Just because you can doesn't mean you should

When I started my career in People Analytics, I was eager to see how far data science could go in the HR space.?For all the mysteries of why people do what they do, all of a sudden we had an opportunity to make people and their behaviors a little more predictable.?We finally had access to data, technology, and expertise to answer the tough questions with statistical models instead of guesswork.?As my career has evolved, however, I realized it's not so much what you can do, but how you do it and why that matters more.

Take predicting attrition, for example.?These capabilities are now commonplace, and can even come straight out of the box from most vendors. The challenge is no longer having a model, but knowing how to use it and creating the right strategy that benefits both employees and the company.

With any project, the starting point has to come from solving business challenges.?In the case of turnover, that's easy - no company wants to see good talent walk out the door when it's preventable.?And for employees, if we can satisfy their needs and keep them happy, everyone wins.?But there is an important question you need to start with - how do you responsibly use a model that can tell you what's expected to happen? There are many things you could do:

• Share names with managers or leaders to intervene at the individual level with specific employees
• Share drivers of turnover (e.g., locations, job tenure or age groups, sentiment, changes in performance) and identify solutions related to each attribute that is associated with higher risk
• Issue early alerts of expected turnover in specific teams where meaningful action can be taken to improve the environment at the group level
• Integrate with strategic workforce plans across teams, skills, or locations
• Target universities and hiring sources associated with lower flight risks
• Provide compensation adjustments for at-risk skill sets more proactively

Some of these ideas are sensible and low-risk, while others could easily have unintended consequences.?Getting a broad and diverse set of stakeholders involved is necessary to examine the plan from all angles and discuss potential good and bad outcomes. The best approach will always take into account the worst case scenarios:

• The data could be used improperly. Knowing "this person looks like they're about to leave” could influence decisions about ratings, rewards, promotions, succession planning, special assignments, or placement on other types of lists with positive or negative implications
• The data could be used without proper context by executives who might give it too much weight (e.g., giving retention agreements based on turnover probabilities without other considerations, including potential adverse impact)
• Employees may respond poorly if word gets out or if messaged inappropriately. What exactly is the plan once a business leader has a list of names? Do managers know how to have “stay” conversations? Do employees trust leaders to do the right thing with this information or will the rumor mill run wild?
• Addressing at the individual level will only be a band-aid solution rather than addressing group-level categories driving turnover (e.g., locations, tenures, career growth concerns)?

There are also risks to the model's integrity to consider:

• Changing behavior. Employees who hear the company is basing retention agreements or other decisions on turnover probability may start responding differently when they see survey questions about intent to stay (which will in turn affect the model’s accuracy)
• Unstandardized interventions that are not systematically tracked will affect the model's accuracy moving forward. So when untracked actions lead some "high attrition risk" employees to stay we won’t know if the model is becoming less accurate or if the interventions are working unless there is a system for capturing interventions and feeding them back into the model (which will also help you evaluate which interventions actually work)

Only after you've planned around all these corners should you invest the time and resources needed to build and deploy a predictive attrition model. But once you reach this stage, there are another set of decisions to slow down and consider - including, what type of data is appropriate to include in your model. There is a lot of data to choose from. There are group-level factors that pose little risk in analyzing and understanding - like global factors (e.g., recessions, pandemics), industry unemployment and hiring trends, major company events (e.g., RIFs, acquisitions), leadership changes, division-wide trends, seasonality of turnover (e.g., stock vesting dates, bonus payouts), and other general inputs.

These will be helpful to include in the model, but on their own insufficient without also including individual-level data which has much more predictive utility. You should divide this type of personal information according to sensitivity of the data and potential for misuse. There is a lot of generally accessible information that can be useful without being overly intrusive - things like hire type, education, job level, job function, location, tenure, time since last promotion, number of manager changes, and perhaps other elements like participation in trainings and membership groups, if available.

Most People Analytics teams have access to more sensitive data points, as well, and these should be governed with even more scrutiny, and in accordance with why the data are being collected in the first place. While some data could be highly predictive, it may or may not be appropriate to use. These include things like age, gender, ethnicity, survey responses, salary and performance data (and changes in ratings and rewards over time), unvested stock or retention agreements, commute distance, organizational network data, or other information companies might be collecting about their employees.

Making a misstep here would be more dangerous than going to Jurassic Park during a power outage. Jeff Goldblum put it best:

"Your scientists were so preoccupied with whether or not they could, they never stopped to think if they should."

So what should you do? All this is not to discourage using data to arrive at answers and inform decisions - we have to be data informed! But we need to move thoughtfully, and in coordination with the right stakeholders, the number of which should naturally increase with the size and importance of the problem you're trying to solve. For a project like this, you should also be leveraging a structured data privacy checklist and impact assessment to evaluate and minimize potential risks of the project against the intended benefits - what data will be collected, how it will be secured, how long it will be retained, who will have access, how it will be used and why. There should be transparency with employees and alignment with informed consent practices and employee guidelines for how personal data will be used, as well as compliance with GDPR and other regulations and requirements.

Overall, what should be evident from this example, is that data science skills are necessary, but not sufficient, for People Analytics teams to be successful. The ability to think strategically, partner cross-functionally, and lead with ethics and data governance are all critical to ensure research is carried out in a way that will actually benefit employees. In the case of predicting attrition, the approach should be scalable, standardized, measurable, and strictly governed. That said, it's not a one-size-fits-all solution, so I'd love to hear from other professionals in this space what has been your experience with applying this type of model? How do you recommend using it or not using it?