Why re-encode ‘Reasons for absence’ ?

There was a mild complication with ‘Reason for absence’ since 0s are present in the data, however Martiniano et al. (2012) do not list ‘0’ as a possible code. Additionally, ‘20’ doesn’t exist in the data, but it exists as a code. Based on other papers on this data, it seems that the codes for 1 through 20 should be applied to 0 through 19 (21 through 28 are fine).

Distribution of Reasons

Data Post-Feature Engineering

Reasons for Absenteeism

Top 4 most common reasons for absence (by frequency):

• Medical consultations

• Dental consultations

• Physiotherapy (PT)

• Genitourinary system diseases

Top 5 most time consuming reasons (by hour):

• Genitourinary system diseases

• External causes of morbidity & mortality

• Medical consultations

• Dental consultations

• Integumentary system (skin & subcutaneous tissue) diseases

Age & Absenteeism

• Age is even distributed from the mid-20s to the early 40s.

• The majority of employees who take leaves of absence do so for 8 hours or less. There is a distinct gap between half-day absences (under 3 to 4 hours) and full-day absences (8 hours). Intervals for absenteeism time then increase by 8 hours at a time.

Original Data set

• Reasons for absence

• Disciplinary failure

• Children (‘Son’)

• Education

• Age

ATW2

• Reasons for absence:

  • Respiratory system diseases

  • External causes of morbidity & mortality

  • Genitourinary system diseases

  • Musculosekeltal system diseases

  • Eye diseases

  • Integumentary system diseases

  • Digestive system diseases

• Children

• Day of the week

ATW2 without ‘Reasons’

• Discipline failure

• Day of the week

• Age

• Season

• Children

‘Absent Type’

• Created a new categorical variable for length of absence (‘One day’, ‘Multi-day’) based on clustering results & EDA (see scatter plot of ‘Age’ & ‘Absenteeism’)

• When cluster assignment was included:

  • \(Accuracy = .905\)

  • \(Specificty = .971\)

  • \(AUC =.808\)

• While accuracy & specificity are high, sensitivity was low. The boosted logit model is good at predicting the positive class (‘Multi-day’), but too cautious and under-predicts the negative class (‘One day’ absences)

Results & Recommendations

• 2 clusters seems appropriate for k-Means clustering
• Classification models were much more successful than regression models. Future models of absenteeism could focus on classifying time periods (rather than treating time like a continuous variable), however further tuning of metrics is required.
• The combination of EDA, data knowledge, and clustering could be an avenue for binning absence length into categories
• Some of reasons for absence may not be controllable, but this company can address the most prevalent reasons for absence by promoting wellness practices & by tailoring benefits to prevent disorders, such as those relating to the genitourinary system (e.g., PT, regular movement in/around the work environment, etc.)

  • In terms of frequency, about 50% of absences are due to medical & dental consultations, PT, and urogenital diseases.

  • Urogenital disease, external causes to mortality, and medical & dental consultation also make up 50% of total hours absent.