Eoghan O’Neill, University of Cambridge
In this paper, we apply some recently developed methods for describing the heterogeneity of treatment effects to an electricity smart metering trial dataset. We use these methods to investigate how household demand response to Time of Use (TOU) electricity pricing schemes varies with survey variables and past consumption data.
There is a need to better understand the distributional implications of energy policies. For example, the impact of TOU tariffs on households depends upon how they currently use electricity, their ability to allocate demand to off-peak periods and the specific structure of the tariffs. Consumers in different socioeconomic groups, with different incomes or behavioural characteristics may react in different ways to the introduction of TOU tariffs. Similarly, customers with distinct historical intra-day load profiles, will respond differently to the introduction of tariffs that charge different prices for electricity at different times of the day.
Household electricity demand response is an example of a policy-relevant application where tree based methods for describing heterogeneity are appealing. A report produced by the Centre for Sustainable Energy for Ofgem (CSE 2012) indicates how policy makers are not simply interested in the average effect of a policy, but are interested in “going beyond the mean” and identifying groups of customers most strongly affected by a policy change.
Heterogeneous treatment effects are described in this paper by estimates of Conditional Average Treatment Effects, which are the expected differences between treated and control households for subsets of the population defined by covariates. First, standard CATE estimates are produced for covariates believed a priori to be informative. Secondly, the method of causal trees (Athey &; Imbens 2016) is applied in order to search across many potential conditioning variables for aspects of heterogeneity that are possibly difficult to hypothesize a priori. Thirdly, we obtain individual-specific estimates from a causal forest.
This application illustrates a number of advantages and disadvantages common to the application of machine learning methods in econometrics, where the interest is in underlying mechanisms and the association between covariates and causal effects. These issues include: the choice of potentially significant, but interpretable, conditioning variables; the choice and interpretation of variable importance measures; and how to present individual treatment effect estimates. Issues related to the chosen methods include the trade-off between the interpretability of a single causal tree and the stability of estimates produced by a causal forest.Individual causal trees estimated with this dataset are found to have instable structure with respect to random subsampling of the data, and therefore do not provide a consistent characterisation of heterogeneity. However, household-specific estimates produced by a causal forest exhibit reasonable associations with covariates that support previous studies. For example, households that consume more electricity and households in which the chief earner has third level education are estimated to respond more to a new pricing scheme. In addition, there is a possibility that some aspects of past consumption information are of greater importance than survey information in producing these estimates.
Athey, S. & Imbens, G. (2016), `Recursive partitioning for heterogeneous causal effects’, Proceedings of the National Academy of Sciences 113(27), 7353-7360.
CSE (2012), “beyond average consumption” – development of a framework for assessing impact of policy proposals on different consumer groups, Final report to ofgem, Centre for Sustainable Energy.