I'll be the first to admit I'm probably in over my head but...trying to perform a forward stepwise regression (using SPSS 20). Have 7 predictor variables (factors comprising items from a survey) and one outcome variable that comprises values from -43.07 to +25.45 where 0.0 is the "best place to be." ENG is the outcome (so trying to find the factors which best predict engagement) but within the engagement model I'm using, engagement is the result of a balance between challenge and capacity so the difference between challenge and capacity gives an engagement score "of sorts"; perfect alignment would be 0.0; more challenge than capacity = positive numbers; less challenge than capacity = negative numbers. Directionality is important within the model as movement away from engagement (0.0) towards positive values results in people being overwhelmed (more challenge than capacity); the other way is underutilized (less challenge than capacity). Hope this makes sense.

  • 1
    To be honest, I don't completely follow your situation, but I can say that stepwise selection is a really terrible thing to do, & is quite unlikely to "find the factors which best predict engagement" in general. If this doesn't make sense, you may want to read my answer here: algorithms-for-automatic-model-selection. – gung Jan 25 '13 at 22:02

If the "best place to be" is 0.00 and the values range from negative numbers to positive ones, then linear regression is not appropriate, because it models a linear (straight line) relationship between the dependent variable and the independent variables. It may be good to take absolute values of the dependent variable first.

On the other hand, you also say that "directionality is important". You may then need some more complex model. One possibility is quantile regression, which lets you use various quantiles of the dependent variable, instead of the mean. Then you could model quantiles that were lower than 0 and higher than 0.

In any case, I can't recommend forward selection as a variable selection method. This has been discussed here and elsewhere many times; searching for "model selection" and "variable selection" should find a lot.

  • Huge thanks; very appreciative! Instinctively I knew that, Peter but darn it anyway...looks pretty good using linear regression :-) I have recoded the raw scores, deciding that directionality probably isn't as important in the prediction but don't get great results. Important to step back and re-group rather than "fish" for the analysis that "looks good" Cheers :-) – user20095 Jan 25 '13 at 22:55

Your Answer

By clicking "Post Your Answer", you acknowledge that you have read our updated terms of service, privacy policy and cookie policy, and that your continued use of the website is subject to these policies.

Not the answer you're looking for? Browse other questions tagged or ask your own question.