Omitted Variable in Margin Order Probit/Logit

[Artur Gai]

Dear all,

I'm having troubles with a simple ordered probit model. Indeed, I am doing a research to know if the risk aversion ratio (varies between 3.5 and 0.9) is perfectly correlated with the investor profile. So I have the following data: Profile type (conservative, moderate, balanced, dynamic, aggressive (all are dummys)), Age (> 39 or <39) dummy, Sex (dummy), respondent has a partner (dummy), and the LossRatio (3.5, 3.3 , 2.5, 2, 1.53, 1.25, 1.1, 1, 0.9) a category of 9 possible values.

When I run the code (I'm not sure it is correct), I get a result for each type of investor profile, except for raggresive, it puts me "omitted"

oprobit LossRatio rconservative rmoderate rbalanced rdynamic ragressive Age_category_39 rpart Nationality_ rfemale c.Age##c.Age , vce(robust)
estimate store Oprobit1

estimates restore Oprobit1
margins, dydx(rconservative rmoderate rbalanced rdynamic ragressive ) predict(pr) post


| Delta-method
| dy/dx Std. Err. z P>|z| [95% Conf. Interval]
--------------+----------------------------------------------------------------
rconservative | -.0454679 .043117 -1.05 0.292 -.1299757 .0390398
rmoderate | -.0377651 .0346429 -1.09 0.276 -.1056638 .0301337
rbalanced | -.0157885 .0148816 -1.06 0.289 -.0449558 .0133788
rdynamic | -.0127741 .014752 -0.87 0.387 -.0416875 .0161393
ragressive | 0 (omitted)

Would anyone have an explanation for this? A big thank you in advance,

Best,

[Clyde Schechter]

Every observation in your data set is characterized by one and only one profile type. That is, exactly one of the variables rconservative, rmoderate, rbalanced, rdynamic, and raggressive is coded 1 (or whatever your numeric code for Yes is) and all the others are coded 0. It follows that the sum rconservative+rmoderate_rbalanced_rdynamic+raggres ive = 1; these variables are colinear. So one of them must be omitted to make the model identifiable. Stata chose ragressive to be the omitted variable. If you prefer to omit one of the others, just do that by leaving it out of the -oprobit- command, and then Stata will include ragressive.

Meanwhile, whichever one is omtited, serves as the reference category for the marginal effects of the others. That is, in the output you show from -margins, dydx()-, the interpretation is that an investor who with a conservative risk profile's expected probability of higher loss ratio is approximatley .045 lower than that of an investor with an agressive risk profile. Analogously for the others.

[Richard Williams]

Clyde does a pretty job of figuring all that out with the information given! I'll just add that I usually don't even try when a post is hard to read. Artur you should read section 12 of the Statalist FAQ, particularly the section on using code tags when you show output.

[John Mullahy]

Artur: Beyond #2 and #3 I'd just add that the margins command you used will give you the result for only the base category's probability, not for the other levels of your dependent variable. If you don't include the predict(pr) but instead use

margins, dydx(rconservative rmoderate rbalanced rdynamic ragressive ) post

you should obtain the marginal effects for all the outcome-category probabilities.

[Artur Gai]

Thank you all for your answers!

[Ashley Keamo]

Clyde, I have now read many of your responses to questions in this forum and I have to say that you are one of the few humans that can translate STATA things into something I can understand. Thank you!