xtbhst 1.1.0

This release corrects the computation of the test statistic and several related defects; the version on CRAN is 1.0.2.

• Unit-specific error variance. The previous versions divided the pooled fixed-effects residual sum of squares of each unit by T minus the number of partialled-out variables, which is neither the estimator of Blomquist and Westerlund (2016) nor that of Pesaran and Yamagata (2008). A new argument variance = c("bw", "py") selects the estimator: "bw" (default) is T^-1 times the residual sum of squares of the unit-specific least-squares regression (Blomquist and Westerlund, 2016, Section 3.1); "py" is the pooled fixed-effects residual sum of squares divided by T - K - 1 (Pesaran and Yamagata, 2008). The chosen estimator is used in the weights of the weighted fixed-effects estimator, in S and in every bootstrap statistic.
• Adjusted Delta. The variance in the denominator was 2K(T - K - Kpartial - 1)/(T - Kpartial + 1), where Kpartial counted the constant; it is now 2K(T - K - 1)/(T + 1) as in Pesaran and Yamagata (2008). The silent fallback to 2K when this was non-positive has been removed; delta_adj is NA with a warning when T - K - 1 <= 0.
• p-values. The bootstrap p-value pval is now based on the statistic S that the paper bootstraps (proportion of S* at least as large as S). Delta and adjusted Delta are fixed rescalings of S, so their bootstrap p-values coincided with each other; the duplicated pval_adj and delta_adj_stars elements have been removed and pval_delta_asy and pval_delta_adj_asy (asymptotic standard normal p-values) are reported instead. Delta and adjusted Delta are always computed with the Pesaran and Yamagata (2008) variance estimator (from the new element S_py), whatever the value of variance, so that they are the statistics of that paper; variance governs only S and the bootstrap. The conclusion in summary() uses the bootstrap p-value only. New elements S, S_py, S_stars, sigma2, variance and n_dropped.
• Lagged cross-sectional averages (csa_lags > 0). Rows with missing lagged averages were previously set to zero (including the constant) without reducing T, which distorted the unit estimates. The first csa_lags periods are now dropped for all units, T is reduced accordingly, and when the dependent variable enters csa its cross-sectional averages are recomputed from the pseudo-data in each bootstrap replication. The documentation states that CSA augmentation is an extension not covered by the paper.
• Block length. The default is now round(2 * T^(1/3)), which gives the block lengths 6, 7 and 9 for T = 25, 50 and 100 stated in Section 4.2 of the paper (previously floor(), giving 5, 7 and 9). A blocklength larger than T is now an error.
• Missing values. Rows with missing values in any variable used are dropped before the balance check, and the check now also detects duplicated (id, time) pairs; previously missing values in the model variables caused a “subscript out of bounds” error.
• Singular unit regressions now stop with an error naming the unit instead of silently applying a ridge correction.
• Citation corrected everywhere to Blomquist and Westerlund (2016), Empirical Economics, 50(4), 1359-1381; the reference to Pesaran and Yamagata (2008) has been added.
• plot(): plot 1 now shows the bootstrap distribution of S and plot 2 that of Delta.
• Removed the unused suggested dependency on ‘plm’.

xtbhst 1.0.2

• Removed Jan Ditzen from the contributors list at his explicit request.
• The acknowledgment of the original Stata xthst command now appears as a narrative reference in README.md only.

xtbhst 1.0.1

• Documentation and metadata updates.

xtbhst 1.0.0

• Initial CRAN release
• Implements bootstrap slope heterogeneity test of Blomquist and Westerlund (2016) [citation corrected in 1.1.0]
• Features:
  • Block bootstrap for panel data
  • Standard and adjusted Delta statistics
  • Support for partialling out control variables
  • Cross-sectional averages for handling cross-sectional dependence
  • Diagnostic plots for bootstrap distributions