2012 HSR&D/QUERI National Conference Abstract
1057 — Who Benefits or Is Harmed by This Intervention Program? Assessing Variation in Intervention Impact in Randomized Field Trials
Toyinbo PA, COE Tampa;
In order to personalize or tailor interventions to maximize impact on a given outcome among different Veteran subgroups, there is a need for a statistical model that is able to describe not only the main effect of an intervention, but also how the effect varies as a function of individual-level baseline risk or protective factors. Immediate challenges faced by candidate statistical models are often produced by nonlinearity and measurement errors which commonly arise in behavioral data. Existing statistical methods are not effective when measurement errors and non-standard nonlinearity jointly feature in the statistical relationship under investigation. With this motivation the objective was to develop a more flexible statistical modeling technique for assessing complex relationships between a distal outcome and 1) baseline characteristics measured with or without errors, and 2) baseline-treatment interaction.
The generalized additive model (GAM) was integrated into a conventional structural equation model (SEM) framework to develop a new hybrid model: Additive Latent Variable (ALV) model. The SEM produces a measurement error-free summary of the outcome from its indicator items. GAM is a semi-parametric model that employs smooth functions to capture nonlinear relationships between the outcome and intervention/baseline, without requiring a priori knowledge of the functional forms of these relationships. A maximum likelihood estimation of the model was implemented by using a Markov Chain Monte Carlo (MCMC) version of the EM algorithm. All programming was carried out in R statistical application.
Simulation studies show that the ALV model performs well and the model estimates are sufficiently close to the population values. Its application was illustrated using real data examples.
The ALV modeling technique allows researchers to assess how an intervention affects individuals differently as a function of baseline risk even when measured with error, and uncover complex relationships in the data that might otherwise be missed. In practice, its users are relieved from the need to decide functional forms for the complex relationships before the model is run.
Investigators engaged in health services or QUERI research can more effectively address a major research question relevant to best practices: who benefits or is harmed by this intervention program?