I am running a linear model with multiple fixed effects. I suspect that there is spatial correlation in my data so I cluster the SEs (46 clusters). But I am worried that the standard errors generated by felm::lfe (and reghdfe in Stata) might be too large. So I want to alternatively also compute wild cluster bootstrap (WCB) based confidence intervals for the main parameters.

This does not seem to be an easy thing to do for panel models with multiple fixed effects. Here's a summary of what I found:

  • Stata: Current procedures to do WCB on stata (boottest, clustse, cgmwildboot) do not work with procedures that implement multiple fixed effects (reghdfe, areg). reghdfe itself does not currently support WCB.
  • R: The felm package that seems to be the best option for linear panel models with multiple FE and clustered SE does not support WCB. In fact, it does not even support predict after the regression.

So, there seem to be two possible solutions:

  1. Workaround: WCB procedures on stata work with one level of FE (for example, boottest). So, converting the reghdfe regression to include dummies and absorbing the one FE with largest set would probably work with boottest. But this is likely to be crazily expensive on memory and processing capacity. If I have access to high-performance servers I could run this code on those servers.
  2. Code own WCB function in R: The second option is to get predictions from felm regression (using this process from a previous answer on Stack Overflow). Then one could write a function to do the wild cluster bootstrap and get the confidence interval for estimated coefficients.

There are probably better solutions out there. What would the community suggest as best practice to get wild cluster bootstrap based CI after a panel regression with multiple fixed effects?

  • $\begingroup$ Hi Peter, Thanks for your response. I can see why this question might be a better fit on CrossValidated. But perhaps people might have tried to code up the wild cluster bootstrap after felm and run into difficulties? That would be useful to know. I will ask this question on CrossValidated as well. Thanks! $\endgroup$
    – Anshuman Tiwari
    May 16, 2020 at 16:42
  • $\begingroup$ As a sidenote: the fixest package comes with a predict method (contrary to lfe). $\endgroup$
    – A.Fischer
    Apr 28, 2021 at 16:20

1 Answer 1


The fwildclusterboot R package implements a wild cluster bootstrap and allows for multiple fixed effects. The package is a port of the boottest package in Stata. fwildclusterboot accepts objects of type felm and fixest where multiple fixed effects are out-projected in the estimation stage. In the bootstrap inference stage, it converts all fixed effects to sets of dummy variables but allows the user to specify one fixed effect to be projected out in the bootstrap - just as in the boottest.ado Stata routine.


# load data set voters included in fwildclusterboot

# estimate the regression model via feols
feols_fit <- feols(proposition_vote ~ treatment + ideology1 + log_income | Q1_immigration , data = voters)
# boottest on an object of type fixest, no fixed effect used for inference
boot_feols <- boottest(feols_fit, clustid = "group_id1", param = "treatment", B = 99999, tol = 1e-08)
# now, project out "Q1_immigration" during bootstrap
boot_feols_fe <- boottest(feols_fit, clustid = "group_id1", param = "treatment", B = 99999, fe = "Q1_immigration", tol = 1e-08)

msummary(list(boot_feols, boot_feols_fe), 
         estimate = "{estimate} ({p.value})", 
         statistic = "[{conf.low}, {conf.high}]", 
         digits = 3, 
         output = "markdown")

#|          |    Model 1     |    Model 2     |
#|treatment | 0.073 (0.001)  | 0.073 (0.001)  |
#|          | [0.033, 0.113] | [0.033, 0.113] |
#|Num.Obs.  |      300       |      300       |
#|R2        |     0.316      |     0.316      |
#|R2 Adj.   |     0.288      |     0.288      |
#|R2 Within |     0.052      |     0.052      |
#|R2 Pseudo |                |                |
#|AIC       |     -84.1      |     -84.1      |
#|BIC       |     -35.9      |     -35.9      |
#|Log.Lik.  |     55.025     |     55.025     |
  • $\begingroup$ Hi @A.Fischer, I did try boottest on stata directly as mentioned in the question. It's nice to have a port to it directly from R so fixest/felm work (and thanks for letting me know about its existence), but the issue with memory will likely come up, given that all other fixed effects are converted to dummies. Thanks, Anshuman $\endgroup$ Apr 28, 2021 at 11:58
  • $\begingroup$ Hi, sorry, I think I read over your comment regarding memory. I believe that the main memory problem with many high-dimensional fixed effects in the algorithm as implemented in fwildclusterboot is due to the inversion of (X'X). With many high-dimensional fixed effects, X should probably be a sparse matrix (which it currently is not), and if that is not sufficiently helpful, solve(crossprod(X)) could also be implemented via the bigstatr package. $\endgroup$
    – A.Fischer
    Apr 28, 2021 at 19:56
  • $\begingroup$ Hi, you're right about the fact that most of the high-dimensional fixed effects are zeros for most observations. I would try to do something but I found another estimation strategy that does not rely on linear regression so I don't quite have the incentive right now. But I think this issue will persist for many use cases. Thanks for your observations! $\endgroup$ Apr 29, 2021 at 9:39

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.