Search code examples
rregressionanalysis

Issue with multiple logistical regresstion code with multiple predictors

I am trying to perform multiple logistical regression with some of the variables that came out as statistically significant for a diseased conditions with univariate analysis. We took the cut off for that as p<0.2 since our sample size was ~300. I made a new dataframe for these variables

regression1df <- data.frame(dgfcriteria, recipientage, ESRD_dx,bmirange,graftnumber, dsa_class_1, organ_tx, transfuse01m, transfuse1yr, readmission1yr, citrange1, switrange, anastamosisrange, donorage, donorgender, donorcriteria, donorionotrope, intubaterange, kdpirange, kdrirange, eptsrange, proteinuria, terminalurea, na.rm=TRUE)

I'm using variables to predict for disease condition, which is DGF (dgfcriteria==1), and non-disease is no DGF (dgfcriteria==0).

Here is structure of the data.

enter image description here

When I tried to run the entire list of variables with the glm code I got:

predictors1 <- glm(dgfcriteria ~.,
data = predictors1df,
family = "binomial" )

Error in contrasts<-(*tmp*, value = contr.funs[1 + isOF[nn]]) : contrasts can be applied only to factors with 2 or more levels.

But when I run it with only some of the variables of the dataframe, there is an output.

predictors1 <- glm(dgfcriteria ~ recipientage + ESRD_dx + bmirange + graftnumber + dsa_class_1 + organ_tx + transfuse01m + transfuse1yr +  readmission1yr +citrange1 +switrange + anastamosisrange+ donorage+ donorgender + donorcriteria + donorionotrope,
data = predictors1df,
family = "binomial" )

This output looks really strange though with alot of NAs.

enter image description here

Where have I gone wrong?

Solution

  • Looking at your data structure, you've got a lot of missing values. Quite a few of your variables look to have only 2 or 3 non-missing values in the first 10 rows. When you run regression on data with missing values, the default is to drop all rows that have any missing values.

    Apparently some of your data has bad overlaps, so that when all the rows with missing values are dropped (see na.omit(your_data) for what is left over), some variables only have one level left and are therefore no longer fit for regression. Of course, when you only use some variables, fewer rows will be dropped and you may be in a better situation.

    So, you'll have to decide what to do with your missing values. This should depend on your goals and your understanding of the reasons for missingness. Common possibilities include omission, imputation, creating new "missing" levels, and taking level of missingness into account in your variable selection.