Skip to main content

Advanced analytics

Advanced Analytics

Machine learning, regression, and statistical analysis for Viva Insights data, including random forest top-performer models, information value, pairwise chi-square tests, difference-in-differences intervention evaluation, meeting engagement drivers, and collaboration by time of day in R and Python.

Advanced Analytics Scripts

This page covers machine learning, regression models, and statistical analysis techniques for Viva Insights data.

The top performers use case is used for understanding the drivers behind top performance, in which top performance is usually provided by a business outcome metric uploaded into Viva Insights. The example scripts below make use of a random forest model, which has the benefits of handling non-linear relationships, providing feature importance rankings, and being robust to outliers and missing values. The same technique can be used for predicting other outcomes, such as high engagement or likelihood to stay (using sentiment surveys).

The information value use case is used for feature selection and understanding which Viva Insights metrics are most predictive of a categorical outcome variable. Similar to the top performers use case, this allows you to identify the most important Viva Insights metrics that differentiate between different groups or categories for a categorical outcome variable. Information Value (IV) is particularly useful for identifying variables with strong predictive power while avoiding overfitting in your models.

When choosing which technique to use, consider if you need to build a predictive model that can handle complex, non-linear relationships and you want to make actual predictions on new data (e.g., predicting which employees are likely to be top performers). Random Forest is ideal when you have sufficient sample size (typically 100+ observations) and want robust predictions with feature importance rankings. Use information value when you need to perform initial feature selection, have limited sample sizes, or want to understand the univariate predictive power of individual variables before building more complex models. IV is particularly valuable for preprocessing large numbers of potential predictors and identifying which variables are worth including in downstream modeling efforts.

The pairwise chi-square tests use case is used for statistical hypothesis testing to determine if there are significant associations between categorical variables - typically organizational attributes or survey attributes - in your Viva Insights data. This technique is particularly valuable when you want to understand relationships between different organizational attributes (such as department, level, or location) and collaboration patterns or behaviors. The scripts include multiple testing corrections to control for false discovery rates when performing many simultaneous comparisons, ensuring reliable statistical conclusions.

The behavioral and program analysis examples move from modelling attributes to answering practical workplace questions. The collaboration by time of day scripts estimate a typical start and end of day from the hourly collaboration metrics, and they show how those hours shift by weekday and by role. The evaluating a workplace intervention scripts set up a treated-versus-control, difference-in-differences design so that a genuine programme effect can be separated from a company-wide or seasonal trend, which makes them directly applicable to measuring the impact of a Microsoft 365 Copilot enablement wave. The meeting engagement drivers scripts model in-meeting messaging as a proxy for disengagement and rank the meeting characteristics that drive it, and they then take a closer look at meeting duration to separate a real effect from simple exposure. Because the sample datasets do not contain the hourly buckets, a real intervention, or enough multi-person meetings, these examples generate small, clearly labelled simulated datasets that share the column names of a real query, so that the same downstream code runs unchanged on your own export.

Machine Learning & Predictive Modeling

Top Performers Modeling (Python)

📓 top-performers-rf.ipynb

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

Top Performers Modeling (R)

📄 top-performers-rf.Rmd

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

Statistical Analysis

Information Value Analysis (Python)

📓 information-value.ipynb

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

Information Value Analysis (R)

📄 information-value.Rmd

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

Pairwise Chi-Square Tests (Python)

📄 pairwise-chisq.py

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

Pairwise Chi-Square Tests (R)

📄 pairwise_chisq.Rmd

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

Behavioral & Program Analysis

Collaboration by Time of Day (Python)

📄 collaboration-by-time-of-day.py

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

Collaboration by Time of Day (R)

📄 collaboration-by-time-of-day.Rmd

  • Purpose: Estimate a typical start and end of day from hourly collaboration metrics
  • Language: R
  • 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

Evaluating a Workplace Intervention (Python)

📄 evaluate-intervention.py

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

Evaluating a Workplace Intervention (R)

📄 evaluate-intervention.Rmd

  • Purpose: Measure a workplace intervention with a treated-vs-control difference-in-differences design
  • Language: R
  • 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

Meeting Engagement Drivers (Python)

📄 meeting-engagement-drivers.py

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

Meeting Engagement Drivers (R)

📄 meeting-engagement-drivers.Rmd

  • Purpose: Rank the meeting characteristics that drive in-meeting messaging as a proxy for disengagement
  • Language: R
  • 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

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

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

2. Top Performers Analysis Workflow

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

3. Statistical Analysis Workflow

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

Prerequisites

Python Environment

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

R Environment

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

Best Practices

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

Need Help?

Last updated: Jul 2, 2026 Edit this page on GitHub