Human Resources Analytics

How do we retain our best and most experienced employees?


Our dataset includes 14,999 observations, with each row representing one single employee.


Fields in the dataset include the following 10 variables for each line:
- Employee satisfaction level
- Last evaluation score
- Number of projects
- Average monthly hours
- Time spent at the company
- Whether they have had a work accident
- Whether they have had a promotion in the last 5 years
- Department
- Salary
- Whether the employee has left

Project Objectives


1) Assess what are the relationship between the 10 variables and what are the significant variables to describe the dataset
2) Undestand who are the employees that have left
3) Focus the analysis on the most valuable employees who have left
4) Devolop a predictive model to assess the likelihood of an employee leaving

The report is divided as follows:


Step 1) Data quality check
Step 2) Basic Data Visualisation
Step 3) Principal Component Analysis
Step 4) Futher comparative analysis on employees that left
Step 5) Prediction Model
Step 6) Conclusion

Step 1: Data quality check


First, we will perform basic statistical analysis and understand the type of factors.

Looking at data we can make the following assumptions related the nature of the 10 variables:


satisfaction_level: A numeric indicator filled out by the employee ranging from 0 to 1
last_evaluation: A numeric indicator filled in by the employee’s manager ranging from 0 to 1
number_project: An integer that indicates the number of projects the employee has worked on
average_monthly_hours: The number of hours employees work in the month
time_spend_company: An integer value indicated the years of service
Work_accident: A dummy variable assessing whether(1) or not (0) they had an accident
left: A dummy variable, leave (1), not leave(0)
promoted_last_5years: A dummy variable, promoted(1), not promoted(0)
sales: A categorical variable assessing the department in which employee is working (sales,technical,support,IT, product,marketing, other)
salary: A 3-level categorical variable (low, medium, high)

Data quality report

First of all we assess that there are no missing data with function is.na(myData) that we would not report in the code and and we perform basic summary statistic of the dataset.

##  satisfaction_level last_evaluation  number_project  average_montly_hours
##  Min.   :0.0900     Min.   :0.3600   Min.   :2.000   Min.   : 96.0       
##  1st Qu.:0.4400     1st Qu.:0.5600   1st Qu.:3.000   1st Qu.:156.0       
##  Median :0.6400     Median :0.7200   Median :4.000   Median :200.0       
##  Mean   :0.6128     Mean   :0.7161   Mean   :3.803   Mean   :201.1       
##  3rd Qu.:0.8200     3rd Qu.:0.8700   3rd Qu.:5.000   3rd Qu.:245.0       
##  Max.   :1.0000     Max.   :1.0000   Max.   :7.000   Max.   :310.0       
##                                                                          
##  time_spend_company Work_accident         left       
##  Min.   : 2.000     Min.   :0.0000   Min.   :0.0000  
##  1st Qu.: 3.000     1st Qu.:0.0000   1st Qu.:0.0000  
##  Median : 3.000     Median :0.0000   Median :0.0000  
##  Mean   : 3.498     Mean   :0.1446   Mean   :0.2381  
##  3rd Qu.: 4.000     3rd Qu.:0.0000   3rd Qu.:0.0000  
##  Max.   :10.000     Max.   :1.0000   Max.   :1.0000  
##                                                      
##  promotion_last_5years         sales         salary    
##  Min.   :0.00000       sales      :4140   high  :1237  
##  1st Qu.:0.00000       technical  :2720   low   :7316  
##  Median :0.00000       support    :2229   medium:6446  
##  Mean   :0.02127       IT         :1227                
##  3rd Qu.:0.00000       product_mng: 902                
##  Max.   :1.00000       marketing  : 858                
##                        (Other)    :2923



From basic statistical analysis we can see that the overall satisfaction of the company is at a low-medium level of 63% and that approximately 24% of the employees have left.


This brings us to the following step: we would like to visualize better who are the employees that have decided to leave.

Step 2: Data Visualization


We will start our analysis looking more deeply at a subset composed of only the employees that have left. In particular we will analyse the distribution of employees across variables:

We cut database in the desired subset composed of 3,571 observations:

## [1] 3571


We plot first of all the most intuitive variables which could provide initial insights into why people leave - Satisfaction Level, Last Evaluation and Average monthly hours.





From the previous histograms we can make the following preliminary observations:


None of the distributions seem normal but we see peaks at the ends of the histograms
Regarding satisfaction level, the distribution of employers that are leaving is quite polarized; employees who left are mostly low (<0.5)or high on satiafaction level (> 0.7).
Regarding employees evaluation, those that leave seems either really good (>.9) or average.
Employees that leave seem to either work a lot( >250 hours) or below average (<150 hours)


We then look at the distribution in the categorical variables:


From a first analysis of these last three histograms we can make the following observations:


- The frequency of work accident per se doesn’t not mean a lot.
- Employees that left seems to have generally low salary.
- Employees that left comes mainly from sales, support and technical departments.

From the preliminary analysis, we would like to focus our analys on employees that we consider most valuable but that are leaving. We decide to set this criteria looking at the median value and choosing those that have worked for the company for more than 3 years, have good last evaluation results >0.72, and have performed more than 4 projects.

This group is composed by 3556 people (23,7% of employees)

## [1] 3556







As of part of the preliminary analysis we then perform an initial correlation analysis for numerical variables for the following group of employees: 1) all the employees 2) the valuable employees identified before 3) the employees that left














As a conclusion we can see how the only singnigicant correlation that we can find are across the subgroup of the employees that have left.

Step 3: Principal Component Analysis


We then perform a more accurate correlation by first taking the entire database, then excluding what we can consider the dependent variable (left) and finally scaling the data.

## [1] "satisfaction_level"    "last_evaluation"       "number_project"       
## [4] "average_montly_hours"  "time_spend_company"    "Work_accident"        
## [7] "promotion_last_5years" "sales"                 "salary"
## [1] "satisfaction_level"    "last_evaluation"       "number_project"       
## [4] "average_montly_hours"  "time_spend_company"    "Work_accident"        
## [7] "promotion_last_5years" "sales"                 "salary"


After having done the first adjustments, we can confirm that data are metric. However, we now need to scale the data to have a more homogeneus dataset.

min 25 percent median mean 75 percent max std
satisfaction_level 0.09 0.44 0.64 0.61 0.82 1 0.25
last_evaluation 0.36 0.56 0.72 0.72 0.87 1 0.17
number_project 2.00 3.00 4.00 3.80 5.00 7 1.23
average_montly_hours 96.00 156.00 200.00 201.05 245.00 310 49.94
time_spend_company 2.00 3.00 3.00 3.50 4.00 10 1.46
Work_accident 0.00 0.00 0.00 0.14 0.00 1 0.35
promotion_last_5years 0.00 0.00 0.00 0.02 0.00 1 0.14
sales 1.00 5.00 8.00 6.94 9.00 10 2.75
salary 1.00 1.00 2.00 1.59 2.00 3 0.64


Here reported the summary statistics of the scaled dataset:

min 25 percent median mean 75 percent max std
satisfaction_level -2.10 -0.70 0.11 0 0.83 1.56 1
last_evaluation -2.08 -0.91 0.02 0 0.90 1.66 1
number_project -1.46 -0.65 0.16 0 0.97 2.59 1
average_montly_hours -2.10 -0.90 -0.02 0 0.88 2.18 1
time_spend_company -1.03 -0.34 -0.34 0 0.34 4.45 1
Work_accident -0.41 -0.41 -0.41 0 -0.41 2.43 1
promotion_last_5years -0.15 -0.15 -0.15 0 -0.15 6.78 1
sales -2.16 -0.70 0.39 0 0.75 1.12 1
salary -0.93 -0.93 0.64 0 0.64 2.21 1


We can now once again perform correlation on those data and as we can see there are few variables showing strong pariwise corrrelation (+0.3)

thecor = round(cor(ProjectDatafactor_scaled),2)
colnames(thecor)<-colnames(ProjectDatafactor_scaled)
rownames(thecor)<-colnames(ProjectDatafactor_scaled)

knitr::kable(round(thecor,2))
satisfaction_level last_evaluation number_project average_montly_hours time_spend_company Work_accident promotion_last_5years sales salary
satisfaction_level 1.00 0.11 -0.14 -0.02 -0.10 0.06 0.03 0.01 0.05
last_evaluation 0.11 1.00 0.35 0.34 0.13 -0.01 -0.01 0.01 -0.01
number_project -0.14 0.35 1.00 0.42 0.20 0.00 -0.01 0.02 0.00
average_montly_hours -0.02 0.34 0.42 1.00 0.13 -0.01 0.00 0.01 0.00
time_spend_company -0.10 0.13 0.20 0.13 1.00 0.00 0.07 -0.03 0.05
Work_accident 0.06 -0.01 0.00 -0.01 0.00 1.00 0.04 0.01 0.01
promotion_last_5years 0.03 -0.01 -0.01 0.00 0.07 0.04 1.00 -0.04 0.10
sales 0.01 0.01 0.02 0.01 -0.03 0.01 -0.04 1.00 -0.06
salary 0.05 -0.01 0.00 0.00 0.05 0.01 0.10 -0.06 1.00


Despite that, we would proceed with PCA ans after trying differerent combination of PCA ( manual, varimax , eigenvalue) we assess that best model is given by “eingenvalue”.

Eigenvalue Pct of explained variance Cumulative pct of explained variance
Component 1 1.83 20.33 20.33
Component 2 1.18 13.06 33.39
Component 3 1.13 12.56 45.95
Component 4 1.00 11.12 57.07
Component 5 0.95 10.56 67.63
Component 6 0.89 9.92 77.54
Component 7 0.84 9.35 86.90
Component 8 0.63 7.02 93.92
Component 9 0.55 6.08 100.00
<img src=“Anal ysisofHR[1]_f iles/figure-html/figure01-1. png" width=“672” />
Component 1 Component 2 Component 3 Component 4
last_evaluation 0.76 0.01 0.21 -0.01
average_montly_hours 0.76 -0.02 -0.05 -0.01
number_project 0.74 -0.03 -0.30 0.02
time_spend_company 0.30 0.31 -0.50 0.07
satisfaction_level 0.08 0.16 0.86 0.10
sales 0.03 -0.45 0.01 0.51
salary 0.00 0.66 0.10 -0.12
Work_accident -0.02 0.15 0.05 0.81
promotion_last_5years -0.04 0.63 -0.10 0.26

Let us nnow focus o		 n the key vari ables in each Component:
Component 1 Component 2 Component 3 Component 4
last_evaluation 0.76 NA NA NA
average_montly_hours 0.76 NA NA NA
number_project 0.74 NA NA NA
time_spend_company NA NA -0.50 NA
satisfaction_level NA NA 0.86 NA
sales NA NA NA 0.51
salary NA 0.66 NA NA
Work_accident NA NA NA 0.81
promotion_last_5years NA 0.63 NA NA


Based on eingvalue we can see that the database can be reduced to three principal components that we try to interpret:


Component 1: Number of projects + monthly hours + last evaluation - we will call this component “commitment”
Component 2: Salary+promotion - we can call this component “renumeration”"
Component 3: Job satisfaction (independent)

From this, we can tell that Principal Component Analisys allows us to easily have a reduced number of factors to describe the dataset, however due to the limited number of dependent variables (10), we can conclude that this analysis is not particularly useful in this case.

Step 4: Futher comparative analysis on employees that left

Let’s dig deeper to analyse employees who left.

vars n mean sd median trimmed mad min max range skew kurtosis se
1 11428 0.7 0.2 0.7 0.7 0.2 0.1 1 0.9 -0.6 -0.2 0.0
2 11428 0.7 0.2 0.7 0.7 0.2 0.4 1 0.6 0.0 -1.0 0.0
3 11428 3.8 1.0 4.0 3.8 1.5 2.0 6 4.0 0.3 -0.4 0.0
4 11428 199.1 45.7 198.0 199.5 56.3 96.0 287 191.0 -0.1 -1.0 0.4
5 11428 3.4 1.6 3.0 3.1 1.5 2.0 10 8.0 2.1 5.2 0.0
6 11428 0.2 0.4 0.0 0.1 0.0 0.0 1 1.0 1.7 0.9 0.0
7 11428 0.0 0.0 0.0 0.0 0.0 0.0 0 0.0 NaN NaN 0.0
8 11428 0.0 0.2 0.0 0.0 0.0 0.0 1 1.0 5.9 33.1 0.0
9 11428 6.9 2.7 8.0 7.2 3.0 1.0 10 9.0 -0.8 -0.6 0.0
10 11428 1.7 0.7 2.0 1.6 1.5 1.0 3 2.0 0.5 -0.7 0.0
vars n mean sd median trimmed mad min max range skew kurtosis se
1 3571 0.4 0.3 0.4 0.4 0.4 0.1 0.9 0.8 0.3 -1.0 0
2 3571 0.7 0.2 0.8 0.7 0.3 0.4 1.0 0.6 0.0 -1.7 0
3 3571 3.9 1.8 4.0 3.7 3.0 2.0 7.0 5.0 0.2 -1.5 0
4 3571 207.4 61.2 224.0 205.7 97.9 126.0 310.0 184.0 0.1 -1.6 1
5 3571 3.9 1.0 4.0 3.8 1.5 2.0 6.0 4.0 0.5 -0.8 0
6 3571 0.0 0.2 0.0 0.0 0.0 0.0 1.0 1.0 4.3 16.2 0
7 3571 1.0 0.0 1.0 1.0 0.0 1.0 1.0 0.0 NaN NaN 0
8 3571 0.0 0.1 0.0 0.0 0.0 0.0 1.0 1.0 13.6 182.8 0
9 3571 7.0 2.8 8.0 7.3 3.0 1.0 10.0 9.0 -0.8 -0.6 0
10 3571 1.4 0.5 1.0 1.4 0.0 1.0 3.0 2.0 0.8 -0.5 0

Here we can have a overview of the data; for example, there are 3571 observations in 1 group, values of mean and median of satisfaction in 1 group is smaller than 0 group’s. The correlation plot show some relationships of first 8 variables. Obviously, with lower satisfaction level, employees will have stronger desire to leave. If someone can deal with more projects and spend more time on work, he or she will obtain higher evaluation.

Predicting Who Will Leave?

we may find the evidence who will leave next.


Let’s directly compare employees who left and those who did not against the 3 ‘intuitive’ variables as earlier mentioned.

## `stat_bin()` using `bins = 30`. Pick better value with `binwidth`.
## `stat_bin()` using `bins = 30`. Pick better value with `binwidth`.
## `stat_bin()` using `bins = 30`. Pick better value with `binwidth`.

These charts illustrate the following:
1) Hours: Employees who work less than 150 hours and more than 250 hours have a higher likelihood of leaving. Perhaps, we can infer that The former group may be underworked and left for more challenging work. On the contrary, the latter group might feel too overworked and hence decide to leave. Indeed, beyond 300 hours, there are no employees who stay on. From the earlier correlation plot, we also saw strong correlation between ‘hours’ and ‘number of projects’ hence we can infer that any more than 4 projects might lead employees to become overworked.


2) Evaluation: Employees who left get lower or higher last evaluation. Sense of satisfaction seems to be another factor; Satisfation level of 71.32 % employees who left is less than 0.5.


2) Satisfaction: Employees who are low on satisfaction are more likely to leave, as we had earlier suspected. We shall take a closer look at Satisfaction level between those who left and those who did not.

## 
##         0         1 
## 0.4707379 0.5292621


The scatter box plot above shows us veyr clearly that there is signigfcant concentration for employees with lowe satisfaction levels. However, we can still see some employees with high sense of satisfaction also left. We will need to further investigate this.


Let’s now dig into the categorical variables of work accidents and salary.


Interestingly, employees with work accidents do not display higher likelihood to leave. Low salary also appeared to contribute to resigning. We did not consider the variable promotion, because the value’s distribution was very uneven; only 319 out of 14999 employees were promoted in the last five years. Hence, we eliminated analysing the variable promotion in this process.

Who Are Valuable Employees And Why Do They Leave?


In our intermediary view, we believe that employees with evaluation >=0.72, more than or equal to 3 years of experience in the company and more than 4 projects can be considered to be experienced and valuable ones for a company. This comprises more than half of employees who left.

Step 5: Prediction Model


It’s time for us to build a model which we can use to preduct the likelihood of a valuable employee leaving. First, we will split the dataset into two parts, train and test. A classification tree will first be built followed by a forest.

## CART 
## 
## 10000 samples
##     8 predictor
##     2 classes: '0', '1' 
## 
## No pre-processing
## Resampling: Cross-Validated (5 fold, repeated 3 times) 
## Summary of sample sizes: 8001, 8000, 7999, 8000, 8000, 7999, ... 
## Resampling results across tuning parameters:
## 
##   cp    Accuracy   Kappa    
##   0.01  0.9689336  0.9131738
##   0.02  0.9629003  0.8973404
##   0.03  0.9629003  0.8973404
## 
## Accuracy was used to select the optimal model using  the largest value.
## The final value used for the model was cp = 0.01.

## Confusion Matrix and Statistics
## 
##           Reference
## Prediction    0    1
##          0 3802   49
##          1   17 1131
##                                           
##                Accuracy : 0.9868          
##                  95% CI : (0.9832, 0.9898)
##     No Information Rate : 0.764           
##     P-Value [Acc > NIR] : < 2.2e-16       
##                                           
##                   Kappa : 0.963           
##  Mcnemar's Test P-Value : 0.0001357       
##                                           
##             Sensitivity : 0.9955          
##             Specificity : 0.9585          
##          Pos Pred Value : 0.9873          
##          Neg Pred Value : 0.9852          
##              Prevalence : 0.7640          
##          Detection Rate : 0.7606          
##    Detection Prevalence : 0.7704          
##       Balanced Accuracy : 0.9770          
##                                           
##        'Positive' Class : 0               
##


Let’s see how we can interpret this tree using a sample employee. Take John, a manager who has the following characteristics:

Satisfaction: 0.55 Years of experience: 4 Evaluation: 0.85 Hours worked: 220 Projects: 3

From the tree above, we can see that he will _____.

So, rf_mdl works better. Another way to look at which variables are important can be seen though an importance chart.

A key takeaway here is that Satisfaction level is a much more important indicator than the others and should be closely monitered to effectively manage talent retention.

Step 6: Conclusion

Through our analysis, managing the level of satisfaction is the key to keep employees with the firm. This is especially important for employees who have been around for more than 3 years. Other than that the employee evaluation and number of projects should also be monitored. This firm’s HR Head would do well to craft programs to keep tabs and these metrics so as to have a successful talent retention policy.