Hi all,

I would like to analyze a two-period survey data with attrition, with both dependent and independent variables being Likert-scale questions. However, I have struggled to find answers to all my questions.
  1. Should I dichotomize Likert-scale variables? I.e. go from a Strongly agree/agree/neither agree nor disagree/disagree/strongly disagree grading to a Not agree/Agree grading? It seems that keeping the variables as they are would allow to better identify changes in attitude (i.e. from not agree to neither agree nor disagree) which would not be captured by dichotomizing the variable. However, I’ve talked to a couple of economists who support dichotomizing variables (I forgot why!)
  2. How to generate composite indicators? Based on what I read, there seem to be two school of thoughts: the “naïve” approach, i.e. just averaging the scores; and using explanatory factor analysis to identify factors and use them instead of the variables. However, I’ve been exploring using Confirmatory Factor Analysis to test whether variables measuring different aspect of a concept share common variance (i.e. test the validity of the theory).
  3. Given the above, I was originally looking at using a Wald Difference-in-differences approach, using indicators for both the dependent and independent variables. However, I’m not sure if instead, I should be using a Wald Ordinary Logit/Probit Difference-in-Differences method. If so, I’m not sure whether such approach exists and how they would be implemented.

Thank you for your help