To make the model more robust to the effects of less predictable seasonality, we first built a core model to control the confounders, including long-term trends, seasonal patterns of ARD admissions and meteorological factors, with natural cubic spline smoothing functions of time, weekly average temperature and relative humidity.
For each of the six PCB exposure time points (e.g., maternal, cord, 6-month infant, 16-month infant, 45-month child, and postnatal average), we first built a core model, before adding potential confounders, that included the corresponding natural log transformed PCB exposure (nanograms per milliliter), frequency, side, and the interaction between side and frequency.
"We have built a solid model.
Firstly, we built a model that included only a limited set of potential confounders (parsimonious model).
We built confounder models separately for the two analysis periods, 1991 2002 (gaseous pollutants and PM10) and 1995 2002 (UFP and PM2.5), without including any air pollutants.
"We actually built a model".
We also built a hidden Markov model (HMM) for comparison.
We built a neural network model using the training database.
We then used forward selection method to build a parsimonious model to adjust for other confounders.
By adjusting for confounders, we were able to build a predictive model in a training cohort that performed well in the validation cohort.
