 |
 |
Hi Helen First, examiners never ask about or comment on what you expect them to ask about. For your examiner to ask about the statistics they need to be fairly confident, and they need to be less interested in the substantive issues than the statistical issues. I'm not your examiner, so you might be OK there. :) Briefly, forward selection in regression is a form of stepwise regression (the terminology isn't really agreed upon, some people say that stepwise is one method, and forward is another, others say that there are several stepwise methods, and that forwards is one). SPSS stepwise regression hasn't changed since 1988 (and probably before then). There are methods that have been developed which try to solve these problems, but they are very complicated and I don't think that they are implemented in SPSS (some of them are implemented in SAS, which tends to be ahead of SPSS). They are called things like least angle regression (LARS), and they try to be less 'greedy' than forms of stepwise regression. However, they still have problems (the wikipedia page is a short explanation.) However, your defence is: What else were you supposed to do? Regular regression wouldn't help, because of the collinearity. You can either run a forward regression model, or you can run no model. You then use the results from the forward regression model to further develop your theory, and the next data collection exercise can be refined so that you can run a better analysis next time. Obviously if you'd known that this was going to be a problem you could have done something about it. But you didn't. Now you did, and so you've learned from the experience. (And all research, including a PhD is about learning from the experience). SEM wouldn't have helped, it's just a form of regression. You would still need to use a variable selection procedure, you would just do it by hand. Final thought: if this is the biggest problem that the examiners can raise, then hooray. It will take you an hour of writing to sort it out. The viva is limited in time to a couple of hours (usually) and if you spend 10 minutes on this, that's 10 minutes they're not going to spend on something that's harder to sort out. :) [I've probably said this on the psych-postgrads list in the past, but I'll say it again. In my viva, I was asked why I'd used a particular approach to analyzing the data which looked weird. I said "I don't remember, but there must have been a good reason, or I wouldn't have done it." The examiners said "Please remember the reason and add it." So I did.] Jeremy On 6 October 2010 03:02, Helen Mann <[log in to unmask]> wrote: > Hello, > > I have my viva in 3 wks time and I'm slightly concerned about my poor > statistical knowledge. I used forwards regression in my studies because > although I was measuring the components of the theory of planned behaviour > (tpb) I wacked in some other measures and then used forwards regression to > see what measures were valid additions to the TPB model. > > Can anyone send me some references to studies that have used forwards > regression (so I can have them in my bag as evidence the technique can be > used). > > I've just read a paper from 1988 about forwards regression (Leigh (1988) > Assessing the importance of an independent variable in multiple regression: > is stepwise unwise? Journal of clinical epidemiology, 41 (7) 669-677) and > it's talking about that there are collinearity problems with data that is > entered into forwards regressions and therefore R2 can never be uniquely > allocated among independent variables. Also if 2 variables move together > then it is likely only 1 out of the 2 will be selected into the regression > model because the other variable will not add to the R2. It also says the > stepwise routine does not indicate the magnitude of an effect. > > Anyway, I pressume that alot has moved on in the world of stats since the > above paper was written and I was wondering therefore if modern SPSS output > from forwards regressions take all these things into account.....In the > output, after providing you with the IV's that should be included in the > model it also lists those excluded and then there's the table where you get > the beta values. The beta values show you the size and direction of the > effect don't they??! So surely that is addressing the problem in the 1988 > paper where stepwise did not indicate the magnitude of the effect? > > Anyone think that my use of forwards regression was an appropriate > technique?? I did not have SEM software (or knowledge!) which is why I ran > regressions in the first place.... > > I really hate stats - can anyone help?! > > H > -- Jeremy Miles Psychology Research Methods Wiki: www.researchmethodsinpsychology.com