Table 2.
Results of the Primary Linear Regression Analysis With IQ at Age 6 Years as the Dependent Variablea
| Effect | F Score | df | Coefficient (95% CI) | P Value |
|---|---|---|---|---|
| Breastfeeding status | 4.1 | 1 | 4.1 (0.1 to 8.1) | .045 |
| AED group across 4 drugs | 7.9 | 3 | Not applicable | <.001 |
| Maternal IQ | 6.8 | 1 | 0.2 (0.0 to 0.4) | .01 |
| Periconception folate use | 8.1 | 1 | 5.7 (1.7 to 9.7) | .005 |
| AED dosage | 6.6 | 1 | −0.1 (−0.2 to 0.0) | .01 |
| Propensity scoreb | 1.1 | 1 | 7.5 (−6.6 to 21.6) | .30 |
Abbreviation: AED, antiepileptic drug.
Linear regression models were used to examine breastfeeding effects adjusting for covariates. The goal was a parsimonious model in which all covariates were significant at the .05 level and a model that was not overfitted. To select covariates for inclusion in the model, we first relied on an approach that considered a priori hypotheses about clinical relevance. Breastfeeding status was included as the primary covariate of interest. Because specific AED, dosage, and maternal IQ were considered important covariates, we included these variables as predictors in the linear model, with child IQ as the outcome. Other covariates were added individually to the model and included if significant (P < .05) and not collinear with existing predictors. We inspected diagnostic plots to ensure that distributional assumptions of the models were met. An automated backward selection method confirmed our selection of covariates. Backward elimination started from the full model, including all possible covariates, which were deleted one by one based on a significance limit of .10. At each step, the covariate showing the smallest contribution was deleted based on the F score.