2011 HSR&D National Meeting Abstract
3091 — A Comparison of Longitudinal Cost Modeling Techniques
Smith VA (COE - Durham), Maciejewski ML
(COE - Durham), Jackson GL
(COE - Durham), Edelman DE
(COE - Durham), Olsen MK
(COE - Durham)
Longitudinal cost analyses are becoming more common in health services research. When the cost variable is semicontinuous (e.g., a high proportion of patients with zero cost), a range of analytic methods exist for modeling these outcomes. In this study (QUERI RRP-09-407), we describe and compare innovative statistical methods for analyzing VA costs from a longitudinal randomized trial of veterans with diabetes and hypertension. We also provide a roadmap for choosing among these analytic methods.
The Group Visits randomized trial (HSRD IIR-03-084) examined the effectiveness of group medical clinics on blood glucose and blood pressure in veterans with comorbid diabetes and hypertension. VA specialty care costs of this cohort were estimated in 6-month intervals over 42-months. A comparison was done of predicted specialty care costs estimated from 1) one-part; 2) uncorrelated two-part; and 3) correlated two-part random effects models developed to determine if specialty care costs differed between treatment and control patients one year prior to trial initiation, in the 12-months during the trial, and in the eighteen months after trial completion.
The VA specialty care costs of 239 veterans over 42-months were drawn from the VA’s DSS National Data Extract. The proportion of veterans with no specialty care costs varied over time from 6% to 19%, which created ambiguity in the appropriate model choice. The statistical significance of the treatment effect varied by modeling strategy.
These three approaches for estimating longitudinal costs produced varying results in analyses of longitudinal cost outcomes among veterans participating in a diabetes self-management trial.
Health services research studies often involve analysis of longitudinal cost outcomes. Results can differ between one-part and two-part models, as well as between correlated and uncorrelated models. Thus, analysts should examine a variety of factors to inform their choice among different analytical approaches when estimating longitudinal costs.