Statistical methods in medical research
-
Stat Methods Med Res · Feb 2012
ReviewA review of causal estimation of effects in mediation analyses.
We describe causal mediation methods for analysing the mechanistic factors through which interventions act on outcomes. A number of different mediation approaches have been presented in the biomedical, social science and statistical literature with an emphasis on different aspects of mediation. We review the different sets of assumptions that allow identification and estimation of effects in the simple case of a single intervention, a temporally subsequent mediator and outcome. ⋯ The understanding of such assumptions is crucial since some can be assessed under certain conditions (e.g. treatment-mediator interactions), whereas others cannot (sequential ignorability). These issues become more complex with multiple mediators and longitudinal outcomes. In addressing these assumptions, we review several causal approaches to mediation analyses.
-
Stat Methods Med Res · Aug 2011
A behavioural Bayes approach to the determination of sample size for clinical trials considering efficacy and safety: imbalanced sample size in treatment groups.
The behavioural Bayes approach to sample size determination for clinical trials assumes that the number of subsequent patients switching to a new drug from the current drug depends on the strength of the evidence for efficacy and safety that was observed in the clinical trials. The optimal sample size is the one which maximises the expected net benefit of the trial. The approach has been developed in a series of papers by Pezeshk and the present authors (Gittins JC, Pezeshk H. ⋯ A Monte Carlo simulation is employed for the calculation. Having a larger group of patients on the new drug in general makes it easier to recruit patients to the trial and may also be ethically desirable. Our results show that this can be done with very little if any reduction in expected net benefit.
-
We wish to deal with investigator bias in a statistical context. We sketch how a textbook solution to the problem of "outliers" which avoids one sort of investigator bias, creates the temptation for another sort. ⋯ Finally, we offer tentative suggestions to deal with the problem of investigator bias which follow from our account. As we have given a very sparse and stylized account of investigator bias, we ask what might be done to overcome this limitation.
-
Stat Methods Med Res · Jun 2008
ReviewRandomized trials for the real world: making as few and as reasonable assumptions as possible.
The strength of the randomized trial to yield conclusions not dependent on assumptions applies only in an ideal setting. In the real world various complications such as loss-to-follow-up, missing outcomes, noncompliance and nonrandom selection into a trial force a reliance on assumptions. To handle real world complications, it is desirable to make as few and as reasonable assumptions as possible. This article reviews four techniques for using a few reasonable assumptions to design or analyse randomized trials in the presence of specific real world complications: 1) a double sampling design for survival data to avoid strong assumptions about informative censoring, 2) sensitivity analysis for partially missing binary outcomes that uses the randomization to reduce the number of parameters specified by the investigator, 3) an estimate of the effect of treatment received in the presence of all-or-none compliance that requires reasonable assumptions, and 4) statistics for binary outcomes that avoid some assumptions for generalizing results to a target population.
-
Stat Methods Med Res · Jun 2007
Multiple imputation of discrete and continuous data by fully conditional specification.
The goal of multiple imputation is to provide valid inferences for statistical estimates from incomplete data. To achieve that goal, imputed values should preserve the structure in the data, as well as the uncertainty about this structure, and include any knowledge about the process that generated the missing data. Two approaches for imputing multivariate data exist: joint modeling (JM) and fully conditional specification (FCS). ⋯ Imputations for these data were created under two models: a multivariate normal model with rounding and a conditionally specified discrete model. The JM approach introduced biases in the reference curves, whereas FCS did not. The paper concludes that FCS is a useful and easily applied flexible alternative to JM when no convenient and realistic joint distribution can be specified.