xtdhcoint 1.0.5

Breaking changes

• The first-difference regression of Westerlund (2008, eq. 12) is now run over t = 2, …, T on T - 1 first differences, the principal components are extracted from the (T - 1) x N residual matrix, and the defactored differences are recumulated from zero (e_i1 = 0). The previous code prepended a zero row before differencing, so the first “difference” was the level y_1 - x_1’b. All values of dhg, dhp, dhg_z, dhp_z and the p-values change; the statistics are now invariant to adding a constant to the response or the regressors.
• The number of factors is selected over k = 0, 1, …, kmax instead of k = 1, …, kmax (Westerlund 2008, Remark 5; Bai and Ng 2002, Theorem 2). nf = 0 means that no factor is removed. nfactors = 0 is accepted.
• Unbalanced panels are aligned on the time variable: the common time window (periods at which every unit is observed) is kept and observations outside it are dropped with a message. The previous code kept the first Tmin rows of each unit regardless of calendar time. A common window with internal gaps, or with fewer than 10 periods, is an error. Duplicate time values within a unit are an error.
• The constant argument has been removed. It had no effect: the intercept was dropped in both branches, and the method has no constant to estimate.
• penalty defaults to NULL (meaning 1 for criterion = "ic" and "pc"). Supplying penalty with criterion = "aic" or "bic" is an error; before, criterion = "aic", penalty = 2 or 3 computed BIC3 and criterion = "bic" overrode the supplied penalty. The returned penalty is NA for "aic" and "bic".
• bandwidth = 0 is used as given (contemporaneous variance only); before it was changed to 1. Values other than -1 or an integer between 0 and T - 2 are an error.
• kmax is reduced to min(N, T - 1) - 1 (before: min(N, T) - 1) because the principal components matrix has T - 1 rows.
• nf_at_kmax is TRUE whenever the criterion selects kmax, including kmax = 1.

Other changes

• Missing values in the response, the regressors, id or time stop the function with an error that reports their count and row positions. Before, an interior NA gave “subscript out of bounds”.
• Exactly collinear first-differenced regressors stop the function with an error naming the unit.
• predet accepts a vector of length K; a scalar is applied to every regressor.
• A warning is issued when N < 5 or N > T, quoting the null rejection rates of a Monte Carlo study of this version (nominal 5 percent, one factor fixed): 38 to 41 percent for DHp with N = 2, 23 to 24 percent for DHp with N = 5 and 57 to 60 percent for DHg with N = 25 and T = 20. The warning is issued when N < 10 and the number of factors is selected automatically (Bai and Ng 2002, p. 203). The N < 10 note was a message before.
• The documentation of bandwidth = -1 now states the rule used, floor(4 (T/100)^(2/9)) from Westerlund (2008, Section 3.4).
• New result elements time_window and n_dropped; print shows the common time window when observations were dropped and marks nf = 0.
• Unused imports arima.sim and rnorm removed.
• inst/CITATION updated to version 1.0.5 and to the author name used in DESCRIPTION.
• New tests: hand rebuild of DHg and DHp with lm.fit, eigen and cumsum (tolerance 1e-8), level invariance, k = 0 selection on a no-factor DGP, NA error, unbalanced-panel alignment, argument validation. The 1.0.4 tests that expected a message for N < 10 and an error for nfactors = 0 were changed accordingly.
• Validation ledger added in inst/validation/ledger_xtdhcoint.md.

xtdhcoint 1.0.4

• The message printed when K > 1 regressors are supplied (“Simulated moments are derived for K=1 … consider running separate bivariate tests”) was wrong and has been replaced. Westerlund (2008, Theorem 1 and Remark 3) shows that the limiting distributions of DHg and DHp do not depend on the regressors, so the simulated moments of Remark 2 apply for any K and there is a single set of critical values. The new note only recalls Assumption 1(d): the regressors must not be cointegrated among themselves. No numerical change.
• New argument nfactors fixes the number of common factors and bypasses the Bai and Ng (2002) information criteria.
• With fewer than 10 cross-sectional units the function now notes that the information criteria are unreliable and reports whether the selected number of factors hit the upper bound kmax (new result elements nf_at_kmax and nfactors; shown by print). With N = 6 the IC criterion selected kmax in every configuration tried, and the standardized statistics changed substantially with the number of factors.
• Added testthat tests covering the three points above.

xtdhcoint 1.0.3

• Removed an empty item from the author list of xtdhcoint-package.Rd, which caused an HTML validation NOTE (“trimming empty
• “). No changes to code.

xtdhcoint 1.0.2

• Corrected the DOI of Westerlund (2008) to 10.1002/jae.967 in DESCRIPTION, README, R and Rd files.
• Authors@R updated; a former contributor entry was removed.

xtdhcoint 1.0.0

• Initial CRAN release.
• Main function xtdhcoint() for Durbin-Hausman panel cointegration tests.
• DHg (group-mean) and DHp (panel) test statistics.
• Automatic factor number selection via information criteria (IC, PC, AIC, BIC).
• Long-run variance estimation using Bartlett kernel.
• Cross-sectional dependence handled through common factor extraction.
• Example dataset fisher_panel for testing Fisher effect.