There is no one-size-fits-all formula for “how many completes per region” in national and regional surveys. If the primary goal is to estimate national indicators, sample allocation should roughly mirror the regional distribution of the target population; if the primary goal is to compare regions, smaller areas may require oversampling; if both outcomes matter, a compromise allocation should be planned in advance, accepting the cost of a reduced effective sample size for national estimates after weighting. Splitting the sample equally across provinces or cities may seem fair but only guarantees equal nominal counts, not national representativeness, nor comparable precision for each region.
The right sequence is: first define the “estimation domains” for which results will be reported separately, then set acceptable uncertainty levels for each domain, and only then allocate the sample. Administrative provinces, sales regions, city tiers, store trade areas, and service locations can all be different domains; only geographic units that truly correspond to decisions and analyses deserve to consume independent sample size.
First, Define National Indicators and Regional Comparisons as Two Deliverables
National indicators answer “what is the average situation in the target population,” where each region’s contribution to the result should be close to its share of the target population. Regional comparisons answer “whether there are actionable differences across regions,” where each region to be compared needs an adequate unweighted base. The United Nations Statistics Division calls population subsets for which separate estimates are planned “estimation domains,” and notes that when similar precision is desired for each domain, equal allocation is often more efficient; for national estimates, allocation proportional to population size is usually more appropriate; when both are desired, a compromise is needed.
Project documents can make the deliverables quite specific:
- National Dashboard: national satisfaction rate and its uncertainty, weighted by target customer or demographic-regional structure.
- Regional Performance Table: satisfaction rate, unweighted base, and issue scenarios for each of the four sales regions.
- City Diagnostics: provide direct estimates only for cities that meet the publication threshold and whose sampling design supports it; list all others as exploratory signals only.
- Action List: assign national commonalities, regional process issues, and site-specific anomalies to their respective owners.
If the business needs to examine results across 30 provinces × 5 customer segments, this is not 150 filter labels, but 150 potential estimation domains. Requiring reliable bases for each cell will quickly inflate the total sample. At that point, priority should be given to reducing published domains, merging similar regions, or acknowledging that some results are exploratory only—do not let an automated map generate numbers that hide insufficient evidence.
The Same 1,200 Interviews: Three Allocations Answer Different Questions
Below we illustrate the trade-offs with a hypothetical customer experience survey. Assume the target customer shares for four regions in the sampling frame are East 50%, South 25%, North 15%, and West 10%, with a total effective sample budget of 1,200. For ease of understanding, the following error margins use only the simple random sampling approximation at a proportion of 50%; real projects with clustering, weighting, nonresponse, or non-probability sources cannot directly apply these numbers.
Option 1: Proportional allocation: The four regions receive 600, 300, 180, and 120 interviews, respectively. This requires no additional weighting for regional quotas, so the national effective sample size is closest to 1,200; but the approximate error margins per region are about ±4.0, ±5.7, ±7.3, and ±8.9 percentage points. Even a six-percentage-point difference between West and East would likely not provide enough information to determine whether the difference exceeds sampling variability.
Option 2: Equal allocation: Each region receives 300 interviews, with approximate error margins around ±5.7 percentage points for each—suitable when the four regions are equally important for comparison. However, for the national aggregate, each East record should receive a relative weight about twice that of East under equal allocation, and each West record about one-third that of East; later calculations show effective sample size is lower. Using the sum of squared weights, the nominal 1,200 corresponds to a weighted effective sample size of about 870, so national precision is lower than for a proportional sample of the same size.
Option 3: Compromise allocation: Apply approximate allocation proportional to the square root of regional size, yielding 444 for East, 314 for South, 243 for North, and 199 for West. Approximate error margins per region are about ±4.7, ±5.5, ±6.3, and ±6.9 percentage points; the national effective sample size after weighting is about 1,105. It does not make precision equal across the four regions but it better covers smaller regions than proportional allocation, and retains more national information than equal allocation.
Square-root allocation is not a mandatory standard answer—it is a transparent starting point. The final allocation should also consider variation in outcomes across regions, cost per completion, incidence rates, expected nonresponse, minimum publishable bases, and the most important—which comparisons matter most. If the West is the only market slated for expansion next quarter, allocating more sample there could be justified; if the national renewal rate is the sole metric, a closer-to-proportional allocation would be more economical.
After Oversampling, National Results Must Restore the Regional Structure
Continue with the compromise allocation. Assume the numbers of satisfied respondents in East, South, North, and West are 329, 214, 173, and 125, respectively, yielding unweighted satisfaction rates of about 74.1%, 68.2%, 71.2%, and 62.8%. Summing the satisfied respondents directly gives 841 ÷ 1,200, or about 70.1%.
But this sample intentionally oversampled the North and West. If the shares 50%, 25%, 15%, and 10% in the sampling frame indeed reflect the target population structure, and other weighting and nonresponse assumptions hold, the national estimate should be aggregated by the four regional shares:
50% × 74.1% + 25% × 68.2% + 15% × 71.2% + 10% × 62.8% ≈ 71.0%.
The raw total differs from the region-weighted result by about 0.9 percentage points. The size of the difference is not the criterion for whether to weight: as soon as selection probabilities are altered deliberately, the design must be restored by weighting; a small difference by chance does not mean weighting is unnecessary. Reports should also provide the unweighted base for each region, the population benchmarks used, the weight range, and the weighted effective sample size. External benchmarks must correspond to the same target population and time window—using the resident population distribution to calibrate a survey of paying customers is not valid, nor is using sales revenue distribution to explain the share of “customers.”
Distinguish Where People Live from Where Events Occur in Geographic Fields
“Region” alone usually cannot support action. A person may reside in Hangzhou, purchase at a store in Shanghai, receive delivery from the South China warehouse, and have after-sales service handled by a national online team. These four locations correspond to population coverage, store experience, fulfillment responsibility, and service ownership, respectively. Research should collect each separately based on decisions, rather than asking respondents to guess which location the researcher means.
We recommend creating a versioned dictionary of geographic data, containing at least:
- Place of residence: used for analyses related to population or living environment, and specify the minimum duration of stay.
- Transaction or experience location: tied to the most recent concrete event, a primary field for store and offline service diagnostics.
- Service responsibility area: derived from orders, stores, or organizational mappings; respondents should not be required to know internal sales regions.
- Geographic hierarchy and code version: preserve the parent–child relationships between province, city, and district/county, along with effective dates, to avoid drift in historical results after boundary adjustments.
- City tier or market group: record the issuing organization, criteria, and version; commercial classifications such as “new first-tier” are not stable administrative codes.
A full address is usually unnecessary for regional analysis. Prioritize letting participants choose a level that supports decisions; collect precise locations only when distance to delivery or trade area accessibility has an analytical use, and restrict access and retention periods. IP addresses and device geolocation are affected by proxies, corporate egress points, permissions, and travel, and should be used only as consented supplementary signals, not to unconditionally override self-reported event locations.
Even with Sufficient Regional Sample, Check for Concentration in a Few Points
Two hundred interviews in West does not mean the region is covered. If 180 of them come from the same city, the same store, or the same partner channel, responses may be similar due to shared circumstances. Clustering makes 200 nominal observations provide less information than 200 independent observations. Suppose the West has a design effect of 1.6; its 199 interviews have a design-effective sample size of about 124, and the approximate error margin around a 50% proportion widens from ±6.9 to about ±8.8 percentage points.
Therefore, for each region, also check the number of cities, the number of sites, the maximum sample per site, the urban–rural or core–non-core market coverage, the sample source, and the field dates. Fewer sites can lower execution costs but increase clustering; more sites typically improve coverage but raise coordination costs per site. The sampling design must choose between number of sites, within-site sample, and budget; it cannot simply fill a total sample quota.
Using different recruitment methods across regions also confounds interpretation. For instance, if East was recruited via emails from a customer list and West via in-store intercepts, then satisfaction differences combine region, channel, in-store eligibility, and response mode. It is best to keep the core source and measurement mode comparable across regions; when that is not possible, at least retain source and mode variables and run within-channel comparisons or sensitivity analyses, noting the limitations in the conclusions.
When Regional Differences Emerge, Rule Out Compositional Effects Before Telling a Geographic Story
A lower satisfaction rate in West does not automatically mean “West customers are harder to satisfy.” It could be driven by a higher share of new customers, longer delivery distances, product stockouts, a promotional failure during the field period, or concentration of the sample in one problematic store. The analysis should first check unweighted bases and uncertainty, then compare customer type, channel, product, event timing, and site structure, and finally interpret mechanisms using open-ended responses or operational records.
If differences shrink after adjusting for observable structure, you can report that “sample composition explains part of the regional gap”; if differences persist, you can only say that the available variables did not explain them—not attribute them directly to culture or regional personality. To infer that a regional policy caused the change would require a stronger design: temporal comparisons, natural experiments, or randomized trials. Geographic variables are usually descriptive and locational clues, not causes.
When Small-Area Samples Are Too Small: Three Honest Options
The first is to combine adjacent waves or similar regions, but confirm that definitions of indicators, market conditions, and questionnaires have not changed materially, and mark the combined time window clearly. The second is to increase sample size or reduce the level of reporting, such as aggregating district/county up to city or region. The third is to use small-area estimation models that combine direct survey data with reliable auxiliary variables.
The U.S. Census Bureau explains that small-area models can improve precision when direct samples are insufficient, but model misspecification—or ignoring correlations between sampling design and outcomes—can produce bias. Therefore, model-based results must be distinguished from direct estimates, with the model, auxiliary data, validation, and prediction error disclosed. Do not present a color-filled map from model predictions as “we surveyed enough people in every district.” If the team lacks model diagnostic capability, reducing reporting granularity is usually safer than outputting precise-looking decimals that are not real.
Regional Sampling Checklist from Recruitment to Reporting
- Define the target population and distinguish residence, event location, responsibility area, and market group.
- List only the estimation domains that truly require separate results; do not spend sample on splits that no one is accountable for.
- Set precision or minimum evidence requirements for the national indicator and for each priority region.
- Compare proportional, equal, and compromise allocations; calculate the weight design effect and expected costs.
- Back out the number of invitations per region using expected eligibility, completion, and quality pass rates.
- Limits on concentration across cities, sites, and sources; pilot-test and then re-check clustering risk.
- Lock the geographic dictionary version and keep both original fields and derived responsibility areas in the data table.
- Monitor the funnel and key variables by region; do not silently relax population criteria at the end.
- Estimate national results according to the target population structure and sampling design; report unweighted bases for regional results.
- Hide or merge small domains with insufficient evidence on maps; label model-based results separately from direct survey estimates.
When you need to determine population boundaries first, combine with methods for defining the target survey audience; when determining how much sample each estimation domain needs, refer to sample size calculation and design effects; after oversampling is done, evaluate weights and effective sample size with the basics of survey weighting. Regional allocation, sample size, and weighting are one design chain; they cannot be fixed at the reporting stage.
Methodological Basis and Further Reading
- United Nations Statistics Division: Household Sample Surveys, Chapter 2: trade-offs among estimation domains and proportional, equal, and compromise allocations.
- UNSD: Handbook of Surveys on Households and Individuals: geographic stratification, planned domains, oversampling, and weight variances.
- U.S. Census Bureau: Small Area Estimation: model use, auxiliary data, and bias risks when direct samples are insufficient.
- AAPOR Disclosure Standards: disclosing unweighted sample bases and replicable methodology when reporting subgroup results.