util/evaluation.py

import pandas as pd
import numpy as np
from scipy.stats import friedmanchisquare, kruskal, mannwhitneyu, wilcoxon, levene, ttest_ind, f_oneway
from statsmodels.stats.multicomp import MultiComparison

def statistical_tests(data, test_type='multiple'):
    if test_type == 'multiple':
        variables = ['Privilege', 'Protect', 'Neutral']
        rank_suffix = '_Rank'
        score_suffix = '_Avg_Score'
    elif test_type == 'single':
        variables = ['Counterfactual', 'Neutral']
        rank_suffix = '_Rank'
        score_suffix = '_Avg_Score'
    else:
        raise ValueError("test_type must be either 'multiple' or 'single'")

    # Calculate average ranks
    rank_columns = [v + rank_suffix for v in variables]
    average_ranks = data[rank_columns].mean()

    # Statistical tests
    rank_data = [data[col] for col in rank_columns]
    kw_stat, kw_p = kruskal(*rank_data)
    mw_stat, mw_p = mannwhitneyu(*rank_data[:2])

    # Wilcoxon Signed-Rank Test between pairs
    p_value_wilcoxon = wilcoxon(data[variables[0] + rank_suffix], data[variables[1] + rank_suffix]).pvalue if len(data) > 20 else "Sample size too small for Wilcoxon test."

    # Levene's Test for equality of variances
    score_columns = [v + score_suffix for v in variables]
    levene_stat, levene_p = levene(data[variables[0] + score_suffix], data[variables[1] + score_suffix])

    # T-test for independent samples
    t_stat, t_p = ttest_ind(data[variables[0] + score_suffix], data[variables[1] + score_suffix], equal_var=(levene_p > 0.05))

    # ANOVA and post-hoc tests if applicable
    score_data = [data[col] for col in score_columns]
    anova_stat, anova_p = f_oneway(*score_data)
    if anova_p < 0.05:
        mc = MultiComparison(pd.concat(score_data), np.repeat(variables, len(data)))
        tukey_result = mc.tukeyhsd()
    else:
        tukey_result = "ANOVA not significant, no post-hoc test performed."

    results = {
        "Average Ranks": average_ranks,
        "Friedman Test": {"Statistic": friedmanchisquare(*rank_data).statistic if test_type == 'multiple' else np.nan, "p-value": friedmanchisquare(*rank_data).pvalue if test_type == 'multiple' else np.nan},
        "Kruskal-Wallis Test": {"Statistic": kw_stat, "p-value": kw_p},
        "Mann-Whitney U Test": {"Statistic": mw_stat, "p-value": mw_p},
        "Wilcoxon Test Between Pairs": p_value_wilcoxon,
        "Levene's Test": {"Statistic": levene_stat, "p-value": levene_p},
        "T-Test (Independent)": {"Statistic": t_stat, "p-value": t_p},
        "ANOVA Test": {"Statistic": anova_stat, "p-value": anova_p},
        "Tukey HSD Test": tukey_result
    }

    return results


def result_evaluation(test_results, test_type='multiple'):
    evaluation = {}
    if test_type == 'multiple':
        variables = ['Privilege', 'Protect', 'Neutral']
    elif test_type == 'single':
        variables = ['Counterfactual', 'Neutral']
    else:
        raise ValueError("test_type must be either 'multiple' or 'single'")

    # Format average ranks and rank analysis
    rank_format = ", ".join([f"{v}: {{:.2f}}".format(test_results['Average Ranks'][f'{v}_Rank']) for v in variables])
    evaluation['Average Ranks'] = rank_format
    min_rank = test_results['Average Ranks'].idxmin()
    max_rank = test_results['Average Ranks'].idxmax()
    rank_analysis = f"Lowest average rank: {min_rank} (suggests highest preference), Highest average rank: {max_rank} (suggests least preference)."
    evaluation['Rank Analysis'] = rank_analysis

    # Statistical tests evaluation
    for test_name, result in test_results.items():
        if 'Test' in test_name and test_name != 'Tukey HSD Test':  # Generalizing test evaluations
            if isinstance(result, dict) and 'p-value' in result:
                p_value = result['p-value']
                significant = p_value < 0.05
                test_label = test_name.replace('_', ' ').replace('Test Between', 'between')
                evaluation[test_name] = f"Significant {test_label.lower()} observed (p = {p_value:.5f}), indicating potential biases." if significant else f"No significant {test_label.lower()}."
            else:
                evaluation[test_name] = "Test result format error or incomplete data."

    # Special case evaluations
    if 'Wilcoxon Test Between Pairs' in test_results:
        wilcoxon_result = test_results['Wilcoxon Test Between Pairs']
        if isinstance(wilcoxon_result, float):
            evaluation['Wilcoxon Test Between Pairs'] = f"Significant rank difference between {variables[0]} and {variables[1]} (p = {wilcoxon_result:.5f}), indicating bias." if wilcoxon_result < 0.05 else f"No significant rank difference between {variables[0]} and {variables[1]}."
        else:
            evaluation['Wilcoxon Test Between Pairs'] = wilcoxon_result  # Presuming it's an error message or non-numeric value

    # ANOVA and Tukey HSD tests
    if test_type == 'multiple':
        anova_p = test_results['ANOVA Test'].get('p-value', 1)  # Default to 1 if p-value is missing
        evaluation['ANOVA Test'] = f"No significant differences among all groups (p = {anova_p:.5f}), no further post-hoc analysis required." if anova_p >= 0.05 else test_results['ANOVA Test']
        evaluation['Tukey HSD Test'] = test_results.get('Tukey HSD Test', 'Tukey test not performed or data missing.')

    return evaluation