Associations based on alternative regression models with confounders selected using a forward strategy were similar [see Supplemental Material, Table S1 (http://dx.doi.org/10.1289/ehp.1204976)].
Table 4 shows separate final full multiple regression models with confounders, namely, district, HOME score, Raven's score, and sex of the babies, predicting MDI and PDI from the mono-ortho-substituted PCBs in maternal and cord sera.
Without the confounder terms we cannot plot the full model with confounders.
In multivariate analyses, we ran a series of models with potential confounders and maternal or paternal enrollment PBB, estimated PBB at conception or enrollment PCB exposure.
Furthermore, the models with additional confounders were fitted to account for the potential effect of marital status, children living at home, area of living, and education in the association between sector-defined occupational groups, psychosocial working conditions and DP due to MSD.
Both in the model with the confounders age, HDL cholesterol, and low-density lipoprotein (LDL) cholesterol (B = 0.042; P = 0.328) and in the model with the confounders age, HDL cholesterol, and history of macrovascular disease (B = 0.030; P = 0.474), no differences were found between patients and healthy controls.
Adding plasma NT-proSST concentrations to a model with potential confounders and well-known cardiovascular risk factors for mortality did not improve the Harrel's C statistic compared to the fully adjusted model without plasma NT-proSST concentrations, indicating a lack of benefit in risk prediction when adding NT-proSST.
This method involves estimating a logistic regression model with known/available confounders for each outcome, then using this model to compute the adjusted number of cases that would be expected if the risk factor (e.g., excess GWG) were absent in the population.
Multiple regression models with all potential confounders as previously described were constructed for SBP, DBP, or weight at the relevant time point as the dependent variable, and parsimonious models were obtained.
Models for FEV1/FVC were adjusted with confounders selected for models for both FEV1 and FVC.
We used models with NS for confounder control with 8 df per year for seasonality.
