Econometrics Journal最新文献

We revisit the classic semi-parametric problem of inference on a low-dimensional parameter θ₀ in the presence of high-dimensional nuisance parameters η₀. We depart from the classical setting by allowing for η₀ to be so high-dimensional that the traditional assumptions (e.g. Donsker properties) that limit complexity of the parameter space for this object break down. To estimate η₀, we consider the use of statistical or machine learning (ML) methods, which are particularly well suited to estimation in modern, very high-dimensional cases. ML methods perform well by employing regularization to reduce variance and trading off regularization bias with overfitting in practice. However, both regularization bias and overfitting in estimating η₀ cause a heavy bias in estimators of θ₀ that are obtained by naively plugging ML estimators of η₀ into estimating equations for θ₀. This bias results in the naive estimator failing to be consistent, where N is the sample size. We show that the impact of regularization bias and overfitting on estimation of the parameter of interest θ₀ can be removed by using two simple, yet critical, ingredients: (1) using Neyman-orthogonal moments/scores that have reduced sensitivity with respect to nuisance parameters to estimate θ₀; (2) making use of cross-fitting, which provides an efficient form of data-splitting. We call the resulting set of methods double or debiased ML (DML). We verify that DML delivers point estimators that concentrate in an -neighbourhood of the true parameter values and are approximately unbiased and normally distributed, which allows construction of valid confidence statements. The generic statistical theory of DML is elementary and simultaneously relies on only weak theoretical requirements, which will admit the use of a broad array of modern ML methods for estimating the nuisance parameters, such as random forests, lasso, ridge, deep neural nets, boosted trees, and various hybrids and ensembles of these methods. We illustrate the general theory by applying it to provide theoretical properties of the following: DML applied to learn the main regression parameter in a partially linear regression model; DML applied to learn the coefficient on an endogenous variable in a partially linear instrumental variables model; DML applied to learn the average treatment effect and the average treatment effect on the treated under unconfoundedness; DML applied to learn the local average treatment effect in an instrumental variables setting. In addition to these theoretical applications, we also illustrate the use of DML in three empirical examples.

We explore the asymptotic properties of strategic models of network formation in very large populations. Specifically, we focus on (undirected) exponential random graph models. We want to recover a set of parameters from the individuals' utility functions using the observation of a single, but large, social network. We show that, under some conditions, a simple logit-based estimator is coherent, consistent and asymptotically normally distributed under a weak version of homophily. The approach is compelling as the computing time is minimal and the estimator can be easily implemented using pre-programmed estimators available in most statistical packages. We provide an application of our method using the Add Health database.

We provide novel, high-order accurate methods for non-parametric inference on quantile differences between two populations in both unconditional and conditional settings. These quantile differences correspond to (conditional) quantile treatment effects under (conditional) independence of a binary treatment and potential outcomes. Our methods use the probability integral transform and a Dirichlet (rather than Gaussian) reference distribution to pick appropriate L-statistics as confidence interval endpoints, achieving high-order accuracy. Using a similar approach, we also propose confidence intervals/sets for vectors of quantiles, interquantile ranges and differences of linear combinations of quantiles. In the conditional setting, when smoothing over continuous covariates, optimal bandwidth and coverage probability rates are derived for all methods. Simulations show that the new confidence intervals have a favourable combination of robust accuracy and short length compared with existing approaches. Detailed steps for confidence interval construction are provided in online Appendix E as supporting information, and code for all methods, simulations and empirical examples is provided.

This paper considers the adaptability of estimation methods for binary response panel data models to multiple fixed effects. It is motivated by the gravity equation used in international trade, where important papers use binary response models with fixed effects for both importing and exporting countries. Econometric theory has mostly focused on the estimation of single fixed effects models. This paper investigates whether existing methods can be modified to eliminate multiple fixed effects for two specific models in which the incidental parameter problem has already been solved in the presence of a single fixed effect. We find that it is possible to generalize the conditional maximum likelihood approach to include two fixed effects for the logit. Monte Carlo simulations show that the conditional logit estimator presented in this paper is less biased than other logit estimators without sacrificing on precision. This superiority is emphasized in small samples. An application to trade data using the logit estimator further highlights the importance of properly accounting for two fixed effects.

Economists are often interested in identifying effective policies or treatments together with subpopulations of individuals who respond positively (or with a sign that is expected) to these treatment interventions. In this paper, we propose an optimal false discovery rate controlling method that is especially useful for such one-sided testing problems. The proposed procedure is optimal in the sense of minimizing the false non-discovery rate while controlling the false discovery rate at a pre-specified level; it uses a deconvolution method based on non-parametric maximum likelihood estimation, which allows for a broader class of treatment effect distributions than existing methods do. The proposed test demonstrates good small-sample performance in Monte Carlo simulations and it is applied to study the effect of attending a more selective high school in Romania. The application reveals strong evidence of treatment effect heterogeneity, in that students who marginally gain access to higher-ranked schools are more likely to benefit if the higher-ranked school has a relatively high admission score cut-off – or, in other words, is more selective.

In this paper, we analyse the time series of 12,000+ networks of traders in the E-mini S&P 500 stock index futures contract and we empirically link network variables with financial variables more commonly used to describe market conditions. We show that network variables lead trading volume, intertrade duration, effective spreads, trade imbalances and other market liquidity measures. Network variables reflect information, information asymmetry and market liquidity and significantly presage future market conditions prior to volume or liquidity measures. We also find two-way Granger-causality between network variables and both returns and volatility, highlighting strong feedback between market conditions and trading behaviour.

Consider estimating the slope coefficients of a fixed-effect binary-choice model from two-period panel data. Two approaches to semiparametric estimation at the regular parametric rate have been proposed: one is based on a sufficiency requirement, and the other is based on a conditional-median restriction. We show that, under standard assumptions, both conditions are equivalent.

GARCH(1,1) models are widely used for modelling processes with time-varying volatility. These include financial time series, which can be particularly heavy tailed. In this paper, we propose a novel log-transform-based least-squares approach to the estimation of GARCH(1,1) models. Within this approach, the scale of the estimated volatility is dependent on an unknown tuning constant. By means of a backtesting exercise on both real and simulated data, we show that knowledge of the tuning constant is not crucial for Value at Risk prediction. However, this does not apply to many other applications where correct identification of the volatility scale is required. In order to overcome this difficulty, we propose two alternative two-stage least-squares estimators and we derive their asymptotic properties under very mild moment conditions for the errors. In particular, we establish the consistency and asymptotic normality at the standard convergence rate of for our estimators. Their finite sample properties are assessed by means of an extensive simulation study.