I have a three-category variable with the sample size for each category being 9600, 2500, and 350. The proportions are way off and on computing the multinomial regression, estimates are not computed for the third category. I was wondering if it would help to develop a model using a random sample of category 1 so as to have similar proportions to category two. My questions are: 1) is this a valid strategy 2) is there a precedence for the sampling.
Thanks
Not enough information. First your proposed sampling idea seems to suggest that you know the answer already. Second, how do you get such weird( sample sizes)? using what I presume was a simple random sample since you don't say. I would have expected to see total sample size and three proportions that add up to one. sent. 3 makes no sense as far as I can see. Perhaps you were thinking of a stratified sample but that is not how to do it. Now how to fix things up.
1. What is your research question? 2. What is your sample design to answer the question? - for this get a copy of this book: https://www.bing.com/search?q=Sampling+Techniques+3rd+W+G+Cochran++&pc=U528&form=QBLH&sp=-1&ghc=1&pq=sampling+techniques+3rd+w+g+cochran+&sc=0-36&qs=n&sk=&cvid=99A8C672D48247FF9A5D883B56FC772E and study it. Now you have to have positive values that add to 1 not frequencies for your DV, If this is really some kind of logistic regression Clearly then define your DV. as it is you don't have a logistic regression. At that point begin to read the attached notes assuming that whatever you have now is like what you will get the next time. and yes there will be a next time, because none of your question seems to make any sense. ALWAYS PLAN THIS STUFF PRIOR TO TAKING DATA otherwise often we find GIGO. after reading the logistic regression parts of the attached notes, google multinomial logistic regression and see if your DV will be ordinal or just categorical.that is how you tell between ordinal and multinomial log. Regn. If only 2 categories it is binomial logistic regression. Now you still need the RHS of your model. BTW if you think you have strange size proportions after you calculate your proportions then google Firth Regression.Your research question should help you with the RHS..Again do this prior to data collection. After discarding the original data set which does not appear salvageable based on your description, Go over all of this again with a statistical consultant. See your University's statistics dept to find one. If s/he says ok THEN collect data. And get more assistance with the analysis. You may find this communication rude. It is not meant to be however you do not see to be aware of what you need to do. Please get competent advice. Best wishes, David Booth
Hello Kamalpreet Banga. First, I think it would be a bad idea to select random samples from the two larger outcome categories just to make all categories have the same number of observations. There is no such requirement. Second, when the outcome variable has k categories, you get parameter estimates for k-1 categories. The kth category is the reference category. If this troubles you, think about what you get when you estimate a binary logit model. In that case, the outcome variable has k = 2 categories, and you get 2-1 = 1 set of parameter estimates. HTH.
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Sampling Design
- Sample Size