Timeline for Addressing multicollinearity when removing or imputing is not an option
Current License: CC BY-SA 4.0
6 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Jan 11, 2023 at 0:46 | comment | added | Graham Wright | ...sometimes missing data means "zero" (if people are missing values for "how many years of college did you attend" because earlier in the survey they said didn't go to college), and sometimes missing data means "not applicable" ("what color is you cat?" if you don't have a cat") and sometimes it means "value unknown" (if you ask someone their income and they just refuse to answer). Each of those kinds of data needs to be dealt with in a different way. You just need to decide whether to treat "missing twitter likes because wasn't promoted on twitter" as "zero" or "not applicable." | |
| Jan 11, 2023 at 0:44 | comment | added | Graham Wright | A high VIF (like higher than 5 or 10) is a measure of multicolinearity yes. There isn't any simple relationship between R2 and multicolineairty one way or the other. But I really don't think multicolinearity is what you should be worried about in this problem. Your problem is missing data, and you have to decide conceptually what that missing data means.... | |
| Jan 10, 2023 at 20:13 | comment | added | Tapio | Thank you again. Would the correct interpretation be that multicollinearity == high VIF? That variance of one variable is excessively explained by another one? The r2 for a model only (tw_likes ~ fb_likes) including those variables is 0.467, which then simply means that there is no multicollinearity issue, end of story. About the rest of the comment, I agree with the sentiment. My goal is to predict and I struggle with making a model which works with "missing data". I guess this is also something I should reconsider? | |
| Jan 10, 2023 at 19:50 | comment | added | Graham Wright | I think you are confusing multicolinearity with just "correlation." We expect that independent variables in a regression will be correlated with one another to some extent - that's why we need regression in the first place. Multicolinearity is just what happens when that correlation becomes SO high that it becomes meaningless to try and talk about the effect of one variable "controlling" for the other, because both are functionally measuring the same thing. That's clearly not the case here, either according to the VIF or conceptually. | |
| Jan 10, 2023 at 16:47 | comment | added | Tapio | About the first point, I've understood that regardless of low VIF, there can be multicollinearity. "The lower precision, switched signs, and a lack of statistical significance are typical problems associated with multicollinearity.". In the OLS model, if both variables are included Twitter follower coefficient turns negative and becomes insignificant. Can I really trust that there is no multicollinearity? | |
| Jan 10, 2023 at 13:56 | history | answered | Graham Wright | CC BY-SA 4.0 |