Econometrics Workshop 2024
This is a website of the Econometrics Workshop at the Department of Economics of the University of Melbourne.
Information
Venue:
The Bishop’s Parlour at Abbotsford Convent
1 St Heliers Street, Abbotsford 3067Time: Friday, 22 November, 2024
Program
9:00 - 9:30 | Morning coffee at Convent Bakery
9:30 - 10:30 | Michael Fan (Chinese University of Hong Kong)
Inference for Nonlinear Endogenous Treatment Effects Accounting for High-Dimensional Covariate Complexity
10:30 - 11:00 | Coffee break at Convent Bakery
11:00 - 11:30 | Sami Stouli
Identification of Treatment Effects under Limited Exogenous Variation
11:30 - 12:00 | Guo Yan
On the Singularity of Machine Learning Methods for Predicting Imbalanced and Noisy Binary Data
12:00 - 13:30 | Lunch at Kappaya
13:30 - 14:30 | Andrés Ramírez Hassan (Universidad EAFIT)
Where do unarrested criminals live? A Spatial one-sided errors model
14:30 - 15:00 | Zheng Fan
Dynamic Shrinkage and Selection in Bayesian Predictive Synthesis for Exchange Rate Forecasting
15:00 - 15:30 | Tea break at Cam’s
15:30 - 16:00 | Chris Skeels
Beyond Hypergeometric Functions For Economists
16:00 - 16:30 | Joe Hirschberg
The impact of Net Feed-in-Tariffs on the Electricity Consumption by Households with Photovoltaic Panels
16:30 - 17:00 | Alexandra de Gendre
Detecting Local Non-Compliance with Random Treatment Assignment in Quasi-Experiments, with an Application to Ability Peer Effects
18:00 | Dinner at Transformer Next Door
Abstracts
Inference for Nonlinear Endogenous Treatment Effects Accounting for High-Dimensional Covariate Complexity
Michael Fan
Nonlinearity and endogeneity are prevalent challenges in causal analysis using observational data. This paper proposes an inference procedure for a nonlinear and endogenous marginal effect function, defined as the derivative of the nonparametric treatment function, with a primary focus on an additive model that includes high-dimensional covariates. Using the control function approach for identification, we implement a regularized nonparametric estimation to obtain an initial estimator of the model. Such an initial estimator suffers from two biases: the bias in estimating the control function and the regularization bias for the high-dimensional outcome model. Our key innovation is to devise the double bias correction procedure that corrects these two biases simultaneously. Building on this debiased estimator, we further provide a confidence band of the marginal effect function. Simulations and an empirical study of air pollution and migration demonstrate the validity of our procedures.
Identification of Treatment Effects under Limited Exogenous Variation
Sami Stouli
Multidimensional heterogeneity and endogeneity are important features of a wide class of econometric models. With control variables to correct for endogeneity, nonparametric identification of treatment effects requires strong support conditions. To alleviate this requirement, we consider varying coefficients specifications for the conditional expectation function of the outcome given a treatment and control variables. This function is expressed as a linear combination of either known functions of the treatment, with unknown coefficients varying with the controls, or known functions of the controls, with unknown coefficients varying with the treatment. We use this modeling approach to give necessary and sufficient conditions for identification of average treatment effects. A sufficient condition for identification is conditional nonsingularity, that the second moment matrix of the known functions given the variable in the varying coefficients is nonsingular with probability one. For known treatment functions with sufficient variation, in triangular models with discrete instrument we find that average treatment effects cannot be identified when the number of support points for the instrument is less than the number of coefficients. For known functions of the controls, we find that average treatment effects can be identified in general nonseparable triangular models with binary or discrete instruments. We extend our analysis to flexible models of increasing dimension and relate conditional nonsingularity to the full support condition of Imbens and Newey (2009), thereby embedding semi- and non-parametric identification into a common framework.
On the Singularity of Machine Learning Methods for Predicting Imbalanced and Noisy Binary Data
Guo Yan
When analyzing binary choice problems with imbalanced classes, where the binary outcome falls into one class more frequently than another, many econometric and machine learning methods tend to underpredict the realization of the minority class. The issue can be more severe when the explanatory variables have a low signal strength for predicting the binary outcome. In this paper, we present a framework to analyze how fast a general method’s predictive performance degenerates in the following sense: It overpredicts the realization of the majority class and approaches a singularity point, at which it predicts only the majority class or purely randomly. We establish the degeneracy rate of the SVM and logistic regression, two widely used machine learning methods for classification, under different settings on the degree of class imbalance and the noise-to-signal ratio. We find that the SVM degenerates at least exponentially fast when both the degree of class imbalance and the noise-to-signal ratio increase, whereas logistic regression has a degeneracy rate of√n. Our results have important implications for selecting appropriate methods to analyze binary choice problems with imbalanced and noisy data.
Where do unarrested criminals live? A Spatial one-sided errors model
Andrés Ramírez Hassan
This paper analyzes the rates of arrested individuals living per neighborhood in Medellín (Colombia) to identify neighborhoods where relatively more uncaptured criminals potentially live. We exploit a novel dataset consisting of confidential reports of residential addresses of individuals at the moment of being arrested by the police. We find that unarrested hot regions depend both on the crime typology and space. Particularly, we find that homicide and drug dealing are strongly linked, whereas there are common links between car and motorcycle thefts. Our model takes into account persistent and transient one-sided error components and spatial effects within a longitudinal data structure. Simulation exercises suggest that our proposal has good finite sample performance. In particular, we find that the predictive intervals have accurate coverage which is robust to endogeneity. Moreover, our proposal has good predictive accuracy, since the hot regions are well-classified even in the presence of endogeneity.
Dynamic Shrinkage and Selection in Bayesian Predictive Synthesis for Exchange Rate Forecasting
Zheng Fan
I propose a novel methodology that consolidates dynamic sparsity and the dynamic shrinkage process within a coherent framework. The approach introduces an adaptive dynamic variable selection mechanism to achieve time-varying sparsity. This enhances the standard dynamic shrinkage process, which emphasizes both global and local shrinkage on coefficient drift, by incorporating a selection mechanism that dynamically identifies relevant predictors based on their evolving importance over time, effectively adapting to shifting economic conditions and reducing overfitting. To achieve optimal predictions, the Bayesian predictive synthesis approach is also employed to balance weights across forecasts generated by multiple monetary models. I further introduce a contemporaneous shrinkage prior over the vector space of dynamic weights, structured as a VECM, which encourages the weights to sum toward one without imposing strict constraints. This methodology effectively addresses model instabilities and dynamically adjusts the relevance of each forecast, resulting in more reliable exchange rate predictions. Empirical analysis using exchange rate data demonstrates the effectiveness of this approach, showing significant improvements in forecast performance compared to conventional methods.
Beyond Hypergeometric Functions For Economists
Chris Skeels
In this paper we explore a class of functions that has not, to the best of our knowledge, previously appeared in the econometric literature. The so-called Meijer G functions extend the generalized hypergeometric functions in situations where the latter functions are not well defined. Our interest in these functions arose during an exploration of certain generalized empirical likelihood estimators which we shall use as an econometric illustration of the use of these functions.
The impact of Net Feed-in-Tariffs on the Electricity Consumption by Households with Photovoltaic Panels
Joe Hirschberg
TBA
Detecting Local Non-Compliance with Random Treatment Assignment in Quasi-Experiments, with an Application to Ability Peer Effects
Alexandra de Gendre
We develop a method to detect local non-compliance with a random treatment to recover valid quasi-experiments from existing data. Our approach combines simulation-based methods and latent class modelling in an intuitive way, can detect and characterize non-compliant risk sets, and is computationally undemanding. To illustrate its usefulness, we use it to estimate ability peer effects in Taiwan, where we have unusually rich data and a national mandate to randomly assign students to classrooms within schools, but with partial non-compliance. After recovering a valid quasi-experiment, we first estimate ability peer effects in line with other studies. We then document precisely estimated null effects on 18 potential mechanisms, many of which have been hypothesized but never tested. Our application shows how to use our method to expand the causal evidence base using existing datasets.