Intuition

Fantastic Inc. has 500 employees and virtually all of them love working there. When asked on a scale of 1-5 if they enjoy their work, 80% reported a 4 or 5. For the sake of simplicity, let’s assume these reflect honest responses.

Our survey results reveal a pleasantly low 5% annual turnover rate for those answering 4 or 5. In contrast, those reporting a 3 or lower on the job satification measure leave at an annual rate of 10%.

Here’s where things get interesting.

If we look only at the probability of leaving voluntary GIVEN our job satisfaction measure, the meaning is clear: those reporting 3 or less are twice as likely to leave (10% v. 5%).

That’s useful information and a whoppingly large difference.

But wait! Recall our basic distributions for job satisfaction. Only 20% of our 500 people provided a score of 3 or lower.

That’s 100 people. With a 10% voluntary turnover rate, we will lose 10.

What about our very satisfied 4s and 5s? That 80% of our 500 works out to 400 employees. With a voluntary turnover rate of 5%, we will be losing 20 of them this year.

That is twice the raw number of those leaving from the low satisfaction group. This difference arises even though people in the low satisfaction group are actually twice as likely to leave their jobs voluntarily.

Probabilities, Proportions, and Group Size

Now we can see the issue. On the one hand, those with low job satisfaction have a higher probability of leaving.

One the other hand, low satisfaction employees represent a smaller overall proportion of Fantastic Inc.

Ironically then, most of those voluntarily leaving Fantastic actually turn out to have high job satisfaction. Group size matters.

To summarize our scenario,

1. Those with lower job satisfaction are twice as likely to leave voluntarily (10% v. 5%) BUT
2. Those who left voluntarily were twice likely to have high satisfaction (20 v. 10) because the high satisfaction group is much larger.

What Have We Learned?

When considering leading indicators for turnover, we need to look at two things:

1. The PROBABILITY of turnover within each group we are comparing
2. The SIZE of those groups

As an extreme example, even if I were 100% certain that people with low job satisfication will leave within the next year, that information is of limited use if only 2 of my 200 employees fall into that category.

Bottom line? When it comes to risk factors for voluntarily turnover, pay attention to both turnover rates within the groups AND to the sizes of the groups.

Beyond Intuition: The Relationship Between Risk Factors and Population

Time for data. Let’s say we have two job satisfaction questions (1-5 scale) and a single measure of turnover. Note that I am just using job satsifaction as an example, but the idea applies to any leading indicator for retention risk (or really any other measure for that matter)

In the R code below, I create some data using the “sample”” function.

Note that the list of probabilities in the code corresponds to the list of the sample answer values (1-5), where I have 45% chance of answering 5, 35% of answering 4, etc.

For the sampled voluntary turnover vector, I set different probabilities of leaving according to the sum of the individual job satisfaction measures.

set.seed(996)
jobsat1 <- sample(c(5,4,3,2,1), size = 1000, replace = T, prob = c(.45, .35, .1, .07, .03)) #creating question 1 data

jobsat2 <- sample(c(5,4,3,2,1), size = 1000, replace = T, prob = c(.45, .35, .1, .07, .03)) # creating question 2 data

jobsat_sum <- jobsat1 + jobsat2 # creating an overall measure

vol <- integer(0) # setting up an empty voluntarily leave flag to fill up

# cycling through the job sat sum measure and selecting stay or leave based on that sum for each individual.

for (i in seq_along(jobsat_sum)){ # go through the whole list
    
    vol[i] <- ifelse(jobsat_sum[i] >=7, sample(c(0,1), 1, prob = c(.95, .05)), 
                                          sample(c(0,1), 1, prob = c(.90, .10)))
}

par(mfrow = c(1,3)) # multiple histograms in 1 view
hist(jobsat1)
hist(jobsat2)
hist(jobsat_sum)

cols <- ifelse(vol == 1, "red", "black") #voluntary terminations in red

sizes <- ifelse(vol ==1, 1, .6) # larger dots for voluntary turnover

plot(jitter(jobsat1), jitter(jobsat2), col = cols, pch = 19, cex = sizes)

# making a single large data frame with all of our data first
vol_df <- data.frame(js1 = jobsat1, js2 = jobsat2, jssum = jobsat_sum, 
           pt_clr = cols, pt_size = sizes)

#filtering down to just those who left voluntarily
vol_df <- subset(vol_df, vol ==1)

# Plotting the data
plot(jitter(vol_df$js1), jitter(vol_df$js2), col = vol_df$pt_clr,
     pch = 19, cex = vol_df$pt_size, 
     xlab = "Job Satisfaction Measure 1", 
     ylab = "Job Satisfaction Measure 2", 
     main = "Voluntary Turnover From Highly Satisfied Employees")

Final Thoughts

Leading indicators are extremely valuable, whether we are talking about HR issues or broader economic factors like labor market participation or existing home sales.

But when looking at group differences for these leading indicators, be sure to take account of both the probabilities and group size.

Visualizations that capture group size (like the scatterplots here) can help you spot meaningful differences that you might miss if you focus only on probabilities and rates.

Looking at both will help you determine where you need to focus your intervention and measurement efforts.

If we ignore group sizes, we might be surprised to find that most of our problems arise from the least likely places.