visStatistics - Automated Selection and Visualisation of Statistical Hypothesis Tests

Automated test selection, visualised. 'visStatistics' automatically selects and visualises statistical hypothesis tests comparing two vectors, based on their class and distribution. Visual outputs, including box plots, bar charts, regression lines with confidence bands, mosaic plots, residual plots, and Q-Q plots, are annotated with relevant test statistics, assumption checks, and post-hoc analyses where applicable. The algorithmic workflow shifts attention from ad-hoc test selection to visual diagnostic assessment and statistical interpretation. It is particularly suited for server-side R applications, where end users interact solely through a web interface to select data groups and receive a complete visual statistical analysis automatically. The same automation makes it useful in time-constrained contexts such as statistical consulting, where it reduces effort spent on test selection and leaves more room for interpretation. The implemented tests cover the most frequently applied inferential methods in biomedical research (Hayat et al. (2017) <doi:10.1371/journal.pone.0179032>). The test selection algorithm proceeds as follows: Input vectors of class numeric or integer are considered numerical; those of class factor are considered categorical; those of class ordered are considered ordinal. Assumptions of residual normality and homogeneity of variances are considered met if the corresponding test yields a p-value greater than the significance level alpha = 1 - conf.level. (1) When the response is numerical and the predictor is categorical, a test comparing central tendencies is selected. In the default setting (group_test = NULL), residual normality is assessed at every group size using shapiro.test() applied to the standardised residuals of lm(). If normality is not met, wilcox.test() is used when the predictor has two levels and kruskal.test() followed by dunn.test() otherwise. If normality is met, levene.test() assesses variance homogeneity. For two-level predictors, Student's t.test(var.equal = TRUE) is applied if variances are homogeneous and Welch's t.test() otherwise. For predictors with more than two levels, aov() followed by TukeyHSD() is applied if variances are homogeneous, and oneway.test() followed by games.howell() otherwise. Setting group_test to "welch" or "rank" bypasses these assumption tests and fixes the analysis to Welch-type or to rank-based tests, respectively. (2) When both vectors are numerical, lm() is fitted by default (correlation = FALSE). If correlation = TRUE, Spearman rank correlation is performed. (3) When the response is ordinal, it is converted to numeric ranks and the non-parametric path from (1) is followed (Wilcoxon or Kruskal-Wallis). When both variables are ordinal and correlation = TRUE, Kendall's tau_b is used instead. (4) When both vectors are categorical, Cochran's rule (Cochran (1954) <doi:10.2307/3001666>) is applied to test independence either by chisq.test() or fisher.test().