stat significance nr times ranked 1st

XX_statistical_significance.script.py

@@ -329,4 +329,42 @@ S.plot_significance_heatmap(
     metadata=_meta_male_top_vis,
     title="Male Voices (Excl. Bottom 3): Total Mentions Significance"
 )
+
+# %% [markdown]
+"""
+# Rank 1 Selection Significance (Voice Level)
+
+Similar to the Total Mentions significance analysis above, but counting
+only how many times each voice was ranked **1st** (out of all respondents).
+This isolates first-choice preference rather than overall top-3 visibility.
+"""
+
+# %% Rank 1 Significance: All Voices
+
+_pairwise_df_rank1, _meta_rank1 = S.compute_rank1_significance(
+    top3_voices,
+    alpha=0.05,
+    correction="none",
+)
+
+S.plot_significance_heatmap(
+    _pairwise_df_rank1,
+    metadata=_meta_rank1,
+    title="Statistical Significance: Voice Rank 1 Selection"
+)
+
+# %% Rank 1 Significance: Male Voices Only
+
+_pairwise_df_rank1_male, _meta_rank1_male = S.compute_rank1_significance(
+    df_male_voices,
+    alpha=0.05,
+    correction="none",
+)
+
+S.plot_significance_heatmap(
+    _pairwise_df_rank1_male,
+    metadata=_meta_rank1_male,
+    title="Male Voices Only: Rank 1 Selection Significance"
+)
+
 # %%

utils.py

@@ -1701,6 +1701,121 @@ class QualtricsSurvey(QualtricsPlotsMixin):
         
         return results_df, metadata
 
+    def compute_rank1_significance(
+        self,
+        data: pl.LazyFrame | pl.DataFrame,
+        alpha: float = 0.05,
+        correction: str = "bonferroni",
+    ) -> tuple[pl.DataFrame, dict]:
+        """Compute statistical significance for Rank 1 selections only.
+
+        Like compute_mentions_significance but counts only how many times each
+        voice/character was ranked **1st**, using total respondents as the
+        denominator.  This tests whether first-choice preference differs
+        significantly between voices.
+
+        Args:
+            data: Ranking data (rows=respondents, cols=voices, values=rank).
+            alpha: Significance level.
+            correction: Multiple comparison correction method.
+
+        Returns:
+            tuple: (pairwise_df, metadata)
+        """
+        from scipy import stats as scipy_stats
+        import numpy as np
+
+        if isinstance(data, pl.LazyFrame):
+            df = data.collect()
+        else:
+            df = data
+
+        ranking_cols = [c for c in df.columns if c != '_recordId']
+        if len(ranking_cols) < 2:
+            raise ValueError("Need at least 2 ranking columns")
+
+        total_respondents = df.height
+        rank1_data: dict[str, int] = {}
+
+        # Count rank-1 selections for each voice
+        for col in ranking_cols:
+            label = self._clean_voice_label(col)
+            count = df.filter(pl.col(col) == 1).height
+            rank1_data[label] = count
+
+        labels = sorted(list(rank1_data.keys()))
+        results = []
+        n_comparisons = len(labels) * (len(labels) - 1) // 2
+
+        for i, label1 in enumerate(labels):
+            for label2 in labels[i+1:]:
+                count1 = rank1_data[label1]
+                count2 = rank1_data[label2]
+
+                pct1 = count1 / total_respondents
+                pct2 = count2 / total_respondents
+
+                # Z-test for two proportions (same denominator for both)
+                n1 = total_respondents
+                n2 = total_respondents
+
+                p_pooled = (count1 + count2) / (n1 + n2)
+                se = np.sqrt(p_pooled * (1 - p_pooled) * (1/n1 + 1/n2))
+
+                if se > 0:
+                    z_stat = (pct1 - pct2) / se
+                    p_value = 2 * (1 - scipy_stats.norm.cdf(abs(z_stat)))
+                else:
+                    p_value = 1.0
+
+                results.append({
+                    'group1': label1,
+                    'group2': label2,
+                    'p_value': float(p_value),
+                    'rank1_count1': count1,
+                    'rank1_count2': count2,
+                    'rank1_pct1': round(pct1 * 100, 1),
+                    'rank1_pct2': round(pct2 * 100, 1),
+                    'total1': n1,
+                    'total2': n2,
+                    'effect_size': pct1 - pct2,
+                })
+
+        results_df = pl.DataFrame(results)
+
+        p_values = results_df['p_value'].to_numpy()
+        p_adjusted = np.full_like(p_values, np.nan, dtype=float)
+
+        if correction == "bonferroni":
+            p_adjusted = np.minimum(p_values * n_comparisons, 1.0)
+        elif correction == "holm":
+            sorted_idx = np.argsort(p_values)
+            sorted_p = p_values[sorted_idx]
+            m = len(sorted_p)
+            adjusted = np.zeros(m)
+            for j in range(m):
+                adjusted[j] = sorted_p[j] * (m - j)
+            for j in range(1, m):
+                adjusted[j] = max(adjusted[j], adjusted[j-1])
+            adjusted = np.minimum(adjusted, 1.0)
+            p_adjusted = adjusted[np.argsort(sorted_idx)]
+        elif correction == "none":
+            p_adjusted = p_values.astype(float)  # pyright: ignore
+
+        results_df = results_df.with_columns([
+            pl.Series('p_adjusted', p_adjusted),
+            pl.Series('significant', p_adjusted < alpha),
+        ]).sort('p_value')
+
+        metadata = {
+            'test_type': 'proportion_z_test_rank1',
+            'alpha': alpha,
+            'correction': correction,
+            'n_comparisons': n_comparisons,
+        }
+
+        return results_df, metadata
+
 
 
 def process_speaking_style_data(