Intersectional Discrimination and Candidate Selection: Evidence from a Lab in the Field Experiment in Kenya

Jonah Foong, Macartan Humphreys, Kimuli Kasara

2024-02-01

Outline

  • Motivation
  • Experiment hypotheses
  • Study design
  • Results
  • Discussion/Conclusion

Preview of Findings

  • Conducted a lab experiment in Nairobi to understand how gender and ethnic discrimination interact
  • Individuals select a candidate (individually or collectively) to perform a task
  • Three treatments
    • a gender quota
    • information on a candidate’s past performance
    • collective discussions in heterogeneous groups.
  • Findings
    • Baseline discrimination against women and outgroups, but no double-discrimination
    • Moreover no double discrimination in any treatment condition
    • Discussions remove biases for outgroups, but not so much for women

Intersectional discrimination

  • Seminal work by Crenshaw (1989); discrimination is not merely additive but more than the sum of its parts. This was first studied in relation to Black women’s employment and lack of legal recourse. It has since been extended to other areas across the social sciences.
  • Political scientists have focused on
  • Interactive interpretation: People who are discriminated against on one dimension are more or less likely to be discriminated on another dimension.
  • Agents: We focus on discrimination by individual actors, which aggregate to systemic discrimination. This lets us assess intersectional discrimination even on unranked dimensions.
  • A key point is that it is not obvious under what circumstances we would expect interactions.

Burying S.M.: An illustration

  • S.M. Otieno a prominent Luo Kenyan criminal lawyer whose burial dispute became famous in 1987
  • Wambui Otieno, Kikuyu, was taken to court in 1987 over the right right to bury him (and to his estate)
  • Can the fact that she lost the case be attributed to gender discrimination being more powerful than ethnic discrimination? Or to an interaction?

Burying S.M.: Possible counterfactual conditions

Surviving spouse is
Male Female
Surviving spouse is Kikuyu Counterfactual case: surviving spouse wins Actual case: surviving spouse loses
Luo Counterfactual case: surviving spouse wins Counterfactual case: surviving spouse wins

Burying S.M.: What this case taught us

  • Because they study ranked ethnic groups, most scholars tend to assume systemic intersectional biases reflect the preferences of men from dominant groups.
  • What we learn from this case, is that it is unclear who wanted to or felt able to support Wambui.
    • Luo women
    • Kikuyu men (generally)
    • Members of the Kikuyu political elite
  • In this case, systemic biases are harder to identify

Institutional conditions

We focus on conditions to assess the presence of intersectional discrimination and to assess whether double discrimination can be addressed, or exacerbated, by institutional features:

  • Gender quotas
  • Information on a candidate’s likely performance
  • Collective decision making

Quotas: Potential Mechanisms

  • Quotas produce double discrimination through constrained optimization
  • Outgroup biases are proportional in each group
  • No interaction without constraints
  • Intuition: When forced to be gender-egalitarian people are more likely to choose people from their own ethnic group.
Group A Group B Average Difference
Male 0.35 0.25 0.3 -0.1
Female 0.25 0.15 0.2 -0.1
Average 0.3 0.2 0.25 -0.2
Difference -0.1 -0.1 -0.2 Interaction: 0
Group A Group B Average Difference
Male 0.29 0.21 0.25 -0.08
Female 0.31 0.19 0.25 -0.12
Average 0.3 0.2 0.25 -0.02
Difference 0.02 -0.02 0 Interaction: -0.04

Information

Information

  • If double discrimination is statistical, information on quality limits discrimination (Unless updating on an identity dimension is biased)
  • Effects depend on whether we are concerned with overall inequality in treatment or inequality given type

Discussions

Multiple mechanisms

  1. they are informative
  2. they invoke egalitarian norms
  3. they lead to bargaining

Discussions

We focus on (3), in this case between group bargaining can lead to, or relieve, double discrimination

e.g. Male A and Male B bargain. They are both double discriminators. Nash barganining solution yields preferential outcomes for males but removes double discrimination

Table 1: Prejudices for male \(A\) prior to Nash bargaining.
Group A Group B Average Difference
Male 0.4 0.3 0.35 -0.1
Female 0.3 0.0 0.15 -0.3
Average 0.35 0.15 0.25 -0.2
Difference -0.1 -0.3 -0.2 Interaction: -0.2
Table 2: Prejudices for male \(B\) prior to Nash bargaining.
Group A Group B Average Difference
Male 0.3 0.4 0.35 -0.1
Female 0.0 0.3 0.15 -0.3
Average 0.35 0.15 0.25 -0.2
Difference -0.1 -0.3 -0.2 Interaction: -0.2
Table 3: Nash bargaining solution from negotiation between an intersectionally discriminating A man and an intersectionally discriminating B man.
Group A Group B Average Difference
Male 0.35 0.35 0.35 0
Female 0.15 0.15 0.15 0
Average 0.2 0.2 0.2 0
Difference -0.2 -0.2 -0.2 Interaction: 0

Hypotheses

Double discrimination:

  • Exists
  • Is exacerbated by quotas on one dimension
  • Is curbed by information
  • Is curbed by collective decision making procedures

How did we test these hypotheses?

Why a Lab Experiment in Kenya?

  • Ethnicity is highly socially salient and people have in-group biases
  • Gender discrimination occurs
  • Gender quotas are politically salient since 2010 constitutional reforms
  • Many studies (including lab experiments) have explored intersectional discrimination in contexts with ranked ethnic identities. Mostly in the US, but also in India.

Design - treatments

In total there are three fully crossed treatment arms and participants make choices across two vignettes:

  1. T1 Quota: electors are asked to choose at least one woman

  2. T2 Information: vignettes display ability measures of all candidates (test scores)

  3. T3 Discussion: triads are assigned to group discussions in either, neither, or one of the vignettes, after which they will be asked to make a group choice

Design - sample

  • We target a sample of 900 participants distributed equally across ethnic (Luo/Kikuyu/Other) and gender (M/F) dimensions

  • Luo and Kikuyu groups were chosen as these are the most populous in the region; the “other” group is heterogeneous, defined simply as non-Luo and non-Kikuyu

Male Female
Kikuyu 150 150
Luo 150 150
Other 150 150

Design - Individual survey

  • Participants first take individual survey with standard questions about personal characteristics as well as a raven’s test, which we use to generate scores and determine electors’ compensation.

Raven’s matrices

Design - experiment and group discussion

  • All tasks are incentivized; participants are allocated payments the more accurate their ratings were and if they chose a successful candidate. Candidates who are chosen are paid out regardless of whether they’re successful
  • Participants are shown 2 vignettes in total

Vignette in non-information treatment

Lab set-up

Demographic breakdown

  • Gender imbalance - we over-sample men (insufficiently) as show up rates are lower

  • Ethnicities are largely balanced

Test scores breakdown

Test is not gender neutral but seems evenly distributed across ethnicity

Test scores breakdown

Implications for updating in the information condition

Information reduces everyone’s beliefs but:

  • increases the gap in expectations for men and women candidates
  • removes the gap in expectations of men and women electors

Luos show a smaller ingroup bias in ratings which remains (is strengthened) when information provided

What do we find?

Marginal means

What are the selection probabilities of different groups at baseline?

  • Vertical line corresponds to \(P = 1/3\), which is the average probability of selection

How do our treatment conditions affect how well chosen candidates perform?

Main results (selection)

  • Treatments are not demeaned but everything else is
Term No score control With score control
Beta SE Beta SE
outgroup -0.08*** 0.017 -0.08*** 0.017
female -0.03* 0.014 -0.03* 0.015
female * outgroup 0.05 0.037 0.05 0.037
quota * female 0.14*** 0.016 0.14*** 0.016
information * female -0.05** 0.020 -0.01 0.020
discussion * female 0.00 0.037 0.00 0.037
quota * outgroup 0.02 0.021 0.02 0.021
information * outgroup 0.02 0.024 0.03 0.023
discussion * outgroup 0.18* 0.079 0.18* 0.079
quota * female * outgroup -0.01 0.048 -0.01 0.048
information * female * outgroup -0.04 0.053 -0.05 0.051
discussion * female * outgroup -0.08 0.174 -0.08 0.174
  • Everything is demeaned including treatments
Term No score control With score control
Beta SE Beta SE
outgroup -0.05*** 0.010 -0.05*** 0.010
female 0.02*** 0.006 0.04*** 0.006
female * outgroup 0.00 0.021 0.00 0.020
quota * female 0.16*** 0.010 0.15*** 0.010
information * female -0.03** 0.012 0.01 0.012
discussion * female -0.02 0.015 -0.02 0.015
quota * outgroup 0.02 0.015 0.01 0.015
information * outgroup 0.01 0.021 0.01 0.020
discussion * outgroup 0.09* 0.043 0.09* 0.040
quota * female * outgroup -0.05 0.035 -0.05 0.034
information * female * outgroup -0.06 0.043 -0.06 0.041
discussion * female * outgroup -0.14 0.097 -0.13 0.090

Main results (ratings)

  • Treatments are not demeaned but everything else is
Term No score control With score control
Beta SE Beta SE
outgroup -1.8* 0.852 -1.8* 0.854
female -1.9*** 0.524 -1.8*** 0.539
female * outgroup -1.3 1.17 -1.3 1.17
quota * female -0.04 0.428 -0.05 0.428
information * female -3.2*** 0.923 1.3 0.833
discussion * female 1.6 1.24 1.6 1.24
quota * outgroup -0.34 0.680 -0.33 0.679
information * outgroup 0.56 1.40 1.2 1.27
discussion * outgroup -2.1 2.72 -2.1 2.72
quota * female * outgroup 0.15 0.983 0.15 0.983
information * female * outgroup -0.32 2.11 -0.88 1.86
discussion * female * outgroup 0.43 5.82 0.43 5.81
  • Everything is demeaned including treatments
Term No score control With score control
Beta SE Beta SE
outgroup -1.8** 0.616 -1.7** 0.551
female -3.4*** 0.393 -1.3*** 0.353
female * outgroup -1.3 0.974 -1.4 0.862
quota * female -0.01 0.340 -0.26 0.299
information * female -3.5*** 0.804 0.79 0.713
discussion * female 0.28 1.19 0.29 1.08
quota * outgroup -0.22 0.521 -0.57 0.480
information * outgroup 0.99 1.25 1.3 1.11
discussion * outgroup -1.6 2.41 -1.6 2.03
quota * female * outgroup 1.0 0.929 0.97 0.793
information * female * outgroup -0.36 1.98 -0.60 1.73
discussion * female * outgroup -1.6 4.54 -0.49 4.02

Structural estimation

Strategy

We assume a particular specification of the utility model of the form:

\[ u(\beta) = \lambda \sum_{\theta}\alpha_\theta\left(\sum_{j|\theta_j = \theta}\beta_{j}\right)^{0.5} + (1-\lambda)\sum_{\theta}\left(\sum_{j|\theta_j = \theta}\beta_{j}\hat\pi_\theta\right) \qquad(1)\]

where utility is defined over choice vectors:

  • \(\theta\) are identity types
  • \(\hat\pi_\theta\) denote beliefs about ability
  • \(\hat\pi_\theta\) denote welfare preferences over types
  • \(\lambda\) denotes weights on welfare versus ability
  • \(\rho\) is a rationality parameter

\[P_h = \frac{e^{\rho u(h)}}{\sum_ke^{\rho u(k)}} \]

Strategy

We assume a particular specification of the utility model of the form:

\[ u(\beta) = \lambda \sum_{\theta}\alpha_\theta\left(\sum_{j|\theta_j = \theta}\beta_{j}\right)^{0.5} + (1-\lambda)\sum_{\theta}\left(\sum_{j|\theta_j = \theta}\beta_{j}\hat\pi_\theta\right) \qquad(2)\]

where

  • \(\hat\pi_\theta\) are measured and depend on ability and preferences
  • quotas impose a constraint on the choice set
  • discussion are assumed to yield outcomes according to a Nash bargaining score

Estimates

Parameter estimates from maximum likelihood estimation
term estimate std.error statistic p.value
\(\lambda\) 0.366 0.024 15.342 0.00
\(\alpha_{00}\) 0.653 0.093 7.020 0.00
\(\alpha_{01}\) 0.534 0.090 5.933 0.00
\(\alpha_{10}\) 0.425 0.089 4.767 0.00
\(\alpha_{11}\) 0.321 0.090 3.557 0.00
\(\rho\) 3.409 NA NA NA
\(\Delta\alpha_{female}\) -0.221 0.035 40.805 0.00
\(\Delta\alpha_{outgroup}\) -0.112 0.028 15.974 0.00
\(\Delta\alpha_{intersection}\) 0.014 0.053 0.071 0.79

Biases seen again, with very modest positive interaction

Experiments

Change in probability of selection smoothed across all choices and varying values of \(\lambda\)

Interpretation: More weight on benefits of office exacerbate double privilege from taste based discrimination

Experiments

Figure 1: Change in probability of selection smoothed across all choices and varying values of \(\lambda\)

But we are seeing this in the no quota cases only

Conditions

Figure 2: Interaction smoothed across all choices for varying \(\lambda\)

Conclusion

Summary results

Quotas

  • We find baseline discrimination against females and outgroups, but not double discrimination
  • Quota effects on double discrimination are weak but signs are correct
  • Quota offset baseline discrimination

Information

  • Discrimination against women disappears when condition on scores
  • Again little evidence for an effect on double discrimination with suggestive worsening due to information

Discussion

  • Discussions remove biases against outgroups but not so much for women
    • Incidentally, baseline discrimination against outgroups is also much stronger than against women
  • Again little evidence for an effect on double discrimination with suggestive worsening due to discussions

Next steps: This project

This is still work in progress.

  • group composition (on ethnic or gender dimensions) on discrimination in the discussion condition
    • When making collective decision, presence of more ingroup or cogender electors may exacerbate discrimination
  • Decomposition of biases in preferences: effect of rating on choice outcome
    • How much of one’s choice is driven by priors about ability?
  • Treatment interactions

Next steps: Potential modifications

  • Change information treatment
  • Allow individuals to be strategic
  • Choose a more gender-neutral test
  • Make more room for taste-based discrimination in the design

Thank you! Questions/comments?

Appendix

Results - fairness perception

  • Fairness determined by; 1 if candidate unfairly disfavored by elector, -1 if unfairly favored, 0 otherwise

term estimate std.error statistic p.value conf.low conf.high
score -0.004 0.002 -1.789 0.074 -0.009 0.000
info -0.009 0.003 -2.641 0.008 -0.016 -0.002
female_c 0.019 0.013 1.549 0.122 -0.005 0.044
female_e 0.004 0.005 0.758 0.449 -0.006 0.015
outgroup 0.051 0.014 3.706 0.000 0.024 0.078
quota -0.015 0.004 -4.156 0.000 -0.023 -0.008
score:info -0.008 0.005 -1.648 0.100 -0.017 0.001
female_c:female_e -0.042 0.025 -1.694 0.091 -0.090 0.007
female_c:outgroup -0.045 0.028 -1.605 0.109 -0.100 0.010
female_e:outgroup -0.044 0.027 -1.630 0.104 -0.097 0.009
female_c:quota -0.110 0.017 -6.288 0.000 -0.144 -0.075
female_e:quota -0.007 0.007 -0.889 0.374 -0.021 0.008
outgroup:quota 0.025 0.020 1.269 0.205 -0.014 0.063
female_c:female_e:outgroup 0.066 0.056 1.178 0.239 -0.044 0.175
female_c:female_e:quota 0.024 0.033 0.705 0.481 -0.042 0.089
female_c:outgroup:quota 0.060 0.041 1.483 0.139 -0.019 0.140
female_e:outgroup:quota 0.088 0.039 2.232 0.026 0.011 0.166
female_c:female_e:outgroup:quota -0.055 0.084 -0.657 0.511 -0.220 0.110

Double advantage and quotas

Attitudes toward gender

Ethnic salience

Respondent education

Manipulation checks

We conduct two manipulation checks, one for our information treatment and the other for the quota treatment. For each we ask these questions:

Information MC

Earlier in the side rooms you were asked to choose two candidates. Think about the first group of candidates (vignettes) you saw. Did their profiles tell you how much they scored on the task?

Only about half of respondents passed this test.

Quota MC

Earlier in the side rooms you were asked to choose two candidates. Think about the first group of candidates (vignettes) you saw. Were you asked to pick at least one woman?

About 63 percent passed this test.

Quotas and intersectional discrimination; what does the evidence say?

  • Hughes (2011) shows in cross-national research shows that as standalone policies, gender and minority quotas tend to benefit primarily majority women and minority men
  • Little evidence for the intersectional effect of quotas in Europe and Southeast Asia (Tan (2014); Celis et al. (2014); Huang (2012); Folke, Freidenvall, and Rickne (2015))
  • Research on India, which has both gender and caste quotas shows mixed results (Cassan and Vandewalle (2021); Jensenius (2016); Karekurve-Ramachandra and Lee (2020); Brulé and Toth (2022)))

Individual sources of intersectional discrimination

  • Social dominance theory For reasons grounded in evolutionary biology men in marginalized ethnic groups face more discrimination than women (Sidanius and Pratto (1999))

  • Subgroup-specific stereotypes (e.g. Kennelly (1999);

  • Intersectional invisibility Ideologies favoring the advantaged define who is seen as a “prototypical” member of a marginalized group (e.g Purdie-Vaughns and Eibach (2008);

  • Gender norms vary by ethnic group Greater gender discrimination in some groups may reduce the potential supply of candidates for jobs or political office

Collective Sources of Intersectional Discrimination

  • Strategic decisions by political gatekeepers (Mügge (2016); Celis and Erzeel (2017);
  • Political mobilization and advocacy by (or for) marginalized groups (Cohen (1999b); Dara Z. Strolovitch (2006b))
  • Responses to institutional remedies to discrimination on one dimension
    • Claim intersectionality in the law (Crenshaw (1989); Best et al. (2011))
    • Electoral quotas for women and marginalized groups change the behavior of voters, potential candidates, and political gatekeepers

How binding is the quota constraint?

Quota effects on base chance a women is chosen.

  • Theoretically possible increase from 0 floor to 1/3 floor.
  • Actual increase much more modest.

References

Bejarano, Christina E. 2014. The Latina Advantage: Gender, Race, and Political Success.
Best, Rachel Kahn, Lauren B. Edelman, Linda Hamilton Krieger, and Scott R. Eliason. 2011. “Multiple Disadvantages: An Empirical Test of Intersectionality Theory in EEO Litigation.” Law & Society Review 45 (4): 991–1025. https://doi.org/10.1111/j.1540-5893.2011.00463.x.
Brulé, Rachel, and Aliz Toth. 2022. “Do Quotas In Two Dimensions Improve Social Equality? Intersectional Representation & Group Relations.” SSRN Working Paper; GDP Center Human Capital Initiative Working Paper.
Cassan, Guilhem, and Lore Vandewalle. 2021. “Identities and Public Policies: Unexpected Effects of Political Reservations for Women in India.” World Development 143: 105408.
Celis, Karen, and Silvia Erzeel. 2017. “The Complementarity Advantage: Parties, Representativeness and Newcomers’ Access to Power.” Parliamentary Affairs 70 (1): 43–61. https://doi.org/10.1093/pa/gsv043.
Celis, Karen, Silvia Erzeel, Liza Mügge, and Alyt Damstra. 2014. “Quotas and Intersectionality: Ethnicity and Gender in Candidate Selection.” International Political Science Review 35 (1): 41–54.
Cohen, Cathy J. 1999b. The Boundaries of Blackness AIDS and the Breakdown of Black Politics /. Chicago : University of Chicago Press,. http://www.columbia.edu/cgi-bin/cul/resolve?clio14645680.
———. 1999a. The Boundaries of Blackness AIDS and the Breakdown of Black Politics /. Chicago : University of Chicago Press,. "http://www.columbia.edu/cgi-bin/cul/resolve?clio14645680".
Crenshaw, Kimberle. 1989. “Demarginalizing the Intersection of Race and Sex: A Black Feminist Critique of Antidiscrimination Doctrine, Feminist Theory and Antiracist Politics.” The University of Chicago Legal Forum 140: 139–67.
Folke, Olle, Lenita Freidenvall, and Johanna Rickne. 2015. “Gender Quotas and Ethnic Minority Representation: Swedish Evidence from a Longitudinal Mixed Methods Study.” Politics & Gender 11 (2): 345–81. https://doi.org/10.1017/S1743923X15000069.
Huang, Chang-Ling. 2012. “Falling Through the Cracks? Indigenous Women’s Political Participation in Taiwan.” Indigenous Women’s Political Participation in Taiwan.
Hughes, Melanie M. 2011. “Intersectionality, Quotas, and Minority Women’s Political Representation Worldwide.” The American Political Science Review 105 (3): 604–20. http://www.jstor.org.ezproxy.cul.columbia.edu/stable/41480860.
Jensenius, Francesca R. 2016. “Competing Inequalities? On the Intersection of Gender and Ethnicity in Candidate Nominations in Indian Elections.” Government and Opposition 51 (3): 440–63. https://www.jstor.org/stable/26349814.
Kao, Kristen, and Lindsay J Benstead. 2021. “Female Electability in the Arab World: The Advantages of Intersectionality.” Comparative Politics 53 (3): 427–64.
Karekurve-Ramachandra, Varun, and Alexander Lee. 2020. “Do Gender Quotas Hurt Less Privileged Groups? Evidence from India.” American Journal of Political Science 64 (4): 757–72. http://www.jstor.org/stable/45295347.
Kennelly, Ivy. 1999. “"That Single-Mother Element": How White Employers Typify Black Women.” Gender and Society 13 (2): 168–92. https://www.jstor.org/stable/190387.
Mügge, Liza M. 2016. “Intersectionality, Recruitment and Selection: Ethnic Minority Candidates in Dutch Parties.” Parliamentary Affairs 69 (3): 512–30.
Purdie-Vaughns, Valerie, and Richard P. Eibach. 2008. “Intersectional Invisibility: The Distinctive Advantages and Disadvantages of Multiple Subordinate-Group Identities.” Sex Roles: A Journal of Research 59 (5-6): 377–91. https://doi.org/10.1007/s11199-008-9424-4.
Scola, Becki. 2013. Gender, Race, and Office Holding in the United States: Representation at the Intersections. Routledge.
Sidanius, Jim, and Felicia Pratto. 1999. Social Dominance: An Intergroup Theory of Social Hierarchy and Oppression. Social Dominance: An Intergroup Theory of Social Hierarchy and Oppression. New York, NY, US: Cambridge University Press. https://doi.org/10.1017/CBO9781139175043.
Strolovitch, Dara Z. 2006a. “Do Interest Groups Represent the Disadvantaged? Advocacy at the Intersections of Race, Class, and Gender.” The Journal of Politics 68 (4): 894–910.
Strolovitch, Dara Z. 2006b. “Do Interest Groups Represent the Disadvantaged? Advocacy at the Intersections of Race, Class, and Gender.” The Journal of Politics 68 (4): 894–910. https://doi.org/10.1111/j.1468-2508.2006.00478.x.
Tan, Netina. 2014. “Ethnic Quotas and Unintended Effects on Women’s Political Representation in Singapore.” International Political Science Review 35 (1): 27–40.