Changes made in response to peer review of the package description paper.
anova_engine(), non_parametric = "auto",
independent groups: when the normality test rejects (p < .05),
Levene’s test (Brown-Forsythe) p <= .10 now routes to Welch’s ANOVA,
and Kruskal-Wallis is used only when Levene p > .10. In 1.2.0, Levene
p between .01 and .10 still led to Kruskal-Wallis. Messages now state
that a non-significant Levene test is insufficient evidence of unequal
variances, not confirmation of equal variances or of genuine
non-normality. When Welch’s ANOVA is chosen with Levene p between .01
and .10, the max/min group variance ratio is reported.anova_engine(), non_parametric = "auto",
paired designs: fixed a 1.2.0 regression where a Levene-based switch to
Welch’s ANOVA could lead to a paired t-test being run while the message
reported Welch’s ANOVA. Paired designs with a normality rejection now
use the Wilcoxon signed-rank test, as before 1.2.0.ols_engine(),
anova_engine()): the Kolmogorov-Smirnov test on
standardized residuals, whose standard p-value is invalid when the mean
and variance are estimated from the data, is replaced by a
Lilliefors-corrected Kolmogorov-Smirnov test with the Dallal-Wilkinson
(1986) p-value approximation, implemented in base R. The method is
reported as “Lilliefors”. For N > 5000 a note reports the residual
skewness and excess kurtosis, because normality tests reject trivial
departures at that size.did_engine(): the pre-treatment comparison of treated
and control mean outcomes is now described as descriptive. A significant
difference is reported as a NOTE (no longer a WARNING), stating that DiD
differences out time-invariant level gaps and that assessing parallel
trends requires multiple pre-treatment periods. The recommendation to
include covariates or change the identification strategy was removed.
Citations are now Roth
(2022) and Rambachan & Roth (2023). The t-test itself is
unchanged.iv_engine(): the F < 10 rule of thumb is now
attributed to Staiger & Stock (1997); Stock & Yogo (2005) are
cited for critical values. Messages state that the first-stage F is the
conventional statistic assuming homoskedastic errors and that F > 10
does not rule out weak identification. With more than one endogenous
regressor, a note states that the minimum of per-regressor first-stage F
statistics is not a valid joint weak-identification test. The Sargan
messages state the homoskedasticity assumption and no longer describe a
non-rejection as the restrictions being “satisfied”.ols_engine()
(previously “Homoskedasticity confirmed” / “assumption is met”), the
normality messages in anova_engine() (previously “Normality
assumption met” / “is met”), the Hausman message in
panel_engine() (previously “supports random effects”), and
the paper_engine() fallback message (previously “The model
meets all diagnosed assumptions”).paper_engine(), README, vignette) now
states that the IV engine treats every right-hand-side variable as
endogenous and does not support included exogenous controls, and
describes the updated ANOVA routing and DiD pre-treatment
comparison.iv_engine(): fixed an O(n^2) memory/time bug caused by
explicitly forming the full n x n projection matrix when computing the
second-stage variance. The offending step now computes
X_tilde via associativity instead, giving O(n*k) memory.
Previously this failed with “cannot allocate vector” errors around n =
100,000; it now runs in a fraction of a second at that size.anova_engine(): fixed a hard crash for n > 5,000,
where shapiro.test() errors out (“sample size must be
between 3 and 5000”). It now falls back to a Kolmogorov-Smirnov test
above that threshold, matching the behavior ols_engine()
already had.panel_engine(): fixed the internal Hausman test
statistic, which did not match plm::phtest(). The between
estimator used to construct the GLS transformation parameter
(theta) was missing its intercept and was not isolating the
idiosyncratic variance component, and the covariance-matrix
positive-definiteness safeguard did not match the reference algorithm
used by plm::phtest() (which inverts directly and takes the
absolute value of the resulting statistic, without a
positive-definiteness check). In some scenarios this could select the
theoretically inconsistent estimator (Random Effects when Fixed Effects
was actually warranted). See the “Statistical correctness note” in this
release’s cran-comments.md for scope and mitigation.model = "panel" in
paper_engine()): fixed effects and random effects
estimators with automatic Hausman-test-based selection
(method = "auto", "fe", or "re"),
plus balanced/unbalanced panel diagnostics.model = "iv"):
two-stage least squares (2SLS) in pure base R, with first-stage
weak-instrument diagnostics.model = "did"): treatment x post-period interaction model
with parallel-trends diagnostics.plot_engine() gains matching plot types for the three
new models: a forest plot for panel estimates, a forest plot for 2SLS
coefficients, and a parallel-trends plot for DiD.academic_salaries example dataset, used in
the new panel/IV/DiD vignette sections.vignette_tutorial.Rmd with worked examples for
panel data, IV, and DiD models.paper_engine() and plot_engine()
documentation for the new parameters (entity_id,
time_id, method, instruments,
treatment_var, time_var,
treatment_level, post_level).