Home › Forums › Methodspace discussion › How to conduct multiple regression on dichotomous and scale variables
 This topic has 11 replies, 5 voices, and was last updated 9 years, 7 months ago by Adam Fominaya.

AuthorPosts

25th May 2011 at 12:56 pm #3459Farrukh JaleelMember
Hi all,
Can any one tell me if i can conduct multiple regression based on below mentioned questions (See Fig.)?
– Below my Independent variable is: facilitate coordination and meetings (Dichotomous)
– While there are three Dependent variables (Likert Scale):
– Project schedule development capability
– Clear definition of project scope capability
– Determination of project budget capability
I want to study effect of my 1 independent variable on these 3 dependent variables.
Farrukh
16th June 2011 at 12:30 am #3470Adam FominayaParticipantHello!
So, Multiple Regression is really only appropriate for a single (1) metric dependent variable.
Multiple dependent variables nixes multiple regression as an option.
You’re most likely in MANOVA territory although I have never run one using only a single independent variable. Still, I think you can do it.
I hope this helps a little!
16th June 2011 at 6:11 am #3469Dr Hareesh N RamanathanParticipantRather than using regression, you can develop a score using the 14th question and use ANOVA for qno 12 and independent sample T test for q no 13 to know about the difference in the mean. That will give a clear explanation.
or can go for logit regression
16th June 2011 at 1:21 pm #3468Adam FominayaParticipantTypically, one wouldn’t want to use multiple tests to analyze a single dataset because of alpha inflation concerns. If possible, I would always suggest using one test to analyze everything in one go.
19th June 2011 at 7:28 pm #3467Jeremy MilesParticipantI don’t think that’s true, even if it might be ideal.
If you have multiple outcomes, you run multiple tests.
Sometimes you run tests with and without control variables.
19th June 2011 at 9:10 pm #3466Adam FominayaParticipantIt depends what you mean multiple outcomes. If you mean multiple dependent variables, you’ll want to use MANOVA to look for an overall effect. That is, does the independent variate have an effect on the dependent variate. In the above example, the independent variate would really only include one predictor.
Once you establish the existence of an overall effect, then and only then can you go on to multiple ANOVAs. Even then, you would still want to apply a Bonferroni correction.
If you mean multiple dependence relationships, SEM might be more appropriate. In either case, running a succession of ANOVAs and TTests is increasingly the likelihood of type I error.
The link below can more clearly elucidate the “multiple comparison problem” as they call it.
Here’s a link to the book’s amazon page: http://www.amazon.com/PrimerBiostatisticsStantonGlantz/dp/0071379460
23rd June 2011 at 4:17 am #3465Farrukh JaleelMemberThanks for your comments.
Well, in fact i have 17 independent variables and 3 dependent variables. So, what i am coming up with the idea is to conduct:
– MANOVA to see the differences of means using 17 independent variables and 3 dependent variables. Only MANOVA can allow me to do so in one go. Right?
– But do you think that MANOVA will be equally authentic/replacement of regression for examining relationship/strength of relationship between variables? Therefore, i think that i’ll have to conduct Multiple regression or logit regression 3 times (1st 17 IV’s and one DV, then same 17 IV’s and other DV then same 17 IV’s and third DV).
2nd Issue:
Same pattern of questions, as the snapshot depicts, continues for all 17 independent variables. For every IV i have 2 multiple choice(3 options) questions and 1 scale question. So, should i still dummy code my question no. 12 & 13 responses?
Farrukh
23rd June 2011 at 4:18 am #3464Farrukh JaleelMemberin fact i have 17 independent and 3 dependent varibles. so, conducting the tests following this structure would be very cumbersome and results difficult to comprehend. Isn’t it so?
24th June 2011 at 2:50 am #3463Adam FominayaParticipantAlright, I’m a little less confident answering this one but let me see if I understand what you want/need.
Let’s start with your 2nd issue. Your 17 independent variables are a mixture of metric (Scale) and nonmetric (categorical) variables. In the screenshot above it looks to me like you have 2 nonmetric independent variables (numbers 12 and 13) and 3 metric dependent variables (Number 14). If you intend number 14 to be a independent (predictor) variable, then I think it is actually 3 separate metric variables.
I just want to make sure I’m looking at everything correctly. Tell me if you disagree with anything I’ve said so far.
For MANOVA or Multiple Regression you would still dummy code your nonmetric IVs.
Now your first concern seems to be that you would like to walk out of the analysis with a regression equation. I’m conflicted about how to answer this and am open to any critique of my suggestion. One thing that might help is to know why you want to know so distinctly about effect size. My hunch is that you are applying this in the real world rather than trying to publish a paper. Is that a correct assumption?
Whatever your reasons, here’s what I suggest:
I think the multiple comparison problem is largely a question of “significance”. So if you want to claim that one variable has a significant effect on another (or a variate having a significant effect on another variate), then you don’t want to run multiple tests.
In that respect, I would strongly suggest that your first analysis should be a MANOVA. The MANOVA can be your significance test.
You would first look at whether the independent variate has a significant effect on the dependent variate.
To refresh, your 17 independent variables would make up your independent variate. Your 3 dependent variables would make up your dependent variate. MANOVAs will test this overall effect.
If you don’t have an overall effect, univariate ANOVAs and/or individual multiple regressions would be a bit inappropriate.
If you do have an overall effect, then check which independent variables have a significant effect on the Dependent Variate and get rid of IVs which are insignificant.
Then, take the remaining IVs and examine significance on each of the DVs onebyone. (You can ask SPSS to give you univariate ANOVAs and post hocs. With that many variables it will be messy.)
If, once you have done all of this, you want to run independent multiple regressions to get a regression equation for each DV, I think that would be okay (it’s basically what SPSS does with Univariate ANOVAs) but I would not report R^2 and adjusted R^2 values from the multiple regression.
So, let me rephrase some of this.
By doing the MANOVA, we’re going to determine which variables are ‘eligible’ for individual multiple regressions.
First, establish that you have an overall effect.
Then, look which IVs onebyone are significant on the Dvariate. (toss out IVs which are insignificant)
Then, look at the effect of the independent variables on the dependent variables onebyone (Ask SPSS for univariate ANOVAS)
Then, look at post hocs
This will let you build a road map of variables which can be analyzed further using a multiple regression if you would like to go this route.
All of this assumes that your Dependent Variables are unidimensional. That is, they are different ways of measuring roughly the same thing (a larger construct). If number 14 above is your DVs, then I think this is the case. If it doesn’t represent a single dimension, we should talk more about which significance tests (Roy’s, Pillai’s Trace, etc.) should be used.
17 independent variables is a lot, I am concerned that sample size may be an issue. How large is your sample? I am not qualified to advise on power analyses but I wouldn’t be surprised if you don’t find an overall effect because of the number of IVs and concerns about sample size. If the overall effect is insignificant, I would guess that part of the issue may be that you just ran out of degrees of freedom.
Also, I have been using the terms “independent” and “dependent”… Strictly speaking, in this situation it would be more appropriate to use the terms “Predictor” and “outcome” variables.
I know this is a lot. I hope I was clear and helpful. If you have additional questions, let me know. I’m attaching a short little worksheet that I hope can show more clearly all of the different steps of the MANOVA.
24th June 2011 at 11:47 am #3462Farrukh JaleelMemberThese comments are really extremely helpful for me.
I have gone through the attached file as well. I have a few more questions. Can you please provide me your email, i want to send some details of my work and comments i have on this discussion.
Thanks,
Farrukh
24th June 2011 at 12:07 pm #3461Adam FominayaParticipantSure! i’m glad it helped. Awfominaya@gmail.com
Thanks,
Adam Fominaya30th June 2011 at 8:14 pm #3460NayanMemberYou may try for multinational logit model. Refer standard text book such as Greene’s book on Econometrics.

AuthorPosts
 You must be logged in to reply to this topic.