03-14-2013 03:29 PM
|Summary of Forward Selection|
|Step||Effect||DF||Number||Score||Pr > ChiSq|
|Association of Predicted Probabilities and Observed Responses|
|Percent Concordant||80.5||Somers' D||0.611|
03-14-2013 03:46 PM
I highly suggest you review the following pages and then try asking your question again. This is a very broad question, in my opinion.
Google search using: proc logistic site:lexjansen.com
03-14-2013 03:53 PM
being a classically trained statistician who was introduced to data mining only later in my career, I consider myself biased against data dredging. Stepwise selection is often brought up as a pragmatic example of using computational power to replace domain knowledge. In data mining is it not uncommon to start with hundreds or thousands of variables. It is just unpractical to analyze one variable at a time. That’s where I tend to use stepwise regression, as an initial variable selection method used in combination with other variable selection methods such as decision trees, IV,…
In your case it seems like you’re starting with a small number of variables. That’s where domain knowledge should come in to help decide what to include and what to exclude, sometimes regardless of their p-value.
03-14-2013 04:09 PM
A quick one you say? Not sure.
You have a large number of cases, which means that you could detect very subtle effects which in fact don't make any practical difference. A statistically significant effect is not necessarily an important effect. One way you can assess the importance of an effect empirically is by looking at the change in Percent Concordant when you remove a variable from the model.
Percent concordant is the proportion of matching prediction changes, i.e. if case i is not an event and case j is an event, then if predicted P(i) < P(j) then the ij pair is concordant, if P(i) > P(j) the pair is discordant and if P(i)=P(j) then the pair is tied.