Timeline for Categorical variable to be predicted from continuous variables with an idea: "maximise boxplots distance"
Current License: CC BY-SA 4.0
13 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Apr 28, 2022 at 7:50 | vote | accept | FluidMechanics Potential Flows | ||
| Apr 27, 2022 at 18:48 | comment | added | Frank Harrell | I was referring to learning about the associations between the many continuous variables and the one categorical variable. | |
| Apr 27, 2022 at 11:37 | comment | added | FluidMechanics Potential Flows | @FrankHarrell I'm not sure I understand what you mean, what are the "features to learn about"? If you can't fit the explanation here, do you have "basic" references / videos to point to? :) | |
| Apr 27, 2022 at 11:29 | comment | added | Frank Harrell | The problem was set up for multinomial logistic regression. Everything else is just a detour. Note that for binary Y the minimum sample size needed if there are zero features to learn about is $n=96$. That is the sample size that yields a margin of error of no more than 0.1 in estimating a probability. If there were binary Y and one binary X the needed sample size is 2 $\times$ 96. For 3-category Y the sample size needs are larger (at least when Y is unordered). | |
| Apr 27, 2022 at 11:07 | history | edited | BlackBear | CC BY-SA 4.0 |
added 17 characters in body
|
| Apr 27, 2022 at 11:07 | comment | added | BlackBear |
@FluidMechanicsPotentialFlows you're welcome! You could also model cat~y but you would need to use either a Binomial GLM (aka binary classification) or a Softmax regression model (aka multiclass classification). But given your description of the problem I think y~cat is more fitting.
|
|
| Apr 27, 2022 at 9:35 | comment | added | FluidMechanics Potential Flows |
Also, why did you not do lm(cat ~ y), is it because the predictor can be a categorical variable but not the predicted variable with lm?
|
|
| Apr 27, 2022 at 9:23 | comment | added | FluidMechanics Potential Flows | thank you so much for your time | |
| Apr 27, 2022 at 8:01 | comment | added | BlackBear |
@FluidMechanicsPotentialFlows yes, the "overall hypothesis" is "no continuous variable is related to the categorical one", and you need to correct because you are conducting a separate test for each variable. The F-test compares the variance of the response variable y with the variance of the residuals of the model y~cat. If the latter is smaller it means that cat has predictive power (ie, it reduces variance)
|
|
| Apr 26, 2022 at 21:45 | comment | added | FluidMechanics Potential Flows | If my boxplot are offset (not the same mean/median) but with the same variance, I would want the test to say that the continuous variable is related to the categorical one. Because you only change the mean of y in your example I guess this isn't the variance we're comparing so I'm a bit confused about this (admitedly I probably should understand it fully because it's the fundamentals but it looks like I haven't mastered them) | |
| Apr 26, 2022 at 21:45 | comment | added | FluidMechanics Potential Flows | Also, I'm not really familiar with the F-test (why does it appear when you're fitting a linear model) - after some research it seems it tests if the variances are the same but what variances are we talking about ? | |
| Apr 26, 2022 at 21:32 | comment | added | FluidMechanics Potential Flows | Thanks for the answer and the examples. I know a bit about adjusted p-value but is it applicable here? It says here (analytics-toolkit.com/glossary/p-value-adjustment) "A p-value adjustment is necessary when one performs multiple comparisons or multiple testing in a more general sense: performing multiple tests of significance where only one significant result will lead to the rejection of an overall hypothesis.". Is our "overall hypothesis" here "no continuous variable is related to the categorical one"? | |
| Apr 26, 2022 at 16:34 | history | answered | BlackBear | CC BY-SA 4.0 |