This Computational Provenance Record documents the statistical computing environment, software dependencies, computational provenance, and bibliographic references associated with the RAISINS Completely Randomized Design (CRD) module. It is intended to support computational reproducibility and software transparency. Detailed statistical methodology, mathematical derivations, and user guidance are provided separately in the official module documentation.
Any issues or updates required please comment here
Pooled MSerror = MSE and DFerror = residual df from the one-way model; alpha = α
Grouping display
Letters shown only when the treatment p ≤ α; group labels ordered with gtools::mixedsort(); grouping is suppressed when the number of treatments exceeds 80
Critical value reported
CD/LSD (LSD), HSD (Tukey), or Duncan critical ranges (DMRT). The trait-wise Individual ANOVA table reports CD at both 5% and 1%
A single CD/critical difference is not defined once comparisons are adjusted, so the CD row is omitted; the adjusted result is carried by the grouping letters
Standard errors
SEM
sqrt(MSE / r) for balanced replication. With unequal replication, the main summary reports “-”; Individual ANOVA provides sqrt(MSE/n_i) and the count for each treatment.
SED
sqrt(2 × MSE / r) (balanced replication only)
Data transformations
(optional, per variable)
Logarithmic
Base-10 logarithm with ordered branches for nonpositive values, positive values below 1, or other positive values; see the transformation table below.
Square-root
sqrt(x); if any value = 0, sqrt(x + 0.5); blocked on negative values
Arcsine
Explicit proportion or percentage units per response; percentage conversion, endpoint adjustment and radians as described below.
MANOVA
(multi-response mode)
Test statistic
Pillai's trace (the stats::manova() default), tidied with broom::tidy()
Requirement
At least two finite numeric responses; residual df must be at least the number of responses, with nonzero residual variation and full residual column rank.
5 Transformation Rules
Apply the selected transformation to the entire response variable across all treatments and replications. Choose only one transformation per response. Rerun the full analysis and recheck assumptions after any change. The uploaded file remains unchanged; table notes identify the analysis scale and formula used.
Transformation
Rule
Settings and validation
Log: any value ≤ 0
log10(x - min(x) + 1) for the whole response
Uses the minimum over all treatments and replications. This branch has priority over the branches below.
Log: all values > 0, with at least one value < 1
log10(x + c) for the whole response, including observations ≥ 1
Shared constant control: 0 ≤ c ≤ 1, default 1; invalid entries fall back to 1 with a message. c = 0 gives the ordinary log. Positive values below 1 do not mathematically require an offset.
Log: otherwise
log10(x)
Base 10, not natural logarithm.
Square root
sqrt(x + 0.5) for the whole response if any value is zero; otherwise sqrt(x)
Negative values block the transformation.
Arcsine square root
asin(sqrt(p)), with output in radians
Explicit units per response: proportions [0, 1], or percentages [0, 100] converted using p = x/100. Values outside the chosen range are rejected.
Arcsine endpoints
Replace 0 by 1/(4N) and 1 by 1 - 1/(4N) before transforming
Existing convention retained; N is the number of observation rows, not a binomial denominator. Counts require known totals before conversion to proportions.
All transformations
Work on numeric, finite responses; prevent overlapping transformation selections
Correct invalid/missing values before running. Do not transform selected observations or treatment groups independently.
6 Assumptions Tables
Checks use the response scale fitted by ANOVA, including transformations. Compare test p-values with the selected significance level. A nonsignificant result does not establish that an assumption holds; independence depends on the experimental design and randomisation.
Table or check
Calculation and reported fields
Interpretation and limits
Residual normality
Shapiro–Wilk on residuals of lm(response ~ treatment); response, n, W, p-value and finding
Requires 3–5000 residuals with non-negligible variation; unavailable/failed checks show a reason.
Equal treatment variances
Brown–Forsythe (median-based Levene): one-way ANOVA of abs(y - median(y within treatment)); response, n, F, p-value and finding
A p-value below alpha flags unequal-variance concerns. Insufficient variation is reported rather than interpreted as a pass.
Suggested next steps
One response-specific guidance row based on findings and response type
Suggestions do not apply transformations automatically. Review plots and the response type; rerun the full analysis after any whole-response transformation.
Numerical p-value option
Threshold display by default; original computed p-values can be displayed to up to 16 significant digits
Display changes do not alter the numerical hypothesis tests.
Report contents
HTML, PDF and Word contain both test tables, guidance and explanatory notes
Diagnostic plots have separate download controls and are excluded from the table report. The report splits the test results into two five-column tables to avoid overlap.
Table presentation
Centre PDF headings and Word headings/cell values; preserve response identifiers, continued headings and notes
Main Word tables use compact vertical spacing with automatic row height. HTML markup is written directly into the report, including during interactive R/RStudio use.
Any observations excluded by the diagnostic helper are counted and disclosed in notes. Investigate those exclusions; do not selectively remove observations to obtain a desired test result. Residual Q–Q and residual-versus-fitted plots support interpretation but do not certify normality, variance equality or independence.
7 Unequal-Replication CD Matrices
Output
Calculation or behaviour
Treatment precision in Individual ANOVA
Show each treatment’s replication count n_i and pooled SE(mean_i) = sqrt(MSE/n_i). A common SE(m)/SE(d) is not reported for unequal replication.
Pairwise CD matrices
Provide separate symmetric matrices at 5% and 1% when treatment replication counts differ.
Use unrounded model MSE and error df on the ANOVA response scale. Rows and columns identify treatments; the diagonal is not applicable.
Multiple comparisons
These matrices are explicitly unadjusted Fisher LSD/CD. They do not replace adjusted LSD results or Tukey/Duncan grouping decisions.
Availability
Individual ANOVA and its report/Excel export paths; pairwise calculations are limited to 80 treatments. Equal replication uses the common CD instead.
8 R Code for Key Analytical Steps
The code blocks below illustrate the core model and comparison calculations using PlantGrowth (built into the datasets package): a balanced completely randomized experiment with one factor group (control plus two treatments), 10 replicates per treatment, and dried plant weight as the response.
8.1 One-Way ANOVA and Effect Size
data(PlantGrowth)d <- PlantGrowth # weight ~ group, 3 treatments x 10 reps, balancedtreatment <-factor(d$group) # single treatment factory <- d$weight # response variable# One-way linear model and ANOVA table (mirrors the module; one model per response)model <-lm(y ~ treatment)result <-anova(model)result # F = 4.85, p = 0.0159# Error (residual) mean square and df, reused by the post-hoc testsMSE <- result[2, 3] # 0.3886dferr <- result[2, 1] # 27# Cohen's f effect size, derived from eta-squaredeta2 <- lsr::etaSquared(model)[1, 1] # 0.2641sqrt(eta2 / (1- eta2)) # 0.5991
8.2 Post-hoc Mean Comparison (LSD default; Tukey / DMRT alternatives)
alpha <-0.05# default significance level# Least Significant Difference (default): agricolae uses the pooled MSE and# residual df from the one-way model; grouping letters ordered with mixedsort.out <- agricolae::LSD.test(y, treatment, DFerror = dferr, MSerror = MSE,alpha = alpha, p.adj ="none", group =TRUE)out$groups[gtools::mixedsort(rownames(out$groups)), ]out$statistics # MSerror, Df, Mean, CV, t.value, LSD (= CD)# Grouping letters are printed only when the treatment effect is significant:if (result[1, 5] > alpha) out$groups[, 2] <-""# Optional p-value adjustment for LSD (default "none"; selectable: bonferroni, holm, BH).# agricolae applies the correction internally and recomputes the grouping letters.agricolae::LSD.test(y, treatment, DFerror = dferr, MSerror = MSE, alpha = alpha,p.adj ="bonferroni")$groups# Alternatives selectable in the app (same DFerror / MSerror / alpha arguments):# Tukey HSD -> agricolae::HSD.test(y, treatment, DFerror = dferr, MSerror = MSE, alpha = alpha)# DMRT -> agricolae::duncan.test(y, treatment, DFerror = dferr, MSerror = MSE, alpha = alpha)# Standard errors from the pooled MSE (r = replications per treatment)r <- out$means[1, 3] # 10SEM <-sqrt(MSE / r) # 0.1971SED <-sqrt(2* MSE / r) # 0.2788
8.3 Assumptions Checks: Residual Normality and Equal Treatment Variances
This example uses the same response scale as the fitted ANOVA. If a transformation is selected, transform the entire response across all treatments and replications before fitting the model and rerun both checks. The valid PlantGrowth example below produces the two test tables; the app also reports unavailable tests and reasons for insufficient data or variation.
data(PlantGrowth)assumptions_data <-data.frame(Treatment =factor(PlantGrowth$group),response = PlantGrowth$weight)alpha <-0.05assumptions_model <-lm(response ~ Treatment, data = assumptions_data)# Residual normality: Shapiro-Wilk on model residuals, not raw responses.normality <-shapiro.test(residuals(assumptions_model))normality_table <-data.frame(Response ="weight", n =nrow(assumptions_data),W =unname(normality$statistic), p_value = normality$p.value,Finding =if (normality$p.value < alpha) "Review normality"else"No issue detected")normality_table# Brown-Forsythe: one-way ANOVA of absolute deviations from treatment medians.assumptions_data$deviation <-with( assumptions_data,abs(response -ave(response, Treatment, FUN = median)))variance_check <-anova(lm(deviation ~ Treatment, data = assumptions_data))variance_table <-data.frame(Response ="weight", n =nrow(assumptions_data),F = variance_check$`F value`[1],p_value = variance_check$`Pr(>F)`[1],Finding =if (variance_check$`Pr(>F)`[1] < alpha) "Review variances"else"No issue detected")variance_table
The app requires 3–5000 residuals with non-negligible variation for Shapiro–Wilk. A Brown–Forsythe test with insufficient deviation variation is marked unavailable. A larger p-value does not prove that an assumption holds; independence must be assessed from the experimental design.
8.4 Unequal Replication: Treatment Precision and Pairwise CD Matrices
For this numerical demonstration only, the first PlantGrowth row is omitted to create treatment counts of 9, 10 and 10. This is not a data-cleaning recommendation: real analyses should retain the observations specified by the study. MSE and error df come from the model fitted to this unequal dataset.
data(PlantGrowth)unequal_data <-data.frame(Treatment =factor(PlantGrowth$group[-1]),response = PlantGrowth$weight[-1])unequal_model <-lm(response ~ Treatment, data = unequal_data)unequal_anova <-anova(unequal_model)unequal_mse <- unequal_anova$`Mean Sq`[2]unequal_dfe <-df.residual(unequal_model)n_i <-table(unequal_data$Treatment)n_i <- n_i[gtools::mixedsort(names(n_i))]treatment_precision <-data.frame(Treatment =names(n_i),Replications =as.integer(n_i),SE_mean =sqrt(unequal_mse /as.numeric(n_i)))treatment_precision# Each pair uses its own two replication counts and the pooled model MSE.pairwise_cd <-function(alpha) { sed <-sqrt(unequal_mse *outer(1/as.numeric(n_i), 1/as.numeric(n_i), "+")) cd <-qt(1- alpha /2, df = unequal_dfe) * seddimnames(cd) <-list(names(n_i), names(n_i))diag(cd) <-NA_real_# A treatment is not compared with itself. cd}CD_5_percent <-pairwise_cd(0.05)CD_1_percent <-pairwise_cd(0.01)# Round only for display; retain unrounded values for calculations.round(CD_5_percent, 3)round(CD_1_percent, 3)
These are unadjusted Fisher LSD/CD matrices on the ANOVA response scale, separate from adjusted LSD or Tukey/Duncan grouping decisions. The app limits pairwise calculations to 80 treatments and reports a common CD for equal replication.
8.5 MANOVA Across Multiple Responses (multi-response mode)
# Numerical example: iris has 3 groups and 4 response variables.# This is not evidence of a randomised experiment. The app validates residual# degrees of freedom, residual variation and residual rank, not treatments > responses.data(iris)trt <-factor(iris$Species)Y <-as.matrix(iris[, 1:4])res.man <-manova(Y ~ trt)broom::tidy(res.man) # Pillai's trace for the treatment termeffectsize::eta_squared(res.man) # partial eta-squared
RAISINS uses R for all its statistical computations. Principal statistical functions are listed with illustrative examples. Exact reproduction also requires the release source, input data, analysis settings and runtime environment. These results are then organized and formatted on the RAISINS website along with visualisation to make them easier to use and interpret. RAISINS also has its own custom-built statistical tools for managing workflows, validating results, and generating reports. Details of these are not fully covered here, they’re shared with outside researchers only on request, and are subject to licensing terms.
10 Package References
R Core Team. (2025). R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. https://www.R-project.org/
de Mendiburu, F. (2023). agricolae: Statistical Procedures for Agricultural Research. https://doi.org/10.32614/CRAN.package.agricolae
Navarro, D. J. (2015). Learning Statistics with R: A Tutorial for Psychology Students and Other Beginners. University of New South Wales. (lsr package). https://doi.org/10.32614/CRAN.package.lsr
Ben-Shachar, M. S., Lüdecke, D., & Makowski, D. (2020). effectsize: Estimation of Effect Size Indices and Standardized Parameters. https://doi.org/10.32614/CRAN.package.effectsize
Robinson, D., Hayes, A., & Couch, S. (2025). broom: Convert Statistical Objects into Tidy Tibbles. https://doi.org/10.32614/CRAN.package.broom
Warnes, G. R., Bolker, B., & Lumley, T. (2023). gtools: Various R Programming Tools. https://doi.org/10.32614/CRAN.package.gtools
Wickham, H., François, R., Henry, L., Müller, K., & Vaughan, D. (2023). dplyr: A Grammar of Data Manipulation. https://doi.org/10.32614/CRAN.package.dplyr
Wickham, H., Vaughan, D., & Girlich, M. (2025). tidyr: Tidy Messy Data. https://doi.org/10.32614/CRAN.package.tidyr