Discussion Paper
No. 2015-44 | June 16, 2015
Stijn Baert
Field Experimental Evidence on Gender Discrimination in Hiring: Biased as Heckman and Siegelman Predicted?

Abstract

Correspondence studies are nowadays viewed as the most compelling avenue to test for hiring discrimination. However, these studies suffer from one fundamental methodological problem, as formulated by Heckman and Siegelman (The Urban Institute audit studies: Their methods and findings. In M. Fix, and R. Struyk (Eds.), Clear and convincing evidence: Measurement of discrimination in America, 1993), namely the bias in their results in case of group differences in the variance of unobserved determinants of hiring outcomes. In this study, the authors empirically investigate this bias in the context of gender discrimination. They do not find significant evidence for the predicted bias.

JEL Classification:

J16, J71, M51, J41, C93

Links

Cite As

[Please cite the corresponding journal article] Stijn Baert (2015). Field Experimental Evidence on Gender Discrimination in Hiring: Biased as Heckman and Siegelman Predicted? Economics Discussion Papers, No 2015-44, Kiel Institute for the World Economy. http://www.economics-ejournal.org/economics/discussionpapers/2015-44


Comments and Questions



Anonymous - Referee Report 1
June 17, 2015 - 10:27
See attached file
[ Download attached file ]

Stijn Baert - Response to Referee Report 1
June 20, 2015 - 19:13 | Author's Homepage
Please find attached my response to Referee Report 1.
[ Download attached file ]

Anonymous - Comments
July 01, 2015 - 09:57
The paper deals with an interesting topic of identifying evidence of employer discrimination by gender/caste/ethnic group by controlling over these factors using correspondence analysis. It tries to address a major critique of this analysis pointed out by Heckman and Siegelman (HS) (1993) who pointed out that the group differences in the variance of unobserved productivity can lead to impossibility of identifying discrimination in the existing methodology. Neumark (2012) proposed a methodology to address the HS critique using a heteroscedastic probit model with a proper modification of the data. The author uses Neumark procedure to a dataset along with a new variable to identify group specific variance in observable determinants of productivity, namely, the distance between the candidate's living place and the workplace. It is claimed that this variable is in some sense better than those used in earlier studies. No significant evidence for HS bias was found. Thus the main contribution of the paper is the variable used in identification and it is shown empirically that it satisfies the requirement for the given data set. For general application one needs to know whether its performance is robust to the choice of the dataset. Like Neumark, it would have been better if some selection criterion and corresponding tests were developed before choosing the variable for identification. References: Heckman, J.J., and P. Siegelman (1993). The Urban Institute audit studies: Their methods and findings. In M. Fix, and R. Struyk (Eds.), Clear and convincing evidence: Measurement of discrimination in America. Washington DC: Urban Institute Press. Neumark, D. (2012). Detecting discrimination with audit andcorrespondence studies. Journal of Human Resources 47 (4):1128–1157.

Stijn Baert - Response to anonymous comment
July 10, 2015 - 18:40 | Author's Homepage
I sincerely thank the commenter for her/his interest in my study and for her/his enthusiasm for its relevance. I agree that the identifying variable we propose is -- besides the application of Neumark's approach to the context of gender discrimination -- the major contribution of the study. I propose this variable based on theoretical grounds and afterwards test its validity in an empirical way. I hope that, after publication of my study, many scholars using correspondence tests will adopt Neumark's framework when analysing their data and will identify it based on the variable I propose to use, which is easy to construct in each setting.

Anonymous - Comments
July 03, 2015 - 08:07
I have read the paper with interest. The problem is nuanced and technical, so a close reading was well worthwhile. I am happy with the content but there are two comments on the presentation:The writing is not very happy, there are too many long and complex sentences.Also, this is actually a companion paper of Baert et al. (forthcoming); so much so that no equations were needed to be stated. This fact may be highlighted in the abstract/title itself.

Stijn Baert - Response to anonymous comment
July 10, 2015 - 19:00 | Author's Homepage
I sincerely thank the commenter for her/his interest in my study and for her/his enthusiasm for its content. Concerning her/his first point, in the case of acceptance, I will keep these language comments in mind when revising the manuscript. In addition, if this is a condition for acceptance, I can have the study proofread by a professional language editing service. Concerning her/his second point, already in the initial version of the manuscript I acknowledge that I use data from the correspondence test conducted by Baert et al. (Forthcoming, ILR Review). Baert et al. (Forthcoming, ILR Review) is cited four times throughout the manuscript. In my opinion, the title and the abstract of the present study should be focussed on its own content and contribution and not on acknowledging the study from which it uses data. However, I am ready to highlight the data source already in the abstract if this is a condition for acceptance.

Anonymous - Table 1
July 06, 2015 - 12:56
On one hand, Table 1 is very concise which is very appreciable; on the other hand, something is / remain unclear. 1) From the table it might seem that you use interactions between female candidate and promotion / not promotion, instead (if I have well understood) you are re-running the estimates on two different sub-samples. 2) Is there any particular reason why the estimate for "distance between the workplace and candidate's residence" is not included in the table? 3) I do not understand panel D, for Model(2) and (4). You are decomposing the gender coefficient for both the sub-sample with promotion and for the sub-sample without promotion. To which of these sub-samples the statistics (log ratio of standard deviations and the two Wald test statistics) in the second box refer? There should be two sets of such statistics, for the two sub-samples. 4) Since also the variable (no)promotion varies among applications, it should be possible to estimate also its coefficient and add it in the identification assumption test, for Model(1) and (3).

Stijn Baert - Response to anonymous comment
July 10, 2015 - 20:11 | Author's Homepage
I sincerely thank the commenter for her/his interest in my study and for her/his comment on Table 1. However, the interpretation of the commenter is wrong. The regression models are explained in the main text of the initial manuscript at page 8 (from line 8 on): "On the one hand (models (1) and (3)), we regress positive callback on a dummy indicating female sex of the candidate and the distance between the workplace and the residence of the applicant. On the other hand, for models (2) and (4), the effect of female sex is broken down by whether the vacancy indicates a job implying a promotion in occupational level compared with the current job of the candidate." In other words, in models (2) and (4) we run regressions on the full dataset but instead of including one dummy indicating female sex as an explanatory variable, we include two dummies "female sex * no promotion job" and "female sex * promotion job" (in one regression). By doing that, we measure the effect of revealing female sex in the access to the first kind of jobs respectively in the access to the second kind of jobs (in one and the same regression model). This set-up is also indicated by the fact that the number of observations is constant across the regression models (as can be seen in the last row of Table 1). Having said that, in case the manuscript gets accepted, I will clarify the set-up of the regression models in a greater extent in the final version of the manuscript. In addition, I will rephrase the sentence "However, if the research sample is broken down by the occupational level of the posted job, we find that a female name lowers the probability of positive callback by four to five percentage points when they apply for jobs implying a promotion in this respect." This sentence might have been the reason for the commenter's confusion. Lastly, the variable "distance between the workplace and candidate's residence" is not included in the table as, for reasons of conciseness, only the variables of main interest (the total gender effect and its decomposition) are presented. Concerning the variable "distance between the workplace and candidate's residence" only the Wald test indicating that this variable is rewarded equally for males and females is important. Therefore, the p-value for this test is included in the table.

Anonymous - Referee Report 2
July 30, 2015 - 08:46
see attached file
[ Download attached file ]

Stijn Baert - Response to Referee Report 2
July 30, 2015 - 13:43 | Author's Homepage
Please find attached my response to Referee Report 2.
[ Download attached file ]