I conducted a survey on college students to assess what demographic variables affected their views of and knowledge towards climate change. There are 11 variables total:

  • Gender
  • Religion
  • Political affiliation
  • Hometown size (rural or urban)
  • Hometown region
  • Did they have a childhood pet (yes/no)
  • Degree of discussion of climate change in high school
  • Type of high school
  • High school's graduating class size
  • University region
  • University major

I ran a Chi-squared test of independence (X2) between each variable pairing to determine which showed a significant (P ≤ 0.05) association, then calculated Cramer's V (V) for each significant pairing. Interpretations of effect size were made according to Cohen 1988 (synopsis borrowed from the R Handbook):

r=2: Small = 0.10-0.30, Medium = 0.30-0.50, Large ≥ 0.50
r=3: Small = 0.07-0.20, Medium = 0.20-0.35, Large ≥ 0.35
r=4: Small = 0.06-0.17, Medium = 0.17-0.29, Large ≥ 0.29
r=smallest number of rows or columns in the contingency table for each pairing

I have 35 pairings with significant chi-squared values, but most of these have small effect sizes. For example, the pairing Gender and Religion have a significant association (X2=7.5, P=0.02), but a small effect size (V=0.06, r=2). Now I'm trying to build my models with uncorrelated variables, but this will be difficult given the number of significant pairings.

My question is this: is there a basis for ignoring significant correlations that have small effects? Again using the example with Gender and Religion, can I include these variables in the same model since the Cramer's V value is so small? Being able to do this would make model building much easier, but I haven't found any basis for it and don't want this to bite me down the road. Any thoughts or suggestions would be appreciated.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.