while running an ordinal regression analysis of life satisfaction data SPSS warning is: There are 148 (65.8%) cells (i.e., dependent variable levels by combinations of predictor variable values) with zero frequencies. Unexpected singularities in the Fisher Information matrix are encountered. There may be a quasi-complete separation in the data.Validity of the model fit is uncertain." what can be done to overcome this problem.Can i still use the results? .Also when should a location or a scale model be used?
Life satisfaction data was obtained through COMQOL questionnaire from 100 respondents aged 60+ with seven indicators of life satisfaction and some demoghraphic questions. And life satisfaction was recoded into three categories.
I had something similar happen to me. I had to eliminate several of my variables until the issue was resolved. Certain combinations of variables gave me this error.
This issue is more common with binary outcomes, but given that ordinal regression is an extension of logistic regression, it is plausible to see it occur here as well. The problem is likely a result of empty cells for a given outcome level, conditional on the covariates.
I suggest that you run some contingency tables with the dependent variable vs covariates. This will help you find the empty cells.
You may want to have a look at the two answers I gave to this question
https://www.researchgate.net/post/How_to_deal_with_quasi-complete_separation_in_a_data_while_using_GEE_models
The underlying problem is that 100 respondents is very low to use these types of models as the data will be very sparse for estimating the many parameters involved. So I would
1) give up on the ordered multinomial model and treat the outcome (1-7? ) as a continuous score
2) or continue to use multinomial logit and group the 7 point score into something meaningful like Low Mid and High and getting roughly equal numbers of respondents in each category will help estimation
3 ) or collect a lot more respondents!
- Badges
- Science topic