Re: st: RE: RE: re: simultaneous equations

 From Austin Nichols <[email protected]> To [email protected] Subject Re: st: RE: RE: re: simultaneous equations Date Fri, 27 Feb 2009 12:01:05 -0500

```Molina et al.--
required to diagnose the problem...  but note that IV is consistent,
not unbiased, and 198 observations is usually not close enough to
infinity to say much in any case.

For example: below the est coef is more than a million times as big as
the true effect in absolute magnitude, and has the wrong sign (these
kinds of examples are easy to come by):

mat c = (1,.5,.01\.5,1,0\.01,0,1)
drawnorm x e z, corr(c) n(10000) seed(1) clear
g y=.5+x/1000+e/10
forv i=1/9 {
g x`i'=x+invnorm(uniform())/1000
}
ivregress 2sls y x? (x=z) in 1/198

On Fri, Feb 27, 2009 at 11:18 AM, Schaffer, Mark E
<[email protected]> wrote:
> An alternative explanation (similar to Nick's point) is that the scaling
> of the variables is crazy.  Stata can get grumpy when your variables are
> wildly different orders of magnitude.
>
> Try rescaling your variables (including the instruments) - i.e.,
> multiply or divide each one by the appropriate power of 10 - so that
> they all have a mean of about the same order of magnitude, and then see
> what happens.
>
> --Mark
>
>> -----Original Message-----
>> From: [email protected]
>> [mailto:[email protected]] On Behalf Of Nick Cox
>> Sent: Friday, February 27, 2009 3:43 PM
>> To: [email protected]
>> Subject: st: RE: re: simultaneous equations
>>
>> With coefficients of the order of 1e8 there seems to be a massive
>> problem with the data. Look (again?) at the data, e.g. using a scatter
>> plot matrix. There could be a substantial outlier, or some built-in
>> dependency, or something else. Whatever it is, it is not obvious that
>> any model can make sense of the predictor set you are using.
>>
>> Nick
>> [email protected]
>>
>> [email protected]
>>
>> As you adviced me I tried ivregress 2sls ... and obtained:
>>
>> ivregress 2sls expshare b2b l1 age l4a (normscore = l11a l10),
>> note: l10 dropped because of collinearity

<big coefs, big SEs>

>> Instrumented:  normscore
>> Instruments:   b2b l1 age l4a l11a
>>
>> I agree with and don't know why the constant term is dropped.
>> I was looking for some causality link between normscore and expshare,
>> and
>> supossed Simultaneous Equations may help me.
>>
>> Kit Baum <[email protected]>

>> I find it very strange, if expshare is really a fraction, that you get
>> coefficients on the order of 10^8 on any regressor. Would you try
>> running
>>
>> ivreg expshare b2b l1 age l4a (normscore = l11a l10), first
>>
>>
>> and show us what you get from that regression? (you could also use
>> ivregress 2sls ... , first if you are using Stata 10)
>>
>>
>> Also, is there a good reason (given definitions of the variables) that
>> the
>> constant term is dropped in your first equation?
>>
>>
>> This system is triangular, in that normscore can be estimated with OLS
>> as
>> it does not contain expshare. You may still want to estimate it with
>> 3SLS,
>> but something fishy is going on here. See what the single- equation
>> estimates of your second equation reveal.

*
*   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/
```