Statistical inference in complex sampling designs
Presenter: Tracey Marsh
When: Monday, July 11, 2016 Time: 2:00 PM - 3:30 PM
Room: Oak Bay 1-2 (Level 1)
Session Synopsis:
Efficient inference for an additive gene-treatment interaction from a nested two-phase study
In the evaluation of effect-modifying genetic variants, it is well known that the precision of interaction estimates can be considerably improved by exploiting so-called gene-environment independence. For an interaction study nested within a clinical trial, gene-treatment independence is guaranteed by randomization. In practice, the relatively expensive genetic information is obtained in a second phase of data collection, often for a case-control subsample. A focus on estimation from case-control samples has produced numerous methods for estimating approximate multiplicative and additive scale interactions in terms of parameters in a logistic regression model. In this context, adjusting for potential confounders introduces a risk of model misspecification and the interpretation of estimands becomes conditional. Direct estimation of additive measures of interactions, in particular by means that adjust for covariates and provide a prospective, population-level interpretation, are particularly relevant to public health and policy decision-making yet have received relatively little attention. In this talk we will propose such a target parameter and present a nonparametric approach to its estimation when the genetic variant is partially missing by design. We will also describe a natural modification to the estimator in order to utilize known independence for efficiency gain. This methodology will build upon concepts of semiparametric efficiency theory and targeted learning.
Statistical inference in complex sampling designs
Presenter: Peter Gilbert
When: Monday, July 11, 2016 Time: 2:00 PM - 3:30 PM
Room: Oak Bay 1-2 (Level 1)
Session Synopsis:
Estimation of Stratified Mark-Specific Proportional Hazards Models for Competing Risks Outcomes under Two-Phase Sampling with Application to HIV Vaccine Efficacy Trials
In preventive HIV vaccine efficacy trials, HIV uninfected volunteers at high risk for acquiring HIV infection are randomized to receive vaccine or placebo, and are followed for the primary endpoint of HIV infection. Because HIV is genetically diverse, trial participants are exposed to a variety of genetic strains of HIV, and it is of interest to assess the mark-specific hazard of HIV infection in vaccine recipients, where the mark is the genetic distance of an infecting HIV sequence to an HIV sequence represented inside the vaccine. Most HIV infected participants have unique mark values, such that the mark is treated as a continuous random variable rather than as a discrete competing risk. It is also of interest to assess whether and how this mark-specific hazard of HIV infection varies over covariate subgroups defined by a biomarker measuring an immune response to the vaccine. The stratified continuous mark-specific proportional hazards model has been previously studied to evaluate mark-specific relative risks of HIV infection where the baseline hazards may vary with strata and the mark-specific relative risks are nonparametrically specified (e.g., Sun and Gilbert, 2012, Scand J Stat; Gilbert and Sun, 2015, JRSS-C). This talk extends this model to include covariates collected under a two-phase sampling design, where the phase-two covariates of interest are immune response biomarkers. New estimation procedures are developed based on augmented inverse probability weighting and smoothing in the mark. The asymptotic properties of the proposed estimators are derived, and their finite-sample performances examined in simulations. The methods are applied to the RV144 HIV vaccine efficacy trial to assess whether anti-V2 loop antibody responses correlate more strongly with HIV infection with strains close to the vaccine in their V2 sequence than with HIV infection with strains farther from the vaccine in their V2 sequences. This method improves the search for immunological correlates of vaccine efficacy by integrating antigen-specific response correlates with antigen-specific failure time outcomes.
Statistical inference in complex sampling designs
Presenter: Thomas Lumley
When: Monday, July 11, 2016 Time: 2:00 PM - 3:30 PM
Room: Oak Bay 1-2 (Level 1)
Session Synopsis:
Multiple Imputation and/or calibration in twophase designs
Multiple Imputation was originally proposed for small amounts of data missing due to nonresponse. Recently there is increasing interest in MI as a highly efficient analysis approach for large amounts of data missing by design. The traditional sampling approach to using population/cohort information is calibration of weights. I will consider the combination of MI and calibration of weights, and discuss what this illuminates about gains and tradeoffs in the two approaches.
Statistical inference in complex sampling designs
Presenter: Edward Kennedy
When: Monday, July 11, 2016 Time: 2:00 PM - 3:30 PM
Room: Oak Bay 1-2 (Level 1)
Session Synopsis:
Semiparametric causal inference in matched cohort studies
Odds ratios can be estimated in case-control studies using standard logistic regression, ignoring the outcome-dependent sampling. In this paper we discuss an analogous result for treatment effects on the treated in matched cohort studies. Specifically, in studies where a sample of treated subjects is observed along with a separate sample of possibly matched controls, we show that efficient and doubly robust estimators of effects on the treated are computationally equivalent to standard estimators, which ignore the matching and exposure-based sampling. This is not the case for general average effects. We also show that matched cohort studies are often more efficient than random sampling for estimating effects on the treated, and derive the optimal number of matches for a given set of matching variables. We illustrate our results via simulation and in a matched cohort study of the effect of hysterectomy on the risk of cardiovascular disease.
