Bug: CI not calculated with "R-lean" engine and not working with "R" engine

[kv9898]

Hi fwildclusterboot team!

I'm trying to replicate Eckhouse (2021) with the fwildclusterboot package. I've attached the relevant code and data. However, with the default "R" engine, it fails to generate a result and reports an error "Error in solve.default(crossprod(weights_sq * X)) : system is computationally singular: reciprocal condition number = 1.92859e-17"

However, it does generate a result when I use the "R-lean" engine. Nevertheless, no confidence intervals are calculated. I noted that I could generate confidence intervals using boottest in Stata, so theoretically this should work with R! Also, I'm looking forward to the confidence interval/weighted regression functionalities for heteroskedastic wild boot!

Many thanks!

bugreport.zip

[kv9898]

Also, every time I run heteroskedastic (no cluster) wild bootstrap with the Julia engine, I get an error of "Error in integer(nrow_df) : invalid 'length' argument" - I tried it with different datasets and this error repeated.

[s3alfisc]

Hi @kv9898 - some of this is a bug! Thanks for reporting =)

If you try to run the heteroskedastic bootstrap via the R engine, it should throw an error - the heteroskedastic bootstrap is only implemented via the "R-lean" algo. It is also not supported via the Julia engine, even though WildBootTests.jl can clearly run a heteroskedastic wild bootstrap. Sorry! If I find the time, I will add support for both.

Confidence intervals are not supported via the heteroskedastic bootstrap. If you really need CIs, and your sample is not too large, there is a quick 'hack' that consists of creating singletons clusters (as the CRV1 and HC1 bootstrap collapse when all clusters are singletons):

library(fwildclusterboot)
data(voters)
voters$clustid = 1:nrow(voters)

fit = lm(proposition_vote ~ treatment, data = voters)

hc = boottest(fit, param = ~ treatment, B = 999)
hc_hack = boottest(fit, param = ~ treatment, clustid = ~clustid, B = 999, ssc = boot_ssc(cluster.adj = FALSE))

summary(hc)
# boottest.lm(object = fit, param = ~treatment, B = 999)
# 
# Hypothesis: 1*treatment = 0
# Observations: 300
# Bootstr. Type: rademacher
# Clustering: 0-way
# Confidence Sets: 95%
# 
# term estimate statistic p.value conf.low conf.high
# 1 1*treatment = 0    0.088     3.144       0       NA        NA

summary(hc_hack)
# boottest.lm(object = fit, param = ~treatment, B = 999, clustid = ~clustid, 
#             ssc = boot_ssc(cluster.adj = FALSE))
# 
# Hypothesis: 1*treatment = 0
# Observations: 300
# Bootstr. Type: rademacher
# Clustering: 1-way
# Confidence Sets: 95%
# Number of Clusters: 300
# 
# term estimate statistic p.value conf.low conf.high
# 1 1*treatment = 0    0.088     3.144   0.002    0.032     0.143

As you can see, both non-bootstrapped t-statistics are identical. P-values differ slightly as different random number generators are used. Both the second option produces CIs! Note that you need to set the cluster.adj = FALSE argument to stop boottest() from setting small sample adjustments for clustering.

I hope this helps!

[kv9898]

Thank you so much!!!