|
Abstract:
|
Distributed lag models (DLMs) find application when the influence of a time varying explanatory variable is spread over a finite set of time lags. DLMs are widely applicable, and are frequently used in time series studies of air pollution exposure and human health. A key technical challenge lies in flexibly estimating the shape of the lag-curve, which expresses the functional form of dependence between explanatory and outcome variables over successive time lags. Since lag-curves can be complex, it is common to avoid restrictive parametric forms and instead enforce smoothness penalties on the curve estimator. However, some existing approaches require the modeler to make important decisions that may be difficult to justify without strong prior knowledge, including the degree of lag-curve smoothness and the total number of lags to consider, which can lead to misspecification and biased estimates of the lag curve. In this talk, a novel class of DLM is proposed that is straightforward to implement and robust to such forms of user misspecification. The new DLM will be compared to some existing approaches, and is applied to data from hydrology and air pollution epidemiology.
|
