Marginal accuracy measures how well a synthetic population matches the target distributions of individual variables (like age, gender, or education) without considering their combinations.
Given these target distributions for a population of 1,000:
Variable | Category | Target Count | Target % |
---|---|---|---|
Age | 18-25 | 300 | 30% |
26-35 | 700 | 70% | |
Gender | Male | 450 | 45% |
Female | 550 | 55% |
Method | Marginal Accuracy | Mechanism | Visual |
---|---|---|---|
Deterministic Reweighting | Perfect for 1-2 dimensions | Direct weight adjustment per variable |
Perfect
|
Iterative Proportional Fitting (IPF) | High for all dimensions | Iterative proportional adjustments |
High
|
Conditional Probabilities | Approximate | Sampling from conditional distributions |
Medium
|
Simulated Annealing | Variable (depends on parameters) | Optimization toward targets |
Variable
|
Methods that prioritize perfect marginal accuracy often sacrifice realistic joint distributions:
Method | Marginal Accuracy | Joint Distribution Quality |
---|---|---|
Deterministic | ✅ Perfect | ❌ Poor |
IPF | ✅ High | ⚠️ Moderate |
Conditional | ⚠️ Approximate | ✅ Best |
Common metrics include: