# Ordinal Logistic Regression in R - BIC calculation

I am working on an Ordinal logistic regression model. I have an outcome variable with 4 categories and they are ordered. My predictor or explanatory variables are 11 in number with the 11th predictor being a categorical variable with 4 levels.

I am taking a stepwise addition approach where in I am starting by adding one predictor in the model and proceeding to include or exclude further predictors depending on BIC statistics. I am including or excluding variables i.e. predictors in the model depending on the comparison of individual model BIC values.

However, as I build my model using 11th predictor, I get an error that "attempt to find suitable starting values failed" and a warning message that "glm.fit: algorithm did not converge".

The output is pasted below for your reference:

model2.11 <- polr(formula = Q13 ~ Q11, data = mydata13, method = "logistic")
Error in polr(formula = Q13 ~ Q11, data = mydata13, method = "logistic") :
attempt to find suitable starting values failed
In addition: Warning message:
glm.fit: algorithm did not converge


Please let me know where I might be going wrong. All suggestions are welcome. Thanks in advance.

## 1 Answer

I haven't use polca recently - In ordinal regression, issue with starting values (SV) typically happens when the SV for the threshold parameter corresponding to a lower category is higher than SV for the threshold parameter of a higher category. Example: You model a 3-categ dependent variable (Y = {Low; Medium; High}) with "Low" as reference categ. You will then estimate 2 threshold parameters (one for "Medium" and one for "High") and the SV for Medium has to be < SV for High (Usually I simply specify 0, 1 , 2, ...).

• what do you mean by threshold parameter? Can you be a little more detailed please, and provide an example. Thank you. – Shawn Gu Jun 9 '17 at 21:18
• I took a quick look at "polr" and don't think your issue comes from specification of starting values. The key error message is "algorithm did not converge". It seems that there is an issue with your 11th variable. Did you check correlation with your dependent variable? Do you have any observations in each category (You might have 4 levels for your 11th variable but 99% observations clustered on 1 particular level). – Umka Jun 9 '17 at 21:47
• I think this should be the issue in my case, thank you so much. – Shawn Gu Jun 12 '17 at 14:44