Meeting of 10-13-16 Public Meeting on Big Data in the Workplace
Chair Yang and distinguished Commissioners, thank you for the opportunity to speak to you today on the use of big data in employment settings. I am honored and excited to be here to discuss this timely and important topic. My remarks today will focus on (a) defining big data in an employment context (b) current and potential uses of big data in employment settings (c) greater historical and developmental context and (d) discussing opportunities and concerns going forward.
'Big data' means different things to different people. One issue that I would like to clarify immediately is that this is not simply about very large datasets, with many columns and rows. Although the size of these datasets is typically quite large this is not what defines big data. Rather, what makes data 'big' has to do with the nature and the source of the data and how it is collected, merged, transformed and utilized. In the employment context, I would define big data as follows: big data is the combination of nontraditional and traditional employment data with technology-enabled analytics to create processes for identifying, recruiting, segmenting and scoring job candidates and employees.
Nontraditional employment data is stored outside of the traditional personnel data landscape. It comes from places like operations and financial data systems maintained by the employer, public records, social media activity logs, sensors, geographic systems, internet browsing history, consumer data-tracking systems, mobile devices, and communications metadata systems. This list is by no means complete, and every day it grows. Even our faces and voices can be reduced to a stream of code so that a computer system can recognize and analyze the information. This is the sea change that we are here today to talk about-everything is data. Everything that we do and say can be coded, quantified and utilized for analytic purposes. For example, written remarks and testimony from this very meeting can be thought of as data as it will be published to EEOC.gov and thus made public. Our written words can then be scraped from the website, tagged, coded, classified and organized into a matrix which will then be available for analysis. The value in doing this would come not from quantifying information about this meeting alone, but from linking it to other information about each of us coded across the internet or within disparate company, vendor, public information or consumer data bases. As more information is collected and organized about each of us, and as it is linked to outcomes of interest observed over time, predictions can be made about our future behaviors.
Employers may utilize their own resources to collect and analyze this type of nontraditional employment data, or they may purchase the data, or insights gleaned from the data, from brokers or vendors. When this type of information is quantified and brought together with traditional employment data like performance appraisals, job tenure, attendance, absenteeism, and salaries, it can be used to uncover patterns of behaviors and outcomes for workers. Those patterns of behaviors and outcomes can be distilled into profiles that can then be used to predict outcomes for similarly-profiled groups of job candidates, applicants, and employees.
In practice, it appears that the primary motivation behind utilizing big data is the ability to profile employees and job seekers. Data scientists, computer scientists, and analysts generally use traditional and nontraditional employment data to create algorithms or statistical models which predict, classify, or cluster workers on outcome variables like job tenure, turnover, satisfaction, performance appraisals, absenteeism and culture fit. Generally speaking, the algorithm is given a training dataset containing information about a group of people, typically current or former employees, from which it uncovers characteristics that can be correlated with some measure of job success. Given the nature of the data included in the training dataset, the factors that emerge as strong predictors of success may be of the traditional (self-report of previous work experience or education) or nontraditional variety (passively-recorded information about choice of internet browser or number of professional connections outside of one's area of expertise), but they are likely to be some combination of the two. The successful profile can then be used in a number of ways, including seeking out passive job candidates, screening active job applicants, or allocating training resources or incentives for current employees.
Employers might develop a profile of the ideal candidate and search for 'similar' people on social media sites or specialized online communities, then encourage these passive candidates to apply for open positions. Employers or vendors might also develop a test or screen based on the ideal profile and apply it to applicants at any stage in the hiring process. They might use the ideal profile to identify current employees of high potential and target them for training opportunities or even pay increases or bonuses. Keep in mind that 'job success' can be operationally defined in multiple ways, including actual job performance ratings, quantified worker output, tenure, or 'culture fit.'
At the opposite end of the success spectrum, employers can use this profiling technique to identify employees who are likely to have excessive absences, safety incidents, or to turn over within a specified time frame and use that information in conjunction with 'worth' and 'cost' estimates to make employment decisions or choose other subsequent actions. Some specialty vendors have also come onto the scene more recently offering 'matching' type services, where the vendor develops the ideal employee profile for the employer, and creates profiles for job-seekers based on some combination of actively or passively supplied information, then notifies each when a 'match' is made. Finally, some employers have developed talent communities where job seekers can engage with one another, and with current employees of the company, to get to know one another over a period of time. During this time the employer develops a profile for the community member and uses it in a similar manner to that described above.
The types of big data analytics that we are seeing in the employment context seem to have naturally developed from other areas of business like marketing and operations. In marketing, analysts seek to segment and identify groups of people for targeting advertisements. The training dataset utilized to develop the algorithm might include information about people who purchase products, and their personal characteristics. This is the type of process that led to Target's now-famous pregnancy prediction score. Data scientist Andrew Pole and his team were able to develop an algorithm that could predict when a shopper was pregnant, as well as her rough due date. This was useful to Target because it allowed the company to focus their advertisement efforts for items that pregnant women need on the right demographic and at the right time (it turns out that gaining the market loyalty of a pregnant woman in her second trimester is considered by some to be the 'holy grail'). The algorithm was trained using data from previous Target shoppers with baby-shower gift registries. Pole and his colleagues were able to determine, by looking backwards in time at shopping behaviors, that women in the early stages of pregnancy tend to purchase certain specific items (toiletries and vitamins) more often than otherwise-similar women. Armed with this knowledge and going forward, the researchers were able to identify subsequent groups of women with a high pregnancy prediction score before the women set up their baby gift registries or purchased necessities. These women were then delivered the relevant advertisements. Andrew Pole started discussing this work in public in 20101. Prior to that, Target had been tracking purchases and demographic information about customers to use for marketing purposes for decades, and Target is just one example. Given that fact that the vast-majority of people move about while carrying 'tracking devices' at all times (mobile phones) it is increasingly possible to accurately predict our next movements, as well as our physical locations at specific set points in the future2 and retailers, years ago, began to use this type of location and movement data to target the right consumers with the right ads at the right time3.
It was somewhat inevitable that this type of work would spill over from marketing to employment, particularly when employers have, or are able to collect, so much information about worker characteristics and performance. Why not optimize hiring and talent management in the same way that we've optimized advertising; particularly when return on investment can be quantified and reported to senior management? Furthermore, the types of software, hardware and skill sets required to do this type of statistical and analytic work are becoming more attainable for the masses thanks to open-source software, cloud computing options, and free online and in-person training opportunities.
This is all happening within a larger context of flourishing artificial intelligence and cognitive computing. Machine learning and natural language processing are already commonly utilized in areas like medicine, banking, wealth-management and even in the criminal justice system. It is expected that within the next five to ten years these types of technologies will impact every important decision that we make in our work and personal lives. Within that time frame self-driving cars are expected to proliferate, 25% of all job tasks will be offloaded to software and 13.6 million jobs will be created for people who know how to work with artificial intelligence tools4. All of this is to say that the proliferation of machine learning techniques and predictive analytics in the employment landscape has been coming for some time and its development is expected to continue and accelerate.
Of course employers want to optimize their selection and talent management strategies to best service the goals of the company. To the degree that this optimization leads to innovations that promote objectivity and equal opportunity, those efforts should be commended. However, employers should not lose sight of the fact that when criteria affecting employment decisions-- including those identified by machine-developed algorithms-- have an impact based on characteristics like race, gender, age, national origin, religion, disability status, and genetic information, those criteria require careful scrutiny. It is the employer's responsibility to utilize vendor tests and screens responsibly, to understand the selection products that they are utilizing or purchasing, and to determine whether these screens result in adverse impact on particular demographic groups. Where the use of these algorithms evidence adverse impact, it is the employer's responsibility to maintain validity evidence that supports their use. Part of the validity assessment should be whether the employer can use the selection procedure in a way that would reduce its disparate impact, or whether another procedure would have less disparate impact.
I hope that the issues raised in today's meeting will serve as an important reminder to vendors and employers, especially given that many of the people who build and maintain these algorithms may not be familiar with equal employment opportunity law. Computer and data scientists transitioning from marketing into employment algorithm development, for example, may lack the regulatory and legal background required to make complex decisions about EEO compliance. Employers who choose to purchase or adopt these strategies must be warned to not simply 'trust the math' as the math in this case has been referred to, by at least one mathematician/data scientist, as an 'opinion formalized in code5.'
The primary concern is that employers may not be thinking about big data algorithms in the same way that they've thought about more traditional selection devices and employment decision strategies in the past. Many well-meaning employers wish to minimize the effect of individual decision-maker bias, and as such might feel better served by an algorithm that seems to maintain no such human imperfections. Employers must bear in mind that these algorithms are built on previous worker characteristics and outcomes. These statistical models are nothing without the training data that is fed to them, and within that, the definition of 'success' input by the programmer. It is the experience of previous employees and decision-makers that is the source of that training data, so in effect the algorithm is a high-tech way of replicating past behavior at the firm or firms used to create the dataset. If past decisions were discriminatory or otherwise biased, or even just limited to particular types of workers, then the algorithm will recommend replicating that discriminatory or biased behavior.
As an example of the type of EEO problems that could arise with the use of these algorithms, imagine that a Silicon Valley tech company wished to utilize an algorithm to assist in hiring new employees who 'fit the culture' of the firm. The culture of the organization is likely to be defined based on the behavior of the employees that already work there, and the reactions and responses of their supervisors and managers. If the organization is staffed primarily by young, single, White or Asian-American male employees, then a particular type of profile, friendly to that demographic, will emerge as 'successful.' Perhaps the successful culture-fit profile is one of a person who is willing to stay at the job very late at night, maybe all night, to complete the task at hand. Perhaps this profile is one of a person that finds certain perks in the workplace, such as free dry cleaning , snacks, and a happy hour on Fridays preferable to others like increased child-care, medical and life insurance benefits. Finally, perhaps the successful profile is one of a person who does not own a home or a car and rather appears to bike or walk to work. If the decision-makers at this hypothetical firm look to these and other similar results to assist in the recruiting of passive candidates, or to develop a type of screen, giving preference to those future job-seekers who appear to 'fit the culture', the employer is likely to screen out candidates of other races, women, and older workers. In this situation, not only would the algorithm cause adverse impact, but it would likely limit the growth of the firm.
The use of big data algorithms could also potentially disadvantage people with disabilities. Academic research indicates that social media patterns of usage are related to mood disorders, for example6. If a machine learning algorithm was to uncover a link between absenteeism and social media posting patterns, its result might suggest that a particular employee, who has recently been posting to social media during certain hours of the night, has a heightened 'absenteeism risk' score. Perhaps when it comes time for performance review, this 'absenteeism risk' score might be reviewed, alongside a heightened 'flight risk' score and the employer may avoid offering certain incentives or career development opportunities to the employee, rather offering those to others with more preferable profiles.
Finally, it merits mention that the relationships among variables that are uncovered by advanced algorithms seem, at this point, exclusively correlational in nature. No one argues that the distance an employee lives from work, or her affinity for curly french fries, the websites she visits, or her likelihood to shop at a particular store, makes her a better or worse employee. The variables and outcomes may be correlated because each is also correlated with other variables that are actually driving the causal aspect of the relationship. For example, with regard to distance from work-it isn't likely that the actual distance causes a different score on the success factor but perhaps the time it takes to commute requires the employee to leave earlier than she otherwise would, or perhaps the commuting increases her stress level thereby reducing some aspect of the quality of her work. It would seem to behoove the employer or vendor uncovering this relationship to do some additional, theory-driven research to understand its true nature rather than to stop there and take distance from work into account when making future employment decisions. This is true not only because making selections based on an algorithm that includes distance from work, or some other proxy representing geography, is likely to affect people differently based on their race but also because it is simply an uninformed decision. It is an uninformed decision that has real impact on real people. Rather, perhaps selecting on some variable that is causally related to work quality, in conjunction with offering flexible work arrangement options, might represent both better business and equal opportunity for workers. Thank you.