Causal treatment effects in time series: CO2 emissions and energy consumption effect on GDP

Abstract

This study broadens the research on the use of neural networks in conjunction with causality techniques for economic policy development. Empirical evidence is presented on the effects of energy consumption and carbon emissions on per capita economic growth with unbalanced panel data for 94 countries between 1971 and 2018. We present a methodological framework that employs the transformation of time series into a set of treatment intervals according to potential policies that a country may adopt. Based on treatment intervals, a two-stage model based on recurrent neural networks (Long Short-Term Memory) is proposed together with a multiclass classifier to obtain propensity scores. We then estimate the effects using the dynamic potential outcomes under a framework of treatment-based causality. The results indicate that for both energy consumption and carbon emissions policies, the treatment range that generates the best performance is between −2% and 0.4%; moreover, in both cases, extreme policies such as drastically reducing or increasing energy consumption generates the worst results for economic development. We find that when grouping countries according to income level, mid-low and low levels of income have better development under moderate growth policy with respect to levels of carbon emissions and energy consumption, supporting the Environmental Kuznets Curve hypothesis. This finding could be related to the fact that these countries do not have the level of infrastructure to sustain growth. However, the impact of this growth on pollution would not be comparable to the reduction effect of high-income countries, given that they make the largest overall contribution to global emissions.