6 Discussion
6.1 Interpretation
Three findings emerge. First, fractional displacement is confirmed for Taiwan. The elasticity of \(-0.039\) is negative, consistent with the theoretical expectation that near-zero marginal cost renewables displace fossil generation in the dispatch order, but its magnitude is negligible relative to one-for-one substitution. This aligns with York (2012) and with the finding of Karlilar Pata & Balcilar (2024) that displacing one unit of fossil capacity requires about 1.15 units of renewable capacity.
Second, intermittency is the operative mechanism, not a secondary consideration. The magnitude of the gap between dispatchable and variable renewables quantifies the system integration cost in an isolated grid. Fossil-fuelled spinning reserve must be maintained to compensate for the variability of solar and wind, and that reserve requires imported fuel whether or not it is called upon.
Third, the additionality paradox is confirmed but in a specific form. The electricity consumption elasticity of \(+0.555\) is fourteen times the magnitude of the renewable elasticity. Demand growth is the dominant driver of import volumes, and renewable capacity has functioned largely to serve incremental load rather than to retire existing fossil assets.
6.2 Policy implications
If renewable expansion is chasing demand growth rather than displacing fossil fuels, supply-side capacity targets alone are insufficient for genuine decarbonization, and two priorities follow.
The first is demand-side management. With a consumption elasticity fourteen times the renewable elasticity, curbing load growth is a considerably more efficient lever on import dependence than adding capacity, and the classical case for utility-led demand-side management applies directly (Gellings, 1985).
The second is grid flexibility over headline capacity. Since variable renewables show no measurable displacement effect while dispatchable renewables do, the binding constraint is the ability to convert intermittent output into firm supply. Utility-scale storage, demand response and grid modernization address that constraint; further capacity additions without them will continue to yield elasticities near zero. The policy implication is not that solar and wind investment is misdirected, but that it is incomplete: the complementary flexibility infrastructure determines whether that investment translates into reduced import dependence.
6.3 Findings that warrant caution
Three results should be read with care rather than presented without qualification.
The error correction coefficient exceeds unity in absolute value (\(\lambda_1 = -1.653\), and between \(-1.05\) and \(-1.69\) across specifications). This indicates overshooting rather than instability: since \(|1 + \lambda_1| = 0.653 < 1\), the system converges, though it oscillates while doing so. The behaviour is consistent with the shipment lumpiness documented in the data chapter, where monthly imports overshoot and revert. It is not an artefact of monthly frequency alone, since the quarterly specification also returns \(\lambda_1 = -1.363\).
The fuel price coefficient is positive (\(+0.047\), \(p < 0.001\)). Read literally, more expensive fuel is associated with larger imports, which is counterintuitive. Two explanations are plausible: prices and quantities are jointly driven by demand, producing simultaneity; and long-term contract pricing means the current benchmark price is not the price actually paid for current deliveries. The magnitude is small and the coefficient is a control rather than an object of interest, but it is reported rather than suppressed.
The consumption elasticity is unstable across specifications, at \(+0.555\) in the primary specification, \(+1.362\) over the full sample, and \(-0.054\) in the all-fuel variant. The all-fuel result is explicable as contamination, as discussed in the previous chapter. The difference between the transition and full samples is larger than would be comfortable and suggests the demand–import relationship itself has changed over 26 years.
6.4 Limitations
Several limitations should be acknowledged.
The LNG calorific value is not drawn from the official table, because the published figure is on a volume basis while the import data are on a mass basis, and the conversion factor is unavailable. The international convention of 52 GJ/t is used instead. Since LNG constitutes 52.5% of \(FFI^{power}\) by 2025, this is the largest single measurement uncertainty in the dependent variable. Sensitivity analysis indicates the substantive conclusions are unaffected, but obtaining the official regasification factor would improve precision.
Temperature is not controlled. Seasonal dummies capture the average summer but not an unusually hot one, so the consumption coefficient carries some climate noise.
The analysis uses generation rather than capacity data, which limits direct comparability with Karlilar Pata & Balcilar (2024), whose 1.15 figure is capacity-based.
Nuclear power enters as a share rather than in logarithms, departing from the original research proposal, and its coefficient is therefore a semi-elasticity rather than an elasticity.
Finally, the transition sample spans 121 months, of which 119 are usable after lagging, against 25 estimated parameters. The bounds test critical values are simulated at that effective sample size precisely because asymptotic approximations are unreliable here, but the statistical power to distinguish a small negative elasticity from zero remains limited. This is the reason \(\eta_{RE}\) sits at the boundary of conventional significance while the control variables are estimated precisely, and it is why the substantive interpretation rests on the magnitude of the estimate and the consistency of its sign across specifications rather than on a single \(p\) value.