P-value mismatch between `etable()` and `wald()` with clustering

[Oravishayrizi]

Hey Laurent,
I'm getting different p-values when using etable() versus wald() with clustered standard errors.
The problem is more severe with fewer clusters.

Example 1: Many clusters (base_did dataset)

data(base_did)

est_did = feols(y ~ x1 + i(period, treat, 5) | id + period, 
                cluster="id", base_did)

# IID case - they match
etable(est_did, keep="4", se="iid", coefstat="pvalue")
# Result: p-value = 0.1068

wald(est_did, c("4"), se="iid")
# Result: p-value = 0.1068

# Clustered case - small difference
etable(est_did, keep="4", cluster="id", coefstat="pvalue")
# Result: p-value = 0.1018

wald(est_did, c("4"), cluster="id")
# Result: p-value = 0.0992

Example 2: Few clusters (iris dataset)

res = feols(Petal.Length ~ Petal.Width | Species, cluster="Species", iris)

# IID case - they match
etable(res, se="iid", coefstat="pvalue")
# Result: p-value = 4.41e-10

wald(res, c("Petal.Width"), se="iid")
# Result: p-value = 4.409e-10

# Clustered case - large difference
etable(res, cluster="Species", coefstat="pvalue")
# Result: p-value = 0.1405

wald(res, c("Petal.Width"))
# Result: p-value = 0.0187

Summary:

• IID cases always match
• Clustered cases diverge, and the problem gets worse with fewer clusters
• In example 2 with only 3 clusters, the difference is dramatic (0.1405 vs 0.0187)

Is this a degrees-of-freedom issue in how clustering adjusts the test?

For the record:

R.Version()$version.string
[1] "R version 4.5.2 (2025-10-31 ucrt)"

packageVersion("fixest")
[1] '0.13.2'