How Long Employees Stay

Turnover costs companies more than people realize. The real question is not how many people left. It's who leaves sooner, and why. That is a question about timing, and timing is what survival analysis is built for.

Most companies answer this with logistic regression instead. Logistic regression can only ask yes or no: did this person leave? That throws away everyone who is still working there, or forces a guess about how to label them. Survival analysis doesn't have that problem. "Still employed" is treated as what it is: time that hasn't run out yet. It also shows when people are most likely to leave, not just whether they will.

The data is a public research dataset. It covers 1,129 employees across retail, banking, IT, manufacturing, and other industries. There are 16 variables and no missing values: how long each person stayed, how they were hired, whether they had a coach, how they were paid, how they commuted, and five personality scores.

Two variables stand out. First, coaching: whether the employee has a coach, and who it is (an external coach, no coach, or their own manager). Second, pay: whether the salary is fully declared, or partly paid off the books in what's called a "grey wage" arrangement. The five personality scores look like they come from the Big Five model.

Before building the model, I grouped two variables that had too many small categories to work with. Industry went from 16 levels down to 5. Profession went from 15 down to 7. Any category with enough people kept its own label. The rest got grouped into "other."

Before building the full model, I checked four things on their own.

Gender: I compared survival curves for men and women. No real difference. The curves overlap the whole time.

Age: each extra year of age added about 2% more risk of leaving. Small, but real.

Personality: none of the five personality scores mattered much on their own. Self-control came closest, but still not significant.

Coaching: this is where the sharpest result in the whole study showed up. Among employees with a fully declared salary, people with no coach left at 2.3 times the rate of people coached by their own manager.

Then I built the full model, using every variable and letting stepwise selection pick what mattered. Thirteen effects came out significant. You can see them in the chart above and the table below.

Recruitment method mattered most. Employees the company found first, through a resume on a job site, left at more than twice the rate of people who applied after seeing an ad. A fully declared salary cut the risk of leaving by 40%. Two personality traits, anxiety and self-control, showed up as small but real protective effects, even though neither mattered on its own. That usually means other variables were hiding their effect.

There's an obvious problem with this sample. Two thirds of the employees worked in HR, and three quarters were women. So I reran the model two more ways.

First, on a subsample capped at 100 HR employees, which brings the mix closer to a normal workforce. Six effects held up, and most got stronger. Second, on HR employees only, to see how much the HR-heavy sample was pulling on the results.

Age and both personality effects dropped out of the balanced model. I'd treat those as the least reliable findings. Recruitment method, banking, and pay arrangement held up in every version.

What can a company actually act on here? Not age, industry, or profession. Those are not levers. Recruitment channel, pay transparency, and who does the coaching are the real levers. The strongest signal: when the line manager acts as coach, fully declared employees stay noticeably longer. If that holds up elsewhere, it's a retention fix that costs nothing to run.