Hello,
I am working on a regression problem where I have to estimate the percentage of acceptance of an offer. Although my dependent variable is categorical one. ( I tired to fit the logistic but it didn't turn out very well; despite being 17% of 1s in the whole data set and sample size being more than 100k. My model failed to classify the 1s. Among 4.5k 1s in validation data it classified 20 odds as 1s.
So I took a different way around it and tried to make appear my dependent variable as percentage (rather than 0 or 1) by clubbing 10 rows together. I took the central tendency measure for numerical predictor variables (n = 100 ).
I know that there is a loss of information with this approach. But this manipulated data makes easier to see what my aim of the analysis is.
I am asking this question because I didn't come across anything like this till now? If anyone has worked with similar approach, please give your valuable inputs.
Thanks and Regards,
Irshad
There are many harms in this; see e.g. Jaeger 2008. I'm especially not sure what it means to arbitrarily lump 10 rows together rather than grouping by some meaningful factor. Finally, I'm not sure that failing to classify the 1s means your model has a problem in the first place; if there's not a reliable relationship between any predictor variable and this outcome, then that seems normal.
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2613284/
@Stephen, thank you so much for the response. By meaningful factor you mean should I group them according to the values of one of the most important variables in the predictor set?
A meaningful grouping variable depends on your experiment. e.g., the percentage of correct responses for each participant, the percentage of something for something else, whatever. "The percentage of 1s in ten rows that happen to be adjacent in my data frame" does not sound very meaningful to me, but I don't know the design of your experiment. Anyway, even if you have a meaningful grouping variable, the problems mentioned above still hold, and overall I just don't see why you can't use logistic regression (you haven't explained why there is a problem with using a logistic model).
As I mentioned the problem with logistic regression is that it is not classifying 1s. Despite the proportion of 1s being 16-17% (in both testing and validating data). I tried oversampling and undersampling methods too, this helped but this triggered misclassification rate.
One more thing I forgot to ask you that what kind of problems my model will suffer if i group my 100 rows together and treat it as proportion?
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Regression