Common fairness metrics#
In the sections below, we review the most common fairness metrics, as well as their underlying assumptions and suggestions for use. Each metric requires that some aspects of the predictor behavior be comparable across groups.
Note
Note that common usage does not imply correct usage; we discuss one very common misuse in the section on the Four-Fifths Rule
In the code examples presented below, we will use the following input arrays:
>>> y_true = [0, 1, 1, 1, 1, 0, 1, 0, 1, 0, 0, 0, 1, 1, 1, 1, 1, 1]
>>> y_pred = [0, 0, 1, 0, 1, 1, 1, 0, 0, 1, 1, 1, 1, 0, 0, 1, 1, 0]
>>> sf_data = ['b', 'b', 'a', 'b', 'b', 'c', 'c', 'c', 'a',
... 'a', 'c', 'a', 'b', 'c', 'c', 'b', 'c', 'c']
Demographic parity#
Demographic parity is a fairness metric whose goal is to ensure a machine learning model’s predictions are independent of membership in a sensitive group. In other words, demographic parity is achieved when the probability of a certain prediction is not dependent on sensitive group membership. In the binary classification scenario, demographic parity refers to equal selection rates across groups. For example, in the context of a resume screening model, equal selection would mean that the proportion of applicants selected for a job interview should be equal across groups.
We mathematically define demographic parity using the following set of equations. A classifier \(h\) satisfies demographic parity under a distribution over \((X, A, Y)\) if its prediction \(h(X)\) is statistically independent of the sensitive feature \(A\). Agarwal, Beygelzimer, Dudík, Langford, and Wallach1 show that this is equivalent to \(\E[h(X) \given A=a] = \E[h(X)] \quad \forall a\).
In the case of regression, a predictor \(f\) satisfies demographic parity
under a distribution over \((X, A, Y)\) if \(f(X)\) is independent
of the sensitive feature \(A\).
Agarwal, Dudík, and Wu2 show that this is equivalent to
\(\P[f(X) \geq z \given A=a] = \P[f(X) \geq z] \quad \forall a, z\).
Another way to think of demographic parity in a
regression scenario is to compare the average predicted value across groups.
Note that in the Fairlearn API, fairlearn.metrics.demographic_parity_difference()
is only defined for classification.
Note
Demographic parity is also sometimes referred to as independence, group fairness, statistical parity, and disparate impact.
Failing to achieve demographic parity could generate allocation harms. Allocation harms occur when AI systems allocate opportunities, resources, or information differently across different groups (for example, an AI hiring system that is more likely to advance resumes of male applicants than resumes of female applicants regardless of qualification). Demographic parity can be used to assess the extent of allocation harms because it reflects an assumption that resources should be allocated proportionally across groups. Of the metrics described in this section, it can be the easiest to implement. However, operationalizing fairness using demographic parity rests on a few assumptions: that either the dataset is not a good representation of what the world actually looks like (e.g., a resume assessment system that is more likely to filter out qualified female applicants due to an organizational bias towards male applicants, regardless of skill level), or that the dataset is an accurate representation of the phenomena being modeled, but the phenomena itself is unjust (e.g., consider the case of predictive policing, where a system created to predict crime rates may correctly predict higher crime rates for certain areas, but simultaneously fail to consider that those higher rates may be caused by disproportionate policing and overcriminimalization of those areas). In reality, these assumptions may not be the true. The dataset might be an accurate representation of the phenomena itself, or the phenomena being modeled may not be unjust. If either assumption is not true, then demographic parity may not provide a meaningful or useful measurement of the fairness of a model’s predictions.
Fairness metrics like demographic parity can also be used as optimization constraints during the machine learning model training process. However, demographic parity may not be well-suited for this purpose because it does not place requirements on the exact distribution of predictions with respect to other important variables. To understand this concept further, consider an example from the Fairness in Machine Learning textbook by Barocas, Hardt, and Narayanan3:
“However, decisions based on a classifier that satisfies independence can have undesirable properties (and similar arguments apply to other statistical critiera). Here is one way in which this can happen, which is easiest to illustrate if we imagine a callous or ill-intentioned decision maker. Imagine a company that in group A hires diligently selected applicants at some rate p>0. In group B, the company hires carelessly selected applicants at the same rate p. Even though the acceptance rates in both groups are identical, it is far more likely that unqualified applicants are selected in one group than in the other. As a result, it will appear in hindsight that members of group B performed worse than members of group A, thus establishing a negative track record for group B.”
It’s also worth considering whether the assumptions underlying demographic parity maintain construct validity (see What is construct validity?). Construct validity is a concept in the social sciences that assesses the extent to which the ways we choose to measure abstract phenomena are valid. For demographic parity, one relevant question would be whether demographic parity meets the criteria for establishing “fairness”, itself an unobservable theoretical construct. Further, it’s important to ask whether satisfying demographic parity actually brings us closer to the world we’d like to see.
In some cases, we may observe a trend in data from multiple demographic groups, but that trend may disappear or reverse when groups are combined. Known as Simpson’s Paradox, this outcome may appear when observing disparate outcomes across groups. A famous example of Simpson’s Paradox is a study of 1973 graduate school admissions to the University of California, Berkley 4. The study showed that when observing admissions by gender, men applying were more likely than women to be accepted. However, drilling down into admissions by department revealed that women tended to apply to departments with more competitive admissions requirements, whereas men tended to apply to less competitive departments. The more granular analysis showed only four out of 85 departments exhibited bias against women, and six departments exhibited bias towards men. In general, the data indicated departments exhibited a bias in favor of minority-gendered applicants, which is opposite from the trend observed in the aggregate data.
This phenomenon is important to fairness evaluation because metrics like
demographic parity may be different when calculated at an aggregate level and
within more granular categories. In the case of demographic parity, we might
need to review
\(\E[h(X) \given A=a, D=d] = \E[h(X) \given D=d] \quad \forall a\)
where \(D\) represents the feature(s) within \(X\) across which
members of the groups within \(A\) are distributed.
Demographic parity would then require that the prediction of the target
variable is statistically independent of sensitive attributes conditional
on D. Simply aggregating outcomes across high-level categories can be
misleading when the data can be further disaggregated.
It’s important to review metrics across these more granular categories,
if they exist, to verify that disparate outcomes persist across all levels
of aggregation.
See Intersecting Groups below to see how MetricFrame can
help with this.
However, more granular categories generally contain smaller sample sizes, and it can be more difficult to establish that trends seen in very small samples are not due to random chance. We also recommend watching out for the multiple comparisons problem, which states that the more statistical inferences are made, the more erroneous those inferences will become. For example, in the case of evaluating fairness metrics on multiple groups, as we break the groups down into more granular categories and evaluate those smaller groups, it will become more likely that these subgroups will differ enough to fail one of the metrics. For dealing with the multiple comparisons problem, we recommend investigating statistical techniques meant to correct the errors produced by individual statistical tests.
Fairlearn provides the demographic_parity_difference() and
demographic_parity_ratio() functions for computing demographic
parity measures for binary classification data, both of which return
a scalar result.
The first reports the absolute difference between the highest and
lowest selection rates \(a \in A\) so a result of 0 indicates
that demographic parity has been achieved.
The second reports the ratio of the lowest and highest selection rates,
so a result of 1 means there is demographic parity.
This metric can potentially be used to implement the ‘Four-Fifths’ Rule,
but read our discussion below to understand whether this is an appropriate metric for your use case.
As with any fairness metric, achieving demographic parity does not automatically mean
that the classifier is fair!
>>> from fairlearn.metrics import demographic_parity_difference
>>> print(demographic_parity_difference(y_true,
... y_pred,
... sensitive_features=sf_data))
0.25
>>> from fairlearn.metrics import demographic_parity_ratio
>>> print(demographic_parity_ratio(y_true,
... y_pred,
... sensitive_features=sf_data))
0.66666...
Equalized odds#
The goal of the equalized odds fairness metric is to ensure a machine learning model performs equally well for different groups. It is stricter than demographic parity because it requires that the machine learning model’s predictions are not only independent of sensitive group membership, but that groups have the same false positive rates and and true positive rates. This distinction is important because a model could achieve demographic parity (i.e., its predictions could be independent of sensitive group membership), but still generate more false positive predictions for one group versus others. Equalized odds does not create the selection issue discussed in the demographic parity section above. For example, in the hiring scenario where the goal is to choose applicants from group A and group B, ensuring the model performs equally well at choosing applicants from group A and group B can circumvent the issue of the model optimizing by selecting applicants from one group at random.
We mathematically define equalized odds using the following set of equations. A classifier \(h\) satisfies equalized odds under a distribution over \((X, A, Y)\) if its prediction \(h(X)\) is conditionally independent of the sensitive feature \(A\) given the label \(Y\). Agarwal, Beygelzimer, Dudík, Langford, and Wallach1 show that this is equivalent to \(\E[h(X) \given A=a, Y=y] = \E[h(X) \given Y=y] \quad \forall a, y\). Equalized odds requires that the true positive rate, \(\P(h(X)=1 | Y=1\), and the false positive rate, \(\P(h(X)=1 | Y=0\), be equal across groups.
The inclusion of false positive rates acknowledges that different groups experience different costs from misclassification. For example, in the case of a model predicting a negative outcome (e.g., probability of recidivating) that already disproportionately affects members of minority communities, false positive predictions reflect pre-existing disparities in outcomes across minority and majority groups. Equalized odds further enforces that the accuracy is equally high across all groups, punishing models that only perform well on majority groups.
If a machine learning model does not perform equally well for all groups, then it could generate allocation or quality-of-service harms. Equalized odds can be used to diagnose both allocation harms as well as quality-of-service harms. Allocation harms are discussed in detail in the demographic parity section above. Quality-of-service harms occur when an AI system does not work as well for one group versus another (for example, facial recognition systems that are more likely to fail for dark-skinned individuals). For more information on AI harms, see Types of harms.
Equalized odds can be useful for diagnosing allocation harms because its goal is to ensure that a machine learning model works equally well for different groups. Another way to think about equalized odds is to contrast it with demographic parity. While demographic parity assesses the allocation of resources generally, equalized odds focuses on the allocation of resources that were actually distributed to members of that group (indicated by the positive target variable \(Y=1\)). However, equalized odds makes the assumption that the target variable \(Y\) is a good measurement of the phenomena being modeled, but that assumption may not hold if the measurement does not satisfy the requirements of construct validity.
Similar to the demographic parity case, Fairlearn provides
equalized_odds_difference() and equalized_odds_ratio()
to help with these calculations.
However, since equalized odds is based on both the true positive and
false positive rates, there is an extra step in order to return
a single scalar result.
For equalized_odds_difference(), we first calculate the
true positive rate difference and the true negative rate difference
separately.
We then return the larger of these two differences.
Mutatis mutandis, equalized_odds_ratio() works similarly.
>>> from fairlearn.metrics import equalized_odds_difference
>>> print(equalized_odds_difference(y_true,
... y_pred,
... sensitive_features=sf_data))
1.0
>>> from fairlearn.metrics import equalized_odds_ratio
>>> print(equalized_odds_ratio(y_true,
... y_pred,
... sensitive_features=sf_data))
0.0
Equal opportunity#
Equal opportunity is a relaxed version of equalized odds that only considers conditional expectations with respect to positive labels, i.e., \(Y=1\). 5 Another way of thinking about this metric is requiring equal outcomes only within the subset of records belonging to the positive class. For example, in the hiring example, equal opportunity requires that the individuals who are actually hired have an equal opportunity of being hired in the first place. However, by not considering whether false positive rates are equivalent across groups, equal opportunity does not capture the costs of missclassification disparities.
The Four Fifths Rule: Often Misapplied#
In the literature around fairness in machine learning, one will often find
the so-called “four fifths rule” or “80% rule” used to assess whether a model
(or mitigation technique) has produced a ‘fair’ result.
Typically, the rule is implemented by using the demographic parity ratio introduced
in the Demographic parity section above
(within Fairlearn, one can use demographic_parity_ratio()), with a result
considered ‘fair’ if the ratio exceeds 80% for all identified subgroups.
Application of this threshold is wrong in many scenarios.
As we note in many other places in the Fairlearn documentation, ‘fairness’ must be assessed by examining the entire sociotechnical context of a machine learning system. In particular, it is important to start from the harms which can occur to real people, and work inwards towards the model. The demographic parity ratio is simply a metric by which a particular model may be measured (on a particular dataset). Given the origin of the ‘four-fifths rule’ (which we will discuss next), its application may also give an unjustified feeling of legal invulnerability by conflating fairness with legality. In reality, ‘fairness’ is not always identical to ‘legally allowable,’ and the former may not even be a strict subset of the latter. 8
The ‘four fifths rule’ has its origins in a specific area of US federal employment law. It is a limit for prima facie evidence that illegal discrimination has occurred relative to a relevant control population. The four-fifths rule is one of many test statistics that can be used to establish a prima facie case, but it is generally only used within the context of US Federal employment regulation. A violation of the rule is still not sufficient to demonstrate that illegal discrimination has occurred - a causal link between the statistic and alleged discrimination must still be shown, and the (US) court would examine the particulars of each case. For an example of the subtleties involved, see Ricci v. Stefano which resulted from an attempt to ‘correct’ for disparate impact. Outside its particular context in US federal employment law, the ‘four fifths rule’ has no validity and its misapplication is an example of the portability trap.
Taken together, we see that applying the ‘four fifths rule’ will not be appropriate in most cases. Even in cases where it is applicable, the rule does not automatically avoid legal jeopardy, much less ensure that results are fair. The use of the ‘four fifths rule’ in this manner is an indefensible example of epistemic trespassing. 9 It is for this reason that we try to avoid the use of legal terminology in our documentation.
For a much deeper discussion of the issues involved, we suggest Watkins et al.6. A higher level look at how legal concepts of fairness can collide with mathematical measures of disparity, see Xiang and Raji7.
Summary#
We have introduced three commonly used fairness metrics in this section, which can be summed up as follows:
Demographic Parity
What it compares: Predictions between different groups (true values are ignored)
Reason to use: If the input data are known to contain biases, demographic parity may be appropriate to measure fairness
Caveats: By only using the predicted values, information is thrown away. The selection rate is also a very coarse measure of the distribution between groups, making it tricky to use as an optimization constraint
Equalized Odds
What it compares: True and False Positive rates between different groups
Reason to use: If historical data does not contain measurement bias or historical bias that we need to take into account, and true and false positives are considered to be (roughly) of the same importance, equalized odds may be useful
Caveats: If there are historical biases in the data, then the original labels may hold little value. A large imbalance between the positive and negative classes will also accentuate any statistical issues related to sensitive groups with low membership
Equal opportunity
What it compares: True Positive rates between different groups
Reason to use: If historical data are useful, and extra false positives are much less likely to cause harm than missed true positives, equal opportunity may be useful
Caveats: If there are historical biases in the data, then the original labels may hold little value. A large imbalance between the positive and negative classes will also accentuate any statistical issues related to sensitive groups with low membership
However, the fact these are common metrics does not make them applicable to any given situation. In particular, demographic parity is often misapplied.
References#
- 1(1,2)
Alekh Agarwal, Alina Beygelzimer, Miroslav Dudík, John Langford, and Hanna M. Wallach. A reductions approach to fair classification. In ICML, volume 80 of Proceedings of Machine Learning Research, 60–69. PMLR, 2018. URL: http://proceedings.mlr.press/v80/agarwal18a.html.
- 2
Alekh Agarwal, Miroslav Dudík, and Zhiwei Steven Wu. Fair regression: quantitative definitions and reduction-based algorithms. In ICML, volume 97 of Proceedings of Machine Learning Research, 120–129. PMLR, 2019. URL: http://proceedings.mlr.press/v97/agarwal19d.html.
- 3
Solon Barocas, Moritz Hardt, and Arvind Narayanan. Fairness and Machine Learning. fairmlbook.org, 2019. http://www.fairmlbook.org/.
- 4
P.J. Bickel, E.A. Hammel, and E.W. and O’Connell. Sex bias in graduate admissions: data from berkeley. Science, 187(4175):398–404, 1975. URL: https://doi.org/10.1126%2Fscience.187.4175.398, doi:10.1126%2Fscience.187.4175.398.
- 5
Moritz Hardt, Eric Price, and Nati Srebro. Equality of opportunity in supervised learning. In NeurIPS, 3315–3323. 2016. URL: https://proceedings.neurips.cc/paper/2016/hash/9d2682367c3935defcb1f9e247a97c0d-Abstract.html.
- 6
Elizabeth Anne Watkins, Michael McKenna, and Jiahao Chen. The four-fifths rule is not disparate impact: a woeful tale of epistemic trespassing in algorithmic fairness. 2022. URL: https://arxiv.org/abs/2202.09519, doi:10.48550/ARXIV.2202.09519.
- 7
Alice Xiang and Inioluwa Deborah Raji. On the legal compatibility of fairness definitions. 2019. URL: https://arxiv.org/abs/1912.00761, doi:10.48550/ARXIV.1912.00761.
Footnotes
- 8
For a related example, see the discussion on ‘law’ and ‘justice’ in The Caves of Steel (Asimov, 1953)
- 9
Epistemic trespassing is the process of taking expertise in one field and applying it to another in which one does not have an equivalent (or any) competence. This is not an intrinsically bad thing - one could label all interdisciplinary research a form of epistemic trespassing. However, doing so successfully requires a willingness to learn the subtleties of the new field.