Tag Archives: causal

Fairness and discrimination, PhD Course, #8 Individual fairness

After our post on “group fairness“, it’s time to discuss so-called “individual fairness“.


The first idea is discussed in Dwork et al. (2012)

our approach is centered around the notion of a task-specific similarity metric describing the extent to which pairs of individuals should be regarded as similar for the classification task at hand. The similarity metric expresses ground truth. When ground truth is unavailable, the metric may reflect the “best” available approximation as agreed upon by society. Following established tradition – Rawls (1971) – the metric is assumed to be public and open to discussion and continual refinement. Indeed, we envision that, typically, the distance metric would be externally imposed, for example, by a regulatory body or externally proposed by a civil rights organization


Counterfactual fairness

The second one is related to causal inference. Ensuring fairness using causal methods will produce “counterfactual fairness” (to use the term introduced in Kusner et al. (2017)), based on the idea a decision is fair towards an individual if the outcome is the same in reality as it would be in a ‘counterfactual’ world, in which the individual belongs to the other group (with respect to the sensitive attribute).

Quite naturally, we should compare potential outcomes, either globally (average treatement effect) or a local version, conditional on characteristics \boldsymbol{x} of an individual.

Based on causal graphs (discussed previously) we can define several notions of individual fairness.

Hence, it is possible to use Plečko et al. (2021), based on transport, and quantile regressions,

To illustrate, we can consider some causal graph on our toy dataset

and then, on some specific individuals in the dataset

Here, we can also get a counterfactual version of all individuals with one-to-one matching, and optimal transport


and we can get a counterfactual version, and possibly, a different prediction, using the fairadapt R package

We can also consider the German credit dataset

or the causal graph used in Watson et al. (2021),

Then, those techniques can be used to see compare the predictions of 6 fictious individuals,

A Fair price to pay: exploiting causal graphs for fairness in insurance

Our paper, a fair price to pay: exploiting causal graphs for fairness in insurance, writen with Olivier Côté and Marie-Pier Côté is now available,  on SSRN

In many jurisdictions, insurance companies must not discriminate on some given policyholder characteristics. Omission of prohibited variables from models prevents direct discrimination, but fails to address proxy discrimination, a phenomenon especially prevalent when powerful predictive algorithms are fed with an abundance of acceptable covariates. The lack of formal definition for key fairness concepts, in particular indirect discrimination, hinders the fairness assessment of methodologies. We review causal inference notions and introduce a causal graph tailored for fairness in insurance. Exploiting these, we discuss potential sources of bias, formally define direct and indirect discrimination, and study the properties of fairness methodologies. A novel categorization of fair methodologies into five families (best-estimate, unaware, aware, hyperaware, and corrective) is constructed based on their expected fairness properties. A comprehensive pedagogical example illustrates the practical implications of our findings: the interplay between our fair score families, group fairness criteria, and sources of discrimination.

Fairness and discrimination, PhD Course, #5 Models and Data

For the fifth course, we will discuss machine learning and standard techniques used to get predictive models, and to assess accuracy of those models.

GLM (possibly constrained)

Classically, we use a penalized version of least squares (but this can be adapted to GLMs, when penalizing the negative log-likelihood).  Because of Karush–Kuhn–Tucker conditions, having a constraint on the parameter is equivalent to the following penalized problem, when the constraint is on the \ell_2 norm of \boldsymbol{\beta},

We can also consider the \ell_1 norm of \boldsymbol{\beta},

Those two approaches can be see as a trade-off between accuracy (here the empirical risk on the left) and complexity of the model (on the right). And we can also consider a mixture of the two norms,

As we will see, it will also be possible to consider some penality related to fairness and discriminiation measures (in-processing).

Classifier and ROC Curves

We will also recall metrics used in the context of classification, such as the ROC curve

Each point of the curve can be related to two areas related to the distributions of the scores (in the two groups), for the same threshold – namely the false positive rate and true positive rate

Based on the ROC curve, we can define the AUC, the area under the ROC curve,

But for classifiers, the important challenge is to have calibrated scores, meaning that we want the score to be interpreted as the true underlying probability.


Well-calibration is defined as follows

or (with different notations)

It is a well know properties in several applications.


The plot on the right is the calibration plot,


We can easily get that plot


This concept is related to the question “do probabilities returned by some model represent reals probabilities ?” For instance, below, we have pictures generated as some sort of geodesic between two pictures, with a woman on the top left, and a man in the bottom right, published in the New York Times. And below, “probabilities” given by  https://www.picpurify.com/demo-face-gender-age.html.

We could agree that it is rather strange that probabilities (to have a man) do not increase continuously, but on top, with extremely high confidence, the model predicts that the picture is the one of a woman, and below, also with extremely high confidence, that the person is a man…

Data, observations vs. experiments

Then, after concept and notations related to models, we will talk about data. More specifically, the distinction between observations and experimentations.

Another popular classification is the one discussed by Judea Pearl.

So we will talk about association, correlation, causal inference, and counterfactuals.

“Correlated variables” or proxies

One important issue, is that with massive data, one can easily get a (good) proxy of almost any sensitive variable.

The concept is related to comonotonicity, or perfect correlation.

But this is clearly too strong, so we will discuss depedence measures, too.

Independence properties

Recall that independence is defined as follows


and we can consider a weaker form, based on null-covariance


or null-correlation

(sidenote, this correlation measure is bounded, and those bounds are related to Hardy-Littlewood inequality and optimal transport)

An interesting measure is the maximal correlation


or we can consider a weaker version, without consider all possible transformation, but only a subset


Another important concept is the one of conditional independence


(the later will be used in the context of causal graphs).


Before talking about causality, recall that what non-independence mean…


We can then construct causal graphs, or “directed acyclic graphs”


where nodes are the variables used in the model, and the outcome (usually that the end of the causal graph). Then we define paths


and the concept of d-separation


This concept is related to the statistical property of conditional independence


More precisely, we have the following Markov property on causal graphs


For example, for such a graphical model,

the joint distribution is \mathbb{P}[x_1,x_2,x_3,x_4]=\mathbb{P}[x_1]\times \mathbb{P}[x_2|x_1]\times \mathbb{P}[x_3|x_2]\times \mathbb{P}[x_4|x_3]and for the graphical model below

we have\mathbb{P}[x_1,x_2,x_3,x_4]=\mathbb{P}[x_1]\times \mathbb{P}[x_2]\times \mathbb{P}[x_3|x_1,x_2]\times \mathbb{P}[x_4|x_3]Those graphs can be related to structural models (with idiosyncratic noise denoted U), since


Potential outome

Another important concept is the concept of counterfactuals, and potential outome. In an ideal world, we would have observed the outome in both cases, with and without the treatement


but in real life, it’s only one of them,


And the goal will be, somehow, to estimate what the non-observed outcome would be. And then, classical quantites we wish to estimate are the average treatement effect, and the conditional version, based on some covariates.


This concept will be related to counterfactual fairness actually, when the “treatement” will be the sensitive attribute.

Twin network representation of the counterfactual

Finally, we will consider a so-called “twin network representation”. Consider a DAG, associated with some simple structural model


Based on a structural model, we can get values of idiosyncratic noise component


Then, we use those values on the twin representation, when the treatement is not 0, but 1. Counterfactuals are created by using the same noises


The difference between the two outcomes is the treatement effect, or the disparate treatement


or more generally, we write


This is an idea used in Plecko & Meinshausen, 2019, in the context of fairness, but we will discuss this more, later on…

Probabilistic Foundations of Econometrics, part 4

This post is the fourth one of our series on the history and foundations of econometric and machine learning models. Part 3 is online here.

Goodness of Fit, and Model

In the Gaussian linear model, the determination coefficient – noted R^2 – is often used as a measure of fit quality. It is based on the variance decomposition formula \underbrace{\frac{1}{n}\sum_{i=1}^n (y_i-\bar{y})^2}_{\text{total variance}}=\underbrace{\frac{1}{n}\sum_{i=1}^n (y_i-\widehat{y}_i)^2}_{\text{residual variance}}+\underbrace{\frac{1}{n}\sum_{i=1}^n (\widehat{y}_i-\bar{y})^2}_{\text{explained variance}} The R^2 is defined as the ratio of explained variance and total variance, another interpretation of the coefficient that we had introduced from the geometry of the least squares R^2= \frac{\sum_{i=1}^n (y_i-\bar{y})^2-\sum_{i=1}^n (y_i-\widehat{y}_i)^2}{\sum_{i=1}^n (y_i-\bar{y})^2}The sums of the error squares in this writing can be rewritten as a log-likelihood. However, it should be remembered that, up to one additive constant (obtained with a saturated model) in generalized linear models, deviance is defined by {Deviance}(\widehat{\beta}) = -2\log[\mathcal{L}] which can also be noted Deviance(\widehat{\mathbf{y}}). A null deviance can be defined as the one obtained without using the explanatory variables \mathbf{x}, so that \widehat{y}_i=\overline{y}. It is then possible to define, in a more general context (with a non-Gaussian distribution for y)R^2=\frac{{Deviance}(\overline{y})-{Deviance}(\widehat{\mathbf{y}})}{{Deviance}(\overline{y})}=1-\frac{{Deviance}(\widehat{\mathbf{y}})}{{Deviance}(\overline{y})}However, this measure cannot be used to choose a model, if one wishes to have a relatively simple model in the end, because it increases artificially with the addition of explanatory variables without significant effect. We will then tend to prefer the adjusted R^2,\bar R^2 = {1-(1-R^{2})\cdot{n-1 \over n-p}} = R^{2}-\underbrace{(1-R^{2})\cdot{p-1 \over n-p}}_{\text{penalty}}where p is the number of parameters of the model. Measuring the quality of fit will penalize overly complex models.

This idea will be found in the Akaike criterion, where AIC=Deviance+2\cdot p or in the Schwarz criterion, BIC=Deviance+log(n)\cdot p. In large dimensions (typically p>\sqrt{n}), we will tend to use a corrected AIC, defined by AIC_c=Deviance+2⋅p⋅n/(n-p-1) .

These criterias are used in so-called “stepwise” methods, introducing the set methods. In the “forward” method, we start by regressing to the constant, then we add one variable at a time, retaining the one that lowers the AIC criterion the most, until adding a variable increases the AIC criterion of the model. In the “backward” method, we start by regressing on all variables, then we remove one variable at a time, removing the one that lowers the AIC criterion the most, until removing a variable increases the AIC criterion from the model.

Another justification for this notion of penalty (we will come back to this idea in machine learning) can be the following. Let us consider an estimator in the class of linear predictors, \mathcal{M}=\big\lbrace m:~m(\mathbf{x})=s_h(\mathbf{x})^T\mathbf{y} \text{ where }S=(s(\mathbf{x}_1),\cdots,s(\mathbf{x}_n))^T\text{ is some smoothing matrix}\big\rbrace and assume that y=m_0 (x)+\varepsilon, with \mathbb{E}[\varepsilon]=0 and Var[\varepsilon]=\sigma^2\mathbb{I}, so that m_0 (x)=\mathbb{E}[Y|X=x] . From a theoretical point of view, the quadratic risk, associated with an estimated model \widehat{m}, \mathbb{E}\big[(Y-\widehat{m}(\mathbf{X}))^2\big], is written\mathcal{R}(\widehat{m})=\underbrace{\mathbb{E}\big[(Y-m_0(\mathbf{X}))^2\big]}_{\text{error}}+\underbrace{\mathbb{E}\big[(m_0(\mathbf {X})-\mathbb{E}[\widehat{m}(\mathbf{X})])^2\big]}_{\text{bias}^2}+\underbrace{\mathbb{E}\big[(\mathbb{E}[\widehat{m}(\mathbf{X})]-\widehat{m}(\mathbf{X}))^2\big]}_{\text{variance}} if m_0 is the true model. The first term is sometimes called “Bayes error”, and does not depend on the estimator selected, \widehat{m}.

The empirical quadratic risk, associated with a model m, is here: \widehat{\mathcal{R}}_n(m)=\frac{1}{n}\sum_{i=1}^n (y_i-m(\mathbf{x}_i))^2 (by convention). We recognize here the mean square error, “mse”, which will more generally give the “risk” of the model m when using another loss function (as we will discuss later on). It should be noted that:\displaystyle{\mathbb{E}[\widehat{\mathcal{R}}_n(m)]=\frac{1}{n}\|m_0(\mathbf{x})-m(\mathbf{x})\|^2+\frac{1}{n}\mathbb{E}\big(\|{Y}-m_0(\mathbf{X})\|^2\big)} We can show that:n\mathbb{E}\big[\widehat{\mathcal{R}}_n(\widehat{m})\big]=\mathbb{E}\big(\|Y-\widehat{m}(\mathbf{x})\|^2\big)=\|(\mathbb{I}-\mathbf{S})m_0\|^2+\sigma^2\|\mathbb{I}-\mathbf{S}\|^2so that the (real) risk of \widehat{m} is: {\mathcal{R}}_n(\widehat{m})=\mathbb{E}\big[\widehat{\mathcal{R}}_n(\widehat{m})\big]+2\frac{\sigma^2}{n}\text{trace}(\boldsymbol{S})So, if \text{trace}(\boldsymbol{S})\geq0 (which is not a too strong assumption), the empirical risk underestimates the true risk of the estimator. Actually, we recognize here the number of degrees of freedom of the model, the right-hand term corresponding to Mallow’s C_p, introduced in Mallows (1973) using not deviance but R^2.

Statistical Tests

The most traditional test in econometrics is probably the significance test, corresponding to the nullity of a coefficient in a linear regression model. Formally, it is the test of H_0:\beta_k=0 against H_1:\beta_k\neq 0. The so-called Student test, based on the statistics t_k=\widehat{\beta}_k/se_{\widehat{β}_k}, allows to decide between the two alternatives, using the test p-value, defined by \mathbb{P}[|T|>|t_k|] avec T\overset{\mathcal{L}}{\sim} Std_\nu, where \nu is the number of degrees of freedom of the model (\nu=p+1 for the standard linear model). In large dimension, however, this statistic is of very limited interest, given a significant FDR (“False Discovery Ratio”). Classically, with a level of significance \alpha=0.05, 5% of the variables are falsely significant. Suppose that we have p=100 explanatory variables, but that 5 (only) are really significant. We can hope that these 5 variables will pass the Student test, but we can also expect that 5 additional variables (false positive test) will emerge. We will then have 10 variables perceived as significant, while only half are significant, i.e. an FDR ratio of 50%. In order to avoid this recurrent pitfall in multiple tests, it is natural to use the procedure of Benjamini & Hochberg (1995).

From a correlation to some causal effect

Econometric models are used to implement public policy evaluations. It is therefore essential to fully understand the underlying mechanisms in order to know which variables actually make it possible to act on a variable of interest. But then we move on to another important dimension of econometrics. Jerry Neyman was responsible for the first work on the identification of causal mechanisms, and then Rubin (1974) formalized the test, called the “Rubin causal model” in Holland (1986). The first approaches to the notion of causality in econometrics were based on the use of instrumental variables, models with discontinuity of regression, analysis of differences in differences, and natural or unnatural experiments. Causality is usually inferred by comparing the effect of a policy – or more generally of a treatment – with its counterfactual, ideally given by a random control group. The causal effect of the treatment is then defined as \Delta=y_1-y_0, i.e. the difference between what the situation would be with treatment (noted t=1) and without treatment (noted t=0). The concern is that only y=t\cdot y_1+(1-t)\cdot y_0 and t are observed. In other words, the causal effect of variable t  on t  is not observed (since only one of the two potential variables – y_0 or y_1  is observed for each individual), but it is also individual, and therefore a function of x-covariates. Generally, by making assumptions about the distribution of the triplet (Y_0,Y_1,T) , some parameters of the causal effect distribution become identifiable, based on the density of the observable variables (Y,T) . Classically, we will be interested in the moments of this distribution, in particular the average effect of treatment in the population, \mathbb{E}[\Delta] , or even just the average effect of treatment in the case of treatment \mathbb{E}[\Delta|T=1] . If the result (Y_0,Y_1) is independent of the processing access variable T, it can be shown that \mathbb{E}[\Delta]=\mathbb{E}[Y|T=1]- \mathbb{E} [Y|T=0]. But if this independence hypothesis is not verified, there is a selection bias, often associated with \mathbb{E}[Y_0|T=1]- \mathbb{E} [Y_0|T=0]. Rosenbaum & Rubin (1983) propose to use a propensity to be treated score, p(x)=\mathbb{P}[T=1|X=x] , noting that if variable Y_0\ is independent of access to treatment T conditionally to the explanatory variables X, then it is independent of T  conditionally to the score p(X) : it is sufficient to match them using their propensity score. Heckman et al (2003) thus proposes a kernel estimator on the propensity score, which simply provides an estimator of the effect of the treatment, provided that it is treated.

To be continued next time, we’ll introduce “machine learning techniques” (references mentioned above are online here)