P R ED ICT ION

ESSAY in health policy. They consider the problem of

deciding which otherwise-eligible patients
should not be given hip replacement surgery
Beyond prediction: Using big data through Medicare. They use SML to predict the
probability that a candidate for joint replace-

for policy problems ment would die within a year from other causes,
and identify patients who are at particularly
high risk and should not receive joint replace-
Susan Athey ment surgery. They argue that “benefits accrue
over time, so surgery only makes sense if some-
Machine-learning prediction methods have been extremely productive in applications one lives long enough to enjoy them; joint re-
ranging from medicine to allocating fire and health inspectors in cities. However, there are placement for someone who dies soon afterward
a number of gaps between making a prediction and making a decision, and underlying is futile—a waste of money and an unnecessary
assumptions need to be understood in order to optimize data-driven decision-making. painful imposition on the last few months of

A
life” (p. 493). In this class of problems, the ra-
recent explosion of analysis in science, few assumptions are required for off-the-shelf tionale for focusing on prediction is clear; the
industry, and government seeks to use prediction techniques to work: The environment average effect of an intervention is known to be
“big data” for a variety of problems. In- must be stable, and the units whose behavior is negative in certain states of the world (if the
creasingly, big-data applications make use being studied should not interact or “interfere” with patient will die soon), so that predicting the state
of the toolbox from supervised machine one another. In many applications, SML tech- of the world is sufficient for the decision to forgo
learning (SML), in which software programs niques can be successfully applied by data scien- the surgery. However, the authors highlight the

Downloaded from http://science.sciencemag.org/ on June 30, 2021

take as input training data sets and estimate or tists with little knowledge of the problem domain. fact that pure prediction methods do not an-
“learn” parameters that can be used to make For example, the company Kaggle hosts predic- swer the more complex question of which patients
predictions on new data. In describing the po- tion competitions (www.kaggle.com/competitions) should be given the highest priority to receive
tential of SML for clinical medicine, Obermeyer in which a sponsor provides a data set, and con- surgery, among those who are likely to survive
et al. (1) have commented that “Machine learn- testants around the world can submit entries, more than a year. The full resource allocation
ing…approaches problems as a doctor progress- often predicting successfully despite limited con- problem requires estimates of heterogeneity in
ing through residency might: by learning rules text about the problem. the effect of surgery, for example, because some
from data. Starting with patient-level observa- However, much less attention has been paid patients may have higher rates of surgical com-
tions, algorithms sift through vast numbers of to the limitations of pure prediction methods. plications than others. The question of optimally
variables, looking for combinations that reliably When SML applications are used “off the shelf” allocating a scarce resource (hip replacement sur-
predict outcomes…where machine learning shines without understanding the underlying assump- gery) to the patients for whom the causal effect
is in handling enormous numbers of predictors— tions or ensuring that conditions like stability of the surgery on patient welfare is highest is a
sometimes, remarkably, more predictors than are met, then the validity and usefulness of much harder problem, one that generally re-
observations—and combining them in nonlinear the conclusions can be compromised. A deeper quires answering counterfactual questions: What
and highly interactive ways.” question concerns whether a given problem would happen under a variety of alternative as-
SML techniques emerged primarily from com- can be solved using only techniques for pre- signment policies, policies that have never been
puter science and engineering, and they have dif- diction, or whether statistical approaches to implemented before?
fused widely in engineering applications such as estimating the causal effect of an intervention In another resource allocation example, it is
search engines and image classification. More re- are required. common in industry to use SML to predict the
cently, the number of applications of SML to sci- Kleinberg et al. (11) highlight a case in which probability of customer “churn,” in which a cus-
entific and policy problems outside of computer off-the-shelf SML techniques can partially, but tomer abandons a company or service, and the
science and engineering has grown. In the public not fully, address a resource allocation problem company responds by allocating interventions
sector, SML models have been introduced in
criminal justice [e.g., (2)]; for predicting economic
well-being at a granular level using mobile data,
satellite imagery, or Google Street View (3, 4, 5);
and for allocating fire and health inspectors in
cities (6), as well as a variety of other urban appli-
cations (7). The techniques have also been used to
classify the political bias of text (8) or the sentiment
of reviews (9). In medicine, SML-based predictive
algorithms have been implemented in hospitals to
prioritize patients for medical interventions based
on their predicted risk of complications (10), and
PHOTO: RICARDO DEARATANHA/LOS ANGELES TIMES

in a wide variety of additional medical applica-

tions, including personalized medicine (1).
The rapid diffusion of SML methods can be
in part attributed to advances in availability of
data, computational techniques and resources,
data analysis techniques, and open-source soft-
ware. Another factor is the simplicity of the prob-
lems the techniques are designed to solve. Very

Graduate School of Business, Stanford University, Stanford,

CA, USA. Fig. 1. What is the best way to allocate food-safety inspectors?
Email: [email protected]

Athey, Science 355, 483–485 (2017) 3 February 2017 1 of 3

 P R ED ICT ION

(such as outreach by salespeople) to those cus- verify. A large literature in causal inference that found that the true return on investment was
tomers at highest risk of churn. Ascarza (12) spans multiple disciplines (social science, com- –63%. Part of the gap between the naïve anal-
documented firms following this type of prac- puter science, medicine, statistics, epidemiology, ysis and the results from the experiment arose
tice, and then used methods from the causal and engineering) has emerged to analyze this type because many people who clicked on eBay search
inference literature to provide empirical evidence of problem [see Imbens and Rubin (13) for a advertisements would have purchased items
that allocating resources according to a simplis- review]. One approach to estimating causal effects from eBay, anyway. Although a click on an eBay
tic predictive model is not optimal. The overlap using data that were not generated from a ad was a strong predictor of a sale—consumers
between the group with highest risk of churn- randomized experiment is to adjust for factors typically purchased right after clicking—the
ing and the group who would respond most to that led to differential inspection probabilities experiment revealed that a click did not have
interventions was only 50%. Thus, treating the in the past, and then to estimate the effect of nearly as large a causal effect, because the con-
problem of retaining customers as if it were a inspection on restaurant-specific health outcomes sumers who clicked were likely to purchase,
prediction problem yielded lower payoffs to (perhaps using audits). Recent methodological anyway.
the firm. advances focus on adjusting for observed con- Beyond resource allocation problems, the dis-
A public-sector resource allocation problem is founders in big-data applications [e.g., (14–16)]. tinction between pure prediction and causal in-
the question of how a city should allocate build- A theme in this literature is that off-the-shelf ference has been the subject of decades of
ing inspectors optimally to minimize safety or prediction methods from SML lead to biased odological and empirical research in many dis-
health violations. New York City’s ciplines. Economics has placed
Firecast algorithm allocates fire particular focus on this distinc-
inspectors according to the pre- tion, perhaps because some of
dicted probability of a violation the most fundamental economic
being detected upon inspection, questions, such as how con-
and Glaeser et al. (6) developed

Downloaded from http://science.sciencemag.org/ on June 30, 2021

sumer demand varies with price,
and implemented a similar sys- cannot be answered with purely
tem for allocating health inspec- predictive models. For exam-
tors to restaurants in Boston, ple, how much of a product would
with preliminary estimates show- consumers buy at different (hy-
ing a 30 to 50% improvement pothetical) price levels? Although
in the number of violations it might seem straightforward
found per inspection. to use off-the-shelf SML to pre-
The decision problem of how dict the outcome “quantity sold”
to optimally allocate inspectors with the price level as an ex-
would fall squarely in the pre- planatory “feature,” in practice,
diction domain if the following this approach fails badly if it is
simplifying assumptions were used as a method to estimate
true: (i) The behavior of the the causal effect of price on quan-
individual establishments being tity sold. Suppose that an analyst
inspected is fixed; and (ii) when has historical data from hotel
problems are identified, they prices and occupancy rates. Typ-
can be immediately fixed at a ically, prices and occupancy are
low cost that does not vary across positively correlated because the
units. Knowing which establish- existing pricing policy for ho-
ments are more likely to have Fig. 2. What is the causal effect of inspections? tels (often implemented through
violations would be equivalent yield management software) spe-
to knowing which ones should cifies that hotels raise their prices
be inspected. However, a more realistic setting estimates of causal effects, but that consistent as they become more fully booked. Off-the-shelf
incorporates heterogeneity across units: A build- and efficient estimation of causal effects can be applications of SML techniques are designed to
ing may be at higher risk of fire due to old achieved by modifying SML techniques. answer the following type of question: If an
wiring, but other considerations make it diffi- Another approach to estimating causal effects is analyst is told that on a particular day, prices
cult to replace the wiring. Other units may have to make use of designed experiments. Blake et al. were unusually high, what is the best prediction
lower predicted risk, but it may be easy and in- (17) used a city-based difference-in-difference meth- of occupancy on that day? The correct answer
expensive to make substantial improvements. odology (implementing a new policy in a ran- is that occupancy is likely to be high. By contrast,
Another consideration is responsiveness; if vi- domly selected set of “treatment” cities while the question of the effect of changing the pricing
olations entail fines, some firms may be more controlling for time trends by also measuring policy is a causal question, and common exper-
sensitive to the prospect of fines than others. outcomes in a set of “control” cities) to evaluate ience indicates that if the firm implemented a new
Overall, solving the city’s inspection allocation the causal effect of search advertising for eBay. policy to systematically raise prices by 5% every-
problem involves estimating the causal effect of Like many search advertisers, eBay relied on his- where, it would be unlikely to sell more hotel rooms.
inspection policies: What is the expected im- torical data to measure the benefit of search A different set of statistical techniques is required
PHOTO: GREGORY REC/STAFF PHOTOGRAPHER

provement in overall quality of units (e.g., food advertising, but it did not attempt to separate to answer this question, perhaps exploiting “natural
poisoning rates) in the city under a new inspector correlation from causality. Rather, eBay mea- experiments” in the data or an approach known
allocation regime? sured advertising effectiveness with a simple as “instrumental variables” [see (13) for a review
Thus, prediction and causal inference are predictive model in which clicks were used to of these techniques]. Recently, several authors
distinct (though closely related) problems. Out- predict sales, finding that the return on in- have combined advances from SML with this
side of randomized experiments, causal inference vestment for advertising clicks (that is, the traditionally “small data” set of methods, both for
is only possible when the analyst makes assumpt- ratio of eBay sales attributed to clicks to the estimating average causal effects (18) and for
ions beyond those required for prediction methods, cost of the advertising clicks) was about 1400%. personalized estimates of causal effects (19).
assumptions that typically are not directly Using the experimental data to measure the Beyond the distinction between prediction
testable and thus require domain expertise to causal effect of the advertisements, the authors and causal inference, methods optimized solely

Athey, Science 355, 483–485 (2017) 3 February 2017 2 of 3

 P R ED ICT ION

for prediction also do not account for other select among job applicants for interviews; but 7. E. L. Glaeser, S. D. Kominers, M. Luca, N. Naik, Big data and big
factors that may be important in data-driven they might wish to incorporate diversity objec- cities: The promises and limitations of improved measures of
urban life (Technical Report, National Bureau of Economic
policy analysis or resource allocation. For ex- tives in the algorithm, or at least prevent in- Research, 2015).
ample, incentives and manipulability can be equities by gender or race. These issues have 8. J. Grimmer, B. M. Stewart, Polit. Anal. 21, 267–297 (2013).
important. If a building or restaurant owner received recent attention in the literature on 9. J. S. Kang, P. Kuznetsova, M. Luca, Y. Choi, in Proceedings
anticipates a low probability of being inspected SML [e.g., (21)]. of the 2013 Conference on Empirical Methods in Natural
Language Processing (Association for Computational
based on these characteristics, he or she may Overall, for big data to achieve its full po- Linguistics, 2013), pp. 1443–1448.
reduce efforts for safety. tential in business, science, and policy, multi- 10. M. Bayati et al., PLOS ONE 9, e109264 (2014).
In an example of data-driven policy where disciplinary approaches are needed that build 11. J. Kleinberg, J. Ludwig, S. Mullainathan, Z. Obermeyer, Am.
manipulability played a role, the market pricing on new computational algorithms from the Econ. Rev. 105, 491–495 (2015).
12. E. Ascarza, Retention futility: Targeting high risk customers
system (MPS) of British Columbia is used to SML literature, but also that bring in the meth- might be ineffective (2016); available at SSRN.
set prices for harvest of timber from government- ods and practical learning from decades of multi- 13. G. W. Imbens, D. B. Rubin, Causal Inference in Statistics, Social,
owned land that has been allocated to timber disciplinary research using empirical evidence and Biomedical Sciences (Cambridge Univ. Press, 2015).
companies under long-term leases. The MPS to inform policy. A nascent but rapidly growing 14. M. Dudık, J. Langford, L. Li, in Proceedings of the 28th
International Conference on Machine Learning (ICML, 2011),
builds a predictive model using data from tim- body of research takes this approach: For ex- pp. 1097–1104.
ber sold at auctions to predict the prices that ample, the International Conference on Machine 15. A. Belloni, V. Chernozhukov, C. Hansen, Rev. Econ. Stud. 81,
would have been obtained if a tract harvested Learning (ICML) in 2016 held separate work- 608–650 (2014).
under a long-term lease had instead been sold shops on causal inference, interpretability, and 16. S. Athey, G. Imbens, S. Wager, Approximate residual
balancing: De-biased inference of average treatment effects
via auction. However, a lease-holder could po- reliability of SML methods, while multidiscipli- in high dimensions; https://arxiv.org/abs/1604.07125
tentially have an incentive to bid artificially nary research teams at Google (22), Facebook (2016).
low in auctions in order to influence the pre- (23), and Microsoft (24) have made available 17. T. Blake, C. Nosko, S. Tadelis, Econometrica 83, 155–174
(2015).

Downloaded from http://science.sciencemag.org/ on June 30, 2021

dicted prices for timber harvested under the toolkits with scalable algorithms for causal in-
18. A. Belloni, D. Chen, V. Chernozhukov, C. Hansen, Econometrica
long-term lease and thus lower their costs of ference, experimental design, and the estima- 80, 2369–2429 (2012).
harvesting from long-term leases. As part of tion of optimal resource allocation policies. As 19. S. Athey, J. Tibshirani, S. Wager, Solving heterogeneous
the model selection process, the predictive mod- the SML research community and other dis- estimating equations with gradient forests; https://arxiv.org/
el for MPS was subject to simulations to assess ciplines continue to join together in pursuit of abs/1610.01271 (2016).
20. S. Athey, P. Cramton, A. Ingraham, Auction based timber
its manipulability by any single large timber solutions to real-world policy problems using big
pricing and complementary market reforms in British
company (20). The model implemented was data, we expect that there will be even greater Columbia, White Paper; ftp://www.cramton.umd.edu/
not the model with the best possible predictive opportunities for methodological advances, as papers2000-2004/ (2002).
power to achieve the desired robustness against well as successful implementations, of data- 21. T. Kamishima et al., “Fairness-aware classifier with prejudice
remover regularizer,” Joint European Conference on Machine
manipulability. driven policy.
Learning and Knowledge Discovery in Databases (Springer,
A number of additional considerations arise 2012).
RE FERENCES AND NOTES
when using statistical models in practice. It is 22. S. L. Scott, Appl. Stochastic Models Data Anal. 31, 37–45
1. Z. Obermeyer, E. J. Emanuel, N. Engl. J. Med. 375, 1216–1219
sometimes important for stakeholders to un- (2016).
(2015).
derstand the reason that a decision has been 23. E. Bakshy, D. Eckles, M. S. Bernstein, “Designing and deploying
2. R. Berk, Criminal Justice Forecasts of Risk: A Machine
online field experiments,” Proceedings of the 23rd International
made, or decision-makers may need to commit Learning Approach (Springer Briefs in Computer Science,
Conference on World Wide Web (Association for Computing
a decision rule to memory (e.g., doctors). Trans- 2012).
Machinery, 2014).
3. N. Naik, R. Raskar, C. A. Hidalgo, Am. Econ. Rev. 106, 128–132
parency and interpretability considerations might (2016).
24. A. Agarwal et al., https://arxiv.org/abs/1606.03966 (2016).
lead analysts to sacrifice predictive power in favor 4. J. Blumenstock, G. Cadamuro, R. On, Science 350, 1073–1076
AC KNOWLED GME NTS
of simplicity of a model. Another consideration (2015).
5. R. Engstrom, J. Hersh, D. Newhouse, “Poverty in HD: What I am grateful to G. Imbens, S. Mullainathan, M. Luca, and S. Wager
is fairness, or discrimination. Consumer protec-
Does High Resolution Satellite Imagery Reveal about Economic for helpful conversations.
tion laws for lending in the United States pro- Welfare?” (2016).
hibit practices that discriminate on the basis of 6. E. L. Glaeser, A. Hillis, S. D. Kominers, M. Luca, Am. Econ. Rev.
race. Firms might wish to use SML methods to 106, 114–118 (2016). 10.1126/science.aal4321

Athey, Science 355, 483–485 (2017) 3 February 2017 3 of 3

Beyond prediction: Using big data for policy problems
Susan Athey

Science 355 (6324), 483-485.

DOI: 10.1126/science.aal4321

Downloaded from http://science.sciencemag.org/ on June 30, 2021

ARTICLE TOOLS http://science.sciencemag.org/content/355/6324/483

RELATED http://science.sciencemag.org/content/sci/355/6324/468.full
CONTENT
http://science.sciencemag.org/content/sci/355/6324/470.full
http://science.sciencemag.org/content/sci/355/6324/474.full
http://science.sciencemag.org/content/sci/355/6324/477.full
http://science.sciencemag.org/content/sci/355/6324/481.full
http://science.sciencemag.org/content/sci/355/6324/486.full
http://science.sciencemag.org/content/sci/355/6324/489.full
http://science.sciencemag.org/content/sci/355/6324/515.full
REFERENCES This article cites 11 articles, 1 of which you can access for free
http://science.sciencemag.org/content/355/6324/483#BIBL

PERMISSIONS http://www.sciencemag.org/help/reprints-and-permissions

Use of this article is subject to the Terms of Service

Science (print ISSN 0036-8075; online ISSN 1095-9203) is published by the American Association for the Advancement of
Science, 1200 New York Avenue NW, Washington, DC 20005. The title Science is a registered trademark of AAAS.
Copyright © 2017, American Association for the Advancement of Science