✨ Adds printing group differences

analyze.py

@@ -1,7 +1,9 @@
 from typing import List
+import itertools
 import argparse
 
 import numpy as np
+import pandas as pd
 import scipy.stats as stats
 import scikit_posthocs as sp
 import matplotlib.pyplot as plt
@@ -14,35 +16,83 @@ graph = args.graph
 save = args.save
 
 
+# source: mostly ChatGPT (ain't no way i'm writing this shit myself)
 def analyze(name: str, data: List[np.ndarray]):
-    #print(f"Checking if normally distributed for {name}")
-    #for i in range(len(data)):
-    #    _, normal_p = stats.shapiro(data[i])
-    #    if normal_p > 0.05:
-    #        print(f"\tGroup {i}: normally distributed")
-    #    else:
-    #        print(f"\tGroup {i}: NOT normally distributed")
-
     filtered_data = []
+    group_names = []
+    all_values = []
     for index, item in enumerate(data):
-        if len(item) > 5:
+        numeric_data = [x for x in item if isinstance(x, (int, float))]
-            filtered_data.append(item)
+        if len(numeric_data) > 5:
+            filtered_data.append(numeric_data)
+            group_names.append(chr(65 + index))
+            all_values.extend(numeric_data)
         else:
-            print(f"Data group at index {index} removed due to insufficient size ({len(item)})")
+            print(f"Data group at index {index} removed due to insufficient size ({len(numeric_data)})")
 
+    if len(filtered_data) < 2:
+        print(f"Insufficient number of groups for Kruskal-Wallis test in {name}")
+        return None, None
+
+    # Kruskal-Wallis Test
     F, p = stats.kruskal(*filtered_data)
-    print(f"F-stats for {name}: {F}")
+    print(f"\nF-stats for {name}: {F:.8f}")
-    print(f"p-value for {name}: {p}")
+    print(f"p-value for {name}: {p:.8f}")
 
-    if round(p, 4) > 0.05:
+    if p > 0.05:
         print("statistically insignificant\n")
         return F, p
 
     print("statistically significant")
-    #tukey_results = stats.tukey_hsd(*filtered_data)
+
-    #print(tukey_results)
+    # Post-Hoc Dunn Test (Bonferroni-adjusted p-values)
-    ps = sp.posthoc_dunn(filtered_data, val_col='Values', group_col='Group', p_adjust='bonferroni')
+    all_ranks = stats.rankdata(all_values)  # Rank all values together
-    print(ps)
+    group_ranks = [all_ranks[start:start + len(group)] for start, group in
+                   zip(np.cumsum([0] + [len(g) for g in filtered_data[:-1]]), filtered_data)]
+    posthoc_results = sp.posthoc_dunn(filtered_data, p_adjust='bonferroni')
+    # we don't really need to print this, it's contained in the big ahh table
+    # print("\nPost-Hoc Dunn Test Results (Bonferroni-adjusted p-values):")
+    # print(posthoc_results)
+
+    results = []
+    total_sample_size = len(all_values)
+    for group1, group2 in itertools.combinations(group_names, 2):
+        idx1 = group_names.index(group1)
+        idx2 = group_names.index(group2)
+
+        mean_rank_1 = np.mean(group_ranks[idx1])
+        mean_rank_2 = np.mean(group_ranks[idx2])
+        rank_diff = mean_rank_1 - mean_rank_2
+
+        n1 = len(filtered_data[idx1])
+        n2 = len(filtered_data[idx2])
+
+        # Effect size (Rank-Biserial Correlation)
+        z_stat = rank_diff / np.sqrt((n1 + n2) * (n1 * n2) / total_sample_size)
+        effect_size = z_stat / np.sqrt(total_sample_size)
+
+        # Mean difference
+        mean_diff = np.mean(filtered_data[idx1]) - np.mean(filtered_data[idx2])
+
+        # Median difference
+        median_diff = np.median(filtered_data[idx1]) - np.median(filtered_data[idx2])
+
+        # Post-Hoc Dunn p-value
+        p_value = posthoc_results.loc[idx1 + 1, idx2 + 1]
+
+        results.append({
+            "Group 1": group1,
+            "Group 2": group2,
+            "Effect Size": f"{effect_size:.4f}",
+            "Mean Difference": f"{mean_diff:.4f}",
+            "Median Difference": f"{median_diff:.4f}",
+            "Post-Hoc p-value": f"{p_value:.4f}"
+        })
+
+    results_df = pd.DataFrame(results, dtype="object")
+    print("\nSummary Table of Effect Size, Mean, and Median Differences:")
+    print(results_df.to_markdown(index=False, tablefmt="github", disable_numparse=True))
+    print("")
 
     return F, p
 
@@ -63,7 +113,7 @@ def plot_violin(data, labels, Fs, ps, title):
             index = j * 2 + k
             step = 1 if index > 0 else 0.5
 
-            axs[j, k].violinplot(data[index], showmedians=True)
+            axs[j, k].violinplot(data[index], showmedians=True, showmeans=True)
             axs[j, k].set_title(grade_names[index])
             axs[j, k].set_xlabel(title, fontweight="bold")
             axs[j, k].set_ylabel(grade_name_labels[index], fontweight="bold")
@@ -72,13 +122,18 @@ def plot_violin(data, labels, Fs, ps, title):
 
             F = round(Fs[index], 2)
             p = round(ps[index], 4)
-            axs[j, k].text(0.01, 0.99, f"F-stat: {F:.2f}\np-val: {p:.4f}", ha="left", va="top", transform=axs[j, k].transAxes,
+            axs[j, k].text(0.01, 0.99, f"F-stat: {F:.2f}\np-val: {p:.4f}", ha="left", va="top",
+                           transform=axs[j, k].transAxes,
                            fontweight="bold")
 
             medians = list([np.median(a) for a in data[index]])
+            means = list([a.mean() for a in data[index]])
             for l in range(len(medians)):
                 median = round(medians[l], 2)
-                axs[j, k].text(l + 1.05, median + 0.05, f"{median}")
+                mean = round(means[l], 2)
+                # left - mean, right - median
+                axs[j, k].text(l + 1.13, median - 0.05, f"{median}")
+                axs[j, k].text(l + 0.77, mean - 0.05, f"{mean}")
 
     fig.tight_layout()
     if save != "":

requirements.txt

@@ -25,6 +25,7 @@ nvidia-nvjitlink-cu12==12.4.127
 nvidia-nvtx-cu12==12.4.127
 packaging==24.2
 pandas==2.2.3
+pandas-flavor==0.6.0
 patsy==1.0.1
 pillow==11.0.0
 pyparsing==3.2.0
@@ -38,7 +39,9 @@ setuptools==75.6.0
 six==1.17.0
 statsmodels==0.14.4
 sympy==1.13.1
+tabulate==0.9.0
 threadpoolctl==3.5.0
 torch==2.5.1
 typing_extensions==4.12.2
 tzdata==2024.2
+xarray==2024.11.0