Re: st: From: David Jose <davidjosework@gmail.com>

[Richard Goldstein <richgold@ix.netcom.com>]

Hi David,

a while ago I send approximately this question to StataCorp and Isabel
Cannette drafted some code (not tested) for me; my project then changed
and I did not test her code or revise it for my circumstances; I am
sending this with her permission; I note that her code is based on A.
Colin Cameron, Jonah B. Gelbach, and Douglas L. Miller (2008),
"BOOTSTRAP-BASED IMPROVEMENTS FOR INFERENCE WITH CLUSTERED ERRORS", The
Review of Economics and Statistics, 90(3): 414-427 (and I had pointed
her specifically to appendix B in that article)

good luck,
Rich

Here is the code: (watch for wrapped lines; you can identify such a line
if the first character on the left is not a ">"; if any, join that line
to the end of the preceding line)

>   *Let's assume that we have the regression
>
>   cscript
>   sysuse auto
>   drop if rep == .
>   regress mpg disp turn, cluster(rep)
>
>   *and the null hypothesis H0: _b[displacement] = 0.1
>
>   ***********
>   * Step 1: *
>   ***********
>   *Step 1-1) Compute w (if my t_0 was 0, it would be the t from the
table).
>
>   regress mpg disp turn, cluster(rep)
>   local w = (_b[displacement] - .1)/ _se[displacement]
>   di "`w'"
>
>   *Step 1-2) fit the restrained constraint
>
>    constraint 1 displacement = .1
>    cnsreg mpg disp turn, constraint(1)
>    mat bR = e(b)
>    predict uR , residuals
>    predict xbR, xb
>   *I have one long vector of residuals uR
>   *I also have the restricted linear prediction xbR
>
>
>
>   ************
>   * Step 2a: *
>   ************
>   *Form a sample of clusters
>
>   *Step 2a-1: generate the weights a_g, which are unique to each
>   *cluster
>
>   sort rep
>   gen ag = cond(runiform()<.5, -1, 1)
>   by rep: replace ag = ag[1]
>   list rep ag, sepby(rep)
>
>   *setp 2a-2: generate uRstar
>   gen uRstar = uR*ag
>
>   *step 2a-3: generate yhatstar
>   gen yhatstar = xbR + uRstar
>
>   ***********
>   * Step 2b: *
>   ***********
>   *compute wstar, the same way as in step 1, but using yhatstar
>   *instead of y (mpg).
>
>   regress  yhatstar disp turn, cluster(rep)
>   local wstar = (_b[displacement] - .1)/ _se[displacement]
>
>
>   ***********
>   * Step 3  *
>   ***********
>
>   *I need to repeat the previous process B times (you can check
>   *if there is a recommended number of replications in the article)
>   *I will use B = 100
>
>   program getwstar, rclass
>     *I am using external variables rep, uR, xbR, disp and turn
>     tempvar ag uRstar yhatstar
>     gen `ag' = cond(runiform()<.5, -1, 1)
>     bysort rep: replace `ag' = `ag'[1]
>
>     gen `uRstar' = uR*`ag'
>     gen `yhatstar' = xbR + `uRstar'
>     regress  `yhatstar' disp turn, cluster(rep)
>     return scalar wstar = (_b[displacement] - .1)/ _se[displacement]
>   end
>
>   getwstar
>   return list
>
>   *now I need to get the 100 values of wstar, and compare w with
>   *quantiles of wstar.
>
>   simulate wstar = r(wstar), reps(100) saving(wstar_file, replace):
getwstar
>   use wstar_file, clear
>
>   _pctile wstar, p(2.5, 97.5)
>   return list
>
>   display `w'
>
>   if (`w'<r(r1))| `w'> r(r2){
>      local mytest "Reject H0"
>   }
>   else{
>      local mytest "Don't reject H0"
>   }
>
>   display "******** `mytest' **********"


On 3/28/13 8:19 AM, owner-statalist@hsphsun2.harvard.edu wrote:
> Hi all,
> 
> I'd like to implement a stata routine that computes SEs based on  a
> "wild cluster bootsrap" procedure (Cameron, Gelbach, Miller (ReStat
> 2008)).
> 
> Any ideas?
> 
> I have found a procedure here http://www.ifs.org.uk/publications/6231,
> but it does not compute SEs, but rather only p values.
> 
> Thanks in advance for any 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/