Advanced Analytics Scripts

This page covers machine learning, regression models, and statistical analysis techniques for Viva Insights data — from predicting top performance and selecting predictive features, to testing for significant associations, to measuring the real-world impact of a program or intervention.

Pick a card to jump straight to a technique, or read the short intro above each section for guidance on which one fits your question.

🌳 Top Performers Modeling Random forest model identifying what characteristics drive top performance, or any other business outcome you upload. Jump to section → 🔍 Information Value Analysis Rank which Viva Insights metrics are most predictive of a categorical outcome, for feature selection before a bigger model. Jump to section → 📐 Pairwise Chi-Square Tests Test whether organizational attributes are significantly associated with collaboration behaviors, with multiple-testing correction built in. Jump to section → 🕒 Collaboration by Time of Day Estimate a typical start and end of day from hourly collaboration metrics, and see how it shifts by weekday and role. Jump to section → 🧪 Evaluating a Workplace Intervention A treated-vs-control, difference-in-differences design that separates a genuine program effect from a company-wide trend. Jump to section → 📅 Meeting Engagement Drivers Rank the meeting characteristics that drive in-meeting messaging, used here as a proxy for disengagement. Jump to section →

Machine Learning & Predictive Modeling

Top Performers Modeling

Understand the drivers behind top performance, where top performance is usually a business outcome metric uploaded into Viva Insights. The scripts below use a random forest model, which handles non-linear relationships, provides feature importance rankings, and is robust to outliers and missing values. The same technique can predict other outcomes too, such as high engagement or likelihood to stay (using sentiment surveys). Random Forest is best when you have sufficient sample size (typically 100+ observations) and want robust predictions with feature importance rankings — for smaller samples or a first pass at feature selection, see Information Value below.

📄 top-performers-rf.Rmd

Purpose: Identify characteristics of top performers using Random Forest
Format: R Markdown
Prerequisites: vivainsights R package, randomForest, dplyr
Key Features: Feature importance analysis, model validation, performance metrics
📥 Download
🌐 View HTML Output

📓 top-performers-rf.ipynb

Purpose: Identify characteristics of top performers using Random Forest
Format: Jupyter Notebook
Prerequisites: vivainsights Python package, scikit-learn, pandas
Key Features: Feature importance analysis, model validation, performance metrics
📥 Download

Statistical Analysis

Information Value Analysis

Feature selection: understand which Viva Insights metrics are most predictive of a categorical outcome variable, identifying variables with strong predictive power while avoiding overfitting. Use Information Value (IV) for initial feature selection, with limited sample sizes, or to understand the univariate predictive power of individual variables before building a more complex model like the Random Forest above — it’s particularly valuable for preprocessing large numbers of potential predictors.

📄 information-value.Rmd

Purpose: Calculate Information Value (IV) for feature selection and variable importance
Format: R Markdown
Prerequisites: vivainsights R package, Information, dplyr
Key Features: IV calculation, binning strategies, feature ranking
📥 Download

📓 information-value.ipynb

Purpose: Calculate Information Value (IV) for feature selection and variable importance
Format: Jupyter Notebook
Prerequisites: vivainsights Python package, pandas, numpy
Key Features: IV calculation, binning strategies, feature ranking
📥 Download

Pairwise Chi-Square Tests

Statistical hypothesis testing for significant associations between categorical variables — typically organizational attributes (department, level, location) or survey attributes — and collaboration patterns or behaviors. Multiple testing corrections control for false discovery rates when running many simultaneous comparisons, keeping the conclusions reliable.

📄 pairwise_chisq.Rmd

Purpose: Perform pairwise chi-square tests for categorical variables
Format: R Markdown
Prerequisites: vivainsights R package, stats
Key Features: Multiple testing correction, p-value adjustment, significance testing
📥 Download
🌐 View HTML Output

📄 pairwise-chisq.py

Purpose: Perform pairwise chi-square tests for categorical variables
Prerequisites: vivainsights Python package, scipy, pandas
Key Features: Multiple testing correction, p-value adjustment, significance testing
📥 Download

Behavioral & Program Analysis

These examples move from modelling attributes to answering practical workplace questions. Because the sample datasets don’t contain the hourly buckets, a real intervention, or enough multi-person meetings needed below, each script generates a small, clearly labelled simulated dataset that shares the column names of a real query — so the same downstream code runs unchanged on your own export.

Collaboration by Time of Day

Estimate a typical start and end of day from hourly collaboration metrics, and show how those hours shift by weekday and by role.

📄 collaboration-by-time-of-day.Rmd

Purpose: Estimate a typical start and end of day from hourly collaboration metrics
Format: R Markdown
Prerequisites: vivainsights R package, tidyverse, lubridate
Key Features: Hourly activity matrix, two-stage aggregation, cuts by weekday and role
📥 Download
🌐 View HTML Output

📄 collaboration-by-time-of-day.py

Purpose: Estimate a typical start and end of day from hourly collaboration metrics
Prerequisites: vivainsights Python package, pandas, numpy
Key Features: Hourly activity matrix, two-stage aggregation, cuts by weekday and role
📥 Download

Evaluating a Workplace Intervention

Set up a treated-versus-control, difference-in-differences design so a genuine programme effect can be separated from a company-wide or seasonal trend — directly applicable to measuring the impact of a Microsoft 365 Copilot enablement wave.

📄 evaluate-intervention.Rmd

Purpose: Measure a workplace intervention with a treated-vs-control difference-in-differences design
Format: R Markdown
Prerequisites: vivainsights R package, tidyverse
Key Features: Before/During/After windows, difference-in-differences, two-stage aggregation, displacement checks
📥 Download
🌐 View HTML Output

📄 evaluate-intervention.py

Purpose: Measure a workplace intervention with a treated-vs-control difference-in-differences design
Prerequisites: vivainsights Python package, pandas, numpy
Key Features: Before/During/After windows, difference-in-differences, two-stage aggregation, displacement checks
📥 Download

Meeting Engagement Drivers

Model in-meeting messaging as a proxy for disengagement and rank the meeting characteristics that drive it, then take a closer look at meeting duration to separate a real effect from simple exposure.

📄 meeting-engagement-drivers.Rmd

Purpose: Rank the meeting characteristics that drive in-meeting messaging as a proxy for disengagement
Format: R Markdown
Prerequisites: vivainsights R package, tidyverse, randomForest
Key Features: Meeting-level modelling, random forest permutation importance, rate-vs-exposure duration analysis
📥 Download
🌐 View HTML Output

📄 meeting-engagement-drivers.py

Purpose: Rank the meeting characteristics that drive in-meeting messaging as a proxy for disengagement
Prerequisites: vivainsights Python package, scikit-learn, pandas, numpy
Key Features: Meeting-level modelling, random forest permutation importance, rate-vs-exposure duration analysis
📥 Download

Sample Datasets

Simulated Person Query

📄 simulated_person_query.csv

Purpose: Simulated person-level data for analysis
Format: CSV
Contents: Weekly collaboration metrics, meeting data, email patterns

Analysis Workflows

1. Feature Selection Workflow

Load Data: Import your Viva Insights query results
Information Value: Run IV analysis to identify important variables
Statistical Testing: Use chi-square tests for categorical relationships
Model Building: Apply selected features to predictive models

2. Top Performers Analysis Workflow

Data Preparation: Clean and prepare performance data
Feature Engineering: Create relevant collaboration metrics
Model Training: Train Random Forest model
Interpretation: Analyze feature importance and model results
Validation: Test model performance on holdout data

3. Statistical Analysis Workflow

Exploratory Analysis: Understand data distributions
Hypothesis Testing: Test relationships between variables
Effect Size: Calculate practical significance
Reporting: Generate analysis reports

Prerequisites

R Environment

install.packages(c("vivainsights", "dplyr", "tidyr", "ggplot2", "scales", "purrr", "randomForest", "fixest", "Information", "rmarkdown"))

Python Environment

pip install vivainsights pandas numpy scikit-learn linearmodels matplotlib seaborn jupyter

Best Practices

Data Quality: Always validate your data before analysis
Feature Selection: Use IV analysis to identify meaningful variables
Model Validation: Always test models on holdout data
Statistical Significance: Consider both statistical and practical significance
Documentation: Document your analysis methodology and assumptions

Causal Inference in Copilot Analytics: move beyond correlation to measure the true impact of an intervention
Network Analysis: organizational network analysis (ONA) as a complementary advanced technique
Copilot Analytics: adoption metrics and Power/Habitual user segmentation
Essentials: utilities and visualizations to prepare your data
Getting Started: environment setup and first steps

Need Help?

Machine Learning: Scikit-learn Documentation
Statistical Analysis: R Stats Documentation
Viva Insights: Package Documentation
Sample Data: Example datasets