Victor Yaneng Wang
Hi there! I'm a first-year PhD student in Information Technology at MIT.
My primary research interests are in econometrics, machine learning and AI. I am also interested in behavioral and experimental economics.
My research explores how intelligent agents can use scientific evidence to learn about the world around them, and how this may subsequently inform the way they communicate and make decisions. I believe this has fruitful applications to a broad range of areas, from policymaking to AI.
Previously, I was a Predoctoral Research Fellow in Economics at the Global Priorities Institute, University of Oxford. I also completed an MPhil in Economic Research and a BA in Economics at the University of Cambridge.
If you have any thoughts or questions I'd love to hear from you! My academic email is vyw20 [at] mit [dot] edu.
WORKING PAPERS
Finding Subgroups with Significant Treatment Effects
With Jann Spiess (Stanford) and Vasilis Syrgkanis (Stanford).
Abstract: Researchers often run resource-intensive randomized controlled trials (RCTs) to estimate the causal effects of interventions on outcomes of interest. Yet these outcomes are often noisy, and estimated overall effects can be small or imprecise. Nevertheless, we may still be able to produce reliable evidence of the efficacy of an intervention by finding subgroups with significant effects. In this paper, we propose a machine-learning method that is specifically optimized for finding such subgroups in noisy data. Unlike available methods for personalized treatment assignment, our tool is fundamentally designed to take significance testing into account: it produces a subgroup that is chosen to maximize the probability of obtaining a statistically significant positive treatment effect. We provide a computationally efficient implementation using decision trees and demonstrate its gain over selecting subgroups based on positive (estimated) treatment effects. Compared to standard tree-based regression and classification tools, this approach tends to yield higher power in detecting subgroups affected by the treatment.
Numbers Tell, Words Sell
With Michael Thaler (UCL) and Mattie Toma (Warwick).
Abstract: When communicating numeric estimates with policymakers, journalists, or the general public, experts must choose between using numbers or natural language. We run two experiments to study whether experts strategically use language to communicate numeric estimates in order to persuade receivers. In Study 1, senders communicate probabilities of abstract events to receivers on Prolific, and in Study 2 academic researchers communicate the effect sizes in research papers to government policymakers. When experts face incentives to directionally persuade instead of incentives to accurately inform receivers, they are 25-29 percentage points more likely to communicate using language rather than numbers. Experts with incentives to persuade are more likely to slant language messages than numeric messages in the direction of their incentives, and this effect is driven by those who prefer to use language. Our findings suggest that experts are strategically leveraging the imprecision of language to excuse themselves for slanting more. Receivers are persuaded by experts with directional incentives, particularly when language is used.
Estimating Long-Term Treatment Effects without Long-Term Outcome Data
With David Rhys Bernard (Open Philanthropy) and Jojo Lee (Oxford).
Abstract: The surrogate index method allows policymakers to estimate long-run treatment effects before long-run outcomes are observable. We meta-analyse this approach over nine long-run RCTs in development economics, comparing surrogate estimates to estimates from actual long-run RCT outcomes. We introduce the M-lasso algorithm for constructing the surrogate approach’s first-stage predictive model and compare its performance with other surrogate estimation methods. Across methods, we find a negative bias in surrogate estimates. For the M-lasso method, in particular, we investigate reasons for this bias and quantify significant precision gains. This provides evidence that the surrogate index method incurs a bias-variance trade-off.
OTHER WRITINGS
Heterogeneous Effects of US Medicaid Expansions on Health Insurance Coverage
MPhil dissertation. Published in Cambridge student journal Finance and Economics Vision.
Abstract: We use cross-sectional data from the 2014 American Community Survey to estimate the heterogeneous treatment effects of expanding Medicaid on health insurance coverage. In doing so, we provide robust evidence which can help policymakers target future Medicaid policies towards particularly responsive individuals. Medicaid expansions were optional by state, allowing us to identify treatment effects by comparing expansion and non-expansion states. We then estimate heterogeneous treatment effects using a non-parametric machine learning algorithm called a causal forest, which offers several advantages over prior methods in the literature. Most notably, it provides a systematic means to discover, from the data, which variables are most relevant for modelling heterogeneity. We find strong evidence of heterogeneity in our estimated treatment effects. Furthermore, we find that individuals with the largest treatment effects were typically aged 25-34; had a high school diploma as their highest level of education; spoke a language other than English at home; and/or were in private for-profit employment.
Finnegans Wake as an AI's stream-of-consciousness
Abstract: In this article, I argue that Finnegans Wake can be read as a poetic attempt to imagine the hypothetical stream-of-consciousness of a Large Language Model (LLM), a modern type of AI algorithm underlying popular chatbots like ChatGPT. I concentrate on four particular characteristics of the Wake. Firstly, its extensive use of puns and portmanteaus parallels the neuron polysemanticity observed in LLMs, where neurons can encode multiple concepts at once. Secondly, the Wake draws upon many world languages and literatures, similar to an LLM’s training corpus. Thirdly, the Wake’s characters do not have a fixed personality, but shift between multiple personas, just as LLMs like ChatGPT can adopt different personalities based on the user’s promptings. Finally, I argue that both Finnegans Wake and modern LLMs adopt a deeply cyclical approach to language.