Covariate
Class Size
Relative importance (LMG) OLS estimate (clustered SE)
Teachers 10.7% -1.096*** (0.114)
Students 13.8% 0.030*** (0.002)
Over-age 0.1% -0.059 (0.046)
No fee 1.3% -0.046 (0.904)
Scenario 1 1.0% 0.115 (0.200)
Scenario 2 0.7% -0.175 (0.133)
Scenario 4 39.6% 0.776*** (0.104)
Scenario 5 6.6% 0.170 (0.141)
Grade (base: R) 12.4% NULL
Grade: 1 / 4.993*** (0.564)
Grade: 2 / 5.947*** (0.555)
Grade: 3 / 6.585*** (0.590)
Grade: 4 / 11.546*** (0.671)
Grade: 5 / 11.605*** (0.647)
Grade: 6 / 11.573*** (0.679)
Grade: 7 / 11.306*** (0.680)
Gender (base: female) 0.0% NULL
Gender: male / 0.087** (0.041)
Race (base: black) 5.3% NULL
Race: coloured / 0.144 (0.640)
Race: indian / -0.641 (0.999)
Race: other / 1.859 (2.458)
Race: white / -2.050*** (0.751)
Quintile (base: 1) 4.3% NULL
Quintile: 2 / 0.215 (0.365)
Quintile: 3 / -0.115 (0.520)
Quintile: 4 / 0.983 (0.922)
Quintile: 5 / 0.439 (0.936)
Province (base: ec) 4.4% NULL
Province: fs / -2.931** (1.484)
Province: gt / -3.935* (2.164)
Province: kz / -2.884 (2.480)
Province: lp / -2.943 (2.363)
Province: mp / -4.270 (3.088)
Province: nc / -1.300 (1.721)
Province: nw / -3.295 (2.620)
Observations 551,707 551,707
R² 0.599 0.599
Adj. R² 0.599 0.599
Clusters (emis) 499 499
  • Relative importance (LMG) via calc.relimp with 10 bootstrap replicates (95% percentile CIs).
  • OLS coefficients with heteroskedasticity-consistent (HC1) SEs clustered at emis (reported as b(SE)).
  • Class size (ECS) model includes educators, learners, and school/learner covariates.
  • Scenario models control for class size, educators, learners, and the same covariates; prov excluded when DV is prov_class_size_reduc.
  • Categorical regressors (Province, Race, Quintile, Grade, Gender) are treated as factors; LMG is reported at the group level, OLS at factor-level coefficients.
  • Significance: *** p<0.01, ** p<0.05, * p<0.10.
  • Scenario 3 (classroom_reduc) excluded from all analyses due to high (90%+) missingness.
  • Scenario 6 (prov_class_size_reduc) is excluded from the class-size specification to avoid collinearity with province effects when provinces share pooling constraints.
  • Province coefficients may be NA where province effects are collinear with province-driven structure, likely because provinces share a binding factor when educators are pooled provincially.