Announcement

Collapse
No announcement yet.
X
  • Filter
  • Time
  • Show
Clear All
new posts

  • dual clustering using cgmwildboot gives confidence interval with 1 p values

    I have a regression of some mutual fund returns on ETF, x_family, ETF#X_family as well as time and benchmark fixed effect. the significance of the ETF#X_family variable is of my interest.
    I was not sure if I have to use both time and benchmark clustering. I can assume that fixed effects capture the time and benchmark effects by demeaning. However, the returns at time t might be correlated across (12) benchmarks (I m confident about this) and even across all observation (48) at time t. for example, observation in benchmark 1 might be correlated with the ones in benchmark 2 (at t). Based on this justification, I applied cgmwildboot as

    Code:
    cgmwildboot return Xfamily ETF ETF#Xfamily benchmark_* date*, cluster(benchmark date) bootcluster(benchmark) seed(999)
    the results are weired. All variables (including the FEs) p-values are 1 but I have some confidence intervals that I can calculate the standard errors (~CI/4).

    I also tried cgmreg and it gives me an error "variance matrix is non-symmetric or highly singular"

    My question is that if I need to time clustering for this problem and if I should, how should I interpret the cgmwildboot 1 pvalues.

    Thank you in advance
    Click image for larger version

Name:	cgmwilboot.png
Views:	1
Size:	155.3 KB
ID:	1405740

    Last edited by Rumi; 09 Aug 2017, 16:16.

  • #2
    I would definitely try my boottest command. You would do something like:

    Code:
    regress return ETF##Xfamily benchmark_* date*
    boottest 1.ETF#1.Xfamily, cluster(benchmark date) bootcluster(benchmark) seed(999)

    Comment


    • #3
      David, Great thanks for your response.
      I created a dummy of x_etf = ETF*Xfamily and used in both stages.

      After I tried the boottest command, I got a p-value for x_etf variable but it gives me an error that the confidence interval could not bounded.
      I wanted to make sure that
      1) I have the right p-value for the coefficient (it is 0.06 and making conclusion is difficult) and 2) whether I can have the p value/ standard errors for other variables like etf and Xfamily.


      Great thanks again.

      Screenshot:

      Click image for larger version

Name:	boottest_results.png
Views:	1
Size:	199.3 KB
ID:	1405762

      Comment


      • #4
        Both programs are indicating to you that there is something wrong with your identification. I think the best thing for you to do is simplify the model, e.g. by dropping dummy sets, and observing how the results behave in order to try to pinpoint the source of trouble.

        Comment


        • #5
          Thanks again for your response, #David Roodman .

          I found out that reghdfe package has a multi-clustering and it reports the standard errors for all variables.

          Code:
           reghdfe  return  Xfamily etf X_etf  , a(date bechmarkid ) cluster(date bechmarkid)

          Comment


          • #6
            ivreg2 does too. The reason to use cgwildboot or boottest would be that the number of clusters is small so that the usual approach to computing p values and constructing confidence intervals is inaccurate. You appear to have 12 clusters, which is rather small.

            Comment

            Working...
            X