The deconfounder uses the multiple causes as a signal for unobserved confounders, combining unsupervised machine learning and predictive model checking to perform causal inference.
Further adjustment for unobserved confounders resulted in an estimated difference of 3.7 and 4.2% (HMO lower), depending on the specification of geographic-fixed effects.
The last concern, the potential for unobserved confounders, is always present in observational research.
Whereas propensity score weighting can effectively adjust for observed differences, it can adjust for unobserved confounders only if they are associated with observed confounders.
31 The values thus obtained were compared with estimates of odds ratios for unobserved confounders, based on another study.
This enabled us to control for potentially unobserved confounders that vary between local authorities, as well as controlling for observed differences in economic trends.
Our study used longitudinal panel data rather than cross-sectional data, which allows us to control for unobserved time-constant confounders, therefore improving the robustness of results compared with previous research.
Another significant finding was the importance of controlling for unobserved environmental confounders confirming a correlation between those factors and the peer measures.
Although some studies examined trajectories over time in either exposure measures or outcome measures, most longitudinal studies have not investigated the impact of changes in both social and economic exposure measures and changes in drinking behaviors within young adulthood using methods that control for time-invariant unobserved confounders [ 14– 16].
On the individual level, however, we found no indication of a causal effect between maternal heavy alcohol use and toddler behavior problems, as the increased levels of toddler behavior problems found in the population were completely explained by controlling for observed and unobserved confounders in a fixed effect model.
