Ordinal regression model and interacting variables

[Mengo Elena]

Hello,

I am doinf a reseacrh aimed at investigating how and if decision making power is affected by gender. I have a dataset with 310 observations across 3 different sites. In order to investigate decision making respondenst have been asked to reply 4 questions about what have beed assumed to be aspects of decision making power which could be observed using likert scales.
Then I calculated the factor score on the latent factor "decision making" and ran an regression model using decision making as dependent variable and socio demgraphic variables as independent explanatory variables.
The independent variables are: Gender (male=0 female= 1), Income level (1=low, 2=medium, 3= high) , Age (numenrical), Years of experience (numerical) job- role (categorical), number of people in the household (categorical), civil status (categorical)


Command: regress DMACT gender age edu civ_status n_ho lev_inc role y_exp

The result is in the attachment below



Now I have tried to compute gender differences in prediction with interactions (e.g. gender*education, gender*years of experience) with the following outcome:
Command: regress DMACT gender age edu civ_status n_ho lev_inc role y_exp genderXed genderXY_exp

The result is in the second attachment below

After controlling for interactions the variables gender and years of experience are no longer significant and also the new interacting variables are no longer significative.
I wonder, Have I used the right command and if so how should I interpret the results?
If I have done something wrong could you please suggest how to amend the mistake?

Many thanks,

Elena

[Richard Williams]

There are lots of issues here.

The significance of the main effects means little once interactions are added. To understand why, see https://www3.nd.edu/~rwilliam/stats2/l53.pdf. If you want to make the main effects more interpretable, you could try centering education and years, as explained in the handout.

Your interaction effects are far from being significant. Unless there is some compelling reason to have them there, you could drop them. You have a fairly small sample and many variables that have little or no effect, so adding additional unnecessary terms doesn't help you any.

This is NOT an ordinal regression. If it were, you would be using something like ologit. You created a scale out of your ordinal variables and are now using OLS regression.

You computed the interactions yourself. Unless you have an ancient version of Stata, that is usually a bad (or at least sub-optimal) idea. Use factor variable notation instead. See -help fvvarlist-. You could also see https://www3.nd.edu/~rwilliam/stats3/Margins01.pdf.

Posting images of output is discouraged on the list. Use code tags instead. See pt. 12 of the FAQ.

Overall, though, I don't see anything wrong. Gender and its two interactions are all insignificant, but it would be a mistake if you tossed all 3 out of the model.

[Carlo Lazzaro]

Elena:
welcome to this forum.
Please note that real given and family name are preferred on this forum (and, admittedly, due to recent gossip echos, nicknames like yours can make interested listers raise their eyebrows in US as well as abroad).
That said:
1) you should by far better off with using -fvvarlist- for categorical variable and interactions creation (see -margins- and -marginsplot-, too);
2) hunting for the regression model with the highest number of statistical significant predictors is not the way to go; try to give a fair and true view of the data generating process, instead;
3) via -estat ovtest- you shoud check whether the specification of your regression model can be improved; for instance, it is quite usual to search for turning points for continuos predictors such as age, education and experience (which, I presume, are expressed in years);
4) your model may suffer from endogeneity, as personal ability, which I do not see among predictors, can influence both decision-making power and income.

PS: crossed in the cyberspace with Richard's helpful list of items to be considered for improving your model and the way to share it with interested listers.

[Mengo Elena]

Many thanks Mr WIlliams,

Indeed you are correct, it is not an ordinal regression model, although I have started with an ologit model and used each of the observable variables used to ascertain whether or not gender has an effect but once I calculated the factor scores I tried the OLS regression aven though it has not helped.
Thanks fr your suggestion about cenetering education and years, I will try it and see what happens.

Best,

Elena

[Mengo Elena]

Thanks Mr Lazzaro,

I have sent a request to change my namae and use the real one. I tried to search for turning points for both age and years of experience but they do not seem to be statistically significant.
I will surely use fvvarilst and see whether it will improve the model.

Many thanks,

Elena

[Carlo Lazzaro]

Elena:
did you consider endogeneity, too?
Thanks for replacing your nickname with your real one (see https://www.statalist.org/forums/help#realnames).
Please, call me Carlo, as all on (and many more off) this forum do. Thanks.