@article {Walker133785, author = {Jeffrey A. Walker}, title = {A Defense of Model Averaging}, elocation-id = {133785}, year = {2017}, doi = {10.1101/133785}, publisher = {Cold Spring Harbor Laboratory}, abstract = {Model averaging of partial regression coefficients has been criticized for averaging over a set of models with coefficients that are either incommensurable or describe fundamentally different things if there is any correlation (multicollinearity) among the predictors. It is easy to show that partial regression coefficients conditional on different covariates are commensurable. And, partial regression coefficients from different models can have the same meaning if they are used to estimate the effects in a causal model, which derive their meaning from the specified paths and not from the set of covariates. A multiple regression model implicitly specifies a causal model with direct, causal paths from each predictor to the response. Consequently, the partial regression coefficient for a predictor has the same meaning across all sub-models if the goal is estimation of the causal effects that generated the response. In order to clarify the effects of multicollinearity on model-averaged estimates, I compare effect estimates using a small Monte-Carlo simulation. The simulation results show that model-averaged and ridge estimates have increasingly better performance, relative to model selection and full model estimates, as multicollinearity increases.}, URL = {https://www.biorxiv.org/content/early/2017/05/03/133785}, eprint = {https://www.biorxiv.org/content/early/2017/05/03/133785.full.pdf}, journal = {bioRxiv} }