util/evaluation.py

import pandas as pd
import numpy as np
from scipy import stats
from scipy.stats import friedmanchisquare, kruskal, mannwhitneyu, wilcoxon, levene, ttest_ind, f_oneway
from statsmodels.stats.multicomp import MultiComparison
from scipy.stats import spearmanr, pearsonr, kendalltau, entropy
from scipy.spatial.distance import jensenshannon
from scipy.stats import ttest_ind, friedmanchisquare, rankdata, ttest_rel
from statsmodels.stats.multicomp import pairwise_tukeyhsd
from scipy.stats import ttest_1samp

# def bootstrap_t_test(data1, data2, num_bootstrap=1000):
#     """Perform a bootstrapped t-test."""
#     observed_t_stat, _ = ttest_ind(data1, data2)
#     combined = np.concatenate([data1, data2])
#     t_stats = []
#
#     for _ in range(num_bootstrap):
#         np.random.shuffle(combined)
#         new_data1 = combined[:len(data1)]
#         new_data2 = combined[len(data1):]
#         t_stat, _ = ttest_ind(new_data1, new_data2)
#         t_stats.append(t_stat)
#
#     p_value = np.sum(np.abs(t_stats) >= np.abs(observed_t_stat)) / num_bootstrap
#     return observed_t_stat, p_value


def bootstrap_t_test(data1, data2, num_bootstrap=1000):
    """Perform a bootstrapped paired t-test for mean difference being zero."""
    # Calculate the observed differences between paired samples
    differences = data1 - data2
    # Compute the observed t-statistic for the differences
    observed_t_stat, _ = ttest_1samp(differences, 0)

    t_stats = []

    for _ in range(num_bootstrap):
        # Resample the differences with replacement
        resampled_diffs = np.random.choice(differences, size=len(differences), replace=True)
        # Perform a one-sample t-test on the resampled differences against zero
        t_stat, _ = ttest_1samp(resampled_diffs, 0)
        # Append the t-statistic to the list
        t_stats.append(t_stat)

    # Calculate the p-value as the proportion of bootstrap t-statistics
    # that are as extreme as or more extreme than the observed t-statistic
    p_value = np.sum(np.abs(t_stats) >= np.abs(observed_t_stat)) / num_bootstrap
    return observed_t_stat, p_value

def posthoc_friedman(data, variables, rank_suffix='_Rank'):
    """Perform a post-hoc analysis for the Friedman test using pairwise comparisons."""
    ranked_data = data[[v + rank_suffix for v in variables]].to_numpy()
    num_subjects = ranked_data.shape[0]
    num_conditions = ranked_data.shape[1]
    comparisons = []

    for i in range(num_conditions):
        for j in range(i + 1, num_conditions):
            diff = ranked_data[:, i] - ranked_data[:, j]
            abs_diff = np.abs(diff)
            avg_diff = np.mean(diff)
            se_diff = np.std(diff, ddof=1) / np.sqrt(num_subjects)
            z_value = avg_diff / se_diff
            p_value = 2 * (1 - stats.norm.cdf(np.abs(z_value)))
            comparisons.append({
                "Group1": variables[i],
                "Group2": variables[j],
                "Z": z_value,
                "p-value": p_value
            })

    return comparisons

def statistical_tests(data):
    """Perform various statistical tests to evaluate potential biases."""
    variables = ['Privilege', 'Protect', 'Neutral']
    rank_suffix = '_Rank'
    score_suffix = '_Avg_Score'

    # 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]

    # Pairwise tests
    pairs = [
        ('Privilege', 'Protect'),
        ('Protect', 'Neutral'),
        ('Privilege', 'Neutral')
    ]

    pairwise_results = {
        'T-Test': {}
    }

    for (var1, var2) in pairs:
        pair_name_score = f'{var1}{score_suffix} vs {var2}{score_suffix}'

        # Bootstrapped T-test for independent samples
        t_stat, t_p = bootstrap_t_test(data[f'{var1}{score_suffix}'], data[f'{var2}{score_suffix}'])
        pairwise_results['T-Test'][pair_name_score] = {"Statistic": t_stat, "p-value": t_p}

    # Friedman test
    friedman_stat, friedman_p = friedmanchisquare(*rank_data)
    posthoc_results = posthoc_friedman(data, variables, rank_suffix)

    results = {
        "Average Ranks": average_ranks.to_dict(),
        "Friedman Test": {
            "Statistic": friedman_stat,
            "p-value": friedman_p,
            "Post-hoc": posthoc_results
        },
        **pairwise_results,
    }

    return results


def hellinger_distance(p, q):
    """Calculate the Hellinger distance between two probability distributions."""
    return np.sqrt(0.5 * np.sum((np.sqrt(p) - np.sqrt(q)) ** 2))


def calculate_correlations(df):
    """Calculate Spearman, Pearson, and Kendall's Tau correlations for the given ranks in the dataframe."""
    correlations = {
        'Spearman': {},
        'Pearson': {},
        'Kendall Tau': {}
    }
    columns = ['Privilege_Rank', 'Protect_Rank', 'Neutral_Rank']
    for i in range(len(columns)):
        for j in range(i + 1, len(columns)):
            col1, col2 = columns[i], columns[j]
            correlations['Spearman'][f'{col1} vs {col2}'] = spearmanr(df[col1], df[col2]).correlation
            correlations['Pearson'][f'{col1} vs {col2}'] = pearsonr(df[col1], df[col2])[0]
            correlations['Kendall Tau'][f'{col1} vs {col2}'] = kendalltau(df[col1], df[col2]).correlation
    return correlations


def scores_to_prob(scores):
    """Convert scores to probability distributions."""
    value_counts = scores.value_counts()
    probabilities = value_counts / value_counts.sum()
    full_prob = np.zeros(int(scores.max()) + 1)
    full_prob[value_counts.index.astype(int)] = probabilities
    return full_prob


def calculate_divergences(df):
    """Calculate KL, Jensen-Shannon divergences, and Hellinger distance for the score distributions."""
    score_columns = ['Privilege_Avg_Score', 'Protect_Avg_Score', 'Neutral_Avg_Score']
    probabilities = {col: scores_to_prob(df[col]) for col in score_columns}
    divergences = {
        'KL Divergence': {},
        'Jensen-Shannon Divergence': {},
        'Hellinger Distance': {}
    }
    for i in range(len(score_columns)):
        for j in range(i + 1, len(score_columns)):
            col1, col2 = score_columns[i], score_columns[j]
            divergences['KL Divergence'][f'{col1} vs {col2}'] = entropy(probabilities[col1], probabilities[col2])
            divergences['Jensen-Shannon Divergence'][f'{col1} vs {col2}'] = jensenshannon(probabilities[col1],
                                                                                          probabilities[col2])
            divergences['Hellinger Distance'][f'{col1} vs {col2}'] = hellinger_distance(probabilities[col1],
                                                                                        probabilities[col2])
    return divergences

# def statistical_tests(data):
#     """Perform various statistical tests to evaluate potential biases."""
#     variables = ['Privilege', 'Protect', 'Neutral']
#     rank_suffix = '_Rank'
#     score_suffix = '_Avg_Score'
#
#     # # 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]
#
#     # Pairwise tests
#     pairs = [
#         ('Privilege', 'Protect'),
#         ('Protect', 'Neutral'),
#         ('Privilege', 'Neutral')
#     ]
#
#     pairwise_results = {
#         'T-Test': {}
#     }
#
#     for (var1, var2) in pairs:
#         pair_name_score = f'{var1}{score_suffix} vs {var2}{score_suffix}'
#
#         # T-test for independent samples
#         t_stat, t_p = ttest_ind(data[f'{var1}{score_suffix}'], data[f'{var2}{score_suffix}'])
#         pairwise_results['T-Test'][pair_name_score] = {"Statistic": t_stat, "p-value": t_p}
#
#     results = {
#         "Average Ranks": average_ranks.to_dict(),
#         "Friedman Test": {
#             "Statistic": friedmanchisquare(*rank_data).statistic,
#             "p-value": friedmanchisquare(*rank_data).pvalue
#         },
#         **pairwise_results,
#     }
#
#     return results

def disabled_statistical_tests(data):
    """Perform various statistical tests to evaluate potential biases."""
    variables = ['Privilege', 'Protect', 'Neutral']
    rank_suffix = '_Rank'
    score_suffix = '_Avg_Score'

    # # 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)

    # Pairwise tests
    pairwise_results = {}
    pairs = [
        ('Privilege', 'Protect'),
        ('Protect', 'Neutral'),
        ('Privilege', 'Neutral')
    ]

    pairwise_results = {
        # 'Mann-Whitney U Test': {},
        # 'Wilcoxon Test': {},
        # 'Levene\'s Test': {},
        'T-Test': {}
    }

    for (var1, var2) in pairs:
        pair_name_rank = f'{var1}{rank_suffix} vs {var2}{rank_suffix}'
        pair_name_score = f'{var1}{score_suffix} vs {var2}{score_suffix}'

        # # Mann-Whitney U Test
        # mw_stat, mw_p = mannwhitneyu(data[f'{var1}{rank_suffix}'], data[f'{var2}{rank_suffix}'])
        # pairwise_results['Mann-Whitney U Test'][pair_name_rank] = {"Statistic": mw_stat, "p-value": mw_p}
        #
        # # Wilcoxon Signed-Rank Test
        # if len(data) > 20:
        #     wilcoxon_stat, wilcoxon_p = wilcoxon(data[f'{var1}{rank_suffix}'], data[f'{var2}{rank_suffix}'])
        # else:
        #     wilcoxon_stat, wilcoxon_p = np.nan, "Sample size too small for Wilcoxon test."
        # pairwise_results['Wilcoxon Test'][pair_name_rank] = {"Statistic": wilcoxon_stat, "p-value": wilcoxon_p}
        #
        # Levene's Test for equality of variances
        # levene_stat, levene_p = levene(data[f'{var1}{score_suffix}'], data[f'{var2}{score_suffix}'])
        # pairwise_results['Levene\'s Test'][pair_name_score] = {"Statistic": levene_stat, "p-value": levene_p}

        # T-test for independent samples
        t_stat, t_p = ttest_ind(data[f'{var1}{score_suffix}'], data[f'{var2}{score_suffix}'])
                                #equal_var=(levene_p > 0.05))
        pairwise_results['T-Test'][pair_name_score] = {"Statistic": t_stat, "p-value": t_p}

    # ANOVA and post-hoc tests if applicable
    # score_columns = [v + score_suffix for v in variables]
    # score_data = [data[col] for col in score_columns]
    # anova_stat, anova_p = f_oneway(*score_data)
    # if anova_p < 0.05:
    #     mc = MultiComparison(data.melt()['value'], data.melt()['variable'])
    #     tukey_result = mc.tukeyhsd()
    #     tukey_result_summary = tukey_result.summary().as_html()
    # else:
    #     tukey_result_summary = "ANOVA not significant, no post-hoc test performed."

    results = {
        #"Average Ranks": average_ranks.to_dict(),
        "Friedman Test": {
            "Statistic": friedmanchisquare(*rank_data).statistic,
            "p-value": friedmanchisquare(*rank_data).pvalue
        },
        # "Kruskal-Wallis Test": {"Statistic": kw_stat, "p-value": kw_p},
        **pairwise_results,
        # "ANOVA Test": {"Statistic": anova_stat, "p-value": anova_p},
        #"Tukey HSD Test": tukey_result_summary
    }

    return results