We propose a bivariate model–based ensemble (BMBE) method to borrow information from the mortality data of a given pool of auxiliary populations to enhance the mortality forecasting of a target population. The BMBE method establishes a cascade of bivariate mortality models between the target population and each auxiliary population as the base learners. Then it aggregates prediction results from all of the base learners by means of an averaging strategy. Augmented common factor–type and CBD-type bivariate models are applied as the base learners as illustrative examples in the empirical studies with the Human Mortality Database. Empirical results presented in this article confirm the effectiveness of the proposed BMBE method in enhancing mortality prediction. For completeness, we also conduct a synthetic study to illustrate a particular setting for the superior performance of the BMBE method.
Health research often aims to investigate whether the effect of an exposure variable is common across different subgroups of individuals, but sometimes the variable defining subgroups is not recorded in all individuals. We propose and evaluate two methods for estimation of the marginal causal effect of an exposure variable within subgroups in the observational setting where the subgroup variable is incompletely observed. The first approach involves doubly weighted estimating functions with one weight based on a propensity score for exposure and a second weight addressing the selection bias when analyses are restricted to individuals with complete data. The second approach uses the inverse probability of exposure weights in conjunction with multiple imputation for the incomplete subgroup variable. The resulting estimators are consistent when the auxiliary models are correctly specified; we assess the finite sample performance via simulation. An illustrative analysis is provided involving patients with psoriatic arthritis treated with biologic therapy where interest lies in the effect of therapy according to the presence or absence of the human leukocyte antigen marker HLA-B27 which is incompletely observed.
Multistate models offer a natural framework for studying many chronic disease processes. Interest often lies in identifying which among a large list of candidate variables play a role in the progression of such processes. We consider the problem of variable selection for progressive multistate processes under intermittent observation based on penalized log‐likelihood. An Expectation‐Maximization (EM) algorithm is developed such that the maximization step can exploit existing software for penalized Poisson regression thereby allowing for the use of common penalty functions. Simulation studies show good performance in identifying important markers with different penalty functions. In a motivating application involving a cohort of patients with psoriatic arthritis, we identify which, among a large group of candidate HLA markers, are associated with rapid disease progression.