Complex models need an explanation of exactly what data they use to answer a question. Black boxes are bad.

Clean, unconfounded variation in non-experimental data is rare. So we need to rely on well-founded models to do causal analysis. These models can be complex, and the tie between the data and model’s conclusion can get lost.

But all model results are ultimately some function of the data. Understanding what the function looks like, and whether it makes sense that these data objects inform the ultimate causal question of interest is critical to evaluating whether the model and analysis are reasonable.