I am currently doing my masters thesis which involves data analysis using SPSS, the dataset has over 135000 items and numerous variables.
I performed a test of multicollinearity for my independent variables (day of the week, year group, free school meal or not and breakfast, break or lunch). All the VIF and Tolerances seemed to be okay although one was VIF=1 and tolerance=1. I also performed a Durbin watson test in the process however some of them came out as low as 0.002 (shouldnt it be between 1.5 and 2.5?)
I then performed binary logistic regression on a new variable (dish of the day or not, 0 or 1) and the independent variables above. The results came out significant however the -2 log likelihood was over 30000, the r squared as low as 0.031 and the hosmer-lemeshow goodness of fit =0.00.
I understand this may meant the model does not fit however I don't know what else to do and my supervisor is away for 3 weeks, its 4 and half weeks until hand in and I am really struggling. Any Advice?
Your problem is likely a combination of a very large sample size and a large number of IVs in your model. Large n means that you can make fine distinctions with hypothesis tests so your hypotheses must be chosen VERY carefully. The large number of variables that are probably in your model isn't helping either. If you are building a predictive model you can use lasso variable selection (see attached papers for descriptions of these things) to remove some unimportant variables from your model. Alternatively you can have a subject matter specialist help you remove unlikely IVs if your model is explanatory as well as help you set your hypotheses finely enough that you can get around the massive sample size. Unfortunately, SPSS is a really poor package to implement any of this mostly because it is limited both in data set size and in statistical procedures available. If you ever do this again think R (book attached). Start with some of these things. It may seem overwhelming at the moment but it does at some time for all thesis students. All this will pass and you will get your degree. Contact me if I can be of further service. Best wishes, David Booth
Clearly, your Durbin-Watson statistics are not appropriate. This indicates that your data is auto correlation.
Also, the most important finding that shows you that the fitted model is good or not is R-square statistic. This number is down to you. So you do not have a good model.
- Badges
- Science topic