STRATEGY
Let’s Predict Who’s Going to Quit
How Can Leaders use Data In HR The Right Way?
When Singapore became independent in 1965, DBS was founded three years later as a local bank before serving the region in the 1980s.
As ASEAN’s “Most Valuable Bank Brand” for the 7th consecutive year, the S$12bn brand built a complex, intricate algorithm that predicts attrition rates from 600 data points within the 11000-strong organization.
Using data to predict employee attrition is nothing new: there a plethora of startups offering software and Mercer’s Global Talent Trends Study 2019 revealed that using artificial intelligence in predicting employees’ risk of leaving remains on the wishlist of many companies, even though the technology is still nascent. HR professionals are involving analytics and data are a lot more over the past few years.
According to LinkedIn, that will continue to grow.
It is an uphill struggle to continue preserving competitive advantages and protecting intellectual capital — retaining the best talents is always the top of many lists of HR challenges.
Data scientists have been creating machine learning models and algorithms for a long time: the only variables are the industry and the purpose. Some companies use data to predict sales and others who use data to predict an employee’s fit to a company.
However, there are many dynamics and nuances that go into how HR professionals can use data to help them create the right action plans. How can numbers and spreadsheets point towards the right general direction to driving employee retention?
The Matrimony with Data
Big data and machine learning have been dominating many industries over the past few years, be it in sales, marketing or creative. The reality is that such technologies have become a fundamental and data-driven decision making in leaders is almost a requirement in today’s world.
The Turnover Propensity Index by Holtom & Allen
Working with a talent intelligence firm, Prof. Brooks Holtom and Prof. David Allen embarked on researching into how big data and machine-learning algorithms could predict whether an employee is likely thinking about quitting — the turnover propensity index (TPI) is created as a real-time indicator, grounded on predictive models in academic research.
The TPI is built around turnover factors that include personal and organizational. Another contributing factor is job embeddedness, which is how deeply connected people are to an organization. Along that line, it is tied very closely to core components in employee engagement.
Based on their assessments of the turnover factors, they used machine learning to identify how likely an individual is to be receptive to new job opportunities. A TPI score is given to the individuals in their sample and they ran two studies to see how accurate their predictions are.
Through a logical conclusion, one would expect that the higher someone is on the TPI, the more likely he is to quit.
Hence, they decided to send out e-mail invitations to a sample of 2,000 employed individuals. These individuals have been identified by the algorithm on how likely they are to open an invitation to view available jobs tailored to their specific skills and interests.
Those who were rated as most likely to be receptive to opening the e-mail invitations were twice as likely to open than, in comparison to those rated as being the least likely to be receptive.
They measured the click-through and realized that amongst those who opened the email, those rated as most likely to be receptive were significantly more likely to click through it.
Using the remained of their sample, over three months, they realized that people with high TPI scores were “63% more likely to change jobs” and they were “40% more likely to quit”, compared to the people with low TPI scores.
Employee Attrition & Performance Analytics by IBM
Using a data set from an employee survey from IBM, a team of four used classification models to predict if an employee “is likely to quit could greatly increase the HR’s ability to intervene on time and remedy the situation to prevent attrition”.
Taking into account the mixed employee architectures in IBM, many people attempted to create an ML model through Python. According to the Performance Analysis, the model is successful in effectively classifying 89.12% unknown sets.
Using data for HR analytics has hit the mainstream so much so that there are even courses on building it. There are even articles that go deep into how this can be possible using Python or R.
Having data creates new dynamics for leaders — today, managers can quantify a person’s propensity to leaving a job within seconds. However, the key comes in balancing between quant-based and qual-based analysis: focusing only on the quantitative is a surefire way to introduce bias.
While data is objective, the margin of errors and algorithmic stability is what causes the results to be inaccurate as scale. Qualitative analysis is subjective, which can be skewed by biases.
There are inherent flaws in both, but mustering the strengths and pulling from ends is what makes a great HR leader. Simply relying on data and numbers is not enough, although many do make that mistake.
Garbage In, Garbage Out
Organizations are actively predicting and managing employee flight risk today, with data-driven companies using predictive analytics as a mission-critical tool — all companies would love to know when talent would leave, and how they can do their best to retain them.
It is all about clarity.
However, like any predictive analytics, it is also as good as the data it is fed — garbage in, garbage out, as many in data analysis would say. The conventional method of creating a predictive model is to first use employee data tracked and stored in an organization’s HR system, such as length of tenure, distance from home to work and date of last promotion.
Regardless of the data points HR leaders have on hand, developing the first predictive model is often a confusing and intimidating project. How can they accurately pinpoint what data points hold the highest weight and influence on the turnover probability?
HR leaders must work with data analysts and HR experts to look at getting the ‘right’ data in — quality in, quality out.
- Accuracy is important, but not paramount. It may not be a good evaluation metric if there are variables that are split unequally.
- False positives and false negatives. Organizations need to decide on the errors that would cost them the least. An employee that isn’t looking to leave but yet predicted to is a false positive: retention resources may be wasted). An employee that is looking to leave but yet predicted not to is a false negative: no retention effort is spent on the employee and the company loses a talent.
- Adding predictor variables. False-negative error rates can be decreased when predictor variables come in, such as job satisfaction rating, engagement score, and the number of previous jobs held.
- Adding third-party variables. External data can also influence employees: a poor labor market may be a reason why an employee stays in the company. Such data points can be the length of tenure at a previous job(s) and relative compensation to other employers for similar positions.
- Go specific and deep in your model. Different departments, different nuances. Cultures may be run differently in various parts of the organization — the product development team may function differently from the people in finance. It is possible to run individually.
- Go deep into clusters of employees. If there are factors that can cause flight risk in an employee that are not traditionally considered in HR, predictive in nature or not, a model should be created specifically to run against them separately.
Spreadsheet Mentality
Poor leaders hide behind spreadsheets and review numbers as though they can entirely represent an employee’s intention. Helicopter view leadership is great when there is a need to strategize and go back to the core mission, but it is certainly not welcome when companies make number-driven decisions against humans.
Behind those bar charts and graphs are teams of human beings, each with their own unique identities and opinions.
One of the biggest decisions made using numbers is layoffs. Uber laid off 400 employees this year. American Express did the same in 2016 as well. Companies often cite the need to improve profitability but studies have shown that layoffs often do otherwise — in the long-term, companies will suffer.
Short-termism is the player here. “So, if you think profit is not where you want it to be, you say, ‘I can pull this lever and the costs will go down’,” said Wharton management professor Adam Cobb.
While it is also dependent on the industry (e.g. the digitization of print has resulted in many printing jobs going away, but a transition into digital and tech-centric jobs in the print industry), the effectiveness of layoffs has been overstated over the past few decades, employers often underestimate how much it will cost them exactly in the long run. Like a chimera that comes back to bite them, the costs form up in lower productivity, hiring and related HR costs and others that are not exactly easy to put on a P&L.
Others would be responding to employee engagement survey results. Employers are interested in the idea of engagement because they believe it tells them whether people are working hard, and along the same line, an indication of job performance.
The problem is, employee surveys — not just on engagement — are notoriously looked down upon by employees. With an average of 7 out of 10 employees not answering the survey and almost a third of employees thinking that the survey is pointless, it is an indication that there are clear biases that will affect the responses.
Engagement surveys can’t take into account most things that influence employee behaviors and job performance, such as the tasks people are given, what their supervisors are doing and the progress of the company. Hence, it is clear the predicting job performance through self-reported statements would be unrealistic — how can companies make decisions on them without having insofar to speak to them personally?
Connecting the Quant with Qual
Quantitative-based analysis like surveys is great at signaling areas for change. However, they are not enough to base solutions on: leaders need to sit down with their employees and understand them deeply to seek affirmation and opposing points for their data.
Stay Clear-Headed: Don’t Assume
Most managers, when they realize certain employees are likely to have turnover, will be biased towards them thereby neglecting those who are not. The over-reliance on data gives managers an excuse to not put in as much as of an effort on everyone.
Understanding that data can give incomplete but objective views is the key: managers can direct more resources to those who are likely to have turnover but they must observe the ground, speak to employees and form connections with them to understand whether they have a propensity for flight risk.
Data Doesn’t Speak
Performance metrics are often tied to strategy — often an accepted best practice over the past few decades. Like a building, strategies form the blueprint, metrics the concrete, wood, and bricks.
Metrics are inherently imperfect at some level and most businesses often fail to capture some underlying intangible goal behind the metrics. This is a case for surrogation where managers think about ways to achieve the objectives set up for them in a way that the company may not want, mostly due to an incorrect or vague correlation between the objective and the method of measurement.
For example, tying the objective “Have happy customers” to “Number of Customer Returns” may cause managers to use creative ways to avoid returns, or take gambles and hope that customers may not return slightly lower quality products instead.
Employees Speak, But It May Be Biased
It depends on who they are speaking to: dependent on the department and then on the individual. When asked about an opinion about the company, not everyone is keen to open up with their CEO — many might with their closest superior instead. Others might open up with someone that they rarely worked with, using it as a connection point.
Any distinct, solid brand image of a manager can create bias in the employee’s feedback as well: an authoritarian leader will never fish out the most honest of opinions and a kind and caring leader may not receive the harshest of criticisms. That is a gateway for bias, which also skews the qualitative-based analysis in the micro, which managers need to keep in mind.
Everything is Individual
Every employee has a unique blend of traits that ultimately shows how a one-size-fits-all approach fails to work in the micro.
While in the macro, an overarching strategy can be applied (e.g. “We want engaged employees who feel that their work contributes to the bigger picture), managers in the micro need to ensure that this strategy is applied in various methods. Every employee responds differently.
For instance, the way employees receive feedback varies as well. Some are receptive to instantaneous feedback, others might be more receptive to social comparison (i.e. your colleagues are doing better in this aspect than you). Some might be receptive to delayed feedback.
Managers ought to experiment with informed assumptions, be in the way they give feedback or the way they give words of affirmation. What occurs in the micro is what will execute the strategy well.
In business, managers and execs need to consider every angle and point to drive business results.
Relying solely on data removes the human aspect of human resources and places the business in a vulnerable position: if the data is wrong, so will the assumptions and solutions be.
When resources run scarce, it is important to stay lean and identify the right points to strike at, then staying patient with it — many things take time.
Execs can be disillusioned by the prowess of AI, believing that it can deliver amazing, game-changing results that they believed it could. The ROI comes from the long-term, even if that means spending millions in infrastructure, people and research.
Hence, managers need to double down on the human aspect — it is more than just data. Meshing data and personal insights are what will keep a business constantly progressing and virtually impregnable, which many execs need to understand so urgently today.