I'm looking for some insight on regression.
I have a nominal variable called educational level with 4 categories.
1. Illiterate.
2.2ndary education
3.university education.
4. Post graduate education.
I'm conducting logistic regression with anxiety and depression as dependent variables in a sample of 712.
But the first category has only 2 responses, and my spss tells me that the test model fit is uncertain.
I'm thinking of making it as system missing, and conduct the analysis with the remaining groups.
Is my approach correct or not?
Thanks
Your best bet is multiple linear regression, I would have suggested you use STATA but the result wouldn't be valid.
Hello Tariq,
Yes, categories with very few instances can make estimation of model parameters more challenging (and likely, less accurate). I'd try either: (a) dropping the category as you propose; or (b) combining that category with the next category and keeping the cases (e.g., illiterate + secondary), then re-estimating your model.
Whatever approach you take, do describe it in your write-up. As well, if the model coefficients turn out to be markedly different from those in the original analysis, make mention of that (and, possibly, present both).
Good luck with your work.
It is obviously preferable to combine the categorical variables using PCA than dropping any of the variables. This is because you will miss the pertinent attributes of part of your initial research variables. It will no longer be addressing the initial study objective.
Correlation coefficients could be used for ORDINAL SCALE data (Spearman's correlation coefficient -the most advisable-, Kendall's Tau-c and Somers's d); as all these measures take values between -1 and +1, and reach extreme values when there is perfect agreement or discordance and values close to 0 indicate absence of association ... that is, as the Pearson coefficient for INTERVALS level, then a REGRESSION COEFFICIENT can be made in a similar way to how it is done with Pearson's r (better with Spearman's) ... if it is not possible, the most advisable would be a MULTIPLE LINEAR REGRESSION
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Regression