| Background. Here we introduce the generalized geoadditive mixed model (GGAMM), a combination of generalized additive model and linear mixed model with unified model structure for more flexible applications, to alternatively examine the influence of air pollution to human health.;Methods. Extant air pollution and mortality data came from the National Morbidity, Mortality, and Air Pollution Study for 15 U.S. cities in 1991-1995. The PM10 main model, distributed lag model and four co-pollutant models used the GGAMM approach to analyze the effect of PM10, lag effects and co-pollutants on several mortalities, adjusting for day-of-week, calendar time and temperature.;Objectives. First, the effects of PM10 on mortality are preliminarily examined; second, a jackknife-bootstrap method and a principal component analysis are proposed to handle potential convergence problems; third, some missing data imputation methods are evaluated in the GGAMM; fourth, the issues of multicollinearity and concurvity in our models are examined; fifth, comparisons of the GGAMM and 2-stage Bayesian hierarchical model are performed; sixth, three simulations are accomplished for investigating the influence of concurvity, multicollinearity and missing data imputation methods on estimates and smoothing functions.;Results. First, the effects of PM10 on mortality are preliminarily examined; second, a jackknife-bootstrap method and a principal component analysis are proposed to handle potential convergence problems; third, some missing data imputation methods are evaluated in the GGAMM; fourth, the issues of multicollinearity and concurvity in our models are examined; fifth, comparisons of the GGAMM and 2-stage Bayesian hierarchical model are performed; sixth, three simulations are accomplished for investigating the influence of concurvity, multicollinearity and missing data imputation methods on estimates and smoothing functions.;Conclusions. The GGAMM provides an integrate model structure to concern national average estimates, city-specific estimates, smoothing and spatial functions simultaneously. Geographical data can immediately be used in the GGAMM without being affected by missing data, and nation-level smoothing functions can be fitted well by enough valid observations from all cities. These properties are not offered by 2-stage Bayesian hierarchical models, and recommended by using spatio-temporal data. |
