Essays on Tax Policy Optimization under Heterogeneous Treatment Effects

Abstract

This dissertation studies tax compliance and enforcement in developing-country settings, combining causal inference methodology with field evidence from Indonesia. It consists of three essays that address, respectively, a large-scale door-to-door enforcement campaign, a randomized messaging experiment, and the econometric methodology that underpins the first two empirical analyses.The first essay, Whom Should the Taxman Visit?, examines the effects and optimal design of a citywide door-to-door property tax enforcement campaign in Kota Gorontalo, Indonesia, in which field officers visited more than 30,000 registered properties between 2020 and 2021. To estimate average and heterogeneous treatment effects, the essay proposes a pseudo-treatment causal forest estimator that extends the R-learner framework to multi-period settings with time-invariant unobserved confounding. Verified visits raised on-time payment rates by 7.8 percentage points in the first post-campaign billing cycle, an effect that persists for at least four subsequent years and is especially pronounced among lower-value properties and taxpayers near the compliance margin. Individualized treatment-effect estimates are then used to evaluate three targeting strategies derived from explicit public-finance objectives: a participation-maximizing rule, a revenue-maximizing rule, and a conventional risk-based rule. The participation-maximizing strategy delivers 17.3 percentage-point gains at the top decile above random assignment; the revenue-maximizing strategy generates substantial additional fiscal returns; and the risk-based rule---the standard operational heuristic in tax administration guidance---performs near zero on both margins. Concentration-curve analysis shows that these rules differ not only in efficiency but in the distributional composition of whom they reach, making the equity--efficiency trade-off in algorithmic enforcement explicit and measurable.The second essay, Does Message Framing Matter for Tax Compliance?, reports the results of a randomized controlled trial conducted with the municipal government of Gorontalo City. Delinquent property owners were randomly assigned to receive a soft-tone WhatsApp message emphasizing civic duty and public benefits, a hard-tone message highlighting penalties and consequences, or no message. The soft-tone nudge raised current-year compliance by 9 to 11 percentage points from two weeks through the payment deadline, and the effect remained 9.9 percentage points higher at six months relative to a 38 percent control mean. Effects were entirely concentrated among historically low-compliance taxpayers, for whom the intervention effectively closed the compliance gap with higher-compliance peers. Framing mattered at longer horizons: the soft tone significantly outperformed the hard tone, whose effects dissipated over time in a pattern consistent with intertemporal substitution. The intervention cost less than two cents per message, making it highly scalable for resource-constrained local governments.The third essay, Pseudo-Treatment Debiasing with Orthogonal Scores in Multi-Period Panels, develops the econometric framework underlying the causal analyses in the first essay and contributes it as a general methodology for panel data settings with persistent unobserved confounding. The central insight is that pre-treatment placebo gaps---periods in which treatment has not yet occurred---provide repeated measurements of the latent selection bias in treated--control outcome comparisons. Under a bias-stability restriction equivalent to conditional parallel trends, the bias surface learned from pre-treatment periods can be subtracted from post-treatment gaps to identify the conditional average treatment effect on the treated. The essay implements this through a two-layer procedure combining within-period orthogonal score estimation with across-period pseudo-treatment debiasing, establishes large-sample inference results, introduces placebo diagnostics that directly test the identifying restriction, and proposes a sensitivity-analysis framework that quantifies how conclusions change under controlled violations of bias stability.Together, these essays advance the empirical and methodological foundations for designing, targeting, and evaluating compliance interventions in settings where administrative capacity is limited and causal identification is difficult, with direct implications for tax policy in developing countries.