicaid home care had a value of zero for * ℎ in all waves. Thus, they did not contribute to the difference-in-difference effect estimate; they did, however, contribute to the estimates of overall time trends and other covariates, and, as such, acted as a comparator group for the difference-in-difference estimate. Individuals who already received Medicaid home care in wave 1 also did not contribute to the effect estimate because they had no pre-exposure outcome and dropped out of the model after the first wave, so their change in outcome could not be assessed.
Our event study model was as follows: Where subscript p indexes the event period (number of rounds before or after onset of Medicaid home care), and ℎ _ denotes a series of binary indicators equal to 1 if an observation occurred in time period p for person i. For individuals who lived in households that never received Medicaid home care, ℎ _ was equal to zero for all p. For individuals who lived in households that already received Medicaid home care in round one, it was impossible to determine the event period, because it was unknown how long these households had received Medicaid home care before the survey began. Thus, these individuals also received a value of zero for ℎ _ for all p and, like in the main difference-in-difference model, did not contribute to the effect estimates. All other notation is as above.
Our model to assess effect measure modification was as follows: Where indicates the effect measure modifier of interest (e.g. unemployment.). All other notation is as above.
All results should be interpreted as the within-person change in the outcome associated with the onset of Medicaid home care, above and beyond any changes in the outcome that occurred over time among individuals who did not start Medicaid home care, conditional on all the covariates in the model. In other words, these results reflect the difference between an individual's self-rated health one to six months after the start of Medicaid home care services and their self-rated health in the six-month to two-year period before anyone in their household received Medicaid home care, relative to the difference over this same period for a similar individual whose household members did not begin receiving services in that wave.

Variance estimation
In fixed effects models, it is generally appropriate to use cluster-robust standard errors clustered at the level of the fixed effect -here, at the individual level. 2 However, because the MEPS data derive from a clustered survey design, we also considered clustering the standard errors at the survey cluster level. In order to assess which approach was most appropriate, we tested our primary model three ways: with standard errors clustered at the individual level, with standard errors clustered at the survey primary sampling unit (PSU) level, and with standard errors clustered at the stratum level.
We found that standard errors were largest when we clustered them at the individual level. When we clustered standard errors at the PSU and stratum level, the standard error on the main association decreased by 33% and 3%, respectively. Thus, we proceeded with the most conservative option: standard errors clustered at the individual level.

Sensitivity Analyses
Weighted analyses MEPS provides cross-sectional and longitudinal survey weights to weight the sample population to match national demographics. The cross-sectional survey weights adjust for the oversampling of racial/ethnic minorities and lowincome populations in MEPS. The longitudinal weights also adjust for loss-to-follow-up; however, they have a 97% correlation with the cross-sectional weights, suggesting that they primarily adjust for oversampling.
Our primary analysis did not apply the survey weights, as we had no basis for estimating a national average effect of Medicaid home care. We performed a secondary analysis that was equivalent to our primary analysis but with the survey weights applied. We then performed the same analysis but truncated the weights at the 90 th percentile, in order to evaluate the impact of using or excluding extremely large weights.
Additionally, we performed descriptive analysis of the survey weights to explore the ways in which weighting vs. not weighting our primary models might affect the results. We described the univariate distribution of the survey weights using summary statistics and a histogram. We assessed how the demographics of individuals with large weights (90 th percentile or higher) differed from the rest of the population using chi-squared tests.

Analyses of clinical significance of the outcome
Our primary outcomes of interest were self-rated overall mental and physical health. We used these outcomes rather than more clinically-applicable health outcomes because they were available for all individuals in every wave. Additionally, for our primary analysis, we operationalized these outcomes as continuous, because this allowed us to use conditional likelihood models without dropping individuals whose outcomes remained the same for the entire study period. (Conditional logistic and ordinal regression would drop all individuals who have no variation in their outcome during the study period.) Although these outcomes worked well for modelling purposes, their clinical interpretation is more challenging. To provide readers with a better sense of the clinical meaning of our primary outcome, we assessed the relationship between self-rated mental health and more commonly used screening tools for mental health disorders in our study population.
Beginning in 2004, we had access to two additional measures of mental health: the Patient Health Questionnaire-2 (PHQ-2) and the Kessler-6. These questions were only asked during round 2 and 4 interviews, so they could not be used in our primary model. However, they provide clinically useful metrics of depression (PHQ-2) and severe psychological distress (Kessler-6). Additionally, these questions were asked using a self-administered questionnaire that each adult in the household filled out for themself, rather than by a single household respondent who answered for everyone in the household. Thus, these measures are less prone to measurement error than the global self-rated mental health measure. Details on the PHQ-2 and Kessler-6, as well as the relationship between these measures and self-rated mental health, are available elsewhere. 3 Using data from all members of our study population in all rounds in which the PHQ-2 and Kessler-6 data were available, we assessed the relationship between an 0.01 standard deviation change in the self-rated mental health score and the log odds of being depressed or having severe psychological distress using logistic regression with robust standard errors. The estimating equation was as follows: is the outcome for person i at time t (either screening positive for depression on the PHQ-2 scale or screening positive for severe psychological distress on the Kessler-6 scale) and is the standardized selfrated mental health score, in increments of 0.01 of a standard deviation, for person i at time t.
We operationalized the exposure as an 0.01 standard deviation change in the self-rated mental health score in order to make it straightforward for the reader to translate any association estimate from our main models into an approximate estimate of clinical significance using the results of this model. We did not condition on any covariates, person, or round, as we sought to simply assess the bivariate relationship between these variables. We defined depression as a score of 3 or higher on the PHQ-2 and severe psychological distress as a score of 13 or higher on the Kessler-6, typical cut-offs used elsewhere in the literature. 3 Results of this model should be used only to assess the validity of the self-rated mental health measure and to provide the reader with a general sense of the clinical meaning of a certain magnitude change in self-rated mental health. They should not be used to make claims about the relationship between onset of Medicaid home care and either depression or severe psychological distress. Our models cannot provide a definitive assessment of those relationships, as our models do not account for the uncertainty of the relationship between self-rated mental health and depression or severe psychological distress. eResults eTable 1 shows a comparison of the demographics of the 50 adults excluded due to missing data versus the analytic population. It should be noted that people living in households that received Medicaid home care were more likely to be excluded due to missingness because these individuals dropped out of the model after the first round of exposure to Medicaid home care; thus, they had fewer rounds of data available and were more likely to have no rounds with complete data.
Certain individuals were included in the model but did not contribute to effect estimates because they did not have both a pre-exposure period and a post-exposure period. This includes all individuals living in households that were never exposed to Medicaid home care, who had no "post" period. It also includes all individuals living in households that were already exposed to Medicaid home care in round one, as these individuals had no "pre" period. A comparison of the demographics of never-disabled adults (likely caregivers) who were already exposed to Medicaid home care in round one (and thus did not contribute to effect estimates) and those who were exposed after round one (and did contribute to effect estimates) can be found in eTable 3.
Results of the primary analysis shown for self-rated mental health on the original scale (unstandardized) are found in eTable 4. Note that the original scale ranges from 1 to 5, with lower numbers indicating better self-rated mental health: 1 indicates excellent health and 5 indicates poor health. On this scale, the average self-rated mental health of likely caregivers in rounds prior to the onset of Medicaid home care in the household was 2.35 (95% CI: 2.28, 2.41). In fully-adjusted models, onset of Medicaid home care was associated with a -0.090 change in self-rated mental health among likely caregivers (95% CI: -0.168, -0.012; p=0.024).
Results of the weighted analyses are found in eTables 8 and 9. In the weighted analysis, the association between onset of Medicaid home care and self-rated mental health was attenuated (0.019 standard deviations; 95% CI -0.065, 0.103; p=0.66).
Results of the descriptive analysis of the weights are found in eFigure 4 and eTable 10. The distribution of the weights was extremely skewed: the median weight was 14,087.64 and the inter-quartile range was 15,500.761, but the 90 th percentile was 35,436.57 and the maximum weight was 170,898.9. This indicated that a small number of individuals were up-weighted more than ten times the median individual.
Comparison of individuals with the top 10% of weights to the rest of the population (eTable 10) revealed that individuals with the top 10% of weights were disproportionately older, male, White non-Hispanic, highly educated, high-income, employed, and in good health, and were less likely to live with someone with cognitive limitations. Thus, the individuals with extremely high weights typically fell into the demographic categories that our analysis of effect measure modification found benefitted less from Medicaid home care.
Truncating the weights at the 90 th percentile to exclude extreme weights resulted in an estimate of association between onset of Medicaid home care and self-rated mental health similar to our unweighted analysis, though the 95% confidence intervals were slightly wider due to the smaller sample size (0.079 standard deviations; 95% CI: -0.002, 0.160, p=0.06).
Results of the analysis of the clinical significance of self-rated health are found in eTable 11. An 0.01 of a standard deviation improvement in self-rated mental health predicts a 1.09% lower odds of having depression (OR: 0.9891, 95% CI: 0.9886, 0.9896) and a 1.25% lower odds of having severe psychological distress (OR: 0.9875, 95% CI: 0.9869, 0.9882). An 0.075 of a standard deviation improvement in self-rated mental health (the main association between onset of Medicaid home care and likely caregivers' self-rated mental health found in our fully-adjusted primary analysis) thus predicts a 7.92% lower odds of having depression (OR: 0.9208, 95% CI: 0.9174, 0.9243) and a 9.02% lower odds of having severe psychological distress (OR: 0.9098, 95% CI: 0.9057, 0.9146). eTables eTable 1: Comparison of demographic and health characteristics of study participants with and without missing data All cells show n (percent). Population comprises all adults living in households with at least one disabled adult interviewed in 1996-2017. Analytic population includes all adults living in households with at least one disabled adult, 1996-2017. "Medicaid home care onset: Likely caregiver" is the primary difference-in-difference estimate of interest: the association between onset of Medicaid home care in the household and self-rated mental health among likely caregivers.     Analytic population includes all adults living in households with at least one disabled adult, 1996-2017. "Medicaid home care onset: Likely caregiver" is the primary difference-in-difference estimate of interest: the association between onset of Medicaid home care in the household and self-rated mental health among likely caregivers.  Analytic population includes all adults living in households with at least one disabled adult, 1996-2017. "Medicaid home care onset: Likely caregiver" is the primary difference-in-difference estimate of interest: the association between onset of Medicaid home care in the household and self-rated mental health among likely caregivers.