Table 5.

Model diagnostics and robustness checks

TestStatisticp-valueInterpretation
Goodness-of-fit (quadratic model)logLik = 68.186; AIC = −128.37; BIC = −121.62; Pseudo-R2 (FC) = 0.946Overall fit of the quadratic (education-only) specification
Hosmer-Lemeshow testN/a for beta regression (binary-outcome test only)Not applicable; outcome is a continuous proportion, not binary
Residual analysis (Pearson)SD = 1.03; |res| > 2:1 obs; |res| > 3:0 obs; Shapiro p = 0.835No concerning outliers; dispersion is reasonable
Linear vs Quadratic (LR test)LR = 67.17 (df = 1)<0.001Quadratic term improves fit, supporting convexity
Quadratic vs Cubic (LR test)LR = 31.85 (df = 1)<0.001Cubic term improves fit; consider higher-order shape
Wild cluster bootstrap (quadratic coeff.)Estimate = 2.8845; 95% CI = [−0.2817, 6.0507]; B = 999; Webb weights; G = 50.0791Marginally significant (p < 0.10); interpret curvature with caution
Note(s):

Beta regression with logit link on FLFP proportion (education-only, no year fixed effects). The log-likelihood value reported here (68.186) differs from that in Table 3 (75.482) because this diagnostic specification excludes year fixed effects; the difference reflects the reduced parameter count rather than any inconsistency in estimation. Pseudo-R2 (FC) = Ferrari and Cribari-Neto (2004). LR tests compare nested polynomial specifications. The Hosmer-Lemeshow test does not apply to continuous outcomes. Wild cluster bootstrap: Webb weights (B = 999), clustered by education group (G = 5), via betareg working-response linearisation. Cameron et al. (2008); MacKinnon et al. (2023) 

Source(s): Prepared by authors

or Create an Account

Close subscription notice
Close access options