Steve--
My point was about the first stage, not the second. Angrist and Evans
show an effect of samesex on subsequent fertility of .06 but the R
squared is 0.004, meaning that you are throwing away a lot of the
variation in the endog var (for good reason, but that does not change
the effect on the SE), though the first-stage F is over 1000 (so no
weak instr worries). They have sample sizes in the 400 thousand
range, and the IV estimates have SEs 12 to 16 times as large as the
corresponding OLS (e.g. http://www.jstor.org/stable/pdfplus/116844.pdf
table 7).
On Fri, Nov 20, 2009 at 10:35 AM, <[email protected]> wrote:
> I think that Shruti is trying to emulate the analysis in Angrist and
> Evans, 1998, which saw much larger effects of morethan2children.
>
> -Steve
>
> Reference:
> #Children and Their Parents' Labor Supply: Evidence from Exogenous
> Variation in Family Size
> # Joshua D. Angrist and William N. Evans
> # The American Economic Review, Vol. 88, No. 3 (Jun., 1998), pp. 450-477
>
> On Fri, Nov 20, 2009 at 10:08 AM, Austin Nichols
> <[email protected]> wrote:
>> I second Maarten: the large SE reflects the large variance inherent in
>> IV. Note that http://papers.nber.org/papers/w10281 indicates the
>> effect of sex mix on subsequent fertility is about .02 to .04 so you
>> will not be using a lot of the variation in your endog var.
>>
>> However: note two other points--if you have survey data, you should
>> not use [aw= but instead [pw= and you should cluster to get more
>> correct SEs.
>> Also, you have a binary RHS endog var and binary outcome so you may
>> prefer another estimator, e.g. -biprobit- or -cmp- (on SSC).
>>
>> Also, why not consider boyfirst an excluded instrument? Is the worry
>> that some families who observe the sex before birth choose not to have
>> a girl first?
>>
>> On Fri, Nov 20, 2009 at 8:13 AM, Maarten buis <[email protected]> wrote:
>>> --- On Fri, 20/11/09, Shruti Kapoor wrote:
>>>> I am using ivreg for the first time and am not sure if i
>>>> can do anything to improve my results. The biggest problem
>>>> i am facing is that the stnd errors on my endogenous variable
>>>> (morethan2children, even when instrumented) is quite high.
>>>> Which makes them insignificant.
>>>
>>> In general, large standard errors are not a problem, they are
>>> a finding. We may or may not like that finding, but that is
>>> irrelevant.
>>>
>>> Specifically with instrumental variables, I am not surprised
>>> that you find large standard errors. Instrumental variables can
>>> potentially provide you with a very strong argument that the
>>> effect you found is likely to be causal, but there is always a
>>> price to be paid: in the case of instrumental variable the
>>> price is low power (i.e. large standard errors). As the
>>> economists say: there is no such thing as a free lunch.
*
* 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/
