03/04/09 — Intra-class Correlations (ICC) within Dental Practices in a Multi-center Observational Study: A Bayesian Approach for Binary Data

Anna Korpak

The intraclass correlation coefficient (ICC) is a measure of pairwise correlation between two observations within the same group.. Accounting for ICC is important for study design and for preserving inference in group-randomized trials. While analysis can account for ICC without actually estimating it, reasonable estimates for ICC are crucial to sample size calculations when planning cluster-based trials. The focus of this talk is the use of Bayesian methods for estimation of ICC in the case where the outcome variable of interest is binary. Using example data from a cross-sectional observational study in dentistry, with patients clustered within dental practices, we estimate ICC for various treatment-pattern and disease outcomes. Turner, Omar, & Thompson (2001) outline a general approach for Bayesian estimation of ICC for binary data and we adapt this to our scenario. We investigate the performance of this Bayes approach to ICC estimation compared to frequentist methods and examine the sensitivity of results to different choices and different parametrizations of the prior.

02/25/2009 — EEBoost: A general framework for high-dimensional variable selection based on estimating equations

Julian Wolfson

Driven by the availability of new, high-throughput technologies and increased computing capacity, variable selection has become a “hot” research area in recent years. Often, the number of predictors (p) is much larger than the number of available observations (n), and traditional methods either break down or cannot be applied. One of the most popular “modern” approaches to variable selection is based on minimizing a penalized loss function; for example, the well-known LASSO minimizes the sum of the target loss (eg. squared-error, negative log-likelihood) and the L_1 norm of the coefficients. While penalized methods have been extended to allow variable selection to be performed in various models (eg. GLMs, Cox proportional hazards), each extension involves a specialized algorithm for carrying out the penalized minimization. As a result, there are many problem setups where (asymptotically) unbiased, efficient estimation can be performed, usually by solving a set of estimating equations, but no methods for variable selection exist. We propose to close this “estimation-selection” gap by describing EEBoost, a simple and fast algorithm which can be used to do variable selection in any context where useful estimating equations have been developed. When the estimating equations correspond to the derivatives of a closed-form loss function, EEBoost performs very similarly to the LASSO. We illustrate the method by applying it to correlated longitudinal data and proportional hazards regression with missing covariates.

02/18/2009 — Adjustment for Constancy Assumption Violation in Non-Inferiority Clinical Trials

Katherine Odem-Davis

Non-inferiority (NI) trials may be employed when a new experimental treatment (EXP) is compared to a previously approved standard treatment (STD) and when there exists a clinically acceptable minimal loss of the effect of STD relative to placebo (PLA).  Standard statistical methods for assessment of NI either rely on the assumption that the estimate of the effect of standard is unbiased in the setting of the NI trial or they adjust for violation of this constancy assumption using a variance based penalty resulting in a more conservative test.  Possible contributors to violation of the constancy assumption are differences in populations, in standard of care, in measurement of outcome, or in adherence.  We propose methods which adjust directly for the non-constancy bias when there is no information regarding non-constancy, when there is information regarding differences in populations or standard of care, and when outcomes are measured differently in a survival setting due to different observation periods (e.g., 1-year survival versus 6-month survival).  We also consider comparison of estimates of STD effect relative to PLA in the NI trial setting with a similar estimate based on an independent historical PLA control as another source of evidence for or against constancy.  Simulations using industry data will be discussed.

02/11/2009 — Longitudinal Structural Mixed Models and Causal Inference in Surgical Trials with Noncompliance

Colleen Sitlani

Randomized surgical trials with the goal of evaluating the long-term benefit of surgical intervention as compared to a non-surgical treatment are often faced with serious patient noncompliance. There are several statistical challenges associated with longitudinal ‘as-treated’ analyses that seek to estimate average causal effects attributable to surgery. We adopt an underlying longitudinal structural mixed model that is a natural example of a structural nested distribution model, and then compare the performance of analysis methods when endogenous processes lead to patient crossover. Standard linear mixed models may not be valid yet can perform surprisingly well. In contrast, causal estimation methods such as marginal structural models, g-estimation and instrumental variable approaches can be valid and their implementation in this setting will be reviewed.