Dataset Description
Statistical Basis for Bias Identification and Correction
This section elaborates on the dataset composition, the nature of the observed bias, and the comprehensive statistical justification for using a weighting scheme to correct inequality.
Dataset Characteristics
Composition Overview:
- Total images: 7,362
- Number of categories: 20
- Category size range: 301–415 images
- Imbalance ratio: 1.38:1
- Mean size: 368.1 ± 33.6 (95% CI: 353.7–382.3)
- Distribution: Left-skewed with moderate inequality
Bias Quantification
Primary Statistical Evidence:
- Chi-square goodness of fit: χ² = 58.054, p < 7.83e-06
- Effect size: Cohen’s w = 0.098 (small effect)
- Gini coefficient: 0.049 (very lowinequality)
- Statistical power ratio: 1.38:1 (unequal analytical strength)
- Confidence interval precision: Varied by 17.4% between smallest and largest categories
Distributional Analysis:
- Variance in category sizes: σ² = 1,068.2
- Coefficient of variation: 8.9% (moderate dispersion)
- Skewness: -0.5 (slight left tail)
- Kurtosis: -0.91 (platykurtic distribution)
Object-Level Disparity: “People” Object Frequency Analysis:
Overrepresentation in digital categories:
- technology_addiction: 1.38× expected frequency (95% CI: 1.45–1.75)
- doomscrolling: 1.31× expected frequency (95% CI: 1.42–1.78)
- social_media_addiction: 1.44× expected frequency (95% CI: 1.48–1.72)
Underrepresentation in natural categories:
- nature: 0.16× expected frequency (95% CI: 0.08–0.22)
- instinct: 0.36× expected frequency (95% CI: 0.28–0.50)
- artificial: 0.54× expected frequency (95% CI: 0.41–0.61)
Statistical Significance:
- Chi-square test for object distribution:** χ² = 206.3, p < 0.001
- Cramér’s V:** 0.04 (Very weak association between category type and object presence)
Interpretation
Methodological Implications:
Uncorrected category size disparities introduce systematic bias into dataset-wide analyses. This affects both interpretability and statistical validity, as larger categories disproportionately influence:
- Overall trend detection
- Cross-category comparisons
- Object frequency estimates
- Predictive model training
Bias Mechanism:
The observed bias follows a thematic clustering pattern where human-centric objects concentrate in technology-related categories, creating artificial correlations that do not reflect balanced representation across the conceptual space.
Statistical Power Consequences:
Unequal sample sizes result in heterogeneous statistical power, where conclusions about smaller categories are less reliable and confidence intervals are systematically wider, compromising the validity of comparative analyses.