I am currently assessing some results I have from a model I applied on a corpus of text data I have mined.
My problem is that my professor have told me to use a certain method, and I do not really know how to attack this problem the most sensible way.
The main idea is to assess if there is a relationship between months and/or years on my response variable. My data has the following nature:
DV: Binary (Event or non event)
IV1: Month where event occurred/not occurred
IV2: Year where event occured/not occurred
I have a total of 431.000 observations
As it is now I have transformed my data so that I have count data for each time period instead of a binary DV. I also did a logit transformation on defined as ln(DV/(1-IV3)), in order to sort out the effect of activity.
DV(Event): Events in a time period given IV1 and IV2
IV1(Year): Year 1995 to 2012
IV2(Month): Jan to Dec within IV1
IV3(Activity): Events + Non-events in time period given IV1 and IV2 (Neutralized)
I have a total of 176 observations
Right now my model is defined as:
DV = b1*IV1 + b2*IV2 + errorbut I am struggling alot with the intuition of whether this make sense. I have been looking into poisson models and zero inflated models, but until now the most sensible I think is to do the logit transformation in order to neutralize activity and the look at year and month as factors in a normal linear regression, but then again I am thinking that I might as well use a logistic regression.
Does any of you know how to handle such a case?
I am using R, SPSS and rapidminer.