How do we tell if an algorithm is discriminating on the basis of race or other
Public concern over discriminatory algorithms is high.
Books such as Weapons of Math Destruction
and Automating Inequality
detail the ways algorithms further disadvantage the disadvantaged.
New York City’s City Council has passed an
algorithmic accountability bill.
Barocas and Selbst
warn that machine learning
algorithms have the potential to “inherit the prejudices of prior decision makers”,
“reflect widespread biases,” and discover “preexisting patterns of exclusion and inequality.”
Kleinberg, Mullainathan, and Raghavan
show that even without these nefarious factors at play, it is usually
impossible for an algorithm to be fair by all of three seemingly sensible definitions.
They study the case of risk assessments, which they define as
“ways of dividing people up into sets [….]
and then assigning each set a probability estimate that the people in this set
belong to the positive class.”
For concreteness, consider the COMPAS algorithm,
a controversial tool that predicts recidivism.
the “positive class” is those who do recidivize, and the negative class is those who do not.
They define three criteria and show that,
any risk assessment algorithm is unfair by at least one of the criteria,
unless the algorithm makes perfect predictions or the groups have the same true
rate of belonging to the positive class.
Society is increasingly relying on algorithms to make decisions in areas as diverse as the criminal justice system and healthcare, but concerns abound about whether algorithmic decision-making may induce racial or gender bias. This paper formalizes three notions of algorithmic fairness as constraints on the decision rule, and shows what the optimal decision rule looks like subject to these constraints. The authors then apply these rules to the context of bail decisions, and estimate the costs of imposing different notions of algorithmic fairness in terms of the number of additional crimes committed relative to an unrestrained decision rule.
Athey and Imbens (2016) and Wager and Athey (2017) introduced causal trees and causal forests as new methods for identifying treatment heterogeneity that have potential gains over traditional methods. This paper applies the causal forest method to data from two randomized experiments that evaluated the impact of a summer jobs program on disadvantaged youth in Chicago.
Instrumental variables (IV) is one of the most important tools used to identify causal effects in economic research. If we can find a suitable instrument (relevant and plausibly exogenous), we can exploit it to identify the coefficient on the regressor of interest when we are worred about omitted variables bias. As presented in any standard econometrics textbook, the standard IV set-up makes strong, linearity assumptions. In particular, we assume that the endogenous regressor relates to the outcome of interest linearly and the instrument relates to the endogenous regressor linearly.
deep learning instrumental variables causal inference
Artificial intelligence has been able to achieve human-like and, in some cases, human-superior performance in a variety contexts such image recognition and natural language processing. Yet, the most famous achievements in artificial intelligence have occured in board games. Events like IBM’s DeepMind beating chess grandmaster, Gary Kasparov, and Deepmind’s AlphaGo beating world class Go player, Lee Sedol, captured the public imagination about the future of artificial intelligence.
Underlying many machine learning prediction techniques is an implicit
assumption of sparsity. That is, out of the many potential covariates, machine learning
techniques typically assume that only a few are actually relevant for
the prediction task at hand.
Shrinkage reduces the variance of estimators at the cost of some bias.
Whether this regularization improves precision depends on its tuning, i.e. the choice of shrinkage factor.
James and Stein show that for the estimation of at least three Normal means there is a data-driven choice of the tuning parameter that always beats the unregularized estimator.