Eliminating Bias in Treatment Effect Estimation Arising from Adaptively Collected Data
It is well understood that bandit algorithms that collect data adaptively - balancing between exploration and exploitation - can achieve higher average outcomes than the “experiment first, exploit later” approach of the traditional treatment choice literature. However, there has been much less work on how data arising from such algorithms can be used to estimate treatment effects. This paper contributes to this growing literature in three ways. First, a systematic simulation exercise characterizes the behavior of the standard average treatment effect estimator on adaptively collected data: I show that treatment effect estimation suffers from amplification bias and illustrate that this bias increases in noise and adaptivity. I also show that the traditional correction method of inverse propensity score weighting (IPW) can even exacerbate this bias. Second, I suggest an easy-to-implement bias correction method: limiting the adaptivity of the data collection by requiring sampling from all arms results in an unbiased IPW estimate. Lastly, I demonstrate a trade-off between two natural goals: maximizing expected welfare and having a good estimate of the treatment effect. I show that my correction method extends the set of choices regarding this trade-off, yielding higher expected welfare while allowing for an unbiased and relatively precise estimate.
Examining the Effect of Retirement on Cognitive Performance - A Unifying Approach
Several recent works investigate the effect of retirement on cognitive performance, arriving at different conclusions. The key ingredient of the various approaches is how they handle the endogeneity of the retirement decision. In order to examine this issue more deeply, I replicate the results of previous works using three waves from the Survey of Health, Ageing and Retirement in Europe (SHARE). I draw attention to potential biases inherent in the standard instrumental variable identification strategies and assess their magnitudes. Based on the lessons learned, I propose a new instrument that utilizes the panel structure of the data, enabling the comparison of individual cognitive paths. I show that if retirement has any adverse effect on cognitive performance it must be really small in magnitude.
Do Elite Schools Benefit Their Students?
Joint with S. Sóvágó
This paper studies the effects of enrollment in an elite school on elite-school students’ academic achievement in Hungary. Enrollment in a Hungarian elite school entails having academically stronger peers and early switching to a secondary school. We examine effects for elite-school students throughout the outcome distribution using a mild stochastic dominance assumption. We find that enrollment in an elite school decreases female and low-ability students’ mathematics test scores two years after enrollment. However, these negative effects are short-lived, and we obtain estimates that are consistent with substantial positive effects four years after enrollment. School value-added estimates lie within our non-parametric bounds, and confirm the positive effects on the medium run.
The Revolution in Economic Data and Panel Econometrics
Joint with L. Balázsi, G. Kézdi, and L. Mátyás
Közgazdasági Szemle (Economic Review-monthly of the Hungarian Academy of Sciences) 2014/11: pp. 1219-1340.
in Hungarian, original title: A közgazdasági adatforradalom és a panelökonometria.
The main aim of this survey is to present, through the history and development of panel data econometrics, the data revolution currently under way in economics. We are demonstrating that progress is happening from the bottom up: the emergence different panel data sets brings to light different questions and problems, which then inspire and bring new theoretical and methodological results, which in turn may result in new, more informative, more complex data sets. This feedback mechanism is the main engine behind the current dig data revolution, and as a result we are getting a deeper, more complex and nuanced picture of the economy.
Low employment among the 50+ population in Hungary: the role of incentives, health and cognitive capacities
Joint with G. Kézdi
In: A. Börsch-Supan et al. Active ageing and solidarity between generations in Europe, Berlin: de Gruyter. 2013.
Employment rate in Hungary among the 50+ population is among the lowest in Europe. Analyzing the first results of the Survey of Health, Ageing and Retirement in Europe (SHARE) for Hungary we make three main points regarding the employment gap: (1) Earnings in Hungary are very close to retirement income, creating incentives to retire early. (2) Hungarians are in significantly worse health than the people in most other European countries. (3) The employment gap between Hungary and Europe is largest among people with bad health and low skills.
Career or Family: an Inclusive or an Exclusive ’or’?
Forum Scientiarum Oeconomicarum, 2010/9: pp. 2-15. in Hungarian, original title: Munka vagy család: megengedő vagy kizáró vagy?
The Effects of Population Ageing on Pension System and Health Care.
Forum Scientiarum Oeconomicarum, 2007/2. pp. 83-120. in Hungarian, original title: Elöregedő társadalom és szociális rendszerek – merre tovább?