From the Fair Distribution of Predictions to the Fair Distribution of Social Goods: Evaluating the Impact of Fair Machine Learning on Long-Term Unemployment

TL;DR

It is argued that addressing this problem requires a notion of prospective fairness that anticipates the change in the distribution of social goods after deployment, and the way predictions change policy decisions and the causally downstream distribution of social goods is focused on.

Abstract

Deploying an algorithmically informed policy is a significant intervention in society. Prominent methods for algorithmic fairness focus on the distribution of predictions at the time of training, rather than the distribution of social goods that arises after deploying the algorithm in a specific social context. However, requiring a "fair" distribution of predictions may undermine efforts at establishing a fair distribution of social goods. First, we argue that addressing this problem requires a notion of prospective fairness that anticipates the change in the distribution of social goods after deployment. Second, we provide formal conditions under which this change is identified from pre-deployment data. That requires accounting for different kinds of performative effects. Here, we focus on the way predictions change policy decisions and, consequently, the causally downstream distribution of social goods. Throughout, we are guided by an application from public administration: the use of algorithms to predict who among the recently unemployed will remain unemployed in the long term and to target them with labor market programs. Third, using administrative data from the Swiss public employment service, we simulate how such algorithmically informed policies would affect gender inequalities in long-term unemployment. When risk predictions are required to be "fair" according to statistical parity and equality of opportunity, targeting decisions are less effective, undermining efforts to both lower overall levels of long-term unemployment and to close the gender gap in long-term unemployment.

Figures (12)

• Figure 1: The left hand side shows the pre-deployment causal graph $G_{\text{pre}}$ inducing a joint probability distribution $P_{\text{pre}}$ over sensitive attributes $A$, features $X$, risk score $R$, decision $D$, and outcome variable $Y$. The risk score $R$ is the output of a learned function from $A$ and $X$. Since this graph represents the situation after training, but before deployment, there is no arrow from the risk score $R$ to the decision $D$. Retrospective fairness formulates constraints $\varphi(G_{\text{pre}}, P_{\text{pre}}, M)$ on the pre-deployment arrangement alone. The right-hand side represents the situation after the algorithmically informed policy has been deployed, with predictions $R$ now affecting decisions $D$. Prospective fairness requires comparing the consequences of intervening on the structure of $G_{\text{pre}}$ and moving to $G_{\text{post}}$. In other words, comparing $\varphi(G_{\text{pre}}, P_{\text{pre}}, M)$ with $\varphi(G_{\text{post}}, P_{\text{post}}, M)$.
• Figure 2: Swiss Long-Term Unemployment Rates by Gender. Data for the period 2010-2022 are from Eurostat eurostat_swissLTUGap; the gender rates in long-term unemployment are computed as the share of all unemployed men/women aged 20-64 who are unemployed for more than a year. Data for the period 1991-2007 are from the 2012 Swiss Social Report buhlmann2013swiss, where age information is not available. Data for 2008-9 is not readily available.
• Figure 3: Estimated (Individualized) Average Treatment Effects for six labor market programs with "no program" as the baseline.
• Figure 4: We plot the gender gap in long-term unemployment (LTU) against program capacity for each combination of prioritization and assignment scheme. The level of transparency shows the gender gap for the corresponding fairness constraint: none, statistical parity, or equal opportunity. The unconstrained risk scores (lowest transparency) result in the smallest gender gap. This effect is especially pronounced as program capacity is increased and program assignments are individualized (optimal).
• Figure 5: We show overall long-term unemployment and the gender gap against program capacity for each combination of prioritization and assignment scheme. For clarity, results are shown only for fairness-unconstrained risk scores. Regardless of the assignment scheme, the Belgian prioritization results in slightly lower overall rates of long-term unemployment (blue line) and a smaller gender gap. Individualized program assignments (optimal) are markedly more effective.

Theorems & Definitions (2)

• Theorem 1
• proof : Proof of Theorem \ref{['thm:Identification']}