2016-Biostatistics: Causal Inference and Measurement Error


Biostatistics: Causal Inference and Measurement Error 
Chair: Mireille Schnitzer (Université de Montréal) 
[PDF]

ASMA BAHAMYIROU, Université de Montréal
Comparison of Causal Inference Approaches for Estimating the Effect of Exposure to Inhaled Corticosteroids on Mean Birth Weight and Gestational Age  [PDF]
 
Regression and propensity score methods to deal with confounder adjustment for the estimation of a treatment effect can produce biased estimates. Semiparametric methods such as Targeted Minimum Loss-Based Estimation (TMLE) can be used to reduce dependence on parametric model specification. In a large covariate space, Collaborative TMLE (C-TMLE) can further improve upon mean squared error. This work aims to apply adaptive methods to estimate the causal effect of exposure to inhaled corticosteroids on mean birth weight and gestational age in pregnant women with mild asthma. We compare Inverse Probability of Treatment Weighting, TMLE and C-TMLE, all implemented with Super Learner 
 
LEILA GOLPARVAR, McGill University
Causal Structure Learning and Propensity Score Adjustment  [PDF]
 
Mathematical representation of causal dependencies among a set of variables via graphs has been in the statistical literature for more than a century. Predicting the effect of manipulations from non-experimental data involves two steps: first, discovery of causal structures, second identification and estimation of causal parameters. We adopted a frequentist constraint-based approach to discover the causal graph, and use the PC algorithm for causal discovery. To explore the use of the PC algorithm in selection of confounders, we conduct a Monte Carlo simulation study. Results show that PC algorithm works very well when the sample size is moderate to large. 
 
KAMAL RAI, University of Waterloo
Bayesian Inference for Stochastic PK/PD Models [PDF]
 
Differential equations (DEs) occupy a central role in the modeling of pharmacokinetic/pharmacodynamic (PK/PD) processes. To estimate the parameters of these equations, a statistical approach might augment the deterministic DE with a stochastic simulation model and attempt to solve the corresponding inverse problem. We present a Bayesian methodology and its software implementation for PK/PD parameter inference, accounting for three sources of stochastic variability: (1) measurement error, (2) within-subject process fluctuations, and (3) between-subject random effects. Through numerical experiments with a small-sample PK study, we explore the effect of increased model complexity on the bias-variance trade-off. 
 
DI SHU, University of Waterloo
Estimation of Causal Parameters in the Presence of Measurement Error  [PDF]
 
Odds ratio, risk ratio and risk difference are widely used measures for comparing the performance of two treatments. In observational studies where confounders are inevitable, obtaining these measures with causal interpretation is of increasing interest. To reduce the confounding bias, many methods have been proposed, assuming that there is no measurement error. However, measurement error unavoidably occurs for various reasons. It is well known that ignoring measurement error effects can result in severely biased parameter estimates. In this talk, I'll discuss estimating causal parameters with measurement error. Numerical studies will be presented to assess the performance of proposed methods. 
 
DENIS TALBOT, Université Laval
A Graphical Perspective of Marginal Structural Models when Estimating the Causal Relationships Between Physical Activity, Blood Pressure, and Mortality  [PDF]
 
Estimating causal effects requires important prior subject-matter knowledge and, sometimes, sophisticated statistical tools. Marginal structural models (MSMs) are a relatively new class of causal models that effectively deals with the estimation of the effects of time-varying exposures. MSMs have traditionally been embedded in the counterfactual framework to causal inference. In this talk, we use the causal graph framework to enhance the implementation of MSMs. We illustrate our approach using data from a prospective cohort study, the Honolulu Heart Program. 
 
YEYING ZHU, University of Waterloo
A Kernel-Based Approach to Covariate Adjustment for Causal Inference  [PDF]
 
In matching, an important goal is to achieve balance in the covariates among the treatment groups. In this article, we introduce the concept of distributionally balance preserving which requires the distribution of the covariates to be the same in different treatment groups. Meanwhile, we propose a new balance measure called kernel distance, which is the empirical estimate of the probability metric defined in the reproducing kernel Hilbert spaces. Compared to the traditional balance metrics, the kernel distance measures the difference in the two multivariate distributions instead of the difference in the finite moments of the distributions.