Population-Error: Population-Indexed Uncertainty

• Population-error is a family of population-indexed error concepts arising when uncertainty in population distributions, locations, or memberships propagates into inferential targets.
• It captures errors across various settings including misaligned denominators in disease mapping, aggregation over uncertain population distributions, and misclassification in register-based data.
• Applications span transported prediction, spatial aggregation, clinical trial error adjustment, and evolutionary models, emphasizing the need for precise population representation in analysis.

Population-error is used across several research literatures for error notions in which the population itself enters the estimand, weighting measure, latent-state structure, or decision criterion. In the work surveyed here, it covers population-at-risk uncertainty in disease mapping, target-population prediction error under transportability, aggregation error induced by uncertain population distributions, false negative and false positive observation error in register-based population inference, joint mutation–population thresholds in finite quasispecies models, and patient-weighted type I error criteria such as the population-wise error rate (Peterson et al., 2023, Steingrimsson et al., 2021, Paige et al., 2022, Brown et al., 25 Mar 2026, Cerf, 2012, Luschei et al., 2023). Taken together, this suggests that population-error is best understood as a family of population-indexed error concepts rather than a single scalar metric.

1. Population-indexed error as an inferential principle

A recurrent theme is that error is population-specific whenever the estimand is defined by averaging over a population distribution. In transported prediction, the target quantity is the target population mean squared error,

$\psi_{\widehat\beta}=\E[(Y-g_{\widehat\beta}(X))^2\mid S=0],$

and the central claim is that source-population performance generally does not equal target-population performance because the loss is marginalized over $f(X\mid S=0)$ rather than $f(X\mid S=1)$; the paper formalizes this with outcome-model and inverse-odds-weighting identification formulas and introduces “prediction error modifiers” for variables associated with conditional prediction error (Steingrimsson et al., 2021).

A closely related construction appears in spatial aggregation. For a finite population distributed over enumeration areas, prevalence and burden are defined as

$$p(A)=\sum_{i=1}^{M} \frac{N_i}{N}\frac{Z_i}{N_i}, \qquad b(A)=\sum_{i=1}^{M} Z_i,$$

and the aggregation model must account for three sources of “aggregation error”: aggregation weights, fine scale variation, and finite population variation. The paper’s central objection to common practice is that replacing the unknown population distribution with a fixed population density surface suppresses uncertainty in the aggregation operator itself (Paige et al., 2022).

Disease mapping makes the same point from the denominator side. In the Bayesian Spatial Berkson Error model, the population-at-risk is not fixed but latent,

$\omega_{c,a}=\mu_{c,a}+\log(R)+\log(\gamma_{c,a}),$

so population uncertainty becomes measurement error in the log-offset. The authors argue that treating denominators as fixed can bias local risk estimates and under-estimate their associated uncertainties (Peterson et al., 2023). This suggests a unifying interpretation: population-error arises whenever uncertainty in population composition, population location, or population membership propagates into the estimand rather than remaining ancillary.

2. Denominator uncertainty and spatial population measurement

In spatial demography and epidemiology, population-error frequently denotes uncertainty in small-area denominators or in gridded population surfaces. The B-Pop framework was designed around this problem. It integrates Decennial Census, PEP, and ACS in “a single model integrating multiple data sources,” while “accounting for data source specific data generating mechanisms” and “data source specific errors,” and it predicts “estimates for years without USCB reported data” for Georgia counties over 2005–2021 (Peterson et al., 2021). The latent annual county-year-race population is linked to Census through an undercount-adjusted normal model, to PEP through cumulative annual-change variance, and to ACS through a multivariate normal model for overlapping 5-year estimates; uncertainty is therefore source-specific rather than pooled into a single denominator variance (Peterson et al., 2021).

The same logic is sharpened in the BSBE opioid-mortality model. ACS denominators use reported margins of error via

$sd=\frac{MOE}{1.645},$

PEP uses latent spatially structured variance, and WorldPop uses that local variance plus an additional source-level component. The simulation study showed that accounting for population error “does not greatly impact estimates of the covariate coefficients, but is consequential in the estimation of smoothed relative risk estimates,” with local-risk MAE improving for PEP and WorldPop denominators under the BSBE formulation (Peterson et al., 2023).

Spatial population products provide a direct error-assessment perspective. For deprived-community polygons in metropolitan São Paulo, WorldPop estimates were “less than 20% off for 67% of the polygons” and the overall error for the full study area was “only -5.9%,” whereas LandScan produced satisfactory results for very few polygons and was non-applicable for most of them because coarse $1\text{ km}$ pixels interact badly with small, irregular polygons (Mattos et al., 2020). In satellite-only high-resolution mapping, POPCORN was reported to produce Rwanda maps at “100m GSD based on less than 400 regional census counts”; in Kigali, those maps reached an $R^2$ score of $66\%$ with “an average error of only about 10 inhabitants/ha” (Metzger et al., 2023). These examples show that, in spatial applications, population-error may be geometric, source-specific, or model-based, but it is always tied to how population is represented in space.

3. Population distributions, transportability, and aggregation error

Population-error also denotes the gap between performance in one population and performance in another. In transported prediction, the paper’s key message is that “population error is a population-specific functional,” because performance measures such as MSE, absolute error, and Brier score average conditional loss over the covariate distribution of the deployment population (Steingrimsson et al., 2021). Under conditional outcome-distribution invariance and positivity, target-population MSE is identifiable even without target outcomes, but the paper is explicit that source-sample performance “generally does not equal target-population performance” and that model and tuning-parameter selection should optimize target, not source, risk (Steingrimsson et al., 2021).

The notion of a “prediction error modifier” makes the mechanism explicit. If

$\E[(Y-g_{\widehat\beta}(X))^2\mid Z=z,S=1]$

varies with $f(X\mid S=0)$0, then any shift in the distribution of $f(X\mid S=0)$1 between source and target populations changes the target error. The simulations make this concrete: source MSE severely underestimated target MSE, while inverse-odds weighted estimation of target MSE was approximately unbiased (Steingrimsson et al., 2021). Population-error here is therefore not measurement error in $f(X\mid S=0)$2 or $f(X\mid S=0)$3, but error induced by deploying a model in the wrong population measure.

Spatial aggregation theory reaches an analogous conclusion from the opposite direction. The proposed sampling-frame model treats the full population distribution as random and distinguishes aggregation-weight error from fine-scale variation and finite population variation. The paper reports that “undercoverage/overcoverage depends arbitrarily on the aggregation grid resolution” for the traditional approach, whereas the proposed approach exhibits low sensitivity, and that differences between the two increase as the population of an area decreases (Paige et al., 2022). This suggests that population-error is often an error in the integration measure itself: not merely in the response surface $f(X\mid S=0)$4, but in the population distribution $f(X\mid S=0)$5 with respect to which $f(X\mid S=0)$6 is aggregated.

4. Latent population membership and register observation error

A different branch of the literature treats population-error as misalignment between administrative visibility and true population membership. In the register-based HMM framework, an individual may be truly present yet leave no trace in a given year, generating false negative observation error, or may be truly absent yet appear in a register because of administrative or household-level processes, generating false positive observation error. The model accounts for temporary emigration, an arbitrary number of possibly interacting registers subject to both error types, and observation probabilities that vary with individual characteristics and unobservable heterogeneity (Brown et al., 25 Mar 2026).

The latent-state architecture makes the population-error interpretation explicit. The observable is the vector of register indicators $f(X\mid S=0)$7, while the inferential target is the latent state $f(X\mid S=0)$8, such as alive and present, alive but abroad, or dead. Overcoverage becomes a state-classification problem rather than a deterministic data-cleaning problem. In the Swedish application, the authors refined the state space to distinguish “abroad with known absence” from “abroad with unknown absence,” so overcoverage was identified with an explicit latent state rather than with a heuristic rule (Brown et al., 25 Mar 2026).

Multiple-system estimation with covariates having missing values and measurement error extends the same idea to latent subpopulation size. In the Māori application, the approach was extended to four registers, some individuals had missing ethnicity, some registers covered subsets of the population by design, and the Māori indicator in each register was treated “as a variable measured with error,” with a latent class model embedded in MSE “to estimate the population size of a latent variable, interpreted as the true Māori status” (Heijden et al., 2020). Here population-error is not only failure to observe people, but failure to observe their true subpopulation membership.

5. Mutation–population thresholds in finite evolutionary models

In quasispecies theory, population-error has a distinct meaning: the interaction between mutation pressure and finite population size in maintaining genetic information. In the finite Moran sharp-peak model, the main phase boundary is

$f(X\mid S=0)$9

where $f(X\mid S=1)$0 is the effective genomic mutation parameter and $f(X\mid S=1)$1 is the population-size-per-sequence-length ratio. If $f(X\mid S=1)$2, then the equilibrium master fraction tends to $f(X\mid S=1)$3; if $f(X\mid S=1)$4, it tends to

$f(X\mid S=1)$5

with variance tending to $f(X\mid S=1)$6 in both regimes (Cerf, 2012). The paper’s central claim is that the classical error threshold becomes a joint population-error threshold: low mutation is not sufficient unless the population is large enough.

Cerf’s Wright–Fisher analysis gives the corresponding finite-population phase boundary

$f(X\mid S=1)$7

If $f(X\mid S=1)$8, the stationary master fraction vanishes; if $f(X\mid S=1)$9, it converges to the same sharp-peak fixed point $$p(A)=\sum_{i=1}^{M} \frac{N_i}{N}\frac{Z_i}{N_i}, \qquad b(A)=\sum_{i=1}^{M} Z_i,$$0. The interpretation is that “mutation pressure alone does not determine the error threshold” in a finite population: the metastable master cloud must persist longer than the exponentially hard discovery time needed to regenerate it from the neutral background (Cerf, 2012).

Later work makes the finite-size correction explicit. In the Moran model, the “first term after $$p(A)=\sum_{i=1}^{M} \frac{N_i}{N}\frac{Z_i}{N_i}, \qquad b(A)=\sum_{i=1}^{M} Z_i,$$1” in the error threshold expansion “scales as $$p(A)=\sum_{i=1}^{M} \frac{N_i}{N}\frac{Z_i}{N_i}, \qquad b(A)=\sum_{i=1}^{M} Z_i,$$2” (Berger, 2019). A separate extension replaces the classical exact-copy threshold with a bounded-error threshold,

$$p(A)=\sum_{i=1}^{M} \frac{N_i}{N}\frac{Z_i}{N_i}, \qquad b(A)=\sum_{i=1}^{M} Z_i,$$3

for replication schemes in which lethality occurs only at the $$p(A)=\sum_{i=1}^{M} \frac{N_i}{N}\frac{Z_i}{N_i}, \qquad b(A)=\sum_{i=1}^{M} Z_i,$$4-th copying error (Velten et al., 2024). Taken together, these results show that in evolutionary dynamics population-error is a threshold phenomenon: mutation load, tolerated error count, and finite population scale jointly determine whether population-level information can be maintained.

6. Population-weighted type I error and finite-population testing

In clinical trials with multiple overlapping target populations, population-error becomes a patient-weighted type I error criterion. The population-wise error rate is defined as “the probability that a randomly selected, future patient will be exposed to an inefficient treatment,” equivalently a prevalence-weighted average of stratum-specific family-wise error rates. Formally,

$$p(A)=\sum_{i=1}^{M} \frac{N_i}{N}\frac{Z_i}{N_i}, \qquad b(A)=\sum_{i=1}^{M} Z_i,$$5

so the error criterion is indexed by the population prevalences $$p(A)=\sum_{i=1}^{M} \frac{N_i}{N}\frac{Z_i}{N_i}, \qquad b(A)=\sum_{i=1}^{M} Z_i,$$6 rather than solely by the hypothesis family (Luschei et al., 2023).

Because the prevalences are usually unknown, the same literature studies plug-in and predictive corrections. Estimating $$p(A)=\sum_{i=1}^{M} \frac{N_i}{N}\frac{Z_i}{N_i}, \qquad b(A)=\sum_{i=1}^{M} Z_i,$$7 by the multinomial MLE $$p(A)=\sum_{i=1}^{M} \frac{N_i}{N}\frac{Z_i}{N_i}, \qquad b(A)=\sum_{i=1}^{M} Z_i,$$8 did not “substantially inflate the true PWER” in the studied settings; the expected PWER was “almost perfectly controlled,” while study-specific values conditioned on realized subgroup sizes varied within a narrow range for moderate to large sample sizes (Luschei et al., 2023). A subsequent paper derived an asymptotic prediction interval for the resulting true PWER via the delta method, with

$$p(A)=\sum_{i=1}^{M} \frac{N_i}{N}\frac{Z_i}{N_i}, \qquad b(A)=\sum_{i=1}^{M} Z_i,$$9

and prediction interval

$\omega_{c,a}=\mu_{c,a}+\log(R)+\log(\gamma_{c,a}),$0

thereby treating the realized true PWER as a random variable induced by uncertainty in population composition (Luschei et al., 6 Feb 2026).

Finite-population trial inference pushes this population weighting one step further. When participants are viewed as “a unique, finite population,” the relevant error criterion is finite-population type I error under the actual randomization design. Under Complete Randomization, RBI and ANOVA are asymptotically equivalent for the difference-in-means test, but “randomization restrictions strongly impacted ANOVA Type I error, even for trials with 1,000 participants.” Adjusting for block randomization corrected type I error, whereas a corresponding correction for MTI designs was left open; RBI, by contrast, directly accounts for the restrictions and ensures correct finite-population type I error (Chipman et al., 8 Oct 2025). This closes the circle: in the most literal sense, population-error may be the error rate attached to the finite population itself rather than to an abstract superpopulation.