Energy and human development as a long run correlate of carbon dioxide emissions in Saudi Arabia

Abstract We assess the long-run correlates of Saudi Arabian CO 2 emissions over 1970–2023 in a setting that combines three difficulties: severe multicollinearity among proximate development indicators, structural breaks that distort unit-root and cointegration inference, and a short annual sample. We aggregate energy supply, urban population and human capital into a single principal-component composite (LNEHD) and estimate its long-run association with per-capita emissions using a battery of cointegration estimators. Because the composite is built from integrated series, we address the Phillips 1 critique directly: we compare loadings estimated on levels with loadings estimated on first differences, benchmark the explained-variance share against a random-walk placebo distribution, and verify that the headline coefficient is invariant to the construction method. The levels composite captures 91.94% of joint sub-component variance, a share exceeded by only 3.0% of draws from three independent drifting random walks. It reduces variance inflation factors from above 200 to below 6. Evidence on cointegration is mixed by construction and we report it as such: the Pesaran–Shin–Smith bounds F-statistic does not reject the null of no level relationship against finite-sample critical values simulated for T = 54 (F = 3.19; bootstrap p = 0.17); the Gregory–Hansen regime-shift test rejects at the 5% level against critical values simulated for this sample size, and the Hatemi-J two-break test does not reject. We therefore describe the long-run relationship as an association that the cointegration battery does not firmly establish, rather than as established cointegration. Across six long-run estimates, the response of emissions to the composite is positive and significant, between 0.03 and 0.22. Because the composite is a standardised index rather than a logarithm, we describe this as the response of log emissions to a unit movement in the index and not as an elasticity. The nonlinear specification yields β⁺ = 0.097 and β⁻ = 0.224, but the long-run symmetry restriction is not rejected at the 5% level under any covariance estimator ( p = 0.058 conventional, 0.079 HAC, 0.197 heteroscedasticity-robust). A Monte Carlo experiment shows the symmetry test has roughly 40% power at the estimated effect size, so we describe the ratchet pattern as suggestive rather than an established result.