Social Work Staffing and Use of Palliative Care Among Recently Hospitalized Veterans

Key Points Question Is additional social work staffing in primary care associated with a higher use of palliative care among recently hospitalized patients? Findings In this cohort study involving 43 200 veterans in the Department of Veterans Affairs health system, the addition of social workers to primary care teams was associated with nearly doubled use of palliative care among veterans who had had inpatient hospital care compared with before the increase in social work staffing. Meaning These findings suggest that social workers in primary care teams may facilitate access to palliative care for recently hospitalized patients.


= + + + +
where is a year-month fixed effect, is fixed effect for sites (i.e., primary care clinic or community-based outpatient clinic), is a dummy variable representing participation in the program at time t, are individual Veteran characteristics, are time varying unobservable variables that are mean independent of everything else, and is the effect of the staffing intervention. This approach is a standard in applied work and historically was considered a strong quasi-experimental design because controls for systematic differences between Veterans at sites that are consistent over time and controls for secular changes over time. In our sample, all sites eventually participate in the program. Therefore, in our TWFE estimators newly treated sites are compared to "not yet treated" sites and "already treated" sites. It requires some key assumptions, including that 1) there are no exogenous events other than the program that affect the outcome and coincide with the treatment intervention, and 2) the trends in outcome for treatment and comparison groups would be similar, or parallel, in the absence of the intervention. In this study, the staggered timing of the program start dates decrease the likelihood of a time-varying confounder that coincided with the intervention. We address the second parallel trends assumption by examining pre-trends in the outcome (see below).

Difference-in-difference: alternative estimators
Recent work has shown that the traditional TWFE estimators can result in hard-to-interpret weights and even negative weights for some units because of dynamic treatment effects. 1 To address this problem, we also estimate our program impact using Sant'Anna and Callaway's approach with group-time-average treatment effects. Sites are grouped by the month when they started the staffing program. The effect of the program is estimated for each group, using the notyet-treated sites as the comparison sample. We report the aggregated group-time-averaged effects: the program effect is calculated for each group of clinics that began the intervention at the same time, and these effects are then weighted by the size of the groups.

Sensitivity analyses and checks
Parallel trends assumption We created and event study plot using the DID package to examine trends in the outcome relative to the comparison in the months leading up to treatment (figure S1). The pre-treatment dummy coefficients are statistically not different from zero, and suggest that prior to the intervention the trend in palliative care may, if anything, trend downward relative to the trend of the comparison sites.

Collider bias: sample selection after the social work staffing intervention
Previous work has shown that the VA Social Work Staffing Program may decrease the risk of hospitalization for high-risk Veterans, which could lead to "collider bias" because the intervention could be associated with selection into the sample in ways that are systematically correlated with use of palliative care. 2 While we cannot entirely rule out possible collider bias, we show that the characteristics of Veterans hospitalized in the month before and two months after the intervention are substantially similar (Error! Reference source not found.). Because o ur study cohort is an unbalanced panel, the table shows only hospitalizations in the month before and the third month after the intervention. The table shows only very small differences between the groups for the demographics and comorbidities in our sample, with all SMD<0.05 (a conservative rule of thumb for covariate balance is that variables have SMD<0.1). Although this check does not rule out the possibility that selection on unmeasured characteristics could bias selection into hospitalization, it is suggestive that systematic differences before and after intervention in the characteristics of hospitalized Veterans are unlikely.
Nonetheless, it is useful to explore the likely extent and direction of collider bias from previous work. In previous work, we found that the social work program intervention was associated with a decrease of approximately 4% decrease in hospitalizations among high-risk Veterans, and no significant effect among the population at large. If those "averted" hospitalizations were among Veterans that would have had low or no palliative care use, the sample difference in the denominator could bias our estimate of the program's effect on palliative care in the positive direction. In a back-of-the-envelope example, suppose we have a group of 1000 hospitalized Veterans, 14.5 who use palliative care, prior to program intervention. After the social work intervention, this group would be reduced by 4.4%, so the new denominator is =956. In the absence of any true program effect on palliative care use, the new rate is 14.5/956 = 15.2. This example puts an upper bound of 0.7 Veterans per 1000 that we are over-estimating due to collider bias. Because the lower bound of the 95% confidence intervals were >2 per 1000 for the overall sample, even accounting for this bias we would still see an impact of the social work program on primary care.
Competing risk: Check for change in 30-day mortality associated with the social work program Increased survival and time at risk could explain an increase in palliative care referrals. We checked for a change in mortality associated with the social work staffing intervention (Error! R eference source not found.). We found small and not statistically significant associations of the intervention with 30-day mortality, suggesting that our results were unlikely to be driven by changes in survival.

Falsification test with palliative care inpatient consults
Because it is possible that other changes in health system delivery that were correlated over time with the staffing intervention may have influenced use of palliative care, we investigated the association of inpatient palliative care consults with the intervention (Error! Reference source n ot found.). We identified consults through orders in the health record with associated stop codes of 351 or 353.