I have data on number of counselling sessions attended by music students in higher education (counts) for a period of 15 years. I also have data on their gender (male/female), level (undergraduate/postgraduate), main instrument (strings/keyboard/wind, brass, percussion/singers) and nationality (UK/EU/Overseas). I don't have a specific hypothesis (as this data was already there), so I want to explore whether the number of sessions attended is predicted by any of the other variables. Do I look at all of them at the same time as part of a regression analysis and report just the significant ones?
I already know that I can either use multiple regression analysis (as I have counts as dependent variable), or I can also use logistic regression analysis once I apply a median split to the dependent variable so that instead of counts I have a binary variable of "low number of sessions" and "high number of sessions".
However, I'm not sure whether to look at all independent variables together, or separately. I don't have any hypothesis or model that I want to promote. I'm just exploring the data (although I am interested in whether the undergraduate/postgraduate variable might predict the number of sessions attended...)
Alternatively, as I am only using SPSS, I could perhaps investigate the association between counts (although I'd need to transform them into a categorical variable) and each of the independent variables through chi-square tests for association?
Thanks so much!
You need a research question to guide your thinking. a median split is a bad idea because it throws data information away. If you remain with counts as a DV you need special methods for count data. See these links.
https://www.google.com/search?q=regression+with+count+data+R&ie=utf-8&oe=utf-8&client=firefox-b-1. R is probably your best choice of statistical package for this situation. A beginning manual is attached. If you are going to use regression you will build either a predictive or explanatory model(see attachment) Your research question will help you decide which. You have lots to think about, but don't hesitate to ask. Best wishes, David Booth
Thanks a lot, David! I'm simply interested in exploring the relationships between the number of attended sessions (counts) and any of those independent variables. I am using SPSS and am not sure which tests might be the most appropriate. Perhaps I could simply investigate the association between them...so, maybe chi-square tests for association (but that means I have to transform the counts into a categorical variable)?
I would suggest sticking with regression. you get more information. Use the count methods I suggested. SPSS should have some basic stuff. Google what I did but replace R with SPSS. Lot to think about but feel free to ask. Best wishes, David
You shouldn't report only the significance ones, as statistical significance has little meaning in the context of exploratory comparisons without a priori predictions. If you want a way to determine which ones look worthwhile to focus on, it would be better to look at their magnitudes directly -- e.g., putting all the predictors on the same scale so you can look at their standardized coefficients and see what's biggest. (If you do look at statistical significance, it is also best to limit your interpretation to just "statistically significant ones are ones that are worthy of looking into in a future experiment", not assuming anything about statistically significant ones representing "real" or "important" effects.)
Raluca -
It seems to me that you do have an idea as to what you want to do: You said "...I want to explore whether the number of sessions attended is predicted by any of the other variables." So I think that you can guess at a few models, each containing one or more independent variables, and compare the performances of these models. - I have not actually used poisson regression (used for counts), so I don't know much about that, but you could look into it. - As one way to compare model performances, if you can save out some data, not used in estimating the regression coefficients for your candidate models, you might compare model performances in predicting for those data.
Best wishes - Jim
PS - Note that the graphical (estimated) residual analyses at the bottom of the following might be of interest to you:
https://onlinecourses.science.psu.edu/stat501/node/377/
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Statistics
- Statistical Software
- SPSS