I would to thank you very much for all what you do, it is really a very high-quality teaching matter. I have a small question, but I’m not really sure about its pertinence in this case. As we performed several statistic tests on the same dataset, don’t we need to consider the effect of the multiplicity of tests on the “P values” in some way.
I have been following your videos for quite a time now, I like them and appreciate your work. For this one though: are we not supposed to check for the effect modification (interaction) first then in its absence check for confounding? if effect modification is present, then checking for confounding would not be necessary right?
Because here we are trying to estimate the effect of one variable (smoke) on the outcome (lie birth weight), and so we want to identify and remove any confounded, etc. automated selection procedures like forward/backward selection are intended for building predictive models, but would end up with a lot of bias in the estimated effect of smoking on low BW. I’ll also add that even for predictive models, I’m not a big fan of automated selection procedures, unless there is a very large number of potential predictor, in which case automation may be necessary. I feel that they rely too much on an algorithm to select variables, without user input/knowledge
Thank you for sharing knowledge. This is exactly a learner needs when fitting a model for the first time.
Tight! Keep them coming!
awesome video and explanation! Got my sub !
I would to thank you very much for all what you do, it is really a very high-quality teaching matter.
I have a small question, but I’m not really sure about its pertinence in this case. As we performed several statistic tests on the same dataset, don’t we need to consider the effect of the multiplicity of tests on the “P values” in some way.
I have been following your videos for quite a time now, I like them and appreciate your work. For this one though: are we not supposed to check for the effect modification (interaction) first then in its absence check for confounding? if effect modification is present, then checking for confounding would not be necessary right?
why not using backward or forward selection? and see the AIC/BIC?
Because here we are trying to estimate the effect of one variable (smoke) on the outcome (lie birth weight), and so we want to identify and remove any confounded, etc. automated selection procedures like forward/backward selection are intended for building predictive models, but would end up with a lot of bias in the estimated effect of smoking on low BW.
I’ll also add that even for predictive models, I’m not a big fan of automated selection procedures, unless there is a very large number of potential predictor, in which case automation may be necessary. I feel that they rely too much on an algorithm to select variables, without user input/knowledge