I have a large and broad dataset with many possible covariates that are both categorical and continuous. I planned to run a correlation matrix with the DV and all possible IVs to determine the most fitting covariates. I was thinking to run a Spearman correlation. However, I was unsuccessful in doing so in R, given that the categorical variables are treated as factors and not numeric. All categorical variables are dummy-coded, but I do not see how it is useful to run a correlation when treating the dummy-coded as numeric. Would anyone have insight into how to treat such data to observe the correlative relationships for covariate determination?
Secondly, I would run a similar analysis for a DV that is dichotomous. In this case, I had planned to use point biserial correlation and chi-square tests. However, I did not know if there was a way to compute a matrix out of this, or if each test would have to be individual.
All guidance is greatly appreciated.
You must have to convert categorical variables to numerical variables then you can run the correlation matrix.
You must have to convert categorical variables to numerical variables then you can run the correlation matrix.
You could run a loop in R to do univariate regressions and get the R2 ( squared correlation) and then order the independent variables accordingly from top to bottom.
But please note that the inclusion of variables should be based on knowledge and not solely on correlations: you could run a model with all the variables and assess their R2 contribution with the relaimpo package and information criteria based on likelihood
What i mentioned would also work for a dichotomous dependent variable.
Just use a glm
- Badges
- Science topic
- Similar topics
- Mathematics
- Algebra
- Matrix