Gologit and Marginal Effect

[Valerie Aulia]

I'm Valerie and currently on my last year of bachelor degree, thus I'm in the process for writing my bachelor theses. I wanna estimate using Gologit as my dependent var is in categorical (Keluhan Kesehatan/Health Complaints (Y) 1 (lowest) and 4 (highest or complaining about health, illness, interference with work, to outpatient care)). I choose Gologit as my model doesnt fit the parallel trend assumption for the usual ordered logit. I have an independent var of Car Ownership (binary, 1 refer to owned car, and 0 otherwise).

I have some control variables: married (0 1), female (0 1), educ (0 1(JHS) 2(SHS) 3(Bachelor dst)), working (0 1), AgeCat (0 1(<15) 2 (15-65) 3(>65)), toilet (0 1), Water Source condition (0 1 good quality), smoking (0 1), commuting frequency (0 1 (2-10) 2(>10)), working hour (0 1 (21-40) 2(41-60) 3(>60)), health insurance (0 1), social insurance (0 1), water deprivation (0 1), food deprivation (0 1).

For calculating the marginal effects, my supervisor told me not to use the AME, rather marginal effect at some specific values. My questions are:
1. Do we have to include interaction terms for my nonlinear model to calculate marginal effects?
2. If so, after including the interaction terms, how to calculate the marginal effect? Do I only have to calculate at the value of interaction terms or we have to add the other control variables too?
3. If not, when I only have to calculate the ME at the specific values. That specific values are only for one variable, or for some variables?

THATS SOO MANY QUESTIONS but i'm super confused since last november😭😭😭😭😭😭😭😭😭😭😭😭😭😭😭😭😭😭😭😭😭😭

[Richard Williams]

First off, I assume you are using gologit2, which is much more powerful than the original gologit program.

Look at this handout:

https://academicweb.nd.edu/~rwilliam.../Margins05.pdf (Adjusted Predictions & Marginal Effects for Multiple Outcome Models & Commands (including ologit, mlogit, oglm, & gologit2)

If you need more background on margins first, see

https://academicweb.nd.edu/~rwilliam/xsoc73994/index.html#Adjusted_Predictions_and_Marginal_effec ts

You may especially want to see

https://academicweb.nd.edu/~rwilliam.../Margins01.pdf (Using Stata's Margins Command to Estimate and Interpret Adjusted Predictions and Marginal Effects)

It gives an overview of margins. Also, slide 45 explains why there is no such thing as a marginal effect of an interaction term.

Run margins AFTER estimating your final model, i.e. the model with interaction terms.

If you read these carefully, I think most of your Qs will be answered. Write back if they aren't.

If you haven't already, I also recommend reading

Williams, Richard. 2006. "Generalized Ordered Logit/ Partial Proportional Odds Models for Ordinal Dependent Variables." The Stata Journal 6(1):58-82. The published article is available for free at http://www.stata-journal.com/article...article=st0097

Williams, Richard. 2016. "Understanding and interpreting generalized ordered logit models." The Journal of Mathematical Sociology, 40:1, 7-20, http://www.tandfonline.com/doi/full/...X.2015.1112384

[Valerie Aulia]

https://academicweb.nd.edu/~rwilliam.../Margins01.pdf (Using Stata's Margins Command to Estimate and Interpret Adjusted Predictions and Marginal Effects)

Dear Richard,
Thank you for he handout! I have some things to clarify: it is said that APR and MERs is better to use so i try to run this:

gologit2 KeluhanKesehatan i.CarOwnership i.Married i.Female ib0.EducCat i.Working ib0.AgeCat i.Merokok ib0.FrekBepergianCat ib0.WorkHourCat i.Toilet i.Kondisi_AirMinum i.HealthInsurance i.SocialInsurance i.FoodDepr i.WaterDepr

margins, dydx(CarOwnership) at(Female=(0 1)) vsquish

The result came out like this:

(I dont know why I cant unbold this )
Does my interpretation right?:
For female, owning a car make them have a probability not to complaining about health (KeluhanKesehatan/HealthComplaints = 1) for about 1.9 percentage poin and female who owns car have a 0.01 (0.0193606 - 0.0182425 = 0.0011181) percentage points higher than male who owns car.

I am confused on interpreting the negative result (for HealthComplaints/KeluhanKesehatan at 2, 3, dan 4). Do I have to say like this:
- Female owner car complains about health at 2 for about 0.6 percentage point less likely than those who doesnt own a car?


So, the positive and negative result are telling us the comparison between owning a car (1) with those who doesnt own a car (0)? If so, for the positive result (for HealthComplaints/KeluhanKesehatan at 1), should I just write: For female, owning a car make them have a probability not to complaining about health (KeluhanKesehatan/HealthComplaints = 1) for about 1.9 percentage poin higher than those who doesnt have a car and female who owns car have a 0.01 (0.0193606 - 0.0182425 = 0.0011181) percentage points higher than male who owns car.

Your help means a lot

Regards,
Aulia Valerie