Date: Sat, 17 Sep 2005 13:47:22 -0400
Reply-To: Lisa Stickney <Lts1@enter.net>
Sender: "SPSSX(r) Discussion" <SPSSX-L@LISTSERV.UGA.EDU>
From: Lisa Stickney <Lts1@enter.net>
Subject: Re: Regression analysis with SPSS
Content-Type: text/plain; format=flowed; charset="iso-8859-1";
Please look below for my answers. I hope they help.
Lisa T. Stickney
The Fox School of Business
----- Original Message -----
From: "Karl Koch" <TheRanger@gmx.net>
Sent: Saturday, September 17, 2005 10:48 AM
Subject: Regression analysis with SPSS
> Hello list,
> I have designed, performed, and analysised a 3x3x2 full factorial
> experiment. For the analysis I have done an analysis with repeated ANOVA.
> All main factors and all interactions were stat. significat.
> Now, for a second purpose I need a different form of analysis. I need a
> function that can predict future dependent variables based on the choice
> factorial levels. The dataset of the previously mentioned experiment was
> rich (over 2300 data points) and should have a decent power. I basically
> want a function that allows me to estimate Y based on my three factors
> X2, and X3). My factor levels are either continous or descrete (which will
> be transformed into contineous).
> I think regression analysis is the right tool for my purpose. However, I
> have not worked with this analytical tool before. I have done some reading
> already but not much of it in SPSS.
> Therefore, I have some questions:
> 1) Based on the experimental design, is the multiple linear regression the
> right statistical method to use for this purpose (assuming that the
> relations are linear)? Or are there other approaches in SPSS that could do
> equally well in order to obtain a function that can best explain and
> future values?
If you're trying to predict, regression is the right tool.
> 2) According to my ANOVA results, confirmed by my contrasts, I have
> interaction effects in all my two-way and also in my three-way
> The bottomline is, this dataset is quite interactive. However the
> interaction effects do not appear to be very strong. Is this a problem for
> regression analysis? If yes, how can it be tackled with? Are ther any aids
> that could be used to account for that?
It shouldn't be, if you have enough power in your dataset.
Two things you should probably know about regression. First it's
correlation based, and second it's actually considered an extention of an
ANOVA analysis. So the correlation tables and the ANOVA results should give
you a rough (sometimes very rough) idea of what you might expect to see in
regression results. However, this is not a hard and fast rule, there are
other factors to consider.
> 3) Since I have three IVs, based on my DOE, could overfitting occur. How
> could I compensate that witout taking out IVs? Can overfitting occur by
> putting in too many results?
I'm not sure what you mean by "overfitting."
The standard measure of model fit in regression is the squared multiple
correlations (R-squared). It represents the amount of variance in the DV
explained by the IVs. The higher the better, but it cannot go above 100%.
As far as what is considered "high enough," basically depends on your
field -- there are many different standards.
> 4) Does somebody know a good tutorial for doing this kind of multiple
> regression (based on the 3 IV and the 1 DV) with SPSS? Are there good
> or articles or other online resources that especially cover this?
There are a lot of good web based resource (try a Google search on "SPSS" &
"Regression"). For an easy to follow book, I'd recommend:
Field, A. P. (2000). Discovering statistics using spss for windows :
Advanced techniques for the
beginner. London ; Thousand Oaks: Sage Publications.
> Kind Regards,
> GMX DSL = Maximale Leistung zum minimalen Preis!
> 2000 MB nur 2,99, Flatrate ab 4,99 Euro/Monat: