Logistic regression and SPSS
August 3, 2009 12:38 AM Subscribe
Stats filter: Is logistic regression what I need? If so, can I do this in regular SPSS (v17) or do I need the add-on?
I have data from a vision-related quality of life questionnaire (VRQOL). I'm trying to investigate whether vision impairment is associated with differences in VRQOL scores. To do this I have to investigate associations between the VRQOL score and other factors such as age/gender/education/location.
Everything I've read so far has led me to logistic regression (using the proportional odds model as my scales are ordinal in nature). But logistic regression seems to be only available through an add-on that I don't have access to with my version of SPSS 17. Is this normal?
If logistic regression is appropriate here, is there any other way I can do this in SPSS without an add-on? If not, is R something I should investigate? Any other options?
I have data from a vision-related quality of life questionnaire (VRQOL). I'm trying to investigate whether vision impairment is associated with differences in VRQOL scores. To do this I have to investigate associations between the VRQOL score and other factors such as age/gender/education/location.
Everything I've read so far has led me to logistic regression (using the proportional odds model as my scales are ordinal in nature). But logistic regression seems to be only available through an add-on that I don't have access to with my version of SPSS 17. Is this normal?
If logistic regression is appropriate here, is there any other way I can do this in SPSS without an add-on? If not, is R something I should investigate? Any other options?
Response by poster: Thanks for the help, but the only options I have under Analyse>Regression are: Linear/Curve Estimation/Partial Least Squares/Ordinal. I'm not sure that any of these are for me?
However under Add-ons>Regression my options are: Binary Logistic/Multinomial Logistic/Probit/Nonlinear/Weighted Least Squares/2-stage least squares.
It seems a bit weird that these are add-ons. Everything I've found in the last few weeks about about conducting logistic regression says to follow the Analyse>Regression>Binary Logistic path. All this time I've been thinking that my newer version of SPSS must just have a slightly menu path so have been going crazy trying to figure out how to do this. But now I'm thinking my particular version of SPSS doesn't have full functionality?
I've actually been using that Andy Field book for a while and find it extremely helpful.
posted by bingoes at 2:27 AM on August 3, 2009
However under Add-ons>Regression my options are: Binary Logistic/Multinomial Logistic/Probit/Nonlinear/Weighted Least Squares/2-stage least squares.
It seems a bit weird that these are add-ons. Everything I've found in the last few weeks about about conducting logistic regression says to follow the Analyse>Regression>Binary Logistic path. All this time I've been thinking that my newer version of SPSS must just have a slightly menu path so have been going crazy trying to figure out how to do this. But now I'm thinking my particular version of SPSS doesn't have full functionality?
I've actually been using that Andy Field book for a while and find it extremely helpful.
posted by bingoes at 2:27 AM on August 3, 2009
the only options I have under Analyse>Regression are: Linear/Curve Estimation/Partial Least Squares/Ordinal
Isn't that the one you want? I thought ordinal regression was a variety of logistic used for outcomes like (excellent, good, poor), which is what your survey sounds like.
posted by magic curl at 3:38 AM on August 3, 2009
Isn't that the one you want? I thought ordinal regression was a variety of logistic used for outcomes like (excellent, good, poor), which is what your survey sounds like.
posted by magic curl at 3:38 AM on August 3, 2009
Vision impairment is your dependent variable? And it is a ranked ordinal measure? It sounds like magic curl is right, you need "ordinal". However, the way that you've asked your question makes me think that you might just be looking for correlation (Pearson's R). Are you attempting to explain a causal relationship?
posted by B-squared at 5:55 AM on August 3, 2009
posted by B-squared at 5:55 AM on August 3, 2009
You would want binary logit or probit if your dependent variable is binary.
Because your VRQOL score is ordinal, you want ordered logit or probit. Or, if you think there might be some nonmonotonicity in the responses or otherwise don't want to assume that more X always leads to a higher score or always to a lower score, multinomial logit or probit.
I don't have the slightest idea how you would do that in SPSS. The only thing I know how to do in SPSS is resave my data so that HLM6 will read it.
Unless you have some need for some piece of software unique to SPSS (or SAS), I generally recommend against any of those jerks with annual licenses, especially when R is free and Stata is ~$150 through a gradplan for a perpetual license. But SPSS has become the default in some fields, and there's a certain wisdom is using software you can ask your neighbor about.
In R the command would be (after loading data and doing a "library(MASS)"):
polr.output.1<-polr(as.ordered(depvar)~iv1 iv2 iv3 iv4, data=dataname otheroptions)
summary(polr.output.1)
In stata the command would be (after loading data):
ologit depvar iv1 iv2 iv3 iv4, options
Both R (through R commander) and stata have some facility for point-and-click operation similar to SPSS's usual mode.
posted by ROU_Xenophobe at 6:11 AM on August 3, 2009 [1 favorite]
Because your VRQOL score is ordinal, you want ordered logit or probit. Or, if you think there might be some nonmonotonicity in the responses or otherwise don't want to assume that more X always leads to a higher score or always to a lower score, multinomial logit or probit.
I don't have the slightest idea how you would do that in SPSS. The only thing I know how to do in SPSS is resave my data so that HLM6 will read it.
Unless you have some need for some piece of software unique to SPSS (or SAS), I generally recommend against any of those jerks with annual licenses, especially when R is free and Stata is ~$150 through a gradplan for a perpetual license. But SPSS has become the default in some fields, and there's a certain wisdom is using software you can ask your neighbor about.
In R the command would be (after loading data and doing a "library(MASS)"):
polr.output.1<-polr(as.ordered(depvar)~iv1 iv2 iv3 iv4, data=dataname otheroptions)
summary(polr.output.1)
In stata the command would be (after loading data):
ologit depvar iv1 iv2 iv3 iv4, options
Both R (through R commander) and stata have some facility for point-and-click operation similar to SPSS's usual mode.
posted by ROU_Xenophobe at 6:11 AM on August 3, 2009 [1 favorite]
Response by poster: I've spent half the day thinking that I do need ordinal regression, and the remainder of the day thinking I need something else. These answers have me heading back to OLS again.
B-squared: I'm trying to show that VRQOL scores are explained by the variability in vision impairment. To do this I believe I need to develop a multivariate model that investigates other potential factors? I guess that means a causal relationship?
Thanks for these answers, I'll follow these leads and see what I can discover.
posted by bingoes at 6:57 AM on August 3, 2009
B-squared: I'm trying to show that VRQOL scores are explained by the variability in vision impairment. To do this I believe I need to develop a multivariate model that investigates other potential factors? I guess that means a causal relationship?
Thanks for these answers, I'll follow these leads and see what I can discover.
posted by bingoes at 6:57 AM on August 3, 2009
I'm not sure I understand your methodology or hypothesis so I apologize for being general.
If you could state your hypothesis and null hypothesis, that would be helpful.
That said I do have a tip or two that might help you, I've heard good things about http://www.spssforum.com. I've also used http://www.ats.ucla.edu/stat/ and found it helpful albeit for SAS, not SPSS.
If your still in a academic program and your advisor hasn't been helpful, you may want to use facebook or something and ask your fellow students who locally knows spss well. You may even get even get the name of somebody outside your dept you'd never have thought to talk to. Somebody at the other end of the computer is oftne less helpful then having somebody sit down next to you.
posted by gryftir at 10:02 AM on August 3, 2009
If you could state your hypothesis and null hypothesis, that would be helpful.
That said I do have a tip or two that might help you, I've heard good things about http://www.spssforum.com. I've also used http://www.ats.ucla.edu/stat/ and found it helpful albeit for SAS, not SPSS.
If your still in a academic program and your advisor hasn't been helpful, you may want to use facebook or something and ask your fellow students who locally knows spss well. You may even get even get the name of somebody outside your dept you'd never have thought to talk to. Somebody at the other end of the computer is oftne less helpful then having somebody sit down next to you.
posted by gryftir at 10:02 AM on August 3, 2009
I've spent half the day thinking that I do need ordinal regression, and the remainder of the day thinking I need something else. These answers have me heading back to OLS again.
OLS will give you a ballpark answer.
If this is for anything serious academically, you can expect discussants or reviewers or both to tell you that because your DV is ordinal, you need to be doing ordered logit or probit.
I'm trying to show that VRQOL scores are explained by the variability in vision impairment.
Do you mean that you think the standard deviation of vision impairment, over some group, partly determines the VRQOL scores of individuals within groups? That a group whose vision impairment is more varied will have different VRQOL scores than a group whose vision impairments are more consistent? That sounds like a pretty unpleasant hierarchical model.
To do this I believe I need to develop a multivariate model that investigates other potential factors?
This is usually pretty simple.
The set of variables to include in anything from the multiple-regression world is:
(1) All the variables relevant to your theory (but not with several variables operationalizing the same conceptual variable; that just causes collinearity).
(2) The variables that the existing literature tells you should act as important controls (or a good reason not to include one or more).
If you don't have a theory, you shouldn't be running a regression (or similar).
If you aren't familiar enough with the existing literature to have a good sense of what the important controls might be, you don't have any business running a regression (or similar).
I guess that means a causal relationship?
Doesn't matter. Except in a few narrowly-specified time series applications*, you're not going to be finding, showing, or proving causality. The most you can do is say that some set of empirical relationships are consistent with one causal story, and not with some other(s).
*And even really just get at another kind of empirical relationship consistent with one causal story and not with another.
posted by ROU_Xenophobe at 10:27 AM on August 3, 2009
OLS will give you a ballpark answer.
If this is for anything serious academically, you can expect discussants or reviewers or both to tell you that because your DV is ordinal, you need to be doing ordered logit or probit.
I'm trying to show that VRQOL scores are explained by the variability in vision impairment.
Do you mean that you think the standard deviation of vision impairment, over some group, partly determines the VRQOL scores of individuals within groups? That a group whose vision impairment is more varied will have different VRQOL scores than a group whose vision impairments are more consistent? That sounds like a pretty unpleasant hierarchical model.
To do this I believe I need to develop a multivariate model that investigates other potential factors?
This is usually pretty simple.
The set of variables to include in anything from the multiple-regression world is:
(1) All the variables relevant to your theory (but not with several variables operationalizing the same conceptual variable; that just causes collinearity).
(2) The variables that the existing literature tells you should act as important controls (or a good reason not to include one or more).
If you don't have a theory, you shouldn't be running a regression (or similar).
If you aren't familiar enough with the existing literature to have a good sense of what the important controls might be, you don't have any business running a regression (or similar).
I guess that means a causal relationship?
Doesn't matter. Except in a few narrowly-specified time series applications*, you're not going to be finding, showing, or proving causality. The most you can do is say that some set of empirical relationships are consistent with one causal story, and not with some other(s).
*And even really just get at another kind of empirical relationship consistent with one causal story and not with another.
posted by ROU_Xenophobe at 10:27 AM on August 3, 2009
I'm with ROU_Xenophobe...I think you need to better define your statistical question. If vision impairment if your dependent variable, and it is a YES/NO variable, then you can use a logistic regression with VR-QOL, gender, age, etc. as the independent variables. If you are only trying to look at how VR-QOL is related to gender, age, etc. without regard to vision impairment, then you should probably examine some correlations first. Check out what they did in this manuscript with seemingly similar variables, and you should also look at what others have done with the VR-QOL variable.
posted by gubenuj at 9:55 PM on August 3, 2009
posted by gubenuj at 9:55 PM on August 3, 2009
This thread is closed to new comments.
In the "Analyse" drop down menu, scroll down to "Regression", and it gives you the full range of choices.
However if you are not that familiar with the regression options in SPSS, you would do best to get someone to teach it to you. It looks intuitive, but actually its all too easy to hit some random choices, and come out with plausible but utterly incorrect results.
The Andy Field SPSS Book will get you started in the absence of an actual SPSS sage to teach you.
posted by roofus at 1:45 AM on August 3, 2009