xtpqardl 1.0.5

Breaking changes

• model = "dfe": the standard errors of rho and beta and the cross-quantile blocks of rho_V and beta_V come from one joint covariance matrix, the Koenker and Bassett (1982, Theorem 3.1) form with the scalar sparsity of summary.rq(se = "iid") (new argument dfe_cov, default "iid"). In 1.0.4 the diagonal blocks were the Powell kernel sandwich (se = "ker") and the off-diagonal blocks the Hendricks-Koenker form (se = "nid"). The estimator is stored in dfe_cov and printed by print() and summary(). dfe_cov = "nid" and "ker" use the corresponding summary.rq estimator for every block; the man page documents the Monte Carlo size of each choice.
• wald_test(): the Koenker-Bassett tests in kb_test restrict only the economic slopes (rho, theta, short run) and are computed from the joint matrix of the fit; in 1.0.4 they were quantreg::anova(..., se = "nid") and also restricted the N - 1 panel-intercept contrasts. New argument all = TRUE restricts every non-intercept coefficient; with dfe_cov = "nid" this reproduces the 1.0.4 values when the quantile regression solutions are unique (when they are not, the ordering of the panel contrasts can change the Hendricks-Koenker densities and so the values). kb_test now holds restricted, se and all; the printed labels state which coefficients are restricted.
• Lags of the differenced variables are built by the time index (the value at time - j within the panel), not by row position. Duplicated (id, time) rows are an error. A gap in a panel’s time index is an error unless the new argument allow_gaps = TRUE is given; rows whose lagged period is missing are then dropped. The time variable must be numeric.
• lagsel: the information criteria count the coefficients of all panel regressions (N times the per-panel count); in 1.0.4 only one panel’s coefficients were counted. The man page states that the selection is by least squares.
• model = "pmg": hausman is a data frame with one row per quantile (statistic, df, p-value, number of non-positive eigenvalues, method). The test is NA with a warning when V(MG) - V(PMG) is not positive definite (new argument hausman = "strict", the default); hausman = "ginv" uses a Moore-Penrose inverse with a warning. In 1.0.4 one statistic over all quantiles was computed with the PMG cross-quantile blocks set to zero and negative eigenvalues dropped silently.
• summary() prints the Hausman table; in 1.0.4 the printing code was never reached.
• rq_fits (DFE): 1.0.4 stored rq(y ~ XX) fits with coefficients named (Intercept), XXL_y, …, and panel-contrast levels in character order. 1.0.5 stores rq(y ~ XX - 1) fits whose coefficients are the slopes, then (Intercept), then the panel contrasts, named by the id variable followed by the panel level (for example countryCountry_B), in panel order. Code that indexes coef(fit$rq_fits[[1]]) by name or position must be updated.
• The warning for fewer than three panels (N < 3) is issued for model = "pmg" as well as model = "mg", since both use the same panel code.

Other changes

• Exactly collinear regressors stop with an error naming the columns (and the panel for MG and PMG).
• Formulas with log(), I() and interactions are expanded with model.matrix before the lags are built; the column names of the expanded design are the regressor names.
• Warnings are emitted once when the per-panel covariance estimator differs from summary.rq(se = "iid") (MG, PMG), when a panel is dropped from the PMG pooling, when a DFE covariance block cannot be estimated (the block is NA), and when summary.rq reports non-positive density estimates. panel_cov_status is also returned for DFE.
• model = "mg" with fewer than three panels warns.
• New return elements dfe_coef, dfe_coef_V, dfe_slopes, n_dropped, time_step; summary() prints the number of rows dropped for missing values.
• The rank-one property of the asymptotic covariance of theta (Cho, Kim and Shin 2015, Theorem 1(ii)) and its consequence for joint tests on several theta coefficients are documented in xtpqardl and wald_test.
• The Koenker and Bassett (1982) reference is “Robust Tests for Heteroscedasticity Based on Regression Quantiles”, Econometrica 50(1), 43-61, doi:10.2307/1912528 (1.0.4 cited “Tests of Linear Hypotheses and L1 Estimation”, doi:10.2307/1913398); Hausman (1978) added to the references.
• print.wald_test() no longer writes text arrows.
• Tests: the 1.0.4 test of the quantreg::anova reproduction now uses dfe_cov = "nid" and all = TRUE; the PMG short-panel test accepts the new switch warning. New tests pin the iid joint covariance, the “nid” and “ker” blocks, the restriction count of the joint test, time-aligned lags on a panel with a gap, the collinearity and formula handling, the NA covariance path, the Hausman rule and the lag-selection penalty to hand computations.

xtpqardl 1.0.4

• Bug fix: wald_test(fit, joint = FALSE) failed inside print() with “$ operator is invalid for atomic vectors” whenever the covariance matrix had NA blocks (every DFE and PMG fit in 1.0.3), because the helper returned one flat list instead of one list per variable. Fixed, and print() now guards against malformed elements.
• New: for model = "dfe" the cross-quantile covariance blocks of beta_V and rho_V are estimated with the Koenker and Bassett (1982) formula (all quantiles share the pooled design), so the cross-quantile Wald tests of wald_test() are now available for DFE. wald_test() additionally returns kb_test, the quantreg::anova(..., se = "nid") equality tests (joint and per coefficient) on the pooled quantile-regression coefficients; these reproduce anova(rq_tau1, rq_tau2, ...) on the equivalent regression d_y ~ L_y + L_x + d_x + factor(id). The rq fits are kept in rq_fits.
• PMG with short panels: the per-panel covariance used for the inverse-variance weights comes from summary.rq, which fails when the degrees of freedom per panel are very few (the study data had n - k = 2); in 1.0.3 every panel was then silently dropped and PMG stopped with “needs at least two panels with a usable long-run covariance”. The function now tries the other summary.rq estimators and an iid formula with a kernel sparsity estimate before giving up, reports the status and reason for every panel and quantile (panel_cov_status, also for MG), explains the degrees-of-freedom problem in the error message, and offers pmg_weights = "equal" (unweighted pooling).
• wald_test(): the singularity check det(RVR') < 1e-15 declared the test “Singular variance matrix” whenever the dependent variable had a small scale (the determinant of a matrix of small variances is tiny even when it is well conditioned); it now uses the reciprocal condition number (rcond < 1e-12).
• Input checks: lr must not contain the dependent variable of formula, nor any of its regressors, and all variables must be numeric; the previous behaviour was “subscript out of bounds” or a silent run with a wrong error correction term.
• Tests added for all of the above.

xtpqardl 1.0.3

• Bug fix: the long-run variables in lr were lagged a second time, so with lr = c("L_y", ...) the error correction term used y(t-2). They now enter exactly as supplied.
• Bug fix: model = "pmg" returned the Mean Group estimator. PMG now pools the long-run coefficients by minimum distance (inverse delta-method covariance weights) and re-estimates each panel with the pooled long run imposed; a Hausman test of long-run homogeneity (MG against PMG) is returned in hausman.
• Bug fix: model = "dfe" returned fixed placeholder variances (0.01 on the diagonal). Its standard errors now come from the Powell kernel sandwich covariance of the pooled quantile regression, transformed by the delta method.
• Bug fix: the half-life was ln(2)/|rho|; it is now the exact ln(0.5)/ln(1 + rho), defined for -1 < rho < 0.
• Cross-quantile covariance blocks that are not estimated (PMG and DFE) are now NA, so wald_test() reports them as unavailable instead of assuming independence.
• The MG and DFE point estimates agree with the Stata command xtpqardl (SSC, v1.0.4) on the same data (MG: rho = -0.5123, beta = 0.5946; DFE: rho = -0.4702, beta = 0.5642).
• Added unit tests.

xtpqardl 1.0.2

• Corrected the DOI of Cho, Kim and Shin (2015) to 10.1016/j.jeconom.2015.05.003 (all occurrences).
• Removed a DOI attached to Bildirici and Kayikci (2022) that could not be verified in CrossRef; the citation text is unchanged.
• Fixed a stray brace in xtpqardl-package.Rd; Authors@R updated.

xtpqardl 1.0.1

New Features

• Initial CRAN release
• Panel Quantile ARDL estimation with PMG, MG, and DFE estimators
• Support for multiple quantiles
• Long-run cointegrating parameters estimation
• ECT speed of adjustment computation
• Half-life of adjustment
• Wald tests for parameter equality across quantiles
• Impulse response function computation
• BIC/AIC automatic lag selection

References

• Based on Pesaran, Shin, and Smith (1999) panel ARDL methodology
• Implements quantile cointegration framework of Cho, Kim, and Shin (2015)