Re: st: Average marginal effects in ordered probit models

[Maarten Buis <maartenlbuis@gmail.com>]

On Thu, Sep 13, 2012 at 9:31 AM, Chris Ambrey wrote:
<snip>. However, I want to use predictnl and/or nlcom
> to obtain the average marginal effects for several interaction terms in the
> one model. I'm estimating an ordered probit model. I'm using version 11.2 of
> STATA. Life satisfaction is my dependent variable (0-10) and I have
> greenspace (continuous, percentage of greenspace in the census tract) and a
> number of socio-economic and demographic characteristics as explanatory
> variables (some dichotomous and others continuous).

<snip>
> Furthermore, I have found a very recent article "Interaction Terms in
> Nonlinear Models" by Karaca-Mandic, Norton and Dowd (2012) which has been
> very helpful.

On Statalist we ask you to provide full references. See:
<http://www.stata.com/support/faqs/resources/statalist-faq/#others>

> Additionally, do the average marginal effects apply to a particular outcome,
> such as a self-reported life satisfaction score of 10 and are thus
> interpreted as the probability of reporting say a 10 for a one unit increase
> in the explanatory variable?

To start with your last question: Yes. So, by estimating marginal
effects you have undone the very advantage of ordered regression. The
point of ordered regression is that you can use the ordered nature of
the dependent variable to get one effect for each (or at least some of
the) explanatory/independent/right-hand-side/x-variable. This
simplifies the model a lot, and simplifying is what a model is
supposed to do but it also involves a risk in the sense that the
constraint could be wrong. If you undo the advantage of an ordered
model, by presenting the results in such a way that you need to
present effects for each category in the
explained/dependent/left-hand-side/y-variable, than there is not much
point in incurring the risk without the advantage, and you can just as
well do a multinomial logit/probit.

However, I would not give up so easily. Instead I recommend you use an
ordered logit and interpret the results in terms of odds ratios. You
can find an explanation on how to do that in
<http://www.stata.com/bookstore/regression-models-categorical-dependent-variables/>.

Hope this helps,
Maarten

---------------------------------
Maarten L. Buis
WZB
Reichpietschufer 50
10785 Berlin
Germany

http://www.maartenbuis.nl
---------------------------------
*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/