Re: st: Several questions regarding xtprobit and margins command

[Nick Cox <njcoxstata@gmail.com>]

Sounds good. But making -margins- work is a secondary issue. You told
your model that -Outcome- was a factor variable, which is a different
story from it being treated as a measured predictor.

Nick

On Tue, Nov 13, 2012 at 1:14 PM, Tobias Morville
<tobiasmorville@gmail.com> wrote:
> Hi Nick, and again thanks. I think i figured it out.
>
> If i did a i.Outcome, it did work, which shows that you were right.
>
> My other problem (with constant effects) was that the default for
> margins, is a linear prediction. Adding predict(pu0) solved this.
>
> As i understand, predict(pu0) says that there is some random effect
> from all subjects, and then averages that, and compares each subject
> to that random effect. It does not assume away the random effects.. i
> think!
>
> But this solved my questions, and i hope that the above might help
> someone with the same problem at some point.
>
> T
>
> 2012/11/13 Nick Cox <njcoxstata@gmail.com>:
>> My guess is that, although the error message may be confusing you,
>> -Outcome- does not qualify as an acceptable argument for -margins-
>> because, as said, it does not qualify as a factor variable or
>> interaction.
>>
>> Nick
>>
>> On Tue, Nov 13, 2012 at 8:29 AM, Tobias Morville
>> <tobiasmorville@gmail.com> wrote:
>>> Hey Nick, and thanks for the anwser on Q#1, and the grammar correction too :-)
>>>
>>> I run
>>>
>>> -xtprobit stop_dummy Outcome outcome_lag1 seqEarn seqearn_outcome,re-
>>>
>>> Where seqEarn goes from 0 to approx 800 (in a very few obs). Only in
>>> integers. Outcome is the number of eyes the die shows, and
>>> seqearn_outcome is a interaction between seqEarn and Outcome.
>>>
>>> This adds another question.
>>>
>>> As seqEarn (their accumulated earnings they get every game round) is a
>>> positive monotonically increasing function of Outcome(t-z), is there
>>> anything wrong with NOT adding the interactionterm between them?
>>>
>>
>> <snip>
>>
>>>>> ________________________________________
>>>>> From: owner-statalist@hsphsun2.harvard.edu [owner-statalist@hsphsun2.harvard.edu] on behalf of Nick Cox [njcoxstata@gmail.com]
>>>>> Sent: 12 November 2012 18:20
>>>>> To: statalist@hsphsun2.harvard.edu
>>>>> Subject: Re: st: Several questions regarding xtprobit and margins command
>>>>>
>>>>> I'll comment on your problem #1.
>>>>>
>>>>> The help for -margins- starts
>>>>>
>>>>> "margins [marginlist] [if] [in] [weight] [, response_options options]
>>>>>
>>>>>     where marginlist is a list of factor variables or interactions that appear
>>>>>     in the current estimation results."
>>>>>
>>>>> When you give arguments immediately after the command, the crucial part is
>>>>>
>>>>> "margins marginlist ...
>>>>>
>>>>>     where marginlist is a list of factor variables or interactions that appear
>>>>>     in the current estimation results."
>>>>>
>>>>> So, it would help if you gave the exact and complete -xtprobit-
>>>>> command you used. I suspect that the error message will make sense
>>>>> when we see the exact model you fitted.
>>>>>
>>>>> Nick
>>>>>
>>>>> P.S. "dice" is a strange word even to those for whom English is a
>>>>> first language. "dice" is a plural: the singular is "die". One die,
>>>>> two dice.
>>>>>
>>>>> On Mon, Nov 12, 2012 at 2:58 PM, Tobias Morville
>>>>> <tobiasmorville@gmail.com> wrote:
>>>>>
>>>>>> I have a set of questions regarding the margins command, and marginal
>>>>>> effects in general.
>>>>>>
>>>>>> I have a unbalanced paneldataset of 4124 observations, unevenly
>>>>>> distributed on 18 subjects.
>>>>>>
>>>>>> My model is as follows: P(stop) = Outcome outcome_lag1 seqEarn, which
>>>>>> im estimating in a RE probit setting with xtprobit command.
>>>>>>
>>>>>> Outcome: Outcome of a dice in period t. Lies from 1 to 6
>>>>>> Outcome_lag1: Outcome of the dice in period t-1
>>>>>> seqEarn: Accumulated earnings over each game. Drops to 0 if subject
>>>>>> chooses to stop, or the dice shows a one. Starts at zero, and can only
>>>>>> get more positive as people climb the reward ladder.
>>>>>>
>>>>>> All of these regressors are significant.
>>>>>>
>>>>>> Sooo, now the questions begin:
>>>>>>
>>>>>> 1) If i use -margins Outcome- (followed this guide
>>>>>> http://www.stata.com/stata12/margins-plots/), then i get this
>>>>>> errormessage: "'Outcome' not found in list of covariates", and that
>>>>>> actually is the case for all margins commands, and is my number one
>>>>>> headache.
>>>>>>
>>>>>> The only marginscommand that works, is if i use the -margins,
>>>>>> dydx(Outcome outcome_lag1 seqEarn)-, which leads me to my next
>>>>>> problem:
>>>>>>
>>>>>> 2) When i use a -margins, dydx(Outcome outcome_lag1 seqEarn)- my
>>>>>> marginal effects are exactly the same as my regression coefficients?
>>>>>>
>>>>>> If i change the code to -margins, dydx(Outcome outcome_lag1 seqEarn)
>>>>>> atmeans- they're the same again..?
>>>>>> (So APE = MEM?)
>>>>>>
>>>>>> Im really confused about this, and i've read the ealier post
>>>>>> (http://www.stata.com/statalist/archive/2009-11/msg01517.html), which
>>>>>> covers some of the questions i have, but dosen't really anwser them.
>>>>>>
>>>>>> If i use mfx compute, predict(pu0) they change, but they become very
>>>>>> small. And im guess that pu0 means that Im setting the random effects
>>>>>> slope to 9, which is a bad idea for my data, as there is quite alot of
>>>>>> random variability.
>>>>>>
>>>>>> 3) If i choose to ignore the fact that my marginal effects are the
>>>>>> same as my probit regression coefficients, then im in my next pickle.
>>>>>> That my marginal effects are constant. If i plot the predicted
>>>>>> probability of stopping the game, over seqEarn, its constant, which
>>>>>> suits my data very badly. And im afraid that i've misunderstood
>>>>>> something very basic.
>>>>>>
>>>>>> if i try -margins, dydx(seqEarn) at (Outcome=(2 3 4 5 6)) vsquish-
>>>>>> marginal effects are the same over Outcome size...
>>>>>>
>>>>>> .. and it's basically the same.
>>>>>>
>>>>>> What i would idealy like to see, is that predicted probability changes
>>>>>> both over seqEarn and over Outcome and outcome_lag1 in some systematic
>>>>>> way, but right now my newbieness in Stata problemshooting is driving
>>>>>> me up the wall.
>>>>>>
>>>>>> ------
>>>>>>
>>>>>> Background (just for the interested):
>>>>>>
>>>>>> I'm currently working with a dataset of 18 subjects, playing a virtual
>>>>>> dicegame for 25 mins while in a fMRI scanner. The dice is random
>>>>>> (1-6), and if you roll one, you lose whatever you accumulated this
>>>>>> round. It's a balloon kind of thing: How far do people dare to go up
>>>>>> the exponential reward ladder, before banking their earnings.
*
*   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/