Group-Level Analysis & Variable-Level Drilldown
Group Decomposition Overview
Purpose: Drill into specific contribution groups to see individual variable performance within that group.
Use Case: After seeing "Media" is a major contributor in main decomposition, drill down to compare TV vs. Digital vs. Radio performance.
Running Group Decomposition
Step 1: Select Group
In Decomposition page, below main charts
Dropdown shows all groups from main decomposition
Choose group to analyze (e.g., "Media")
Step 2: Click "Run Group Decomposition"
Calculates individual variable contributions within selected group
Generates variable-level chart
Shows only variables assigned to that group
Step 3: Analyze Results
Stacked bars show individual variables
Black line shows total group contribution
Each variable has distinct color/shade
Group Decomposition Chart
Structure:
Similar to main decomposition but variable-level
Each variable within the group shown separately
Sum of variable contributions = total group contribution
Chart Elements:
Stacked Bars:
One bar per time period
Each segment = one variable's contribution
Height of bar = total group contribution
Total Line (Black):
Shows overall group contribution from main decomposition
Bars should sum to this line
Validates calculation
Colors:
Each variable has unique color
Distinct from other variables in group
Auto-assigned or customizable
Interpreting Variable-Level Results
Identify Top Performers:
Insights:
TV dominates media contribution (50%)
Digital is second (30%)
Print contributes minimally (5%)
Performance Comparison:
Which specific channels drive the group?
Are all channels contributing proportionally?
Any underperformers to cut or optimize?
Timing Analysis:
Do all variables spike together?
Or do they have different timing?
Campaign coordination visible?
Common Group Drill-Down Scenarios
Scenario 1: Media Group Analysis
Question: "Within Media, which channels perform best?"
Process:
Run main decomposition → Media is top contributor
Select "Media" group in dropdown
Run group decomposition
Compare TV, Digital, Radio, Print, etc.
Typical Findings:
TV often largest absolute contributor
Digital often highest ROI
Some channels underperform
Reallocation opportunities identified
Scenario 2: Digital Channel Mix
Question: "Within Digital, is Search or Social more effective?"
Process:
Ensure Search, Social, Display assigned to "Digital" group
Select "Digital" in group dropdown
Run group decomposition
Compare individual digital channels
Typical Findings:
Search drives consistent baseline
Social creates spikes during campaigns
Display supports other channels
Retargeting has high efficiency
Scenario 3: Promotional Analysis
Question: "Which types of promotions drive most lift?"
Process:
Group promotional variables together
Select "Promotions" group
Run group decomposition
Compare different promotion types
Typical Findings:
Deep discounts drive volume but hurt margins
Limited-time offers create urgency
Free shipping has consistent impact
BOGO has highest incremental lift
Variable-Level ROI Calculation
From group decomposition data:
Get variable contributions:
Sum each variable's contribution across time
From group decomposition chart or export
Match with spend:
Get actual spend for each variable
Same time periods
Calculate individual ROI:
Rank by performance:
Sort channels by ROI
Identify best and worst performers
Guide budget reallocation
Example:
Multi-Level Drill-Down Strategy
Level 1: Overall Model
Total KPI decomposition
All groups visible
Strategic overview
Level 2: Group Analysis
Drill into major groups
Compare related channels
Tactical insights
Level 3: Individual Campaigns (if modeled)
Specific campaign performance
Campaign vs. always-on
Granular optimization
Workflow:
Start broad (all groups)
Identify top contributors
Drill into those groups
Analyze variable-level
Make specific recommendations
OLS vs Bayesian Decomposition
Conceptual Difference
OLS Decomposition:
Uses point estimates from OLS regression
Single coefficient value per variable
Deterministic contributions
Frequentist approach
Bayesian Decomposition:
Uses posterior mean from Bayesian inference
Based on MCMC samples
Incorporates prior information
Bayesian approach
Both produce the same chart format - the difference is which coefficient values are used.
When to Use Each
Use OLS Decomposition When:
Standard Analysis:
Default approach for most MMM
Quick and straightforward
Well-understood by stakeholders
Sufficient Data:
Large sample size (50+ observations)
Strong signal-to-noise ratio
Clear relationships
Speed Matters:
Need results quickly
Iterative model building
Exploratory analysis
Simple Reporting:
Stakeholders prefer point estimates
Less need to explain methodology
Focus on actionable insights
Use Bayesian Decomposition When:
Limited Data:
Small sample sizes
Noisy data
Weak signals
Prior Knowledge:
Strong beliefs about effects
Industry benchmarks available
Historical estimates informative
Uncertainty Quantification:
Need credible intervals
Risk assessment important
Conservative estimates preferred
Hierarchical Effects:
Regional models
Multi-product models
Partial pooling beneficial
Comparing OLS and Bayesian Results
Expected Differences:
With Non-Informative Priors:
Results very similar
Bayesian often slightly more conservative
Practical differences minimal
With Informative Priors:
Bayesian pulls estimates toward priors
More stable coefficients
Reduced extreme values
With Limited Data:
Bayesian handles better
OLS may overfit
Bayesian shrinkage helps
Running Both for Comparison
Process:
Fit both OLS and Bayesian models
Run decomposition with OLS coefficients
Run decomposition with Bayesian coefficients
Compare results
What to Check:
Contribution Magnitudes:
Are they similar?
Which variables show largest differences?
Do business conclusions change?
ROI Rankings:
Same channel ranking?
Or different priorities?
Material impact on decisions?
Temporal Patterns:
Same spikes and valleys?
Timing aligned?
Seasonal patterns consistent?
Interpreting Differences
Small Differences (<10%):
Both methods agree
Robust findings
Proceed with confidence
Moderate Differences (10-30%):
Some uncertainty present
Investigate why (data, priors)
Consider both estimates
Report range
Large Differences (>30%):
Significant disagreement
Check model specification
Review priors if Bayesian
Investigate data quality
May need more data
Reporting Considerations
For OLS:
"Point estimate decomposition"
"Based on OLS regression"
Standard confidence intervals
For Bayesian:
"Bayesian posterior mean decomposition"
"Incorporating prior information"
Credible intervals available
When Showing Both:
"Range of estimates"
"OLS vs. Bayesian comparison"
"Sensitivity analysis"
Technical Details
OLS Contribution Calculation:
Bayesian Contribution Calculation:
Both sum to predicted value:
∑Contributions_OLS = Predicted_OLS
∑Contributions_Bayesian = Predicted_Bayesian
Best Practices
Start with OLS:
Faster iteration
Build intuition
Establish baseline
Add Bayesian When:
OLS results unstable
Need uncertainty quantification
Have strong priors
Data is limited
Compare Both:
Validate findings
Understand sensitivity
Build confidence
Inform decisions
Document Choice:
Explain why OLS or Bayesian
Note key differences
Justify for stakeholders
Common Questions
Q: Which is more accurate? A: Neither is universally more accurate - depends on data and context. Bayesian can be more accurate with good priors and limited data.
Q: Can I use both? A: Yes, running both provides sensitivity analysis and validates findings.
Q: Do they give different ROI? A: Yes, potentially. But with sufficient data and non-informative priors, differences are usually small.
Q: Which should I present to executives? A: OLS is typically simpler to explain. Use Bayesian if you need to quantify uncertainty.
Q: How do I know if differences matter? A: If business decisions would change, differences matter. If same actions result, differences are academic.
Summary
Key Takeaways:
Group Decomposition:
Essential for understanding channel-level performance
Enables detailed ROI calculation
Supports tactical optimization
Drill-down reveals variable-level insights
OLS vs. Bayesian:
Both valid approaches
OLS is default, Bayesian for special cases
Small differences usually don't matter
Large differences warrant investigation
Document and justify choice
Workflow Integration:
Run main decomposition (OLS or Bayesian)
Identify top contributing groups
Drill into those groups
Calculate variable-level ROI
Compare across methods if needed
Make data-driven decisions
Last updated