Legal aid eligibility and court outcomes: a designbased double-machine-learning approach

The UN General Assembly recognises the right to equal treatment before the law as a universal human
right. A minimum requirement to uphold this right is to guarantee access to a lawyer independently of
one’s income, and this is what legal aid does. While legal aid is necessary to reduce income-based
inequality before the law, the extent to which this is achieved depends on the quality of its services. In a
context where all defendants have access to a lawyer, I study the impact of refusing legal aid on court
outcomes. This is a conservative estimate of the performance gap between privately and publicly-funded
legal services. I do this by taking a double machine learning approach to a new administrative dataset
linking data from legal aid application forms to their court outcomes in New South Wales, Australia.
Access to the legal aid application forms, which include all the factors used in the decision to provide aid,
allows me to learn the unknown treatment assignment function and identify treatment effects via the
random forests algorithm. I find that aid applicants who fail the means test and hire a private lawyer are 10
p.p. less likely to be incarcerated than if they passed it and relied on legal aid. With an average positive
incarceration length of almost 4 years, this gap is consequential. Finally, I argue that this approach,
combining high quality administrative data with double machine learning, may allow policy-makers to
track the performance of public programs that are hard to evaluate using standard econometric
approaches.