The big dangers of ‘big data’
By Konstantin Kakaes
Updated 12:11 PM ET, Wed February 4, 2015
- Kakaes: Many people and institutions are intoxicated by the potential of big data
- He warns data can mislead us and not every valid judgment can be summed up in a number
Konstantin Kakaes, a program fellow at New America’s International Security Program, is the author of the e-book “The Pioneer Detectives: Did a Faraway Space Probe Prove Einstein and Newton Wrong?” This is the fifth in a series, “Big Ideas for a New America,” in which the think tank New America spotlights experts’ solutions to the nation’s greatest challenges. The opinions expressed in this commentary are solely those of the author.
(CNN)One way to tell the story of human progress is to gawk at our increasing capacity to measure. Precise statements such as “the economy shrank by 2.7%” replace generalities like “times are tough.”
“Big data” and “evidence-based policy” are the dominant ideas of our moment. A May 2014 White House report put it this way: “Big data will become an historic driver of progress, helping our nation perpetuate the civic and economic dynamism that has long been its hallmark.”
The White House report presents big data as an analytically powerful set of techniques. It says the social and economic value created by big data should be balanced against “privacy and other core values of fairness, equity and autonomy.”
But the White House effort to balance the costs and benefits of big data misses the bigger picture. There are limits to the analytic power of big data and quantification that circumscribe big data’s capacity to drive progress.
Data-driven techniques are only one part of how government, industry and civil society should make important decisions. Bad use of data can be worse than no data at all. As a December 2014 New York Times Magazine story about Marissa Mayer, Yahoo’s chief executive, pointed out:
“Mayer also favored a system of quarterly performance reviews, or Q.P.R.s, that required every Yahoo employee, on every team, be ranked from 1 to 5. The system was meant to encourage hard work and weed out underperformers, but it soon produced the exact opposite. Because only so many 4s and 5s could be allotted, talented people no longer wanted to work together; strategic goals were sacrificed, as employees did not want to change projects and leave themselves open to a lower score.”
As the Yahoo example shows, the presumption that quantitative techniques objectively assess “what works” is deeply flawed. Many attempts to collect and interpret data not only miss key factors, but transform for the worse the systems they claim only to be measuring.
A legal test
Sheri Lederman, a fourth grade teacher on Long Island, sued the New York State Education Department in October 2014 in what is perhaps the clearest legal test case of the dangers of big data. Lederman is highly regarded by her peers and superiors, an “exceptional educator” in the words of her school district’s superintendent.
Yet a statistical technique called “value-added modeling” that purports to evaluate teachers based on students’ standardized test scores said Lederman was ineffective. The American Statistical Association has criticized value-added modeling as an ineffective measure. “Ranking teachers by their VAM scores can have unintended consequences that reduce quality,” the statisticians said.
Despite the skepticism of statisticians — the experts best aware of the weaknesses of the tools they created — bureaucrats at the state Department of Education have embraced the use of value-added modeling. Lederman appears to be just one individual among the many who are being hurt by the vogue for data.
The impulse to overuse data is not unique to educational bureaucrats. Quantification centralizes bureaucratic power and gives outsize importance to short-term effects because they are easier to measure. It is not a question of balancing the power of big data against its dangers, but of recognizing the nonobvious limitations of that power.
The central claim of data proponents is that data always has some positive value. This premise is false. Data-gathering that seems innocuous enough to the managerial class often brings with it undue burden on the subjects of the data gathering.
Take reports attributed to Amazon customer service representatives about how each moment of their workday is monitored and measured, or similar practices recalled by people who said they had been Target employees. In both instances, the decentralized, human processes in which supervisors evaluate their subordinates have been replaced by centralized, quantitative metrics. This shift has been taking place across retail, customer service and food preparation sectors, which together account for over 20% of America’s workforce.
As a result, in the words of a person reported by Gawker to have been a manager at Target, “Of course we cheated, as the saying went, if you weren’t cheating you weren’t trying.” According to the former manager’s statement, Target’s corporate management was trying to increase customer satisfaction by measuring customer satisfaction scores, which employees falsified. If those compiling the data cheat, the data won’t be useful to the central office.
The burden of compiling data causes retail employees who previously had some professional autonomy to feel constantly under centralized surveillance.
Data is far more subject to manipulation than its proponents realize. Even the 2011 McKinsey Global Institute report that popularized the term “big data” acknowledged that its central claim that “we are on the cusp of a tremendous wave of innovation, productivity and growth … all driven by big data,” was supposition. “As of now,” McKinsey admitted “there is no empirical evidence of a link between data intensity … and productivity in specific sectors.” In the intervening years, such evidence remains scant, even as the quantification bandwagon has gathered steam.
College rankings and federal sentencing guidelines, for example, are both quantifications of complex social systems that are broadly agreed to have harmed the systems they set out to standardize and order.
What data won’t tell us
Many important questions are simply not amenable to quantitative analysis, and never will be.
Where should my child go to college, or when? How should we punish criminals? Are charter schools a good idea? Should we fund the human genome project, or basic science in general? Should we have preschools? Taking quantitative answers to these questions seriously not only risks getting the answer wrong, but shapes the underlying reality in ways that are detrimental to our collective well-being.
Such questions call for informed judgment that balances values, incentives, context and other factors. It is often difficult to find disinterested individuals who can balance these factors and be trusted. That difficulty is inherent in all social systems.
Settling vital questions on the basis of informed judgment only appears to be more subjective than using quantitative techniques. By laundering their biases and preconceptions into the methodology they use to devise quantitative metrics, policymakers and social scientists can fool themselves and others into believing they are impartial and unbiased.
To take another example, ascertaining the worth of the human genome project ought to depend on one’s view of the value of the knowledge derived from it within the domain of biology and medicine.
There is simply no useful way to assess the long-term economic impact of either the human genome project or of the Parthenon, and to do so is to miss the point. Pericles didn’t build the Parthenon to draw tourists to downtown Athens 2,500 years later. The fact that it is now a tourist attraction does little to explain its value.
Similarly, investments in understanding the human genetic code will be realized over time, and cannot be justified in terms of their short-term economic impact.
To focus on the many methodological flaws in the return-on-investment techniques used by the Battelle Memorial Institute in the study Obama was referring to is to miss the point. (The Battelle study counts money spent on the human genome project as both a cost and a benefit, for example.)
Obama proclaimed the precise numerical return as a totem, which legitimized the money spent. But the strong case for the human genome project rests on the knowledge it created, rather than the economic benefit, which cannot be meaningfully measured.
Of course the fact that quantitative studies of social and economic systems were systematically flawed in the recent past is no proof that future investigations will suffer from the same shortcomings. However, it is reasonable to believe that if the same basic methodology is used — even if more data is gathered — these flaws will persist.
The only way to understand the fact of the matter about whether butter is good or bad for you is to actually understand what happens when you eat butter, not to continue to try to tease out more intricate statistical regressions between health indicators and butter consumption. (Large effects — like the link between smoking and lung cancer — do show up using such techniques, but for more subtle effects, the answers depend very much on how statistics are compiled.)
In the late 1980s, spurred by the publication of James Gleick’s best-selling book, “Chaos: Making A New Science,” there was a wave of popular attention paid to the then-nascent discipline of chaos theory. Gleick introduced the public to the idea that many real-world systems exhibit “sensitive dependence on initial conditions.” Change the inputs slightly, and radically different outputs will emerge. It is impossible to pinpoint with certainty just what causes, say, a hurricane to form.
Human social systems — public primary and secondary schools, universities or the criminal justice system — are complex systems, just like the weather. The vogue of attention to chaos theory passed before policymakers came to take it seriously. Understanding the complexity of social systems means understanding that conclusive answers to causal questions in social systems will always remain elusive. Gathering more data — twice as much, 10 times as much, a hundred times as much — won’t change this.
To effectively debate public policy or corporate strategy, we will have to continue to have debates over principles. In such debates, disagreement among individuals with different ideological presuppositions will continue.
To believe that disinterested, “rigorous” quantitative judgment can be systematically substituted for such debate imperils programs and practices whose costs are direct, but whose benefits are indirect and thus more difficult to measure. Ease of measurement does not correspond with importance. The administrative apparatus of evidence generation does not, as it claims to, merely pursue “good policy” but is itself a self-interested actor pursuing particular political ends.
A December 2014 book published by the Brookings Institution, “Show Me the Evidence: Obama’s Fight for Rigor and Results in Social Policy,” sums up this belief: “The vision of the evidence-based movement is that the nation will have thousands of evidence-based social programs that address each of the nation’s most important social problems and that under the onslaught of these increasingly effective programs, the nation’s social problems will at last recede.”
This grandiose vision of evidence as panacea is dangerous and damaging. Unless the evangelists of evidence are resisted, they will steamroll over what they cannot measure, leaving us poorer as individuals and as a society, buried in a bureaucracy of numbers untethered from reality.