Regression for prediction versus understanding independent associations I see that regression is used for 2 main purposes:

*

*To develop a predictive model for future use


*To understand associations. Specifically, to know which of the predictors have an "independent association" with the outcome or response variable. This is used especially in biomedical field.
How should approach and application of regression be different in above 2 purposes? Thanks for your insight.
 A: It's my limited experience (2nd year undergrad) that the main difference in approach depending on purpose is in the variable selection.
If you're trying to make a predictive model only for the purpose of prediction, it makes sense to include any predictors which correlate significantly with the response (within some reason, it takes a long time to gather data). However, when you're looking for the association of a specific predictor with the response, you want to be pretty careful not to include too many things which correlate with that predictor. This can lead to issues with multicollinearity (when two or more predictors correlate with the response exactly in the same way as another)
For example, let's say the relationship I care about is between how much an indoor plant grows and how much it was watered. I might reasonably make a model like: growth ~ water + sun. It would still make sense to include the variable about how much sun the plant got, as that information is probably not highly related to how much water it got. There would be examples in the data of plants that got a lot of water and sun, a lot of water but no sun, a lot of sun but no water, and not much sun or water. So I (or a regression package) could see how much growth increased due to the water, and how much due to the sun.
However, it would not make sense to include a variable measuring the overall skill of the plant's keeper, since that variable is probably highly correlated with how much water the plant got. In a model with water and grower_skill, each relationship would be obscured somewhat by the other. It would be hard to tell how much taller the plant got due to the water, and how much it got taller due to the grower's skill, since all good growers water their plants and all bad growers don't. There would be no examples in the data of plants with high grower skill and low water or vice versa.
So, if I was making a predictive model, my predictions would probably be slightly better if I included grower_skill (maybe at the high levels they use fertilizer or something). However, if I included it, the association of water and growth would be less clear.
If you're unfamiliar, here's a site with some more on multicollinearity:
https://www.google.com/amp/s/blog.minitab.com/blog/understanding-statistics/handling-multicollinearity-in-regression-analysis%3fhs_amp=true
