The Art and Science of Predictive Analytics in HR – by Helen Friedman, Anna Marley and Wendy Hirsch
By | IHRIM Marketing | ihrim.org
After creating predictive models for drivers of so many HR outcomes in the past 20+ years, there is one fundamental truth that has stood the test of time. The most actionable models require as much art as they do science.
Earlier in our careers in people analytics, we were preoccupied by creating the “best possible” predictions. After all, organizations were asking us where their greatest risks were in hiring, engaging, developing, promoting and retaining talent. What would be more important than the accuracy of those predictions?
Here’s the problem. Generally, albeit not always, the best predictions may not provide sufficient insight to act in a way that allows an organization to change the ultimate outcome.
As a practical example, three common variables driving an outcome like turnover tend to be age, tenure with an organization and career, or pay, level. The more experience someone has with an organization or in their career overall and the more advanced they are professionally, the less likely they are to leave. Of course, we all can think of exceptions to these general patterns, but they are general patterns, nonetheless. As such, we inevitably see these variables as factors driving turnover—recognizing that there can be intercorrelations between these variables in some organizations and situations. That said, these factors typically are not the most actionable from an intervention perspective. We cannot alter someone’s age, for example, although perhaps we have new opportunities for speeding up the aging process in the future with the potential advent of commercial space travel (as there has been some research to suggest that it accelerates aging!)?
All kidding aside (or perhaps not?), the more actionable variables in many statistical models predicting workforce outcomes from an organizational standpoint tend to be less about the individual and more about the teams and organizational structures within which they work.
Think about your organization. When teams get too large, do employees have less support? When there is less promotional opportunity due to the embedded hierarchy or lack of internal movement, is there an unintended perception that it is easier to be successful by leaving your organization? Even if there is promotional opportunity, is there a perception that it is easier to change roles externally vs. internally?
These are but a few examples of team and organizational factors that could impact predictive models in HR for which we potentially could take action. We might categorize these variables as more of the “art” of predictive analytics, since they tend to be derived from the individual-level, or raw, data vs. representative of the raw data itself, which might be more traditionally characterized as the pure “science” of prediction.
As an example, a span of control is a variable that could be calculated and “attached” to the data for both the manager (or “owner” of that span) and the team member (or “participant” in that span). We then can create a statistical model to see the consequences of spans of control on both the owner and the participant. If negative consequences are found when our spans of control are too large (for the owner, the participant, or both), we could consider redefining teams. Now, you could argue: Making spans of control smaller may increase organizational costs. And, you could be right, depending on how these interventions were enacted and the financial effect of any cost implications after accounting for the benefit of any productivity, performance and turnover differentials. While there certainly is a Return on Investment (ROI) that needs to be considered carefully when evaluating any action, the key is to identify impactful actions vs. predictive factors in and of themselves.
Just as art can be both presented and interpreted in vastly different ways by different audiences, so too can the fundamental definition of success vary from organization to organization. Or, what might be deemed actionable in one organization might not hold for another organization in the same way.
As a case in point, we asked Wendy Hirsch, Vice President – HR Technology, Analytics and Services, at Eaton Corporation for how art-and-science has played out in their predictive modeling efforts. Here’s a recent use case that she shared:
As the unpredictable year of 2020 came to a close, the leadership team of one of Eaton’s industrial products divisions was concerned about potential pent-up voluntary turnover among salaried employees. With historically-low voluntary turnover in 2020, business leaders were looking to avoid a rise in turnover costs and lost productivity, if attrition were to return to pre-pandemic levels. Working with the business, Wendy’s team undertook an effort to identify the drivers of voluntary turnover to support actions to maintain the division’s newfound, lower voluntary attrition rate.
While the remit was to apply a scientific approach to uncover the predictors of turnover, Wendy knew that the power was in embedding a combination of both art and science in every process step:
- • Generating hypotheses: Engaging with leaders in the business up-front established hypotheses of what they thought drove people to quit. Their assumptions, anecdotes and perceptions were important inputs into what data might be worth collecting and what statistical models to run. For example, conjectures around work overload and financial uncertainty led to additional data collection beyond the core Human Resource Information System. Part of the art was in finding both ways to collect these data and creating data proxies to test the theories presented by leaders.