什么时候吃够了?评估培训计划时的风险规避和不平等规避外文翻译资料

Evaluation Only. Created with Aspose.Pdf. Copyright 2002-2014 Aspose Pty Ltd.

WHEN IS ATE ENOUGH? RISK AVERSION AND INEQUALITY

AVERSION IN EVALUATING TRAINING PROGRAMS *

Rajeev Dehejia

Columbia University and NBER

rd247@columbia.edu

This paper explores the relationship between the theory and practice of program

evaluation as it relates to training programs. In practice programs are evaluated by mean-

variance comparisons of the empirical distributions of the outcome of interest for the

treatment and control programs. Typically, earnings are compared through the average

treatment effect (ATE) and its standard error. In theory, programs should be evaluated as

decision problems using social welfare functions and posterior predictive distributions for

outcomes of interest. This paper considers three issues. First, under what conditions do

the two approaches coincide? I.e., when should a program be evaluated based purely on

the average treatment effect and its standard error? Second, under more restrictive

parametric and functional form assumptions, the paper develops intuitive mean-variance

tests for program evaluation that are consistent with the underlying decision problem.

Third, these concepts are applied to the GAIN and JTPA data sets.

First version: 4 April 2000

Current version: 27 March 2003

* The author gratefully acknowledges conversations and collaboration with Joshua Angrist that sparked and

helped to refine this research and comments from Alberto Abadie, Gary Chamberlain, Richard Ericson,

Andrew Gelman, Jinyong Hahn, Gur Huberman, Charles Jones, David Kranz, Adriana Lleras-Muney, and

Dale Poirier. Thanks to seminar participants at Brown, Columbia, EUI Florence, UC Irvine, and University

of Wisconsin, Madison.

Evaluation Only. Created with Aspose.Pdf. Copyright 2002-2014 Aspose Pty Ltd.

When is ATE enough? Rules of Thumb vs. Decision Analysis in Evaluating

Training Programs

1. Introduction

Program evaluation is typically carried out by considering the average treatment effect

(ATE) of a new program under consideration (called the treatment) relative to a status

quo program (called the control). This is true both in experimental settings where the

ATE can be estimated by a simple difference in means for outcomes of interest between

the treatment and control groups, and also in non-experimental settings where the ATE is

often a parameter in a much more complicated model. Uncertainty regarding the

treatment impact is summarized by the standard error of the ATE, often through the

statistical significance of the point estimate.

In contrast, decision theory offers a more comprehensive method for evaluating

programs. A decision-theoretic analysis leads to a choice that maximizes (minimizes)

expected utility (loss) given a likelihood model of the data. This result was first

established by Wald (1950) and has led many researchers to formalize and extend the

decision-theoretic framework. Of course, one might argue that this formalized

statement of objectives misses intangible features of the decision problem. It is

clear, however, that the decision-theoretic framework, while adding a layer of

complexity to simple Neyman-Fisher hypothesis-testing, provides a solid

foundation for inference that explicitly links empirical estimates with the broader

framework of rational, economic decision-making.

1

Evaluation Only. Created with Aspose.Pdf. Copyright 2002-2014 Aspose Pty Ltd.

At a more practical level, a decision theory approach is more general along two

dimensions than simply looking at ATE. First, it accounts for uncertainty regarding the

treatment impact in a systematic way, through the decision-makerrsquo;s risk attitude in an

expected utility setting. To the extent that the Neumann-Morgenstern (1944) / Wald

(1950) approach is widely accepted in economics, the decision theoretic approach reflects

how we should account for uncertainty. Second, the decision approach allows for the

decision-maker to exhibit inequality aversion, which would lead him or her to consider

the treatment impact on features of the distribution other than the average.

The aim of this paper is to develop rules of thumb for evaluating programs which

are consistent with the decision framework. Interpreted literally – as adopting a program

when its treatment effect is positive and statistically significant, we show that the

traditional approach to evaluating programs is valid only under very strong assumptions.

We then relax these assumptions to develop simple techniques – rules of thumb – for

evaluating programs that are valid under more general conditions.

The paper proceeds as follows. In Section 2 we set up the general framework for

evaluation. In Section 3 we establish conditions for equivalence between the traditional

approach and the decision approach. In Section 4 we develop several rules of thumb for

evaluating programs. In Section 5 we apply these rules in evaluating the GAIN and JTPA

data sets. Section 6 concludes.

2

Evaluation Only. Created with Aspose.Pdf. Copyright 2002-2014 Aspose Pty Ltd.

2. A Framework for Evaluation

2.1 Decision Theory and Evaluation

In a world without uncertainty, the policy-maker adopts a social welfare function,

S(u1(y1), u2(y2), hellip;, uN(yN)) which, for the population of interest, n<!--

剩余内容已隐藏，支付完成后下载完整资料

资料编号：[502698]，资料为PDF文档或Word文档，PDF文档可免费转换为Word