Re: st: ordered logistic regression with endogenous variable

[Austin Nichols <austinnichols@gmail.com>]

Justina Fischer <JAVFischer@gmx.de>:
Typically, using predicted values in a nonlinear second stage does not
result in a consistent estimator, but including residuals together
with endog variables can.  Some kind of control function approach is
typically feasible--I would want to see some simulation evidence on
small-sample performance before I trusted any conclusions about
consistency though...

On Thu, Oct 11, 2012 at 4:50 PM, Justina Fischer <JAVFischer@gmx.de> wrote:
> Hi Austin,
>
> I was always wondering whether the following two-step-procedure would be a feasible solution:
>
> 1) assuming a valid instrument at hand, run the auxiliary (first stage) regression by hand (OLS) and predict values of the instrumented
>
> 2) run the second stage regression with ologit, using the predicted values as independent variable, _but_ bootstrap standard errors ?
>
> Or does the non-lineratity of the model pose a problem when using this approach?
>
> Thx
>
> Justina
> -------- Original-Nachricht --------
>> Datum: Thu, 11 Oct 2012 16:33:14 -0400
>> Von: Austin Nichols <austinnichols@gmail.com>
>> An: statalist@hsphsun2.harvard.edu
>> Betreff: Re: st: ordered logistic regression with endogenous variable
>
>> Anat (Manes) Tchetchik <anatmanes@gmail.com>:
>> You can also recast your ordinal variable as ranging from 0 to 1
>> with outcomes in {0,.25,.5,.75,1} and use a fractional model
>> as described in e.g.
>> "Inference for partial effects in nonlinear panel-data models using Stata"
>> by Jeffrey Wooldridge, linked from
>> http://www.stata.com/meeting/snasug08/abstracts.html
>>
>> On Thu, Oct 11, 2012 at 2:04 PM, Anat (Manes) Tchetchik
>> <anatmanes@gmail.com> wrote:
>> > Thanks Jay,
>> > Actually this is not our main model (rather it is an "auxiliary" one
>> > aiming to validate some relations) our main model is a count one with
>> > IVs.
>> > I'm not sure I understood what did you mean by: problems with the
>> > residuals, I ran the  IVregress and received the following stats.
>> > (with some of the coefficients signif. as expected )
>> > Instrumental variables (2SLS) regression      Number of obs =  603
>> >                                                        Wald chi2(14) =
>> 169.62
>> >                                                        Prob > chi2   =
>> 0.0000
>> >                                                        R-squared     =
>> 0.3208
>> >                                                        Root MSE      =
>> 1.0537
>> > Anat
>> >
>> >
>> > On Thu, Oct 11, 2012 at 7:10 PM, JVerkuilen (Gmail)
>> > <jvverkuilen@gmail.com> wrote:
>> >> On Thu, Oct 11, 2012 at 12:38 PM, Anat (Manes) Tchetchik
>> >> <anatmanes@gmail.com> wrote:
>> >>> Hi Jay, It is a 5 categories var. however not symmetric (i.e. value 1
>> >>> appears 10%, 2 appears 11%, 3- 22% , 4-27% and 5- 31%) so it doesn't
>> >>> fit into the IV estimator, shell I run gmm?
>> >>
>> >> That's not too bad in terms of skew, but you could have important
>> >> subgroups be skewed, so if for instance males are really positive on
>> >> the measure and females are really negative, the overall measure might
>> >> appear symmetric but not be at the level you want to analyze.
>> >>
>> >> You will get some attenuation of statistical power due to the coarse
>> >> response scale. You can try running an ordinary estimator, but if you
>> >> notice problems with the residuals, I'd switch to -gllamm- for an
>> >> ordinal probit model, or -gmm-. Specifying the model for either is not
>> >> a trivial matter, though, so I totally understand the desire to work
>> >> with a linear estimator!
*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/