Hypothesis testing
H2 Over-correction vs Maslow entropy
H1: Images with a more even spread of Maslow needs (higher entropy) show more over-correction (higher OverCorrection score).
H0: Maslow entropy is not associated with OverCorrection.
Question: Do images that balance more need-levels tend to overshoot more (i.e., show greater over-correction)?
Data
Data subset: Images containing at least one Maslow category and at least one DA category
Variables:
-
- Predictor(s): Maslow entropy (H) per image. (Shannon entropy over the 6 need levels)
- Outcome(s): OverCorrection per image (≈ −1 to +1).
Controls: none (bivariate design for H2).
Normalisation & Weighting: See Weighting and Normalisation Methods reference page.
N: 6,789 (weighted), 7,362 (raw).
Methods
Test type:
-
- Primary: Ordinary Least Squares (OLS) regression of OverCorrection (continuous) on Maslow entropy (continuous), with HC3 heteroskedasticity-robust SEs.
- Robustness (non-parametric): Mann–Whitney U test comparing High-entropy vs Low-entropy images (median split), plus Cliff’s δ as an effect size (with bootstrap CI if desired).
- Visual diagnostic: Hex-scatter of OverCorrection vs entropy with fitted OLS line.
Rationale:
-
- OLS captures the direction and size of the linear association across the full distribution and gives confidence intervals and an R² for model fit. HC3 SEs guard against heteroskedasticity from the bounded, clumpy outcome (spikes near −1, 0, +1).
- Mann–Whitney U does not assume normality and tests a stochastic dominance shift between low- and high-entropy images.
- Cliff’s δ communicates the practical size of any shift on a common, distribution-free scale (−1…+1), which is useful when effects are statistically detectable but small.
Implementation:
-
-
Predictor: Maslow entropy
from L1–L6 proportions.
-
Outcome: OverCorrection (continuous).
-
Clean: drop NaN/inf; bivariate spec (no extra controls).
-
Report: β, SE, 95% CI, p, R²; U, p(MWU), Cliff’s δ.
-
MWU split at sample median entropy; same code/spec for weighted & raw.
-
Results — Weighted
- OLS: β = +0.138, SE = 0.021, 95% CI [0.097, 0.180], p ≈ 6.9×10⁻¹¹, R² ≈ 0.009. → Positive but small association; entropy explains ~1% of variance.
- Mann–Whitney U (High vs Low entropy): p ≈ 1.37×10⁻¹⁸, Cliff’s δ = −0.122 (n₁=3485, n₂=3455). → Very small location shift with opposite sign; magnitude is small (|δ|≈0.12).
Interpretation (weighted): On average, higher entropy is linked to slightly more over-correction, but the effect is tiny.
The non-parametric split shows a small shift the other way, consistent with the heavy, discretised outcome (lots of mass near −1, 0, +1) where medians can move differently than the overall slope.
Results — Raw
- OLS: β = +0.141, SE = 0.021, 95% CI [0.100, 0.183], p ≈ 2.8×10⁻¹¹, R² ≈ 0.009.
- Mann–Whitney U (High vs Low entropy): p ≈ 3.66×10⁻²⁰, Cliff’s δ = −0.127 (n₁=3507, n₂=3433).
Interpretation (raw): Mirrors the weighted results: statistically strong but practically small association; tiny group shift in the opposite direction under a median split.
Combined Interpretation
Across both raw and weighted datasets, OLS shows a positive (β≈0.14) relation between entropy and over-correction with very small R² (~0.9%).
The Mann–Whitney/Cliff’s δ indicates a small high-vs-low entropy shift in the opposite direction (δ≈−0.12). Given the heavily discretised OverCorrection distribution, this split captures median behavior, while OLS captures the overall linear trend (likely influenced by the tails).
Substantively: higher need-spread is not a strong driver of over-correction. Any association that exists is tiny and sensitive to how you summarise the outcome (median split vs continuous trend).
Conclusion
- Reject H₀ on statistical grounds (both OLS models highly significant), but emphasise tiny effect sizes.
- Practical takeaway: Maslow entropy has, at most, a minimal relationship with over-correction. The direction depends on summary (continuous vs split), so treat the effect as weak and not decisive for theory.