vartype = "simu" is accepted for the kernel estimator but crashes

[soodoku]

interflex() accepts vartype = "simu" for every estimator (R/interflex.R:273), but R/kernel.R only implements the "bootstrap" (line 2034) and "delta" (line 2410) branches. With estimator = "kernel" and vartype = "simu", neither branch runs, the output objects are never built, and assembly fails with an internal error that gives the caller nothing to act on.

Reproduction

library(interflex)
set.seed(5); n <- 400
X <- rnorm(n); D <- rbinom(n, 1, 0.5)
Y <- 1 + 2*D + 0.5*X + 1.5*D*X + rnorm(n)
dat <- data.frame(Y, D, X)

interflex(estimator = "kernel", Y = "Y", D = "D", X = "X", data = dat,
          Xunif = FALSE, CI = TRUE, vartype = "simu", bw = 1,
          parallel = FALSE, figure = FALSE)

On master:

kernel vartype=delta      ok
kernel vartype=bootstrap  ok
kernel vartype=simu       Error: object 'diff.output.all.list' not found
linear vartype=simu       ok

Unconditional — independent of nsimu, diff.values, bandwidth and sample size.

Worth noting the default is safe: the signature sets vartype = "delta" (line 20), so ordinary calls are unaffected. Two ways in: passing vartype = "simu" explicitly, or passing vartype = NULL, which line 271 turns into "simu".

Change

Reject the combination during validation, where the other estimator/option constraints already live (e.g. the pcse-requires-linear check a few lines below):

if (estimator == "kernel" & vartype == "simu") {
    stop("\"vartype = 'simu'\" is not supported for estimator = \"kernel\"; use \"delta\" or \"bootstrap\".")
}

I have deliberately not tried to implement a simulation branch for the kernel estimator — that is a feature decision, and if you would rather have it than an error, this patch is the wrong change.

Verification

• kernel + simu now stops with an actionable message; kernel + delta/bootstrap and linear + simu are unaffected.
• Full suite via cd tests && NOT_CRAN=true Rscript testthat.R: FAIL 1 | SKIP 4 | PASS 344 both before and after. The single failure is test-dml.R:19 ("ml_method = "randomforest" requires the 'ranger' package"), which is a missing optional dependency in my environment, pre-existing on master and unrelated.

Separate observation, not addressed here

vcov.type never reaches the kernel estimator — the string does not occur in R/kernel.R, and interflex() forwards it only to the linear (line 1064) and binning (line 1128) estimators. With the bandwidth held fixed so nothing else varies, kernel standard errors are bit-identical under vcov.type = "robust" and vcov.type = "cluster":

KERNEL fixed bw, robust : 0.20552 0.15232 0.13660 0.15257 0.20801
KERNEL fixed bw, cluster: 0.20552 0.15232 0.13660 0.15257 0.20801   (identical)

On the same cluster-assigned data the linear estimator's clustered SEs are about 4.5x its robust ones, so the choice is not immaterial there.

I am raising this rather than patching it because ?interflex does scope vcov.type to "linear models", so ignoring it for the kernel estimator may well be intended. The part that seems worth a second look either way is that interflex() still requires cl when vcov.type == "cluster" (line 301) even for the kernel estimator, so a user is asked for a clustering variable that is then unused, with no warning. Happy to send a small warning patch if that would be useful.

[soodoku]

Correcting my own closing observation above, having dug further — I was partly wrong and would rather say so than leave it standing.

I wrote that a user is "asked for a clustering variable that is then unused". That is only true under the default vartype = "delta". Under vartype = "bootstrap" the kernel estimator does use cl, and clustered inference works correctly:

                                      SEs at the 10/30/50/70/90th pct of X
delta,     vcov.type = "robust"     : 0.2055 0.1523 0.1366 0.1526 0.2080
delta,     vcov.type = "cluster"    : 0.2055 0.1523 0.1366 0.1526 0.2080
bootstrap, cl = "g"                 : 0.6576 0.6783 0.7007 0.7175 0.7406
bootstrap, no cl                    : 0.2034 0.1525 0.1397 0.1594 0.2147

(cluster-assigned treatment, 40 clusters, bandwidth fixed so nothing else varies)

The block bootstrap lands at ~0.66–0.74, against ~0.71 for the linear estimator's clustered SEs on the same data. So clustered inference for the kernel estimator is available and correct — via vartype = "bootstrap" + cl.

I also traced why vcov.type cannot simply be forwarded: the kernel path's analytic variance is fixed to heteroskedasticity-robust in four separate places — vcov(fe_res, vcov = "hetero") at kernel.R:199 and :271, vcov(iv.reg, type = "H2") at :376, and vcovHC(glm.reg, type = "HC2") at :507. There is no vcov dispatch for the argument to reach, so supporting it would mean threading a variance choice through all four sites and the delta-method assembly. That is a feature request, not a missing argument, and ?interflex already scopes vcov.type to linear models.

So please disregard that paragraph — there is no correctness problem there, and I am not proposing any change for it. The only thing that might still be worth a look is that interflex() requires cl when vcov.type == "cluster" even for estimator = "kernel", where the delta path will not use it; a one-line warning would close that, but it is entirely your call and I have not sent one.

None of this affects the vartype = "simu" fix in this PR, which is unchanged.