Step 1 — Upload data

Click Browse .xlsx to load an Excel file. Numeric columns are auto-classified as Numeric; text/factor columns as Ignore.

Step 2 — Classify variables

Each column is assigned one of two roles:
Numeric variable: Included in the pairwise correlation matrix. At least two numeric variables are required.
Ignore: Excluded from the analysis.

Step 3 — Analysis setup

Correlation method:
• Pearson — Measures linear association between two continuous variables. Assumes bivariate normality. Best when the relationship is approximately linear.
• Spearman — A rank-based non-parametric measure of monotonic association. Robust to non-normality and outliers. Recommended when normality cannot be assumed.

Handle missing values:
• Complete observations only — Only rows with no missing values across all selected variables are used. All pairs share the same sample size.
• Pairwise complete — Each pair uses all rows where both variables are non-missing. Maximises data usage but pairs may have different sample sizes.

Significance level (α): The threshold for declaring a correlation statistically significant (default: 0.05). Each pair is tested individually with cor.test() . Significance codes: *** p < 0.001, ** p < 0.01, * p < 0.05, . p < 0.10, ns otherwise.

How to cite in your manuscript

Pearson:
Pearson, K. (1895). Notes on regression and inheritance in the case of two parents. Proceedings of the Royal Society of London , 58 , 240–242.

Spearman:
Spearman, C. (1904). The proof and measurement of association between two things. The American Journal of Psychology , 15 (1), 72–101.

Step 1 — Upload data

Click Browse .xlsx to load an Excel file. Character/factor columns are auto-classified as Factor and numeric columns as Response.

Step 2 — Classify variables

Each column is assigned one of three roles:
Factor (grouping): The categorical column defining groups.
Response variable: A numeric column with measurements.
Ignore: Excluded from the analysis.

Step 3 — Analysis setup

Grouping factor: Which Factor column defines the groups to compare.

Response variable: Which numeric column contains the measurements.

Y-axis label: Custom label for the bar plot Y-axis. Auto-filled with the response variable name; editable to any text (e.g. 'Shoot length (cm)').

Significance level (α): Threshold for ANOVA and post-hoc significance (default: 0.05). Affects the CLD grouping letters and the Tukey/Bonferroni significance flag.

Post-hoc test:
• Tukey HSD — Honestly Significant Difference. Controls the family-wise error rate for all pairwise comparisons. Computed via agricolae::HSD.test() . This is the standard choice for ANOVA post-hoc.
• Bonferroni — Adjusts α by the number of comparisons. More conservative than Tukey HSD; appropriate when controlling for multiple comparisons strictly.

Color palette:
• AgroBio (green) — The institutional green palette.
• Viridis — A perceptually uniform, colourblind-friendly palette.
• Colorblind safe — The Okabe-Ito palette, optimised for deuteranopia and protanopia.

Bar plot order:
• Alphabetical (A → Z / Z → A) — Groups sorted by name.
• Mean (high → low / low → high) — Groups sorted by their mean response value.
• Custom — A text area appears where you define the exact order (one group per line).

How to cite in your manuscript

For One-way ANOVA:
Fisher, R. A. (1925). Statistical Methods for Research Workers. Oliver & Boyd.

For Tukey HSD:
Tukey, J. W. (1949). Comparing individual means in the analysis of variance. Biometrics , 5 (2), 99–114.

For agricolae (CLD implementation):
de Mendiburu, F. (2023). agricolae: Statistical Procedures for Agricultural Research. R package.

For Bonferroni correction:
Dunn, O. J. (1961). Multiple comparisons among means. Journal of the American Statistical Association , 56 (293), 52–64.

Step 1 — Upload data

Click Browse .xlsx to load an Excel file. Numeric columns are auto-classified as Predictor; text/factor columns also as Predictor (CTree handles both numeric and categorical predictors natively).

Step 2 — Classify variables

Each column is assigned one of three roles:
Target (response): The variable to model (numeric or categorical). Exactly one target must be selected.
Predictor: Input variables used to split the tree. Multiple predictors can (and should) be selected. Numeric predictors with ≤ 10 unique values are automatically converted to factors.
Ignore: Excluded from the analysis.

Step 3 — Analysis setup

Target variable: The response variable to model.

Predictors (select multiple): Multi-select list of input variables. All predictors are selected by default; deselect any you wish to exclude.

Significance level (α): The threshold for splitting (default: 0.05). At each node, the algorithm tests independence between each predictor and the response using permutation tests. A split is only performed if the smallest adjusted p-value is below α. Lower values produce simpler trees.

Max tree depth: Maximum number of levels from root to terminal node (default: 4). Depth 3–4 is recommended for biological interpretability. Depth ≥ 6 may produce overly specific splits.

Min observations per node: Minimum number of observations required in a terminal node (default: 10, minimum: 3). Prevents the tree from making splits based on very few data points. Increase for noisy datasets.

Node annotation (continuous target): Controls the statistics displayed inside terminal-node boxplots:
• None — Standard boxplot only.
• Median — Annotates each node with Md = value.
• Mean ± SD — Annotates with x̅ = value ± SD.
• Median + Mean — Shows both Md and x̅ side by side.

How to cite CTree in your manuscript

For the algorithm:
Hothorn, T., Hornik, K., & Zeileis, A. (2006). Unbiased recursive partitioning: A conditional inference framework. Journal of Computational and Graphical Statistics , 15 (3), 651–674. https://doi.org/10.1198/106186006X133933

For the R implementation:
Hothorn, T., & Zeileis, A. (2015). partykit: A modular toolkit for recursive partytioning in R. Journal of Machine Learning Research , 16 , 3905–3909.

Step 1 — Upload data

Click Browse .xlsx to load an Excel file. Numeric columns are auto-classified as Predictor (numeric); text/factor columns as Predictor (categorical).

Step 2 — Classify variables

Each column is assigned one of the available roles:
Target (response): The numeric variable to predict. Exactly one must be selected.
Predictor (numeric): A continuous input variable.
Predictor (categorical): A categorical input variable. Automatically one-hot encoded via model.matrix() (full-rank encoding) before scaling and training.
Ignore: Excluded from the analysis.
Note: text/factor columns can only be assigned as Predictor (categorical) or Ignore; numeric columns can be Target, Predictor (numeric), or Ignore.

Step 3 — Analysis setup

Target variable: The numeric response to model.

Predictors (select multiple): Multi-select list of input variables (both numeric and categorical). All predictors are selected by default.

Hidden layers

Defines the topology of the ANN. Enter comma-separated integers, e.g. 8,5,3 for three hidden layers with 8, 5, and 3 neurons. A single number (e.g. 5 ) creates one hidden layer.
8,5,3 is the default. Simpler datasets may work with 5,3 or even 3 ; more complex datasets may benefit from wider layers. Avoid very deep architectures with small datasets (n < 100).

Activation function

The non-linear function applied at each hidden neuron.
Logistic (sigmoid) — outputs in [0, 1]. The classical default for most applications.
Tanh — outputs in [−1, 1]. Often converges faster than logistic for standardised data.

Learning algorithm

The optimisation algorithm used for training.
Rprop (Resilient Propagation) — the recommended default. Adapts the learning rate per weight, solving convergence issues common with standard backpropagation.
SCG (Scaled Conjugate Gradient) — a second-order method that can be faster for medium-sized datasets.
Backprop + Momentum — classic backpropagation with momentum term. May require tuning the learning rate.
Standard Backprop — basic gradient descent without momentum. Generally slower and less robust than Rprop or SCG.

Max iterations (maxit)

Maximum number of training iterations. The algorithm may converge before reaching this limit. The default ( 1000 ) is appropriate for most plant tissue culture datasets with Rprop. If the model underfits, increase to 2000–5000. Very high values increase computation time.

Training repetitions

Number of independent ANN training runs, each starting from different random initial weights. AgroBioSTAT automatically selects the replicate with the lowest error. More repetitions increase the probability of finding a good solution, at the cost of longer computation time. The default ( 3 ) is adequate for most cases; increase to 5–10 for complex datasets.

Surrogate tree max. depth

Controls the maximum depth of the surrogate decision tree. A deeper tree generates more rules with more conditions (conjunctions). A shallower tree produces fewer, simpler rules. Depth 2–4 is recommended for biological interpretability; depth ≥ 5 tends to generate overly specific rules. Default: 2 .

Surrogate tree complexity (cp)

The complexity parameter for rpart . A split is only performed if it improves the overall fit by at least cp (relative to the root node error). Lower values allow more splits (more complex tree, more rules); higher values produce simpler trees. Default: 0.01 . Range: 0.001–0.1.

Validation (nested)

AgroBioSTAT uses a nested validation strategy: a Train/Test split (external validation) combined with internal cross-validation within the training set.

Training set (%): Slider to set the proportion of data used for training (default: 70%). The remaining observations form the held-out test set. A donut chart shows the split visually. Range: 50–90%.

Internal CV method:
• k-fold CV — Splits the training set into k folds; trains on k−1 folds and validates on the held-out fold, rotating through all k. The default k = 5 is suitable for most datasets; increase to 10 for smaller datasets.
• LOOCV (Leave-One-Out CV) — k = n(train). Each observation is used as validation once. Computationally expensive for large datasets; consider k-fold if n > 50.

Number of folds (k): Only visible when k-fold CV is selected. Default: 5. Range: 3–20.

Random seed: Ensures reproducibility of the Train/Test split, fold assignment, and ANN weight initialisation. Default: 42.

How to cite in your manuscript

For RSNNS (ANN engine):
Bergmeir, C., & Benítez, J. M. (2012). Neural Networks in R Using the Stuttgart Neural Network Simulator: RSNNS. Journal of Statistical Software , 46(7), 1–26.

For NeuralNetTools (variable importance):
Beck, M. W. (2018). NeuralNetTools: Visualization and Analysis Tools for Neural Networks. Journal of Statistical Software , 85(11), 1–20.

For Olden variable importance:
Olden, J. D., Joy, M. K., & Death, R. G. (2004). An accurate comparison of methods for quantifying variable importance in artificial neural networks using simulated data. Ecological Modelling , 178 (3–4), 389–397.

For rpart (surrogate tree):
Therneau, T. M., & Atkinson, E. J. (2019). An introduction to recursive partitioning using the RPART routines. Mayo Foundation.

For the surrogate modelling approach:
Craven, M. W., & Shavlik, J. W. (1996). Extracting tree-structured representations of trained networks. Advances in Neural Information Processing Systems , 8, 24–30.