Users Of Veteran-Directed Care And Other Purchased Care Have Similar Hospital Use And Costs Over Time

The Veteran-Directed Care (VDC) program facilitates independent community living among adults with multiple chronic conditions and functional limitations. Family caregivers value the choice and flexibility afforded by VDC, but rigorous evidence to support its impact on health care costs and use is needed. We identified veterans enrolled in VDC in fiscal year 2017 and investigated differences in hospital admissions and costs after initial receipt of VDC services. We compared VDC service recipients to a matched comparison group of veterans receiving homemaker or home health aide, home respite, and adult day health care services and found similar decreases in hospital use and costs from before to after enrollment in the groups. Further investigation into trends of nursing home use, identification of veterans most likely to benefit from VDC, and relative costs of operating VDC versus other purchased care programs is needed, but our results suggest that VDC remains a valuable option for supporting veterans and caregivers.


Service Duration
The count of months during which a patient has at least one VDC or other purchased care service each month

Mortality
Percentage of patients who died during the 12 months following the index date in VDC or comparison groups SOURCE: VHA Corporate Data Warehouse (CDW) Factbooks and authors' analysis of patient data in VHA.

Differences in characteristics of enrollees in VDC and comparison groups
In Exhibit A3, we present differences in characteristics of enrollees in VDC and the comparison groups. We can write this model as: Y = β0 + β1*TMT + β2*TIMEt + β3*OFFER + β4*OFFER*TIMEt + β5*TMT*TIMEt+ β6*OFFER*TMT + δ*VD + u, where TMT is an indicator for whether a patient belonged to VDC treatment group, OFFER is an indicator that a patient received care (either VDC or other purchase care services) at a site that offered VDC in FY17, TIME represents a set of indicators for month t, and VD is an indicator that the patient received VDC during the "post" period at a site that offered VDC.
We are interested in estimating δ. After accounting for collinearity with the treatment variable and removing terms for time-invariant characteristics; this model simplifies to: Y = β0 + β2*TIMEt + β4*OFFER*TIMEt + δ*VD + u We further simplify the OFFER*TIMEt term to an indicator that equals one if the individual is in a site without VDC in the post period, and zero otherwise: NOOFFER_POST. The model is simplified to: Y = β0 + β2*TIMEt + β4*NOOFFER_POST + δ*VD + u.

Matching on Time-Invariant Covariates
Because our treatment groups differed on some sociodemographic and clinical characteristics (Exhibit A4), we explored whether matching on time-invariant covariates improved precision of estimates. We considered matching on patterns of health care use before VDC, but matching on pre-treatment levels of outcomes increases risk of bias from regression to mean. Because NOSOS risk score and CAN score are used to predict costs and risks over time and are regularly re-calculated, they also were excluded from the match. VDC and comparison groups were matched on the following characteristics: Elixhauser comorbidities, age, dementia diagnosis, and presence of either SCI or TBI. Coarsened exact matching was used to create matched subsets of the data. Analyses were re-run on matched datasets.

Full list of models and covariates
A summary of all models and results is presented in Exhibit A7 and is followed by detailed results for specific models.
To model all-cause hospital admissions, we began with a logistic fixed effects model in our unmatched sample (Exhibit A8). Because fixed effects logistic models can be biased away from the null for panel data, we reestimated this model with population-averaged models, which are biased towards the null (Exhibit A9). We repeated these analyses in unmatched samples within active sites (Exhibits A10 and A11). We repeated these analyses in two different matched samples: VDC and both comparison groups matched (Exhibit A12 -fixed effects, Exhibit A13 -population-averaged), and VDC and comparison group at active sites matched (Exhibit A14 -fixed effects, Exhibit A15 -population-averaged).
We analyzed cost among all individuals in our sample, and excluded one person (two person-months) with unreasonably high costs (cost > $450,000). To account for time-invariant confounders, we began with a linear fixed effects model in our unmatched sample (Exhibit A16). To better approximate the distribution of cost data, we also ran population-averaged panel-data models with gamma distribution and log link (Exhibit A17) . We repeated our fixed effects and population-averaged models using two different matched samples (VDC and both comparison groups matched [Exhibits A18, A19], and VDC and comparison group at active sites matched [Exhibits A22, A23]). We also ran an unmatched fixed effects and population-averaged models on the sample at sites with active programs (Exhibit A20, A21). To explore whether any differences across groups might be obscured by the large number of zeroes in the dataset, we also modeled costs with a denominator of those with at least one month of non-zero cost (Exhibit A40, A41). Standard errors were calculated using bootstrapping in unmatched models.
We analyzed ambulatory care sensitive hospital admissions (Exhibit A28 -A35) using the same specifications as for models of all-cause hospital admissions.
Sensitivity tests: Sensitivity tests included using a different definition of post-period, different selection criteria of VDC patients, a restricted sample of individuals who stayed alive till the end of the study, and different service durations. Most results remained non-significant. Among non-decedents, VDC receipt was associated with 15%-19% lower odds of all-cause hospital admissions, but only in our matched samples (details below). Among non-decedents, VDC receipt was associated with reduced hospitalization costs in both fixed effects and population-averaged models.

Definition of post period
To allow for potential lags in data processing, we re-ran our models with an alternate definition of the post period (beginning two months after the index date).

Sample selection criteria
In our primary analyses, we allowed VDC enrollees into the sample if they did not have any past GEC purchased care services. Our strategy allows us to model the effect of VDC, regardless of whether a patient received other GEC services in the past year, but it may introduce differences among our treatment and comparison groups. We re-ran analyses among individuals who received FY16 purchased care services in all 38 sites. Among this subset, we re-ran our models to compare treatment effect estimates to main analysis.

Results among non-decedents
In our primary analyses, we included the person-months for patients who died during the study period if the observation was for a month before the patient's month of death. Outcomes for decedents were treated as missing and non-informative. We re-ran our models for hospital admissions, restricting the sample to those who remained alive till the end of the study period. Results were substantively similar.

Service duration
We repeated our analyses on subsets of patients who received at least one service per month for at least 3 or 6 months in the follow-up period.