Overdispersion models for correlated multinomial data: Applications to blinding assessment.

Overdispersion models have been extensively studied for correlated normal and binomial data but much less so for correlated multinomial data. In this work, we describe a multinomial overdispersion model that leads to the specification of the first two moments of the outcome and allows the estimation of the global parameters using generalized estimating equations (GEE). We introduce a Global Blinding Index as a target parameter and illustrate the application of the GEE method to its estimation from (1) a clinical trial with clustering by practitioner and (2) a meta-analysis on psychiatric disorders. We examine the impact of a small number of clusters, high variability in cluster sizes, and the magnitude of the intraclass correlation on the performance of the GEE estimators of the Global Blinding Index using the data simulated from different models. We compare these estimators with the inverse-variance weighted estimators and a maximum-likelihood estimator, derived under the Dirichlet-multinomial model. Our results indicate that the performance of the GEE estimators was satisfactory even in situations with a small number of clusters, whereas the inverse-variance weighted estimators performed poorly, especially for larger values of the intraclass correlation coefficient. Our findings and illustrations may be instrumental for practitioners who analyze clustered multinomial data from clinical trials and/or meta-analysis.

Estimating survival probabilities by exposure levels: utilizing vital statistics and complex survey data with mortality follow-up.

We present a two-step approach for estimating hazard rates and, consequently, survival probabilities, by levels of general categorical exposure. The resulting estimator utilizes three sources of data: vital statistics data and census data are used at the first step to estimate the overall hazard rate for a given combination of gender and age group, and cohort data constructed from a nationally representative complex survey with linked mortality records, are used at the second step to divide the overall hazard rate by exposure levels. We present an explicit expression for the resulting estimator and consider two methods for variance estimation that account for complex multistage sample design: (1) the leaving-one-out jackknife method, and (2) the Taylor linearization method, which provides an analytic formula for the variance estimator. The methods are illustrated with smoking and all-cause mortality data from the US National Health Interview Survey Linked Mortality Files, and the proposed estimator is compared with a previously studied crude hazard rate estimator that uses survey data only. The advantages of a two-step approach and possible extensions of the proposed estimator are discussed.

Efficient analysis of case-control studies with sample weights.

Analysis of population-based case-control studies with complex sampling designs is challenging because the sample selection probabilities (and, therefore, the sample weights) depend on the response variable and covariates. Commonly, the design-consistent (weighted) estimators of the parameters of the population regression model are obtained by solving (sample) weighted estimating equations. Weighted estimators, however, are known to be inefficient when the weights are highly variable as is typical for case-control designs. In this paper, we propose two alternative estimators that have higher efficiency and smaller finite sample bias compared with the weighted estimator. Both methods incorporate the information included in the sample weights by modeling the sample expectation of the weights conditional on design variables. We discuss benefits and limitations of each of the two proposed estimators emphasizing efficiency and robustness. We compare the finite sample properties of the two new estimators and traditionally used weighted estimators with the use of simulated data under various sampling scenarios. We apply the methods to the U.S. Kidney Cancer Case-Control Study to identify risk factors. Published 2012. This article is a US Government work and is in the public domain in the USA.

On estimating average effects for multiple treatment groups.

We propose to estimate average exposure (or treatment) effects from observational data for multiple exposure groups by fitting an approximation of the marginal sample distribution of the response variable in each exposure group to the data. The marginal sample distribution is a function of the true distribution of the response variable in the population and the assignment rule governing the allocation of the subjects to different exposure groups. The assignment rule can depend on the response variable in addition to measured covariates, and hence the method is appropriate even when the assumption of ignorable treatment assignment is not justified. We estimate the exposure effects are estimated based on the population expectation (PE) of the outcome variable. We compare the PE approach with an instrumental variable method and with several other methods including propensity score based approaches that assume ignorable assignment mechanisms. We evaluate the robustness of the PE method under model misspecifications and illustrate it using data from a study of the impact of soy consumption on urinary concentrations of estrogen and estrogen metabolites in Asian American women. Published 2012. This article is a US Government work and is in the public domain in the USA.