EXECUTIVE SUMMARY: DATA QUALITY & EXPLORATORY ANALYSIS

Purpose

This section identifies multicollinearity in the dataset by detecting strong linear relationships (|r| > 0.7) between numeric variables. Understanding these dependencies is critical for predictive modeling, as highly correlated features can inflate coefficient estimates, reduce model interpretability, and create instability in parameter estimation.

Key Findings

• High Correlations Detected: 7 pairs of variables exceed the 0.7 threshold, indicating substantial linear dependence
• Correlation Method: Pearson correlation captures linear relationships; notable pairs include YearsAtCompany–YearsWithCurrManager (r=0.77) and YearsInCurrentRole–YearsWithCurrManager (r=0.74)
• Statistical Significance: All 7 high correlations are marked as significant, confirming these relationships are unlikely due to chance
• Overall Correlation Structure: Mean correlation of 0.12 across all 400 variable pairs suggests most features are relatively independent, with multicollinearity concentrated in specific clusters

Interpretation

The dataset exhibits localized multicollinearity rather than systemic redundancy. The strong correlations cluster around tenure-related variables (Years at Company, Years in Current Role, Years with Current Manager), which logically measure related but distinct temporal dimensions of employment history. This pattern suggests these variables capture overlapping information

Purpose

This section identifies and visualizes the distribution of numeric variables across the dataset, with specific focus on detecting anomalous values. The 348 outliers flagged using the IQR method (1.5 × IQR threshold) represent extreme observations that deviate significantly from typical patterns. Understanding these distributions is critical for assessing data quality and determining whether outliers represent genuine business phenomena or measurement errors.

Key Findings

• Outliers Detected: 348 observations (3.48% of 10,000 distribution records) flagged as anomalies across 20 numeric variables
• Detection Method: IQR-based approach identifies values beyond 1.5 × interquartile range, a standard statistical threshold
• Distribution Characteristics: Values range from 0 to 19,999 with high right skew (0.56), indicating right-tailed distributions with extreme high values
• Variable Coverage: Outliers distributed across all 20 numeric variables, with consistent 500-observation samples per variable

Interpretation

The moderate outlier rate (3.48%) suggests the dataset contains legitimate extreme values rather than systematic data quality failures. The high skewness and wide value range (0–19,999) indicate several variables have naturally occurring outliers reflecting real business variation. These outliers warrant investigation to determine whether they represent valid observations (

Purpose

This section assesses data completeness by identifying missing values across all 27 variables in the dataset. Complete data is essential for reliable statistical analysis and modeling, as missing values can introduce bias, reduce statistical power, and complicate feature engineering. Understanding missingness patterns helps determine whether imputation, removal, or alternative analytical approaches are needed.

Key Findings

• Overall Missing Percentage: 0% – The dataset contains no missing values across any variable
• Rows with Missing Data: 0 – All 500 observations are complete with values for every variable
• Variable Coverage: All 27 variables (20 numeric, 7 categorical) have 100% data completeness
• No Imputation Required: Zero variables exceed the 10% missing threshold that would typically trigger imputation or removal decisions

Interpretation

The dataset exhibits perfect data completeness, eliminating a major source of analytical uncertainty. This clean state means all 500 observations can be used without data loss from listwise deletion, and no imputation assumptions are needed. The absence of missing data strengthens the validity of statistical tests (140 conducted, 23 significant) and feature importance rankings, as these analyses operate on the full sample without bias from incomplete cases.

Context

This ideal completeness status simplifies downstream modeling but does not address other data quality concerns identified elsewhere: 348 outliers (69.

Purpose

This section examines how Age varies across the three Department categories using boxplot analysis. By comparing group medians, spreads, and ranges, we identify whether department membership has a meaningful relationship with employee age—a key indicator of whether this categorical variable should be prioritized in predictive modeling or segmentation analysis.

Key Findings

• Human Resources Mean Age: 39.79 years (n=14) - Oldest department on average, though smallest sample size
• Research & Development Mean Age: 37.24 years (n=333) - Largest group with moderate age, median of 36
• Sales Mean Age: 35.88 years (n=153) - Youngest department, median of 34
• Spread Consistency: Standard deviations range 9.16–12.81, indicating similar variability across groups
• Age Range: All departments span 18–60 years, with overlapping distributions suggesting modest departmental differences

Interpretation

The three departments show a slight age gradient (HR > R&D > Sales), with Human Resources employees averaging ~4 years older than Sales. However, the substantial overlap in distributions and comparable standard deviations indicate that department explains only modest variation in age. This aligns with the Kruskal-Wallis test results showing no significant differences for numeric

Purpose

This section evaluates whether numeric variables differ significantly across categorical groups using statistical hypothesis testing. Of 140 tests performed, 23 revealed statistically significant differences (p < 0.05), indicating that certain numeric variables behave differently depending on group membership. This identifies which variables have meaningful predictive or explanatory power relative to categorical factors.

Key Findings

• Total Tests Conducted: 140 across 20 numeric variables and 7 categorical variables
• Significant Results: 23 tests (16.4%) showed p-values below the 0.05 threshold, indicating genuine group differences
• Test Method: Kruskal-Wallis tests used exclusively, appropriate given that 0 of 20 numeric variables follow normal distributions
• P-Value Distribution: Mean p-value of 0.39 with median of 0.32 suggests most variable-group pairs show no significant association
• Strongest Signals: Age and YearsWithCurrManager show multiple significant associations with categorical variables like JobRole

Interpretation

The 16.4% significance rate indicates selective rather than pervasive group differences. Most numeric-categorical pairs (83.6%) show no meaningful variation across groups, suggesting limited discriminatory power for those combinations. However, the 23 significant findings identify specific numeric variables that meaningfully stratify by categorical factors—

Purpose

This section evaluates whether the 20 numeric variables follow normal distributions using the Shapiro-Wilk test. Normality is a critical assumption for parametric statistical methods (t-tests, ANOVA, linear regression). Understanding the distributional properties of your data determines which analytical techniques are appropriate and whether data transformations are needed.

Key Findings

• Variables Tested: 20 numeric variables analyzed
• Normal Distributions Found: 0 (0% of variables pass normality test at p > 0.05)
• Test Used: Shapiro-Wilk with all p-values = 0, indicating strong evidence against normality across all variables
• Pattern Observed: All variables show significant deviation from normality, with Q-Q plots revealing systematic departures from the theoretical normal line

Interpretation

The complete absence of normally distributed variables indicates your dataset exhibits non-normal characteristics across all numeric features. This finding is consistent with the earlier observation of 8 skewed variables and 348 outliers detected via IQR method. The data's departure from normality—evidenced by positive skewness in variables like MonthlyIncome (skew=1.32) and YearsSinceLastPromotion (skew=1.94)—means parametric assumptions are violated, potentially affecting the validity of standard statistical tests already

Purpose

This section visualizes pairwise relationships between numeric variables through scatterplots, revealing both linear and non-linear associations that correlation coefficients alone cannot capture. By examining 6 variable pairs across 500 observations, the analysis identifies which features move together and detects patterns that may inform predictive modeling and feature engineering decisions.

Key Findings

• Variable Pairs Analyzed: 6 combinations examined across 3,000 plotted points
• Age vs. MonthlyIncome: Shows moderate positive relationship (r=0.49), with income ranging 1,102–19,999 across age span 18–60
• High Variability in Income: Standard deviation of 4,814.58 indicates substantial income dispersion independent of age
• Non-linear Patterns: Scattered distributions suggest relationships may not be purely linear; curved or clustered patterns could indicate threshold effects or categorical influences

Interpretation

The pairwise analysis reveals that while some numeric variables exhibit positive correlations (particularly Age with MonthlyIncome), the scatter and high variance indicate these relationships are moderate at best. The wide range of values and skewed distributions (skewness 1.57 for x-values, 1.05 for y-values) suggest outliers and non-uniform patterns that simple linear models may not fully capture. This aligns with earlier findings

Purpose

This section ranks 19 features by their linear correlation strength with the target variable (Age), identifying which predictors have the strongest associations with outcomes. The absolute correlation values quantify predictive power—longer bars indicate stronger relationships. This ranking helps prioritize variables for modeling and reveals which factors most consistently relate to the target in linear frameworks.

Key Findings

• Top Predictor (TotalWorkingYears): Correlation of 0.69 with Age—substantially stronger than other features, indicating career tenure is the most predictive variable
• Secondary Predictors: JobLevel (0.51) and MonthlyIncome (0.49) show moderate correlations, suggesting career progression and compensation relate meaningfully to age
• Weak Predictors: Bottom-ranked features (EnvironmentSatisfaction, JobInvolvement, DistanceFromHome) have correlations near zero with p-values >0.59, indicating negligible linear relationships
• Statistical Significance: Top 10 features show p-values near 0, confirming strong relationships; bottom 9 features are not statistically significant

Interpretation

The feature importance ranking reveals a clear hierarchy: career-related variables (working years, job level, income) dominate predictive power for Age, while satisfaction and environmental factors contribute minimally. The steep drop-off after rank 5 suggests

Purpose

This section evaluates the target variable (Attrition) to understand class distribution, identify which features predict attrition, and assess statistical relationships between predictors and the outcome. This foundation is critical for building reliable classification models and understanding which employee characteristics drive attrition risk.

Key Findings

• Class Imbalance: 84.4% "No" attrition vs. 15.6% "Yes" (5.41:1 ratio, MODERATE severity) — the minority class is substantially underrepresented, requiring careful model evaluation
• Top Predictive Feature: Age ranks first in importance (0.21 point-biserial correlation), followed by JobLevel and JobInvolvement
• Significant Categorical Associations: 5 of 7 categorical variables show statistically significant relationships with Attrition (OverTime, JobRole, MaritalStatus, EducationField, BusinessTravel; p < 0.05)
• Significant Numeric Associations: 10 of 20 numeric variables significantly differ across attrition classes, with JobLevel, YearsWithCurrManager, and MonthlyIncome showing strongest differences

Interpretation

The moderate class imbalance indicates that standard accuracy metrics will be misleading—a model predicting "No" for all cases would achieve 84% accuracy