People Analytics

My first exposure to a People Analyst was seeing Bill Yost’s LinkedIn activity. As far as I’ve seen, he doesn’t post a lot specifically about people analytics, and mostly just posts really funny content, including satire. Anyway, I knew that data analytics was relevant to pretty much any area of a company, but seeing HR analytics being called “People Analytics” was pretty cool.

In this project, I took on the role of “People Data Analyst” for a fictional HR department. My tool was R, which has become the third programming language I have used (after Python and SQL, in that order). I’ve learned that it’s a really powerful and awesome tool to have in one’s belt. My job was to explore issues including accusations of ageism in layoffs. In my analysis, I found that:

• Layoffs were actually more likely for younger employees
• There was no statistically significant difference between firing rates for more recent or more tenured employees
• Age and years worked account for approximately 60% of variance in monthly income

As usual, read below to retrace my steps!

The Dataset

The dataset was created by data scientists at IBM, and can be found here.

Analysis

Correlation

The first step was simply to import the data:

hrdata <- read.csv("HR-Employee-Attrition.csv")

After doing this, the first thing I did was check if any of the demographic characteristics had any correlations. I used the “cor” method on the most relevant columns:

cor(hrdata[, c("Age", "DailyRate", "DistanceFromHome", "Education", "HourlyRate", "MonthlyIncome", "MonthlyRate", "NumCompaniesWorked", "TotalWorkingYears", "TrainingTimesLastYear")])

Shown above is a subset of the correlation matrix. Among the strongest correlations are those between income, age, and years worked/trained, which makes sense - the more experience someone has, the more money they should make. For example, monthly income and total working years have a correlation value of .773; age and monthly income are almost at .500.

It made sense to visualize these relationships with scatter plots, and just like in Python, we have a simple built-in command to get some pair plots:

pairs(~MonthlyIncome + Age + TotalWorkingYears + Education, data = hrdata, main = "Scatterplot Matrix")

As we expect, income/working years and income/age have rather strong correlations. However, I noticed something interesting - the age and working years variables have a very smooth upper/lower range so that the plot looks like it’s inside of a right triangle. It’s like there’s an upper or lower bound defined by a linear equation. In this case, for a given age, there is a maximum possible number of working years. Upon further inspection, I realized this makes sense, since generally speaking, one cannot begin employment until they are a certain age. So, the line defining the maximum possible years worked represents is in fact a hard and fast bound.

Boxplots/Hypothesis Testing

It was then time to answer the accusations of ageism. An employee has sued the company, saying that layoffs were age-based. I checked this by observing the age distributions for those who were let go and those who remained. I did this with a boxplot:

boxplot(Age~Attrition, data = hrdata, main = "Who Got Fired", xlab = "Attrition", ylab = "Age")

The box on the right being slightly lower is promising - the average age for those fired could be lower than for those who stayed on. To be sure about the result, I decided to employ a hypothesis test. Specifically, I used a Welch Two Sample t-test. I created a variable for the ages of those fired and another for the ages of those not fired:

yes_age <- hrdata[(hrdata$Attrition == "Yes"),'Age']
no_age <- hrdata[(hrdata$Attrition != "Yes"),'Age']
t.test(yes_age, no_age)

The p-value is well under .05, indicating that there is a statistically significant difference between our variables. However, we see that the mean value for our first array is lower than that of the second, meaning that those who left were actually younger than those who stayed. This is further confirmed by the negative values under our confidence interval.

There was another complaint, however, that more recent employees were let go at a higher rate than more tenured employees. So, I made boxplots and a t-test, this time based on Employee Number (lower numbers mean more tenured employees):

boxplot(Age~Attrition, data = hrdata, main = "Who Got Fired", xlab = "Attrition", ylab = "EmployeeNumber")

yes_id <- hrdata[(hrdata$Attrition == "Yes"), 'EmployeeNumber']
no_id <- hrdata[(hrdata$Attrition != "Yes"), 'EmployeeNumber']
t.test(yes_id, no_id)

With a very high p-value, we see that there was not statistically significant difference in layoffs by employee number. This should put to bed any suspicion that newer employees were targeted in the layoffs.

Linear Regression

My last task was to predict monthly income based on the age of an employee. To do this, I employed basic linear regression model using the convenient lm function in R:

model1 = lm(MonthlyIncome ~ Age, data = hrdata)
summary(model1)

We have a small p-value, so our R-squared value of about 25% means that about 25% of the variance in monthly income can be explained with our model.

I made one more regression model, this time using another variable that is likely a strong predictor of monthly income - total years worked:

model2 = lm(MonthlyIncome ~ Age + TotalWorkingYears, data = hrdata)
summary(model2)

This model is even better - we still have a small p-value, and this time our R-squared value is a strong .599! We can conclude that while age is a somewhat loose predictor of monthly income, combining age with experience is a much better predictor.

Takeaways

The main concerns of our employer here are averted - in terms of ageism, there was actually evidence for the opposite case - younger employees were more likely to be fired than older ones. In terms of employee tenure, we see that there was no significant difference in firing rates. While we got results that are significant for our boss, we also used some linear models that may be more useful for prospective employees - a prospective employee could use the models to predict their income after they spend a certain number of years at the company. Or, the HR division can use it to provide fair salary adjustments and negotiation for new employees. All things considered, the insight gained from this analysis is incredibly useful.

Avery Smith’s description of Python as a Swiss army knife and R as a hatchet is very accurate, as seen in this project. R makes statistical analysis very simple and straightforward - the lines of code here are much shorter and concise, while Python makes things a bit clunkier.

Let’s Connect!

Thanks for reading my project! Please connect on LinkedIn and send any feedback, or just say hi!