Multiple Logistic Regression Help

[David Harley]

Hi everyone,

I am bit new to STATA so please bear with me, I really really appreciate your help!!!

Do you think it is appropriate/possible to use STATA's logistic regression to accomplish the following:

Objective:
Measure the correlation/affect membership in a network has on the likelihood of being employed.

Dependent variable:

Employment status (employment/unemployed)

Independent variables:

Membership in 9 different types of organizations:

A.Political association
B. Neighborhood association
C. Social service club
etc

To each of these respondents answered either:
1. Yes, actively involved
2. Yes, but hardly involved
3. No

Notes:

I am mostly interested to see an overall correlation, that is if you are more actively involved in more organizations are you more likely to be employed?
My secondary interest is comparing the relative correlation between each independent variable and employment.

Am I on the right path? is a multiple linear regression the correct one to use?

Thank you so much in advance!!!!! I am really struggling!!!!!

-David

[Maarten Buis]

You have one binary dependent/left-hand-side/explained/y-variable, so logistic regression seems appropriate.

The most complete way to describe the relationship you are looking for is interacting all the association variables. This will lead to way too many effects, and the model may very well not converge. So you probably need to simplify your model, which necessarily means loosing details. One way to do that is to create a new variable that counts the number of organisations in which a respondend is actively involved, and use that as an independent/right-hand-side/explanatory/x-variable. That reduces 27 coeffcients to 1 coefficient, because you no longer distinguish between the degree in which respondents are involved and the types of associations in which the respondent is involved. You may want to choose an intermediate solution, in order to keep a bit more details.

[David Harley]

Thank you, yes that makes sense. In fact I did this which gave a statistically significant correlation but negative. So I disaggregated the correlation so I could how each type of association correlates to employment. Which gave a good result for one of them "V5," so then I tried using just V5 and the results changed substantially. Odds ratio goes from 2.28 to 1.61 and the p value from 0.004 to 0.044. Why would this be? I have attached screenshots of both regressions for reference.

Thanks

bryce

[Carlo Lazzaro]

David:
as Maarten's helpful reply implied, simplification comes at a cost (in this case a "worsened" model performance).
That's probably why Maarten's wisely suggested a midway in between your first and last model; in my opinion, the latter is possibly affected by omitted variable bias.
I would skim the literature and see what Others did in the past when presented with the same research topic.

[Maarten Buis]

A possible explanation for the negative effect of aggregated social capital can be that you failed to adjust for the respondents' gender. Women tend to be more active in various organizations, and young women tend to have a smaller chance of being employed (children).

[David Harley]

Thank you both! Very useful again.

Re: Carlo: I found a paper with a quite similar topic and they seemed to only use the latter, wherein just only one variable was compared to employment status. So then it would be best to follow their example?

Re: Maarten: I was also thinking of this, so I controlled for both years of education and gender (since literature suggests these are significant factors affecting employment status) and my regression yields more or less the same results. So this should be a good thing, right? I should be able to safely say that the results suggest participation in neighborhood associations positively affects employment, specifically those young people involved in this type of association were 1.6 times more likely to be employed. Is this accurate?

Again, I'd really like to thank you both for insight and kind help.