To examine the effect of food intakes on the occurrence of a specific disease, it is necessary to take account of numerous measurement errors in dietary assessment instruments, such as the 24-hour recall or the food frequency questionnaire. The regression calibration (RC) method has been widely used for correcting the measurement error. However, the resulting corrected estimator is generally more variable than the naive biased one. Using the Bayesian hierarchical regression models, one can obtain more precise estimates than using ordinary regression models by incorporating additional information into a second-stage regression. In this paper, we propose a hierarchical Poisson regression model, in which multivariate measurement errors are adjusted by RC method. Simulation studies were conducted to investigate the performances of the proposed method, which showed that the proposed estimators were nearly unbiased, and were more precise than the usual RC ones even in the case of a few number of exposure. We also applied the proposed method to the analysis of a large prospective study, JDCS (Japan Diabetes Complications Study), to examine the effect of food group intakes on the occurrence of the cardiovascular disease (CVD) among type2 diabetic patients.
Epidemiologic findings by conventional statistical methods reflect uncertainty due to random error but omit uncertainty due to biases, such as unmeasured confounding, selection bias, and misclassification error. One approach for addressing this problem is to perform sensitivity analyses. We used MCSA (Monte Carlo sensitivity analysis) to analyze data from a large population-based cohort study, Japan Arteriosclerosis Longitudinal Study-Existing Cohorts Combine. The effects of the blood pressure on arteriosclerotic disease were examined among 21,949 subjects accounting for both misclassification of exposure and unmeasured confounding. We used a Poisson regression model to estimate the gender-specific incidence rate ratio (IRR) of each blood pressure category adjusted for several measured risk factors. The prior information on the misclassified blood pressure and the unmeasured diabetes mellitus history was obtained from sub-cohort members. Sequential correction of two biases by the MCSA led to large decrease of IRR among pre-hypertensive men (IRR = 1.79 [95% limits = 0.22−3.78]) and women (1.15 [0.28−2.25]), and large increase of IRR among stage 2 hypertensive men (7.24 [3.50−11.2]) and women (4.12 [2.14−6.89]). Our expanded MCSA provides valuable approach for bias analysis, which makes explicit and quantifies sources of uncertainty.
Differences in some traits between males and females, called sexual dimorphism, are observed among wild and livestock animals. For traits in which variances may be heterogeneous between sexes in some cases, evaluating the relevant genetic parameters, including genetic correlation between sexes, is an important topic requiring estimation of the components of (co)variances. This study developed a Bayesian approach via the Gibbs sampler to estimate the (co)variance components and genetic parameters of sexual dimorphism. As prior distributions, uniform, multivariate normal, two dimensional scaled inverted Wishart and independent scaled inverted chi-square distributions were used for the macro-environmental effects, breeding values, additive genetic (co)variances and residual variances, respectively. This approach was applied to beef carcass trait data, and the estimates of the (co)variance components and genetic parameters (especially the modes of the marginal posterior densities) were generally in agreement with those obtained using the restricted maximum likelihood procedure.
Noncompliance is an important problem in randomized trials. The estimation and bounds of average causal effects (ACEs) have been discussed as a way to address this issue. Previous studies have considered ACEs under the instrumental variable (IV) assumption, which postulates that potential outcomes are constant across subject sub-populations assigned to separate treatment regimens. However, the IV assumption may not be valid in unmasked trials. In the present analyses, the IV assumption is relaxed to the monotone IV (MIV) assumption, which replaces equality in the IV assumption with inequality. We propose bounds on ACEs under the MIV assumption in addition to the other existing assumptions. The results demonstrate that the intention-to-treat effect is an upper or lower bound under one assumption and the per-protocol effect is an upper or lower bound under the other assumption, even using the MIV assumption in place of the IV assumption. These proposed bounds are illustrated using a classic randomized trial.