mvardlurt 1.1.0
This release corrects the bootstrap and the interpretation of the
test so that they follow Sam, McNown, Goh and Goh (2025), Studies in
Economics and Econometrics 49(1), 17-33,
doi:10.1080/03796205.2024.2439101. The version on CRAN is 1.0.2.
Breaking changes
level is now the significance level of the two tests
(default 0.05, the level used in the paper’s simulations); in 1.0.2 it
was a confidence level (default 0.95) that was stored but never used,
the decision always being made at 10 percent. Values of 0.5 or more are
rejected with a message pointing to 1 - level.- The lag convention follows equation (1) of the paper: an ARDL(p, q)
model has p - 1 lagged differences of y and q - 1 lagged differences of
x (p, q at least 1).
fixlag = c(p, q) in 1.0.2 corresponds
to fixlag = c(p + 1, q + 1) in 1.1.0 (plus the
contemporaneous difference of x, see below). The paper itself is
inconsistent: Section 4.1 and Table 4 count lagged differences directly
(the 1.0.2 convention), so the paper’s ARDL(0, 2) is
fixlag = c(1, 3) here. maxlag must be between
1 and 12.
reps was renamed to nboot (default 999,
minimum 99); reps still works with a deprecation warning.
pi_coef, pi_se, delta_coef and
delta_se were renamed to b1,
b1_se, beta2 and beta2_se;
fstat_p was removed (see below).
Corrections
- Bootstrap (Section 4.2 of the paper). Version 1.0.2 simulated y* and
x* as independent Gaussian random walks without any estimated
coefficient and without resampling residuals, so the critical values did
not depend on the data (Monte Carlo size of the t test with drift was
0.010 at the nominal 5 percent level). The t and F statistics are now
bootstrapped separately with the respective null imposed: the restricted
regression (without y[t-1] for the t test, without x[t-1] for the F
test) is estimated, its recentred residuals are resampled with
replacement, y* is generated recursively from the restricted estimated
equation with the observed x held fixed, and the unrestricted regression
is re-estimated on (y*, x). The contemporaneous difference of x, which
equation (24) of the paper omits although equation (1) has it, is
included in the restricted regressions and in the bootstrap, following
equation (1). Bootstrap critical values at 10, 5, 2.5 and 1 percent and
bootstrap p-values (
t_pval, f_pval) are
reported; the bootstrap distributions are returned in
t_boot and f_boot.
- Four cases (Section 3.2 of the paper). The labels were wrong in
1.0.2. They are now: Case I, neither test rejects, y is I(1) with no
cointegration; Case II, only the t test rejects, y is I(0); Case III,
only the F test rejects, y is I(2); Case IV, both reject, y is I(1)
through cointegration with x. The order of integration is stated in the
output (
decision$integration). The wording “degenerate
case”, “spurious” and “H0: delta = 0, no cointegration” was removed; the
F test is for H0: beta2 = 0 (equation (12)).
- Test regression (equation (1) of the paper). The contemporaneous
difference of x is now included in the regression, in both restricted
regressions and in the bootstrap (equation (1); see Breaking changes for
the lag convention).
x may be a matrix with several covariates; the F
statistic is then the joint Wald F test on all lagged levels.
beta2, beta2_se and lr_mult have
one element per covariate, and coef() returns
b1, beta2.<name> and
lr_mult.<name>.- The decision uses
level (see Breaking changes). The
significance codes for 10, 5, 2.5 and 1 percent are still reported.
seed now defaults to NULL. When a seed is
supplied the caller’s RNG state is saved and restored on exit, so a call
no longer resets the global random number stream (1.0.2 called
set.seed(12345) unconditionally).fstat_p, a p-value from the standard F distribution
that is invalid here (Section 3.3 of the paper), was removed in favour
of the bootstrap p-value.- Lag selection compares all candidate models on a common sample (t =
maxlag + 1, …, n); 1.0.2 compared AIC or BIC values computed on
different samples. The IC table is indexed by p = 1, …, maxlag and q =
1, …, maxlag.
- The citation of the paper was corrected to volume 49(1), pages
17-33, 2025 (it read 2024, 1-17).
- The data frame of the test regression is returned as
data; the model component is for inspection,
and can be refitted with
lm(formula(object$model), data = object$data).
mvardlurt 1.0.0
- Initial CRAN release.
- Implements the multivariate ARDL unit root test (Sam, McNown, Goh
and Goh, 2025) [citation corrected in 1.1.0].
- Features:
- Automatic lag selection via AIC/BIC
- Bootstrap critical values for correct size
- Three deterministic cases (none, intercept, intercept + trend)
- Four-case decision framework for inference
- Comprehensive print and summary methods
- Diagnostic plots