We all understand the costs of employee turnover in our business. It's expensive (see Employee Turnover Calculator). Therefore it's vitally important to understand the dynamics of how and when employees exit your business.
Just like death and taxes, in the world of HR, turnover is guaranteed. It sounds a little bleak, but it's not. At some point, all employees will eventually exit the business, it may be within their probation period or it may be after a fruitful twenty years or more.
Once we, as HR professionals come to terms with this fact, it follows that we would want to predict how long we can expect an employee to stay with our business once they're hired. This has utility in many domains, from predicting the return on investment probability of a new hire, to modelling your future workforce by considering future movements in and out of your business.
Survival analysis is a statistical method aimed at determining the expected duration of time until an event occurs. In this instance, the event is an employee exiting the business. Survival analysis answers questions such as: what proportion of our organisation will stay with the business past a certain time? Given they reach a certain tenure, at what rate would we expect them to leave the business.
Survival Analysis - Background
You've probably heard of it before:
Survival Analysis, as the name might suggest was developed in biomedical sciences to analyse the proportion of patients surviving to particular times after the application of a treatment. Since then, it's been applied to many situations where the event of interest is binary, that either it doesn't happen or it does.
- Engineering (Reliability Analysis): how long until a machine or part fails?
- Sales (Churn Prediction): how long until a customer terminates their contract?
- Human resources: how long until an employee terminates?
Generally: “Survival analysis is a set of statistical approaches used to investigate the time until an event of interest.”
In English: “Expected time until an event happens”
Why Survival Analysis?
Turnover calculations are an important metric in most businesses. They are useful to provide a method to track the movement of employee's out of your business and identifying potential risks. However, attrition rates calculated in isolation can sometimes be misleading. They are heavily impacted by reporting periods, i.e. a period of downturn or an acquisition, may significantly skew the output for a particular period.
From a retention strategy perspective, an attrition rate ignores important patterns as it considers time as a chronology, rather than a variable. That is, it treats the turnover of a recent starter the same as a tenured veteran. This results in missing important patterns such as milestone-based turnover trends or staying power of certain employee groups, both of which help inform retention strategy.
Understanding Turnover and Tenure
Turnover and tenure have a complex relationship. A business with a high turnover will logically have a lower average tenure, and at the same time tenure is an important factor in the decision to exit. It follows that any analysis of attrition in isolation from tenure will smooth over important insight.
Considering the above diagram, which visualises the careers of individuals within an organisation (each line represents the start and end date of an employee), a traditional attrition calculation would simply count the number of end-points during a period and divide by the number of lines present at the beginning of the period.
Survival analytics considers all historical and current data points together, including those that remain employed, and groups them by key tenure groups, delivering useful, predictive insights about turnover probability.
Calculating the Survival Function:
The Challenge
The challenge with predicting employee turnover is that for anyone that is still employed at the time of observation, their future behaviour is uncertain. They might resign the next day, continue for another ten years, or anything in between. This uncertainty is called "right-censoring". Each line segment below represents the career of an employee from start to finish date. Survival analytics converts these career lengths to tenures to calculate the probability of reaching each milestone.
To convert time into tenure, which becomes the explanatory variable in the survival function, all start dates are normalised to a 'time zero' (see below).
Survival analytics calculates the probability of a termination occurring, based on the number of employees terminated and the size of the sample still remaining at the time. intelliHR uses the Kaplan-Meier method to estimate the survival function due to its ability to handle right-censored data. This is important for an HR tool as right censored data is so prominent. This method incorporates information from all observations available by splitting tenure into logical milestones (6 months), and considers the probability of reaching the next milestone, assuming all previous milestones were successfully reached.
The Survival Function is the probability that the ‘career’ of an employee will be greater than a particular time.
Reading a Survival Curve:
The probabilities calculated above are plotted on the stepped survival curve (below). Although tenure is based on time and is therefore a continuous variable, the probabilities are calculated by grouping data into logical milestones of six months, giving it the stepped shape that you can see.
Because the probabilities are 'cumulative', meaning that the probability of reaching a given milestone relies on the fact that each previous milestone was achieved, the function is always decreasing. That is, the likelihood of reaching your third year will always be less than or equal to the probability of reaching your second year.
It's expected that each survival curve will start at a probability of 1.0 at t=0. That is, there is a 100% likelihood that an employee will last until their first day (although, we know this isn't always the case, if an employee fails to start). From here, the chart can be read by considering each vertical drop (step) as the change in cumulative probability as tenure advances. As can be seen in the above example:
- Assuming an employee starts (100%) there is an 80% chance they will make it past six months.
- The probability of 'surviving' more than one and a half years is 7% less than reaching one year tenure.
A steep drop off in the curve suggests a greater risk of employees leaving the business at that particular length of tenure. The length of the horizontal line denotes the length of time (tenure) between the event of interest (employee terminating). A long horizontal line means that no employees were terminated across those values of tenure, therefore according to the data, the probability of surviving beyond that point does not decrease.
Key Points:
- Median: your median survival tenure can be read from the chart by drawing a horizontal line from the 0.5 point on the y-axis. The tenure at the point this horizontal intersects the curve is the duration you can expect 50% of your employees to 'survive' until.
- Cumulative break-even: it is useful to consider the cumulative breakeven point for your employees. This can be done on an organisational level, or down to business units or cost centres (see Employee Life-cycle). By calculating the time it takes for employees to start 'paying back' the early investment of onboarding and training, you can calculate the probability of employee's reaching this milestone through your survival curve.
95% Confidence Intervals:
The Kaplan Meier estimate is a statistic and therefore is subject to variance from the true value. The blue shaded areas represent the 95% confidence intervals. The variance in statistics increases with smaller samples. As we expect that there will be less high-tenured employees, certainty of retention probability reduces. This results in a larger blue shaded area at higher tenures, denoting a larger range of 95% confidence. To see the 95% confidence interval at any point on the graph, hover over the marker. This confidence interval should be considered when making any statistical inferences between different data sets.