Model selection bias and Freedman’s paradox

摘要: In situations where limited knowledge of a system exists and the ratio data points to variables is small, variable selection methods can often be misleading. Freedman (Am Stat 37:152–155, 1983) demonstrated how common it select completely unrelated as highly “significant” when number similar in magnitude variables. A new type model averaging estimator based on with Akaike’s AIC used linear regression investigate problems likely inclusion spurious effects bias, bias introduced while using single seemingly “best” from (often large) set models employing many predictor The helps reduce these provides confidence interval coverage at nominal level traditional stepwise has poor inferential properties.