I have been using the R survey package to run multivariable logistic regressions (multiple independent variables, both categorical and continuous, with some of the categorical variables having multiple levels).

I am using the svyglm function.

Is it possible to use the same method if the dependent variable has multiple levels that are ordered, i.e. A < B < C < D < E? What about unordered? Right now, my dependent variable only has two levels (i.e. 0 and 1). If ordered is possible, do I need to convert them into integers?

Thank you!

CodePudding user response：

For an ordered polychotomous DV, you would want ordinal logistic regression, which is available in the MASS package (see MASS::polr()). For an unordered DV, you’d want multinomial regression, available in the nnet package (see nnet::multinom()).

And no, you wouldn’t convert your ordered DV to integers; you would create an ordered factor like this:

df$var1 <- ordered(df$var1, levels = c("A", "B", "C", "D", "E"))

CodePudding user response：

For dependent variables that are categorical where order matters (for example, trying to predict what star rating a movie will receive on a 5-star scale where 1 is the worst and 5 is the best), you will want to run an Oridinal Logistic Regression model - also sometimes called a Proportional Odds model.

Keith McNulty has written a great explanation of this, with tutorials in R! It has a walkthrough example of how to set up your model, as well as how to interpret the results: https://peopleanalytics-regression-book.org/ord-reg.html

When the dependent variable is unordered and there are more than two possible outcomes, you will want to run a mulitnomial regression model, found here: https://peopleanalytics-regression-book.org/multinomial-logistic-regression-for-nominal-category-outcomes.html

If your categories are unordered but you ONLY have two possible outcomes, you will want to run a binary logistic regression: https://peopleanalytics-regression-book.org/bin-log-reg.html