In real-world studies of geriatric palliative care programs, policies, or treatments, treatment initiation may need to be staggered across units in ways that are outside of an investigator's control. If differences across cohorts or in organizational characteristics associated with both treatment timing and outcome are not controlled for in analyses, they may obscure estimates of true treatment effects. Current methods for accounting for treatment effect timing heterogeneity either do not account for confounding or may introduce additional bias due to regression to the mean. A potential solution involves inverse probability of treatment weights (IPTW) to adjust for confounding across treatment groups defined by receipt timing, but IPTWs lead to biased estimates in cross-sectional evaluations comparing multiple treatments. We have developed an alternative method, vector- based kernel weighting (VBKW), that outperforms IPTW in cross-sectional evaluations. The degree to which VBKW reduces bias and improves efficiency over IPTW in longitudinal applications has not yet been explored. We propose to evaluate the bias and efficiency of estimates in data generated from observational designs with staggered data collection and/or staggered treatment timing. We aim to: 1) Compare the bias and efficiency of estimated average treatment effects using VBKW, IPTW, and no adjustment in difference-in- difference analyses of retrospective cohort studies with staggered treatment timing within a cohort, 2) Compare the bias and efficiency of estimates using VBKW, IPTW, and no adjustment on data from longitudinal panel studies where treatment effects may vary across and within waves of data collection, and 3) Identify the degree of treatment effect heterogeneity required for VBKW and IPTW to lead to different inferences. We will use Monte Carlo (MC) and plasmode simulations to obtain estimates of average treatment effects with VBKW, IPTW, or no adjustment of data obtained from observational studies with staggered data collection and/or staggered treatment timing. MC simulations on investigator-generated data (n=900, 9600) will allow us to examine the impact of different analytic scenarios (e.g., sample distribution across treatment timing groups, number of distinct times in which treatment is initiated) on the relative performance of VBKW and IPTW estimates. Plasmode simulations will allow us to verify that our results are robust to data generating process and will be derived from an observational analysis of veterans' participant-directed care services and days in the community (n=848,500 person-months, 38 medical centers) and the Health and Retirement Study (HRS). We will evaluate bias, efficiency, and covariate balance. We will identify when VBKW or IPTW is superior for estimating the effect of a treatment provided at different times. We will summarize our results in practical guidance for investigators. Our results will improve investigators' ability to generate rigorous evidence from studies of geriatric palliative care in which treatment or event timing cannot be controlled.

Treatment effects may vary with time for several reasons, including changes in policies or attitudes over time, interaction with simultaneous delivery system change, and differences between early and late adopters. To the extent these differences are ignored, treatment effect estimates can be biased, leading promising treatments to be overlooked or allowing harmful practices to be sustained. Our results will improve investigators' ability to generate rigorous evidence from studies of geriatric palliative care in which treatment or event timing cannot be controlled.