摘要：In the literature, different shapes of associations have been found between body mass index (BMI) and mortality and some of the findings were opposite to each other. The association of BMI and mortality in a single cohort has been found to be dynamic that can lead to different findings under different settings. The identified dynamic features were consistent with the heterogeneity in the literature. It is meaningful to find out whether such dynamic associations exist in other populations. Data of six different cohorts were used for analysis and comparison. The proportional hazards assumptions for BMI in Cox models were tested to identify dynamic associations in each cohort. Time-dependent covariates Cox model was used to model the association of BMI and mortality risk as functions of follow-up time. The Cox model was applied to the pooled data with survival times censored at 5 to 40 years to show the potential impact of the dynamic association on traditional Meta-analysis. Dynamic associations were identified in six models (4 for men and 2 for women), four of which showed the same changing pattern: the elevated mortality risk for low BMI decreased while that for high BMI increased with follow-up time. When the Cox model was applied to the pooled data excluding the largest and also the shortest cohort, low BMI was but high BMI was not associated with high mortality for men with censoring at 5 years but the association for low BMI became weaker and that for high BMI became much stronger when censoring time was at 40 years. The dynamic association indicated that shorter studies tend to obtain inverse associations between BMI and mortality while longer studies tend to obtain J-shaped associations. Different or even opposite results about body weight and mortality in the literature may be in part due to the underlying dynamic association of BMI and mortality. The dynamic features need to be taken into consideration in future studies.
关键词：Cox model ; Non-proportional hazards ; Time-dependent covariates Cox model ; Length of follow-up
