Department of Biostatistics
Statistical Methods in Epidemiology Working Group
2012 - 2013
Judith Lok, Ph.D.
Assistant Professor, Department of Biostatistics, Harvard School of Public Health
"Defining and Estimating Causal Direct and Indirect Effects: An Intervention Based Approach"
ABSTRACT: Natural direct and indirect effects decompose the effect of a treatment into the part that is mediated by a covariate (the mediator) and the part that is not. The definition of natural direct and indirect effects relies on cross-worlds quantities: the outcomes under treatment with the mediator "set" to its value without treatment. How to set the mediator is usually unspecified, which in practice often renders these cross-worlds quantities undefined. This talk introduces "organic" direct and indirect effects, which can be defined, identified, and estimated without relying on the potential outcomes under all combinations of treatment and mediator. For example, part of the effect of maternal smoking on infant mortality may be mediated by low birth weight. Similarly, only part of the effect of some treatments for HIV/AIDS on mother-to-child transmission of HIV-infection is mediated by the effect of the treatment on the HIV viral load in the blood of the mother. However, setting birth weight or setting HIV viral load to specific values is not possible in practice. The newly defined organic direct and indirect effects are therefore more appropriate in such settings than the well-known natural direct and indirect effects.
Polyna Khudyakov, Ph.D.
Research Fellow, Departments of Biostatistics and Epidemiology, Harvard School of Public Health
"The Impact of Covariate Measurement Error on Risk Prediction"
ABSTRACT: In risk prediction models, key factors are often measured with error, which can affect the quality of prediction. In this work we studied the impact of covariate measurement error on risk prediction based on logit regression models. In particular, we compared the performances of prediction based on a costly true and an error-free covariates (true model) with those from the model based on an inexpensive surrogate and the error-free covariates (surrogate model). The comparison is based on the area under the receiver operating characteristic curve (AUC of ROC), Brier score, and the ratio of the observed and expected number of events. By extensive simulation study we show that: (i) all the models are well calibrated; (ii) a mismeasured covariate can reduce the AUC and Brier score dramatically, in compare to those of the true model; (iii) all other error-free covariates should be included in the prediction model, but in many settings they still cannot compensate for the entire loss in the AUC or Brier score due to mismeasured covariate; (iv) adding, in the prediction model, instrumental variables, can improve the AUC and Brier score dramatically, but not the entire loss. Hence, we conclude that improving precision of measures is of practical importance for improving risk prediction.
This is joint work with Malka Gorfine (Technion - Israel Institute of Technology), David Zucker (Hebrew University of Jerusalem) and Donna Spiegelman (HSPH).
Molin Wang, Ph.D.
Assistant Professor, Department of Medicine, Harvard Medical School / Departments of Biostatistics and Epidemiology, Harvard School of Public Health
"Statistical Methods and SAS Macros for Molecular Pathological Epidemiology (MPE) Analysis"
ABSTRACT: Epidemiologic research typically investigates the associations between exposures and the risk of a disease, in which the disease of interest is treated as a single outcome. However, many human diseases, including colon cancer, type II diabetes mellitus and myocardial infarction, are comprised of a range of heterogeneous molecular and pathologic processes, likely reflecting the influences of diverse exposures. The approach, which incorporates data on the molecular and pathologic features of a disease directly into epidemiologic studies, Molecular Pathological Epidemiology, has been proposed to better identify causal factors and better understand how potential etiologic factors influence disease development. In this talk, I will present statistical methods for evaluating whether the effect of a potential risk factor varies by subtypes of the disease, in cohort studies, case-control studies and case-case study designs. Efficiency of the tests will also be discussed. SAS macros will be presented to implement these methods. The macros test overall heterogeneity through the common effect test (i.e., the null hypothesis is that all of the effects of exposure on the different subtypes are the same) as well as pair-wise differences in exposure effects. In adjusting for confounding, the effects are allowed to vary for the different subtypes or they can be assumed to be the same across the different subtypes. To illustrate the methods, we evaluate the effect of alcohol intake on LINE-1 methylation subtypes of colon cancer in the Health Professionals Follow-up Study, where 51,529 men have been followed since 1986 during which time 268 cases of colon cancer have occurred. Results are presented for all 3 possible study designs for comparison purposes.
This is a joint work with Aya Kuchiba and Donna Spiegelman.
|Back to HSPH Biostatistics||
Maintained by the
Last Update: April 9, 2013