Childhood obesity is associated with increased morbidity and mortality in
adulthood, leading to substantial healthcare cost. There is an urgent need to
promote early prevention and develop an accompanying surveillance system. In
this paper, we make use of electronic health records (EHRs) and construct a
penalized multi-level generalized linear model. The model provides regular
trend and outlier information simultaneously, both of which may be useful to
raise public awareness and facilitate targeted intervention. Our strategy is to
decompose the regional contribution in the model into smooth and sparse
signals, where the characteristics of the signals are encouraged by the
combination of fusion and sparse penalties imposed on the likelihood function.
In addition, we introduce a weighting scheme to account for the missingness and
potential non-representativeness arising from the EHRs data. We propose a novel
alternating minimization algorithm, which is computationally efficient, easy to
implement, and guarantees convergence. Simulation shows that the proposed
method has a superior performance compared with traditional counterparts.
Finally, we apply our method to the University of Wisconsin Population Health
Information Exchange database.