Journal of biometrics & biostatistics最新文献

Intervention effects on continuous longitudinal normal outcomes are often estimated in two-arm pre-post interventional studies with b≥1 pre- and k≥1 post-intervention measures using "Difference-in-Differences" (DD) analysis. Although randomization is preferred, non-randomized designs are often necessary due to practical constraints. Power/sample size estimation methods for non-randomized DD designs that incorporate the correlation structure of repeated measures are needed. We derive Generalized Least Squares (GLS) variance estimate of the intervention effect. For the commonly assumed compound symmetry (CS) correlation structure (where the correlation between all repeated measures is a constantρ) this leads to simple power and sample size estimation formulas that can be implemented using pencil and paper. Given a constrained number of total timepoints (T), having as close to possible equal number of pre-and post-intervention timepoints (b=k) achieves greatest power. When planning a study with 7 or less timepoints, given large ρ(ρ≥0.6) in multiple baseline measures (b≥2) or ρ≥0.8 in a single baseline setting, the improvement in power from a randomized versus non-randomized DD design may be minor. Extensions to cluster study designs and incorporation of time invariant covariates are given. Applications to study planning are illustrated using three real examples with T=4 timepoints and ρ ranging from 0.55 to 0.75.

Background: Intervention effect on ongoing medical processes is estimated from clinical trials on units (i.e. persons or facilities) with fixed timing of repeated longitudinal measurements. All units start out untreated. A randomly chosen subset is switched to the intervention at the same time point. The pre-post switch change in the outcome between these units and unswitched controls is compared using Generalized Least Squares models. Power estimation for such studies is hindered by lack of available GLS based approaches and normative data.

Methods: We derive Generalized Least Squares variance of the intervention effect. For the commonly assumed compound symmetry correlation structure, this leads to simple power formulas with important optimality properties. To maximize power given a constrained number of total time points, we investigate on the optimal pre-post allocation with the local minimization of variance.

Results: In four examples from nursing home and HIV patients, the Toepltiz within-unit correlation of repeated measures differed from compound symmetry. We applied empirical Toeplitz based calculations for variance of the estimated intervention effect to these examples (each with up to seven longitudinal measures). Unlike what happened under compound symmetry, where power was often maximized with multiple observations being pre-intervention, for these examples, having one pre-intervention measure tended to maximize power. Attempts to approximate the Toeplitz variance structures with compound symmetry (to take advantage of the simpler formulas) resulted in overestimation of power for these examples.

Conclusions: While compound symmetry correlation among repeated within-unit measures leads to simple power estimation formulas, this structure often did not hold. There may be strong underestimation of variance of the intervention effect estimate from incorporating short-term within-unit correlation estimates as a common compound symmetry correlation to approximate an unknown Toeplitz correlation without adequately accounting for the correlation between repeated measures declining with time.