Book Read Free

The Economics of Artificial Intelligence

Page 91

by Ajay Agrawal


  grom, and Al Roth (Roth 2002), and even “economist as plumber” (Dufl o

  2017) will move beyond the fi elds of market design and development. As

  digitization spreads across application areas and sectors of the economy, it

  will bring opportunities for economists to develop and implement policies

  that can be delivered digitially. Farming advice, online education, health

  information and information, government- service provision, government

  collections, and personalized resource allocation—all of these create oppor-

  tunities for economists to propose policies, design the delivery and imple-

  mentation of the policy including randomization or staggered roll- outs to

  enable evaluation, and to remain involved through successive rounds of

  incremental improvement for adopted policies. Feedback will come more

  quickly and there will be more opportunities to gather data, adapt, and

  adjust. Economists will be involved in improving operational effi

  ciency of

  government and industry, reducing costs, and improving outcomes.

  Machine- learning methods, when deployed in practice in industry, gov-

  ernment, education, and health, lend themselves to incremental improve-

  ment. Standard practice in the technology industry is to evaluate incre-

  mental improvements through randomized controlled trials. Firms like

  Google and Facebook do 10,000 or more randomized controlled trials

  of incremental improvements to ML algorithms every year. An emerging

  trend is to build the experimentation right into the algorithm using bandit

  techniques. As described in more detail earlier, multiarmed bandit is a term

  for an algorithm that balances exploration and learning against exploiting

  information that is already available about which alternative treatment is

  best. Bandits can be dramatically faster than standard randomized con-

  trolled experiments (see, e.g., the description of bandits on Google’s web

  540 Susan Athey

  site: https:// support.google .com/ analytics/ answer/ 2844870?hl=en) because

  they have a diff erent goal: the goal is to learn what the best alternative is,

  not to accurately estimate the average outcome for each alternative, as in a

  standard randomized controlled trial.

  Implementing bandit algorithms requires the statistical analysis to be

  embedded in the system that delivers the treatments. For example, a user

  might arrive at a web site. Based on the user’s characteristics, a contextual

  bandit might randomize among treatment arms in proportion to the cur-

  rent best estimate of the probability that each arm is optimal for that user.

  The randomization would occur “on the fl y” and thus the software for the

  bandit needs to be integrated with the software for delivering the treatments.

  This requires a deeper relationship between the analyst and the technology

  than a scenario where an analyst analyzes historical data “offl

  ine” (that is,

  not in real time).

  Balancing exploration and exploitation involves fundamental economic

  concepts about optimization under limited information and resource con-

  straints. Bandits are generally more effi

  cient and I predict they will come into

  much more widespread use in practice. In turn, that will create opportunities

  for social scientists to optimize interventions much more eff ectively, and to

  evaluate a large number of possible alternatives faster and with less ineffi

  -

  ciency. More broadly, statistical analysis will come to be commonly placed

  in a longer- term context where information accumulates over time.

  Beyond bandits, other themes include combining experimental and obser-

  vational data to improve precison of estimates (see, e.g., Peysakhovich and

  Lada 2016), and making use of large numbers of related experiments when

  drawing conclusions.

  Optimizing ML algorithms require an objective or an outcome to opti-

  mize for. In an environment with frequent and high- velocity experimenta-

  tion, measures of success that can be obtained in a short time frame are

  needed. This leads to a substantively challenging problem: what are good

  measures that are related to long- term goals, but can be measured in the

  short term, and are responsive to interventions? Economists will get involved

  in helping defi ne objectives and constructing measures of success that can be

  used to evaluate incremental innovation. One area of research that is receiv-

  ing renewed attention is the topic of “surrogates,” a name for intermediate

  measures that can be used in place of long- term outcomes (see, e.g., Athey

  et al. 2016). Economists will also place renewed interest on designing incen-

  tives that counterbalance the short- term incentives created by short- term

  experimentation.

  All of these changes will also aff ect teaching. Anticipating the digital

  transformation of industry and government, undergraduate exposure

  to programming and data will be much higher than it was ten years ago.

  Within ten years, most undergraduates will enter college (and most MBAs

  will enter business school) with extensive coding experience obtained from

  The Impact of Machine Learning on Economics 541

  elementary through high school, summer camps, online education, and

  internships. Many will take coding and data analysis in college, viewing

  these courses as basic preparation for the workforce. Teaching will need to

  change to complement the type of material covered in these other classes.

  In the short run, more students may arrive at econometrics classes thinking

  about data analysis from the perspective that all problems are prediction or

  classifi cation problems. They may have a cookbook full of algorithms, but

  little intuition for how to use data to solve real- world problems or answer

  business or public policy questions. Yet, such questions are prevalent in the

  business world: fi rms want to know the return on investment on advertising

  campaigns,2 the impact of changing prices or introducing products, and so

  on. Economic education will take on an important role in educating stu-

  dents in how to use data to answer questions. Given the unique advantages

  economics as a discipline has at these methods and approaches, many of the

  newly created data science undergraduate and graduate programs will bring

  in economists and other social scientists, creating an increased demand

  for teaching from empirical economists and applied econometricians. We

  will also see more interdisciplinary majors; Duke and MIT both recently

  announced joint degrees between computer science and economics. There

  are too many newly created data science master’s programs to mention, but

  a key observation is that while early programs most commonly have emerged

  from computer science and engineering, I predict that these programs will

  over time incorporate more social science, or else adopt and teach social

  science empirical methods themselves. Graduates entering the workforce

  will need to know basic empirical strategies like diff erence- in-diff erences

  that often arise in the business world (e.g., some consumers or areas are

  exposed to a treatment and not others, and there
are important seasonality

  eff ects to control for).

  A fi nal prediction is that we will see a lot more research into the societal

  impacts of machine learning. There will be large- scale, very important regu-

  latory problems that need to be solved. Regulating the transportation infra-

  structure around autonomous vehicles and drones is a key example. These

  technologies have the potential to create enormous effi

  ciency. Beyond that,

  reducing transportation costs substantially eff ectively increases the supply

  of land and housing in commuting distance of cities, thus reducing housing

  costs for people who commute into cities to provide services for wealthier

  people. This type of reduction in housing cost would be very impactful for

  the cost of living for people providing services in cities, which could reduce

  eff ective inequality (which may otherwise continue to rise). But there are a

  plethora of policy issues that need to be addressed, ranging from insurance

  2. For example, several large technology companies employ economists with PhDs from

  top universities who specialize in evaluating and allocating advertising spend for hundreds of millions of dollars of expenditures; see Lewis and Rao (2015) for a description of some of the challenges involved.

  542 Susan Athey

  and liability, to safety policy, to data sharing, to fairness, to competition

  policy, and many others. Generally, the problem of how regulators approach

  algorithms that have enormous public impact is not at all worked out. Are

  algorithms regulated on outcomes, or on procedures and processes? How

  should regulators handle equilibrium eff ects, for example, if one autono-

  mous vehicle system makes a change to its driving algorithms, and how is

  that communicated to others? How can we avoid problems that have plagued

  personal computer software, where bugs and glitches are common following

  updates? How do we deal with the fact that having an algorithm used by 1

  percent of cars does not prove it will work when used by 100 percent of cars,

  due to interaction eff ects?

  Another industry where regulation of ML is already becoming prob-

  lematic is fi nancial services. Financial- service regulation traditionally con-

  cerned processes, rules, and regulations. There is not currently a framework

  for cost- benefi t analysis, or deciding how to test and evaluate algorithms,

  and determining an acceptable error rate. For algorithms that might have an

  eff ect on the economy, how do we assess systematic risks? These are fruitful

  areas for future research as well. And of course, there are crucial questions

  about how ML will aff ect the future of work, as ML is used across wider

  and wider swaths of the economy.

  We will also see experts in the practice of machine learning and AI col-

  laborate with diff erent subfi elds of economics in evaluating the impact of

  AI and ML on the economy.

  Summarizing, I predict that economics will be profoundly transformed

  by AI and ML. We will build more robust and better- optimized statistical

  models, and we will lead the way in modifying the algorithms to have other

  desirable properties, ranging from protection against overfi tting and valid

  confi dence intervals, to fairness or nonmanipulability. The kinds of research

  we do will change; in particular, a variety of new research areas will open

  up, with better measurement, new methods, and diff erent substantive ques-

  tions. We will grapple with how to reorganize the research process, which

  will have increased fi xed costs and larger- scale research labs, for those who

  can fund it. We will change our curriculum and take an important seat at

  the table in terms of educating the future workforce with empirical and

  data science skills. And, we will have a whole host of new policy problems

  created by ML and AI to study, including the issues experienced by parts of

  the workforce who need to transition jobs when their old jobs are eliminated

  due to automation.

  21.6 Conclusions

  It is perhaps easier than one might think to make predictions about the

  impact of ML on economics, since many of the most profound changes are

  The Impact of Machine Learning on Economics 543

  well underway. There are exciting and vibrant research areas emerging, and

  dozens of applied papers making use of the methods. In short, I believe there

  will be an important transformation. At the same time, the automation of

  certain aspects of statistical algorithms does not change the need to worry

  about the things that economists have always worried about: is a causal

  eff ect really identifi ed from the data, are all confounders measured, what

  are eff ective strategies for identifying causal eff ects, what considerations are

  important to incorporate in a particular applied setting, defi ning outcome

  metrics that refl ect overall objectives, constructing valid confi dence inter-

  vals, and many others. As ML automates some of the routine tasks of data

  analysis, it becomes all the more important for economists to maintain their

  expertise at the art of credible and impactful empirical work.

  References

  Abadie, A., A. Diamond, and J. Hainmueller. 2010. “Synthetic Control Methods for

  Comparative Case Studies: Estimating the Eff ect of California’s Tobacco Control

  Program.” Journal of the American Statistical Association 105 (490): 493– 505.

  Asher, S., D. Nekipelov, P. Novosad, and S. Ryan. 2016. “Classifi cation Trees for

  Heterogeneous Moment- Based Models.” NBER Working Paper no. 22976, Cam-

  bridge, MA.

  Athey, S. 2017. “Beyond Prediction: Using Big Data for Policy Problems.” Science

  355 (6324): 483– 85.

  Athey, S., M. Bayati, N. Doudchenko, G. Imbens, and K. Khosravi. 2017. “Matrix

  Completion Methods for Causal Panel Data Models.” Cornell University Library.

  arXiv preprint arXiv:1710.10251.

  Athey, S., D. Blei, R. Donnelly, and F. Ruiz. 2017. “Counterfactual Inference for

  Consumer Choice across Many Product Categories.” Working paper.

  Athey, S., D. M. Blei, R. Donnelly, F. J. Ruiz, and T. Schmidt. 2018. “Estimating

  Heterogeneous Consumer Preferences for Restaurants and Travel Time Using

  Mobile Location Data.” AEA Papers and Proceedings 108:64– 67.

  Athey, S., R. Chetty, G. Imbens, and H. Kang. 2016. “Estimating Treatment Eff ects

  Using Multiple Surrogates: The Role of the Surrogate Score and the Surrogate

  Index.” Cornell University Library. preprint arXiv:1603.09326.

  Athey, S., D. Coey, and J. Levin. 2013. “Set- Asides and Subsidies in Auctions.”

  American Economic Journal: Microeconomics 5 (1): 1– 27.

  Athey, S., D. Eckles, and G. W. Imbens. 2015. “Exact P-Values for Network Interfer-

  ence.” NBER Working Paper no. 21313, Cambridge, MA.

  Athey, S., and P. A. Haile. 2007. “Nonparametric Approaches to Auctions.” Hand-

  book of Econometrics 6:3847– 965.

  Athey. S., and G. Imbens. 2015. “A Measure of Robustness to Misspecifi cation.”

  American Economic Review 105 (5): 476– 80.

  ———. 2016. “Recursive Partitioning for Heterogeneous Causal Eff ects.” Proceed-

  ings of the National Academy of Sciences 113 (27): 7353– 60.

&
nbsp; ———. 2017. “The State of Applied Econometrics: Causality and Policy Evalua-

  tion.” Journal of Economic Perspectives 31 (2): 3– 32.

  544 Susan Athey

  Athey, S., G. Imbens, T. Pham, and S. Wager. 2017. “Estimating Average Treat-

  ment Eff ects: Supplementary Analyses and Remaining Challenges.” American

  Economic Review 107 (5): 278– 81.

  Athey, S., G. W. Imbens, and S. Wager. 2016. “Approximate Residual Balancing:

  De- Biased Inference of Average Treatment Eff ects in High Dimensions.” Cornell

  University Library. preprint arXiv:1604.07125.

  Athey, S., J. Levin, and E. Seira. 2011. “Comparing Open and Sealed Bid Auctions:

  Evidence from Timber Auctions.” Quarterly Journal of Economics 126 (1): 207– 57.

  Athey, S., M. M. Mobius, and J. Pál. 2017. “The Impact of Aggregators on Inter-

  net News Consumption.” Working Paper no. 3353, Stanford Graduate School of

  Business.

  Athey, S., J. Tibshirani, and S. Wager. 2017. “Generalized Random Forests.” Cornell

  University Library. https:// arxiv .org/ abs/ 1610.01271.

  Athey, S., and S. Wager. 2017. “Effi

  cient Policy Estimation.” Cornell University

  Library. https:// arxiv .org/ abs/ 1702.02896.

  Bai, J., and S. Ng. 2008. “Large Dimensional Factor Analysis.” Foundations and

  Trends® in Econometrics 3 (2): 89– 163.

  Banbura, M., D. Giannone, M. Modugno, and L. Reichlin. 2013. “Now- Casting

  and the Real- Time Data Flow.” ECB Working Paper no. 1564, European Central

  Bank.

  Bastani, H., and M. Bayati. 2015. “Online Decision- Making with High- Dimensional

  Covariates.” SSRN Electronic Journal. https:// www .researchgate .net/ publication

  / 315639905_Online_Decision- Making_with_High- Dimensional_Covariates.

  Belloni, A., V. Chernozhukov, and C. Hansen. 2014. “High- Dimensional Methods

  and Inference on Structural and Treatment Eff ects.” Journal of Economic Perspec-

  tives 28 (2): 29– 50.

  Bickel, P. J., C. A. Klaassen, Y. Ritov, J. Klaassen, J. A. Wellner, and Y. Ritov. 1993.

  Effi

  cient and Adaptive Estimation for Semiparametric Models. Baltimore: Johns

  Hopkins University Press.

  Bjorkegren, D., and D. Grissen. 2017. “Behavior Revealed in Mobile Phone Usage

  Predicts Loan Repayment.” SSRN Electronic Journal. https:// www .research

 

‹ Prev