paneltests 1.0.9
Breaking changes
xtkao(): argument convention,
"kao" (default) or "stata". With
"kao" the fixed-effects regression uses all T0
periods; the long-run covariance matrices are built from the first
differences of the response and the regressors, not demeaned (Kao 1999,
eqs. 10-12); the residual variances of the DF and ADF regressions are
divided by N (T0 - 1); and the modified DF statistics use
sqrt(N) T0 (rho - 1). convention = "stata" is
the computation of 1.0.8.xtkao(), xtpedroni(): stop with an error
naming the panel or the setting when a (panel, time) pair is duplicated;
when a numeric or Date time index is not equally spaced
within a panel; when the response has no variation within panels
(xtkao()) or is constant in a panel
(xtpedroni()); when a regressor is constant within panels
or collinear with the others after removing the fixed effects
(xtkao()) or with the deterministic terms of a panel
(xtpedroni()); when kernel_lags is not a
non-negative integer smaller than the number of differenced observations
per panel; when lags is not a positive integer or leaves no
more observations than coefficients in the ADF regression; and when an
ADF regression has zero residual variance. Up to 1.0.8 duplicated and
unequally spaced periods were treated as consecutive,
kernel_lags above the panel length and collinear regressors
stopped with internal errors, negative kernel_lags was
treated as 0, and an exact ADF fit returned an infinite statistic with
p-value 0 (xtpedroni()) or a statistic of order 1e14
(xtkao(), convention = "stata").xtkao(): the returned T is the number of
periods of the DF regression (T0 - 1 with
"kao", T0 - 2 with "stata"); the
returned list includes T0 and convention. The
print method shows both period counts and the convention.
Other changes
xtkao() man page: the two conventions are described;
the agreement with Stata xtcointtest kao is stated for
convention = "stata".xtkao(), xtpedroni() man pages: the input
checks are listed; the xtkao() page states that the
statistics of convention = "stata" change when a
panel-specific constant is added to a variable.- Tests: the five Kao statistics with
convention = "kao"
rebuilt by hand; the earlier hand rebuild and the Stata agreement test
use convention = "stata"; the input checks; invariance of
both functions to rescaling, panel-specific constants, row and panel
order and time labels.
inst/validation/kao_table4.R and
inst/validation/kao_tables_5_6.R: simulations of the design
of Kao (1999, section 5.2) compared with Kao (1999, Tables 4, 5 and 6).
inst/validation/ledger_xtkao.md updated.xtkao() man page: simulated rejection rates under the
null for N = 6, T0 = 17 with one and four regressors.inst/CITATION version updated to 1.0.9.
paneltests 1.0.8
Breaking changes
xtpedroni(): p-values are one-sided in the tail stated
by Pedroni (1999, p. 668): 1 - pnorm(z) for the panel
variance ratio and pnorm(z) for the other six statistics.
Up to 1.0.7 all seven were 1 - pnorm(abs(z)). The new
element tail records the tail of each statistic and the
print method shows it.xtkao(): p-values are pnorm(stat) (left
tail) for all five statistics. Up to 1.0.7 they were
1 - pnorm(abs(stat)).xtpedroni() with one regressor and
trend = TRUE: the standardization moments are Pedroni
(2004, Corollary 1) case 3, mu = (17.86, -10.54, -2.29, -13.65, -2.53),
nu = (101.68, 39.52, 0.66, 50.91, 0.56). Up to 1.0.7 the Pedroni (1999,
Table 2) row for two regressors with trend was used. The internal “none”
case holds Corollary 1 case 1. The warning attached to these cases is
removed.xtpedroni(): Pedroni (1999, Table 2) cell for three
regressors, heterogeneous intercepts, panel t mean is -2.576 (was
-2.567); it enters the panel PP t and panel ADF t for
k = 3, trend = FALSE.xtkao(): the ADF t statistic uses
s_v^2 = RSS / (N T) with T the number of
periods of the DF step, as in Kao (1999, p. 9-10). Up to 1.0.7 the
residual sum of squares was divided by the number of ADF observations
N (T - p). On the Stata reference panel the standardized
ADF t now differs from xtcointtest kao by -0.009 to -0.19
(was +0.02 to +0.14); the other four Kao statistics and all seven
Pedroni statistics are unchanged and agree with Stata within 1e-6.xtpretest(), module HETEROGENEITY: the Swamy statistic
is centred on the 1/s_ii weighted pooled estimator of Swamy
(1970, eq. 6.2); up to 1.0.7 it was centred on the unweighted
fixed-effects estimator. The returned list gains
beta_hat.xtpretest(): rows with missing values in the response
or regressors stop the function with an error giving their number and
positions; up to 1.0.7 the F statistics were NA but printed
as F = 0.0000, p = 1.0000 [accept] with a
recommendation.xtpretest(): regressor columns exactly collinear with
the others are removed with a warning naming them and K is
the rank of the pooled design (returned as dropped); up to
1.0.7 K counted the aliased column and the degrees of
freedom and F statistics were wrong.xtpretest(): fewer than two panels, or a panel with
fewer than K + 2 observations, stop the function with an
error naming the panels; up to 1.0.7 short panels were skipped in
S1 while the degrees of freedom kept N.xtpretest(), modules HSIAO and HETEROGENEITY: an
individual regression that is rank deficient (a regressor constant
within one panel) stops the module with a message naming the panel,
recorded in failed; up to 1.0.7 lm.fit fitted
the reduced-rank design silently (HSIAO) or the panel contributed 0 to
the Swamy statistic.
Other changes
xtpretest(), xtpedroni(),
xtkao(): formula terms are evaluated with
model.frame() and model.matrix(), so
log(x), I(x^2) and interactions work; each
model-matrix column is one regressor.xtpedroni(), xtkao(): missing values are
reported with their count and row positions; a warning is issued when
N < 2 or when the panel has fewer than 10 periods;
xtkao() warns with more than seven regressors.xtpretest(), module ROBUST: the printed output and the
man page state the simulated size of the HC1 Wald slope test under a
true null (33.4 percent for N = 10, T = 20, from 8000 replications; 64.8
percent for N = 25, T = 20; 15.4 percent for N = 10, T = 50; 11.8
percent for N = 25, T = 100, the last three from 500 replications each;
nominal 5 percent) and refer to the classical F test or a bootstrap. The
estimator is unchanged.- Man pages: the labels F1, F2, F3 of
xtpretest() are
mapped to Hsiao (2014) eqs. (2.2.14), (2.2.15), (2.2.16), there named
F3, F1, F4; the CD module is documented as using fixed-effects residuals
with pairwise-complete overlaps; the Stata conventions kept in
xtpedroni() (demeaned DF residuals, weighted s*^2 in the
panel ADF t) and xtkao() (AR(1) residual innovations, T0 -
1 fixed-effects sample) are stated as package choices.
- Header of
xtpretest() prints the installed package
version.
- Printed reference lines of
xtcsdq() and
xtqsh() use “and” instead of “&” between author
names.
- Arrows removed from printed messages and comments; the 1.0.5 NEWS
entry rewritten without the HTML-escaped arrow.
- Tests: Pedroni (2004) case 1, 2, 3 moments and the Table 2 m = 3
cell pinned to the printed values; p-value tails checked against
pnorm(); Kao ADF t rebuilt with lm.fit();
Swamy statistic rebuilt by hand with lm(); Hsiao F2 with a
collinear column rebuilt with lm(); errors for missing
values, N = 1, short panels and rank-deficient panels; formula terms.
The test .xmp_locf fills forward correctly referenced the
retired namespace xtmispanel::: and now uses
paneltests:::. The Stata agreement test for the Kao ADF t
allows the documented gap (0.02 on the tested specification).
inst/validation/: derivation ledgers for
xtpretest(), xtpedroni() and
xtkao().inst/CITATION version updated to 1.0.8.
paneltests 1.0.7
- New functions
xtpedroni() and xtkao(): the
seven Pedroni (1999, 2004) and five Kao (1999) residual-based panel
cointegration statistics, following the formulas of the Stata manual
entry for xtcointtest and its sample conventions (DF and
ADF regressions without constant, demeaned residuals for the
contemporaneous and Bartlett long-run variances, sums over the
differenced sample; Pedroni moments from Pedroni (2004) for one
regressor and Pedroni (1999, Table 2) for two to seven regressors;
one-sided p-values from the absolute value of the standardized
statistic). With the same kernel bandwidth and ADF lag, all seven
Pedroni statistics and four of the five Kao statistics reproduce Stata
19.5 to at least 1e-6 on a six-country panel (four specifications,
bandwidths 0, 1 and 2). The Kao ADF t is larger in absolute value in
Stata by 0.02 to 0.25; its internal convention could not be identified.
Stata’s default Newey-West (1994) per-panel bandwidth is not reproduced;
pass kernel_lags.
xtpretest(), module ROBUST: the interacted design
matrix included the regressors together with the interactions of all N
panels, which are exactly collinear (the N interactions of a regressor
sum to the regressor itself). solve(crossprod(X)) therefore
failed with “system is computationally singular” on every data set, and
because the error was not caught it aborted the remaining modules when
tests = "ALL". The module now uses deviations from the
first panel (N-1 dummies and (N-1)K interactions), so the slope test has
(N-1)K restrictions as in Hsiao’s F2 and the intercept test N-1
restrictions. Aliased columns (for example a regressor constant within a
panel) are dropped and reported in robust$aliased.- Module ROBUST: the HC1 “meat” was computed as
crossprod(X * e^2), which is sum e^4 x x'; it
is now sum e^2 x x'. The inverse uses a QR
decomposition.
- Each module of
xtpretest() runs inside
tryCatch(); a failing module is skipped with a message and
recorded in the new element failed instead of stopping the
function.
- Documentation now states that the ROBUST module is an extension of
the package and not part of Hsiao (2014), who only notes that the
F-tests assume i.i.d. errors (3rd ed., p. 24).
- New tests for the ROBUST module (clean panel, agreement with a
direct sandwich computation, aliased regressor, no abort of other
modules).
paneltests 1.0.6
- Corrected the DOIs of Galvao, Juhl, Montes-Rojas and Olmo (2017)
(10.1093/jjfinec/nbx016), Little (1988) (10.1080/01621459.1988.10478722)
and Swamy (1970) (10.2307/1913012).
- Demetrescu, Hosseinkouchack and Rodrigues (2023): title corrected to
the published working-paper title and DOI corrected to 10.4419/96973210
(Ruhr Economic Papers 1041); the journal DOI previously given in
DESCRIPTION pointed to an unrelated article and was removed.
- No changes to code.
paneltests 1.0.5
- Resubmission addressing CRAN reviewer (Benjamin Altmann) feedback on
1.0.4:
- Added DOI / ISBN links to all references in the Description field
(Hsiao 2014, Swamy 1970, Pesaran 2004).
- Removed the surrounding single quotes around function names in the
Description text and added trailing parentheses,
e.g.
'xtpretest' became xtpretest().
- Synchronised CITATION version with DESCRIPTION (1.0.5).
- No code or behavior changes.
paneltests 1.0.4
- Resubmission to CRAN.
- Removed UTF-8 BOM (Byte Order Mark) from
inst/CITATION
and from R/xtcsdq.R, R/xtmispanel.R,
R/xtqsh.R. The BOM in inst/CITATIONwas causing
the CRAN incoming auto-check to fail with:
Invalid citation information in 'inst/CITATION': 1:1: unexpected input.
- Synchronised the version string in
inst/CITATION with
DESCRIPTION.
- No code or behavior changes.
paneltests 1.0.3
- Resubmission. Removed redundant “a suite of tools for” from the
Description start. All references use
doi:10.xxx
format.
paneltests 1.0.0
- Initial CRAN release.
- Consolidates four previously separate packages (xtpretest,
xtmispanel, xtcsdq, xtqsh) into a single panel-data pre-testing
suite.
- Implements:
- Hsiao (2014) homogeneity F-tests (F1/F2/F3),
- Swamy (1970) parameter heterogeneity test,
- Pesaran (2004) cross-sectional dependence test
(
xtpretest),
- Missing-data detection, mechanism testing, and imputation for
unbalanced panels (
xtmispanel),
- Demetrescu, Hosseinkouchack and Rodrigues (2023) quantile-regression
cross-sectional dependence tests (
xtcsdq),
- Galvao, Juhl, Montes-Rojas and Olmo (2017) panel quantile-regression
slope homogeneity statistics (
xtqsh).