DSBA 2302PythonLinear RegressionStatistical AnalysisTableau

Tech Industry Salary Demographics Study

A statistical analysis of compensation trends across 62,642 tech industry employees, investigating how demographic factors like education, race, gender, and experience drive differences in base salary. Built a linear regression model to predict total yearly compensation from behavioral and demographic variables.

research question

“How do demographic factors such as education, race, gender, and level of experience affect someone's base salary in the tech industry?”

my role

Team of five (Nick Krassy, Alec Lundy, Christian Chapman, Ezra Moen, Nash Bachman). Contributed to statistical modeling, data cleaning strategy, and correlation analysis, turning raw demographic data into usable model inputs and presenting findings to both technical and non-technical teammates.

dataset overview

62,642

Total Records

Key Variables

Model Features

Data sourced from tech industry salary records covering employees across companies of varying sizes. Key variables include total yearly compensation, years of experience, years at current company, gender, race, and education level. The dataset contained substantial missing values requiring targeted handling prior to modeling.

missing values by variable

Race had the highest missingness at 64.2%, limiting its use as a reliable predictor. Gender (31.2%) and Education (51.5%) were also heavily incomplete. All three were excluded from the final regression model due to insufficient data density.

outlier analysis

Box = interquartile range (Q1–Q3)

Bold line = median

Circle = outlier

correlation matrix

Years of experience showed the strongest correlation with compensation (r = 0.37), making it the primary predictor. Years at company had a weaker direct relationship with pay (r = 0.11), but correlated moderately with experience (r = 0.44), indicating collinearity between the two predictors.

model selection

Linear Regression

Selected

Targets a continuous quantitative variable. Preliminary analysis confirmed linear trends between experience and compensation.

Logistic Regression

Rejected

Designed for categorical outcomes. Not appropriate for predicting a continuous salary value.

Decision Tree

Rejected

Also targets categorical variables. Prone to overfitting on high-variance salary data.

model performance

Baseline: Linear Regression

R²

0.187

explains 18.7% of variance

Durbin-Watson

2.008

no autocorrelation

MSE

6.715 × 10⁹

mean squared error

Tuned: Ridge + Z-Score Filtering

Technique

Ridge Reg.

L2 regularization

Filter

Z-Score

outlier removal

MSE

6.714 × 10⁹

marginal improvement

Variance Explained (R²)

18.7%

0%100%

The model explains ~19% of the variation in compensation. The remaining 81% comes down to factors not captured in the dataset: company size, location, stock grants, and role specifics.

key findings

-Years of experience is the strongest predictor of compensation (r = 0.37), consistently outweighing tenure at a single company.
-Demographic variables (race, gender, education) could not be reliably included due to high rates of missing data (up to 64% for race).
-Ridge regression and Z-score filtering produced only marginal MSE improvement over the baseline, suggesting the data itself is the primary constraint on accuracy.
-A Durbin-Watson statistic of 2.008 confirms residuals are uncorrelated, so the linear model assumptions hold.
-The model is best used as a directional tool for salary benchmarking, not precise prediction.

limitations

Data Quality

Dataset contains significant missingness in key demographic variables, limiting the scope of the analysis.

Model Accuracy

R² of 0.187 indicates the model captures a fraction of compensation variance. Further feature engineering is needed.

Missing Context

Location, company size, and equity/bonus structures are missing, all of which are major drivers of real-world compensation.

Download Full Presentation