The GRI is built on Total Variation Distance (TVD), one of the most fundamental metrics in probability theory. For discrete distributions \(P\) and \(Q\) over a shared set of categories:
\[ \text{TVD}(P, Q) = \frac{1}{2} \sum_{i=1}^{k} |p_i - q_i| \]
TVD has several properties that make it ideal for representativeness measurement:
The GRI is defined as the complement of TVD:
\[ \text{GRI}(p, q) = 1 - \text{TVD}(p, q) = 1 - \frac{1}{2} \sum_{i=1}^{k} |p_i - q_i| \]
This maps to a 0–1 scale where 1.0 = perfect match and 0.0 = complete mismatch.
Inputs:
Process:
The GRI captures how closely the sample matches the population, but not how many population segments are represented at all. The Diversity Score fills this gap:
\[ \text{Diversity} = \frac{|\{i : p_i > 0 \text{ and } q_i \geq X\}|}{|\{i : q_i \geq X\}|} \]
where \(X = \frac{1}{2N}\) is a dynamic relevance threshold that scales with sample size. This measures the fraction of meaningfully-sized population strata that appear at least once in the sample.
The GRI is computed across 13 dimensions organized in three tiers:
These combine geography with demographics, producing the finest-grained assessment:
| Dimension | Strata Count | Benchmark Source |
|---|---|---|
| Country x Gender x Age | 2,699 | UN WPP 2023 |
| Country x Religion | 1,607 | Pew GLS 2010 |
| Country x Environment | 449 | UN WUP 2018 |
Replace countries with 20 world regions, reducing cardinality:
Assess one demographic factor at a time:
Each dimension tells a different story. A survey might achieve strong continental representation (GRI = 0.88) while poorly representing the Country x Gender x Age intersection (GRI = 0.37)—the scorecard makes these trade-offs transparent.
With thousands of strata and finite samples, perfect GRI scores are mathematically impossible. We use Monte Carlo simulation to estimate the maximum achievable GRI at any sample size by generating optimally-allocated samples.
At \(N = 1{,}000\):
| Dimension | Max GRI | Max Diversity |
|---|---|---|
| Country x Gender x Age | 0.792 | 0.925 |
| Country x Religion | 0.938 | 0.970 |
| Country x Environment | 0.950 | 0.977 |
These ceilings enable efficiency ratios—the percentage of the theoretical maximum that a real survey achieves—providing a fairer basis for comparison than raw scores alone.
The GRI measures how far a sample deviates from the population. The design effect answers the next question: what does that deviation cost?
When a survey’s demographic composition differs from the population, analysts must post-stratify—reweight responses so that each stratum contributes proportionally to its population share. Reweighting inflates variance, because over-represented strata get downweighted (wasting collected data) and under-represented strata get upweighted (amplifying noise).
\[ d_{\text{eff}} = \sum_{i \in S} \frac{\hat{q}_i^2}{p_i} \]
where \(p_i\) is the sample proportion and \(\hat{q}_i\) is the population proportion renormalized over represented strata. When \(p = q\) (perfect representation), \(d_{\text{eff}} = 1\). Any deviation from the population distribution pushes \(d_{\text{eff}}\) above 1.
\[ N_{\text{eff}} = \frac{N}{d_{\text{eff}}} \]
This is the number of optimally-allocated respondents that would yield the same precision. A survey of 58,000 with \(d_{\text{eff}} = 29\) has the same statistical power as ~2,000 optimally-allocated respondents.
\[ \text{Precision Retained} = \frac{1}{d_{\text{eff}}} \]
This is the fraction of the sample budget that contributes to inferential precision after reweighting. When precision retained is 10%, ninety percent of the survey effort is consumed by correcting demographic mismatch.
Reweighting can correct estimates of population means but cannot recover the precision lost to poor allocation. A survey that over-represents one country by 10x and under-represents another by 10x can be reweighted to produce unbiased estimates—but the confidence intervals will be much wider than a survey that allocated respondents proportionally. The design effect quantifies exactly how much wider.
GRI and design effect measure complementary aspects of representativeness:
The key distinction is symmetry vs. asymmetry. GRI treats overrepresentation and underrepresentation symmetrically: sampling 5 percentage points too many or too few in a stratum contributes the same amount to the TVD. Design effect is fundamentally asymmetric: underrepresentation is far more expensive. When a stratum has population share \(q_i = 0.10\) but sample share \(p_i = 0.01\), the reweighting ratio \(q/p = 10\) amplifies that stratum’s noise tenfold. The reverse — \(q_i = 0.01, p_i = 0.10\) — merely downweights excess data, wasting budget but preserving precision.
This asymmetry means two surveys with similar GRI scores can have very different design effects. The survey whose misallocation involves severe underrepresentation in a few strata pays a much higher precision cost than one whose misallocation is more evenly distributed. Reporting both metrics gives a complete picture: GRI measures how much the sample deviates; design effect reveals how costly that deviation is.
| Dataset | Source | Year | Use |
|---|---|---|---|
| World Population Prospects | UN DESA Population Division | 2023 | Country x Gender x Age benchmarks |
| Global Religious Landscape | Pew Research Center | 2010 | Country x Religion benchmarks |
| World Urbanization Prospects | UN DESA Population Division | 2018 | Country x Environment benchmarks |
All benchmark data is included in the repository under data/raw/benchmark_data/ with full attribution in Sources.csv.
The religious composition data from Pew dates to 2010 and may not reflect current demographics in rapidly changing regions. Urban/rural classifications use national-level data that can mask within-country variation. These limitations are inherent to the best available global demographic data and should be considered when interpreting GRI scores.