RE: st: Instrumental Variables approach using -xt- for three-level growth model

[David Torres <writeon4truth2@msn.com>]

Billy,

I am following the 2008 kindergarten cohort over five years and there are about 175 schools at level 3. I'm only planning to focus on students who DO NOT change schools.

Since I have dummies for different school types, doing school fixed effects is not optimal. I want to know the between school effects as well as the within school effects.

Is it possible to use xtivreg with the re (random effects) option for three-level data?  That is, since the data are already xtset with the panel and time variable, can I include an option i(school level id) and produce random effects?

Diego


----------------------------------------
> Subject: Re: st: Instrumental Variables approach using -xt- for three-level growth model
> From: william@williambuchanan.net
> Date: Mon, 6 May 2013 10:15:10 -0700
> To: statalist@hsphsun2.harvard.edu
>
> Hi David,
>
> How many schools are you working with? Are students changing schools during the observational period (e.g., cross-classified)? Depending on the number of schools, it may be easier to use school fixed-effects. Most of the work that I saw at AERA recently (for estimating two-level models of students nested within schools) suggested a minimum of 50 higher level units. I would assume that a similarly large number would be needed to generate reasonable estimate of the level-3 parameters as well.
>
> HTH,
> Billy
>
>
> On May 6, 2013, at 10:09 AM, David Torres <writeon4truth2@msn.com> wrote:
>
> > Thanks again, Austin. I do understand that my instrument cannot be correlated with my outcome except via the predictor for which selection bias might be an issue. Given that I have info on students' home addresses, though, I was hoping to use as instruments factors at the tract or sabans boundary level that might affect the likelihood of choosing but not students' scores. Distance to a school of choice could be an instrument, I would think. But so perhaps could population density (if we assume that individual knowledge is highest the greater the number of people by which one is surrounded). I like Hoxby's choice of streams and rivers, but I anticipate issues given my focus is intra- rather than interdistrict.
> >
> > At any rate, once I find my instrument I would like to know what Stata command is best for a three-level growth model.
> >
> > Diego
> >
> > --
> >
> > David Diego Torres, PhD
> > NSF SBE Postdoctoral Research Fellow
> > Houston Education Research Consortium
> > Rice University
> > 6100 Main Street, MS-28
> > Houston, TX 77005
> > Phone: 713-348-2984
> > Email: ddtorres at rice dot edu
> >
> >
> > ----------------------------------------
> >> Date: Mon, 6 May 2013 12:49:15 -0400
> >> Subject: Re: st: Instrumental Variables approach using -xt- for three-level growth model
> >> From: austinnichols@gmail.com
> >> To: statalist@hsphsun2.harvard.edu
> >>
> >> David Torres <writeon4truth2@msn.com>:
> >> On Hoxby, see
> >> http://gsppi.berkeley.edu/faculty/jrothstein/hoxby/documentation-for-hoxby-comment
> >> and note that school choice is affected by parental income, education,
> >> and employment,
> >> but so are test scores, so these are not valid instruments.
> >>
> >> On Mon, May 6, 2013 at 12:39 PM, David Torres <writeon4truth2@msn.com> wrote:
> >>> Hi, Billy,
> >>>
> >>> I'd very much appreciate the full citations of the articles/papers you mentioned.
> >>>
> >>> Thanks,
> >>> Diego
> >>>
> >>> --
> >>>
> >>> David Diego Torres, PhD
> >>> NSF SBE Postdoctoral Research Fellow
> >>> Houston Education Research Consortium
> >>> Rice University
> >>> 6100 Main Street, MS-28
> >>> Houston, TX 77005
> >>> Phone: 713-348-2984
> >>> Email: ddtorres at rice dot edu
> >>>
> >>>
> >>> ----------------------------------------
> >>>> Subject: Re: st: Instrumental Variables approach using -xt- for three-level growth model
> >>>> From: william@williambuchanan.net
> >>>> Date: Mon, 6 May 2013 07:52:55 -0700
> >>>> To: statalist@hsphsun2.harvard.edu
> >>>>
> >>>> Hi David,
> >>>>
> >>>> I don't have access to the full citations at the moment, but there is a paper from Caroline Hoxby that used natural boundaries (e.g., rivers) as a school choice instrument. More recently, Justine Hastings along with Tom Kane and some other folks have done a fair amount of modeling of school choice in the Charlotte-Mecklinburg school district. I know both of the papers I am thinking of are available via NBER and can send full citations later if you'd like. Do you have good parental education/income data? One of the things that Hastings found was that school choice decisions were a function of parental income, education, and employment (e.g., parents with higher incomes tended to choose schools that were higher performing regardless of location while parents with lower incomes tended to choose schools that were closer to the home). If nothing else, those papers could give you a good idea of how the issue has been approached in the past.
> >>>>
> >>>> HTH,
> >>>> Billy
> >>>>
> >>>> Sent from my iPhone
> >>>>
> >>>> On May 6, 2013, at 7:33, David Torres <writeon4truth2@msn.com> wrote:
> >>>>
> >>>>> Thanks for you response, Austin.
> >>>>>
> >>>>> To be clear, though, I am using data from a large school district and am using as my outcome scores on the Stanford 10 math and reading assessments, so no plausible values. My explanatory variable of interest is whether a child's parents exercised choice to send the child to a school other than the zoned school (0 = no, 1 = yes). It's this explanatory variable that I'm focusing on as one that needs to be instrumented, given the differential likelihood of exercising choice deriving from parents of various education levels, density of neighborhood, etc. I'm only using the 2008 kindergarten cohort and am following these children across five years. The boundaries remain constant across the years, as far as I can tell.
> >>>>>
> >>>>> As an aside, I do have addresses for all families in the district and will be geocoding these soon. I then plan to use the -vincenty- command to calculate the distance to the nearest school (which will also be geocoded) to which a student is not zoned. This will be included as an additional control.
> >>>>>
> >>>>> --
> >>>>>
> >>>>> David Diego Torres, PhD
> >>>>> NSF SBE Postdoctoral Research Fellow
> >>>>> Houston Education Research Consortium
> >>>>> Rice University
> >>>>> 6100 Main Street, MS-28
> >>>>> Houston, TX 77005
> >>>>> Phone: 713-348-2984
> >>>>> Email: ddtorres at rice dot edu
> >>>>>
> >>>>>
> >>>>> ----------------------------------------
> >>>>>> Date: Mon, 6 May 2013 10:18:40 -0400
> >>>>>> Subject: Re: st: Instrumental Variables approach using -xt- for three-level growth model
> >>>>>> From: austinnichols@gmail.com
> >>>>>> To: statalist@hsphsun2.harvard.edu
> >>>>>>
> >>>>>> David Torres <writeon4truth2@msn.com>
> >>>>>> You don't specify, but I assume you have 5 plausible values drawn from
> >>>>>> a posterior distribution on achievement from some standardized test as
> >>>>>> your outcome variable, and school type attended as your explanatory
> >>>>>> variable of interest--none of the variables you cite satisfy an
> >>>>>> exclusion restriction, as all would be expected to have direct impacts
> >>>>>> on achievement. If you do in fact have 5 plausible values as your
> >>>>>> outcome measure, you should not be using -xtmixed- IMHO but perhaps
> >>>>>> -mi- since these are 5 imputations for one observation, not 5
> >>>>>> observations. If you have 5 plausible values at two points in time,
> >>>>>> you now have 25 imputations, and so on. You should get similar results
> >>>>>> if you just average across the 5 imputations, but no identical of
> >>>>>> course. The more important challenge is finding good instruments for
> >>>>>> school type attended. Do you know where people live relative to the
> >>>>>> boundary of the "zoned school area" so you can compute that distance
> >>>>>> precisely? Do boundaries change over time?
> >>>>>>
> >>>>>> On Sun, May 5, 2013 at 4:57 PM, David Torres <writeon4truth2@msn.com> wrote:
> >>>>>>> Hey, all,
> >>>>>>>
> >>>>>>> I would like to utilize the instrumental variables approach with hierarchical data that has repeated measures (five) of students nested within schools. I'd prefer to do this in HLM7, but I'm not sure if the program allows for this more complex model setup. Anyway, I know Stata has the -xtivreg- command. Is there something similar for mixed effects models, i.e., -xtivmixed- perhaps? If not, how can I accomplish my aim using what's available in Stata. (I do have the Rabe-Hesketh & Skrondal texts.)
> >>>>>>>
> >>>>>>> I am instrumenting for the likelihood of parents exercising choice to send their children to a school outside the zoned school area. Obviously, self-selection is an issue here, but proxies that might explain the likelihood of choice are, and this list is by no means exhaustive, number of older children in the household, maternal education, or percent of persons over 18 with a bachelor's degree...or something like that. HELP.
> >>>>>>
> >>
> >> *
> >> * 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/
> > *
> > * 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/
>
>
> *
> * 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/ 		 	   		  
*
*   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/