Over the last two weeks I have been asked many times by people what they should do if the 'best performing' model as selected by AIC/VIF analysis/Forward selection doesn't include the effects that represent their hypothesis. My hot take has been to say model selection is almost always a bad idea