```Date: Wed, 29 May 1996 14:59:50 -0500 Reply-To: John R Ogawa Sender: "SPSSX(r) Discussion" From: John R Ogawa Organization: University of Minnesota Subject: Logistic regression weirdness (or ignorance) I have a question that arises from what seems to me to be a strange result from a logistic regression analysis done in SPSS. Let me first say that I know very little about logistic regression, and therefore think about it in regression terms. This may be what is confounding my brain, but we'll see. :-) We are predicting a dichotomous variable, AA, with a fairly robust distribution of 65%-35% (robust meaning that we have split it with several different variables and it still comes out 65-35 almost every time). Our independent variables are 1) SS, a dichotomous variable with distribution 40-60, 2) DF, a factor score from a factor analysis, and 3) SC, a 7-point rating scale with an approximately normal distribution. We are primarily interested in any interactions, and then any main effects. Since we have priorities in our questions, we decided to mimic a normal regression's hierarchical entry scheme, entering the 3-way interaction first, then all of the 2-ways in the second block, and finally all of the main effects in the last block. So far so good, I think. There are two things that confuse me about the results. The first occurs on the very first block. SPSS reports that the 3- way interaction has a significant Change chi-squared. Okay. But then, in the VARIABLES IN THE EQUATION report for that step it lists the 3-way term as non-significant (a change of .01 in the p-value). It is my understanding that the Change chi-squared is the change in the model above and beyond the previous model, which for this step is the constant only. If so, then it seems to me that the unique contribution from the 3-way term reported in the VARIABLES IN THE EQUATION section should be the same as the Change chi-square, at least for the step on which the term was entered. Am I thinking too much in regression terms? Or is it that the Change chi-squared is not what I think it is? The second "anomaly" comes later, with one of the 2-way terms. When the DF by SC interaction goes in along with the other 2-way terms in the second block it is not significant. But then, after the third block in which the main effects are entered, in the VARIABLES IN THE EQUATION section it becomes significant! Now I have always thought of things like this as "enough of the garbage shared variance was taken up by the other variables in the equation that the variable in question became significant." But now I'm not too sure. Especially in light of the first "anomaly." The variable becoming significant in the last VARIABLES IN THE EQUATION section is as if we had just entered everything at once in the same block (the "unique" method), no? So how do we reconcile that with the non-significant entry in the second block when we write this up? Anyone? :-) Like I said, I know very little about logistic regression, so my apologies if these questions are stupid. I am just somewhat confused. :-) BTW, we are following up the 2-way interaction (while we are waiting to find out if we should) by dichotomizing SC (splitting at the mean) and running logistic regressions predicting AA with DF for both the low SC group and the high SC group separately. Is this okay? Many thanks for any answers you can give us. John -- =========================================================================== | John Ogawa | INTERNET: | | Senior Statistician | jogawa@maroon.tc.umn.edu | | Minnesota Mother-Child Project |--------------------------------------| | Institute of Child Development | "Oh boy ..." | | University of Minnesota | -Dr Sam Beckett, "Quantum Leap" | =========================================================================== ```

Back to: Top of message | Previous page | Main SPSSX-L page