Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at statalist.org.
| From | Austin Nichols <austinnichols@gmail.com> |
| To | statalist@hsphsun2.harvard.edu |
| Subject | Re: st: Combining ivregress and heckman |
| Date | Thu, 16 May 2013 15:17:44 -0400 |
Barbara Engels <engels.ba@gmail.com>: I would suggest not using those variables as instruments, as they do not satisfy the exclusion restrictions. You need to find a natural experiment of some kind. On Thu, May 16, 2013 at 2:07 PM, Barbara Engels <engels.ba@gmail.com> wrote: > Well, actually, yes, I am trying to instrument the share of employment time and the share of unemployment time with the year-to-year change (in Stata: d.) of these shares. Which transformation would you suggest - assuming there are no "better" instruments than transformations? > > > On 16.05.2013, at 20:00, Austin Nichols wrote: > >> Barbara Engels <engels.ba@gmail.com>: >> Sounds like no instruments at all. Are you saying you are >> instrumenting for weeks unemployed and weeks employed for those who >> worked at some point during the year with "year-to-year variation of" >> weeks unemployed and weeks employed? Makes no sense to me. >> >> >> On Thu, May 16, 2013 at 1:49 PM, Barbara Engels <engels.ba@gmail.com> wrote: >>> my instruments are year-to-year variation of the endogenous variables, alongside a handful of exogenous RHS variables from the original regression. >>> >>> >>> On 16.05.2013, at 19:44, Austin Nichols wrote: >>> >>>> Barbara Engels <engels.ba@gmail.com>: >>>> So you have a whole list of endogenous RHS variables? What are your instruments? >>>> >>>> On Wed, May 15, 2013 at 8:28 AM, Barbara Engels <engels.ba@gmail.com> wrote: >>>>> it's a standard wage equation. the independent, presumably endogenous variables are employment status shares, i.e. for example the share of months that an individual worked full-time during a year, the share of months that it was unemployed etc. ... . Hence I have observations on x and xe for both observed and missing wages. My selection dummy is "labor force participation" which I define as being =1 if the hours worked are positive and =0 for zero hours worked. I assume that the employment status variables are correlated with the error term since I assume that they are correlated with unobserved individual (fixed) effects. >>>>> >>>>> >>>>> >>>>> On 15.05.2013, at 14:19, Austin Nichols wrote: >>>>> >>>>>> Barbara Engels <engels.ba@gmail.com>: >>>>>> >>>>>> Do you have observations on x and xe for the observed and missing y? >>>>>> E.g. are you worried about selection into the labor market, and your >>>>>> data has workers and nonworkers? Can you specify what the selection >>>>>> process is, and what the nature of the endogeneity in xe is? >>>>>> >>>>>> >>>>>> On Wed, May 15, 2013 at 7:54 AM, Barbara Engels <engels.ba@gmail.com> wrote: >>>>>>> Thanks for this comment. However, I am not sure I understood it correctly. >>>>>>> Say that I want to estimate >>>>>>> >>>>>>> y x xe >>>>>>> >>>>>>> where y is the dependent, x and xe the independent variables, where xe are the independent variables presumed to be endogenous. >>>>>>> >>>>>>> The Heckman two--step is described by >>>>>>> >>>>>>> heckman y x xe (selection= x xe excl) >>>>>>> >>>>>>> where excl are the exclusion restrictions and selection is some binary selection variable. >>>>>>> >>>>>>> the IV would look something like >>>>>>> >>>>>>> ivregress 2sls y x (xe=instrumentals). >>>>>>> >>>>>>> Are you saying that I should use "instrumentals" instead of xe in the selection equation, then predict using predict select_xb, psel (=p_hat) and then just add psel to the instruments in ivregress, so that the latter becomes >>>>>>> >>>>>>> ivregress 2sls y x (xe=instrumentals p_hat)? >>>>>>> >>>>>>> I am not sure this is what you meant. thanks in advance? >>>>>>> >>>>>>> >>>>>>> On 15.05.2013, at 13:33, Austin Nichols wrote: >>>>>>> >>>>>>>> Barbara Engels <engels.ba@gmail.com>: >>>>>>>> There are some bad suggestions in that thread, IMHO. You can use >>>>>>>> -ivregress- or -ivreg2- (SSC) to correct for bias of several >>>>>>>> varieties; excluded instruments used in the selection equation can be >>>>>>>> used in a first stage probit, then predict to make a new generated >>>>>>>> instrument to accompany your other excluded instruments. I.e. use >>>>>>>> predicted probability of y1, call it p_hat, as an instrument for y1 at >>>>>>>> the same time you use z as an instrument for x. >>>>>>>> >>>>>>>> On Wed, May 15, 2013 at 5:33 AM, Barbara Engels <engels.ba@gmail.com> wrote: >>>>>>>>> Dear Statalist users, >>>>>>>>> >>>>>>>>> I want to estimate a model that accounts both for sample selection like Heckman Two-Step, and endogeneity like ivregress. >>>>>>>>> The dependent variable is continuous. The variables to be instrumented appear both in the normal regression and in the selection equation in the Heckman model. >>>>>>>>> I found some hints on Statalist (http://www.stata.com/statalist/archive/2009-08/msg00219.html), but didn't quite understand how to perform the combined "ivheckman" after all. * * 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/