Spatial Epi Public Health GIS

Exposure Measurement Error & Uncertainty

This guide is part of the Environmental Exposure & Risk Surface Modeling section. Everything upstream of it produces a predicted exposure. This page is about what happens when that prediction is handed to a health model as though it were a measurement, and what to do instead.

Concept & Epidemiological Alignment

Let XX be the true personal exposure and ZZ the value your pipeline assigned. The relationship between them determines what happens to the estimated health effect, and there are two canonical structures with opposite consequences.

Classical error has Z=X+UZ = X + U, where the assigned value scatters around the truth with independent noise. This is what a monitor reading with instrument error looks like, and what an address-level assignment from a noisy surface approximates. Its effect on a linear model is attenuation: the estimated slope shrinks toward zero by roughly the reliability ratio σX2/(σX2+σU2)\sigma^2_X / (\sigma^2_X + \sigma^2_U). A surface where a third of the assigned variance is error will shrink a true effect by about a third.

Berkson error has X=Z+UX = Z + U, where the true values scatter around the assigned one. This is what area-level assignment produces: everyone in a tract receives the tract mean, and their individual exposures vary around it. Its effect on a linear model is different and much friendlier — the slope estimate stays approximately unbiased, but the standard error grows, so the study loses precision rather than validity.

Most real exposure pipelines produce a mixture. A kriged surface assigned to a rooftop geocode is mostly classical, because the prediction error is independent noise around the truth at that point. A kriged surface averaged over a tract and assigned to everyone in it adds a Berkson component on top. Which dominates determines whether the honest correction is to inflate the effect estimate or only its interval.

Two further structures deserve naming because they break the friendly result. Differential error — assignment quality that differs between cases and non-cases — biases in an unpredictable direction and is a validity problem rather than a precision one. It arises whenever geocode quality correlates with the outcome, which it frequently does through rurality. Spatially correlated error — neighbouring people sharing the same prediction error because they share the same surface neighbourhood — inflates effective sample size if ignored, producing intervals that are too narrow.

Error Structures and Their Consequences

Structure Where it comes from Effect on the estimate Effect on the interval Correction
Classical Point assignment from a noisy surface Attenuates toward null Slightly narrow Regression calibration, SIMEX
Berkson Area-level or buffer-mean assignment Approximately unbiased Widens Multiple imputation from the predictive distribution
Differential Geocode quality related to outcome Bias of unknown direction Unreliable Restrict, or model the mechanism
Spatially correlated Shared surface neighbourhood Usually mild Too narrow Cluster-robust or spatial standard errors
Mixed, most common Real pipelines Attenuation, magnitude uncertain Too narrow Simulation from the full pipeline
Classical and Berkson Error Do Opposite Things Two scatter panels of true exposure against assigned exposure. In the classical panel the assigned values scatter horizontally around each true value, so regressing the outcome on the assigned value flattens the fitted line and the estimated effect shrinks toward zero. In the Berkson panel the true values scatter vertically around each assigned value, which are discrete because everyone in an area shares one, so the fitted line keeps its slope while the residual spread grows and the confidence interval widens. The identity line is drawn in both panels for reference. Which way the scatter runs decides whether your estimate or your interval suffers Classical: assigned scatters around true fitted slope flattens — effect shrinks Berkson: true scatters around assigned slope holds — interval widens

Spatial Data Prerequisites

  • A prediction variance for every assigned exposure, from the surface itself where the method supplies one.
  • A validation set, ideally a subset with a better exposure measure — personal monitoring, a denser local network, a rooftop geocode where most are coarser.
  • Assignment metadata per record: the method used and the positional error assumed, from Assigning Exposure to Residential Locations.
  • The surface’s own cross-validated error, from the design in Cross-Validating an Exposure Surface With Held-Out Monitors.
  • A coordinate reference system is not needed here — this topic operates on the assigned table, not on geometry — but the spatial structure of the error is, so retain an area identifier for clustering.

Production Implementation

The routine below quantifies the attenuation implied by a known error variance and implements two corrections: regression calibration for classical error, and multiple imputation from the predictive distribution for the Berkson and mixed cases.

# Attenuation diagnostics and two corrections for exposure measurement error.
# Pinned: numpy==1.26.4, pandas==2.2.2, statsmodels==0.14.2
import logging

import numpy as np
import pandas as pd
import statsmodels.api as sm

logging.basicConfig(level=logging.INFO, format="%(asctime)s %(levelname)s %(message)s")
log = logging.getLogger("exposure.error")

SEED = 20260918


def reliability_ratio(assigned: np.ndarray, error_var: np.ndarray | float) -> float:
    """Share of assigned-exposure variance that is signal rather than error.

    Under classical error the naive slope is attenuated by roughly this factor,
    so dividing the naive estimate by it gives a first-order correction.
    """
    var_z = float(np.nanvar(assigned, ddof=1))
    var_u = float(np.nanmean(error_var)) if np.ndim(error_var) else float(error_var)
    var_x = max(var_z - var_u, 1e-12)
    lam = var_x / (var_x + var_u)
    log.info("var(assigned)=%.4f var(error)=%.4f reliability=%.3f "
             "=> naive slope attenuated to ~%.0f%% of the truth",
             var_z, var_u, lam, 100.0 * lam)
    return lam


def regression_calibration(df: pd.DataFrame, outcome: str, exposure: str,
                           error_var_col: str, covariates: list[str]) -> dict:
    """First-order classical-error correction: divide the slope by the reliability."""
    d = df.dropna(subset=[outcome, exposure, error_var_col] + covariates)
    X = sm.add_constant(d[[exposure] + covariates])
    naive = sm.OLS(d[outcome], X).fit()
    lam = reliability_ratio(d[exposure].to_numpy(), d[error_var_col].to_numpy())
    beta = float(naive.params[exposure]) / lam
    se = float(naive.bse[exposure]) / lam       # first-order; ignores uncertainty in lam
    log.info("naive beta=%.4f corrected beta=%.4f (reliability=%.3f)",
             float(naive.params[exposure]), beta, lam)
    return {"naive_beta": float(naive.params[exposure]),
            "naive_se": float(naive.bse[exposure]),
            "corrected_beta": beta, "corrected_se": se, "reliability": lam,
            "n": int(len(d))}


def imputation_correction(df: pd.DataFrame, outcome: str, exposure: str,
                          error_var_col: str, covariates: list[str],
                          n_draws: int = 50, seed: int = SEED) -> dict:
    """Draw plausible exposures from the predictive distribution and pool."""
    rng = np.random.default_rng(seed)
    d = df.dropna(subset=[outcome, exposure, error_var_col] + covariates)
    sd = np.sqrt(np.clip(d[error_var_col].to_numpy(float), 0, None))
    betas, vars_ = [], []
    for _ in range(n_draws):
        z = d[exposure].to_numpy(float) + rng.normal(0.0, sd)
        X = sm.add_constant(pd.DataFrame({exposure: z, **{c: d[c] for c in covariates}}))
        fit = sm.OLS(d[outcome].to_numpy(float), X).fit()
        betas.append(float(fit.params[exposure]))
        vars_.append(float(fit.bse[exposure]) ** 2)
    b = float(np.mean(betas))
    within = float(np.mean(vars_))
    between = float(np.var(betas, ddof=1))
    # Rubin's rules: the between-draw variance is the exposure uncertainty.
    total = within + (1.0 + 1.0 / n_draws) * between
    log.info("imputation beta=%.4f se=%.4f (within=%.5f between=%.5f draws=%d)",
             b, np.sqrt(total), within, between, n_draws)
    return {"beta": b, "se": float(np.sqrt(total)), "within_var": within,
            "between_var": between, "n_draws": n_draws, "n": int(len(d))}

The reliability ratio is worth computing even when no correction will be applied, because it converts an abstract concern into a number. A surface whose cross-validated error variance is half the variance of the assigned exposures has a reliability near 0.5, which means the naive estimate is roughly half the true one — a difference that changes whether a finding is reported as substantial or marginal.

The imputation correction is the more honest of the two for real pipelines, because the between-draw variance it produces is an actual propagation of exposure uncertainty rather than a first-order algebraic adjustment. Its assumption is that the predictive distribution is correctly specified, which for a kriged surface with a fitted variance is reasonable and for a land-use regression with no variance at all is impossible — one more reason to prefer methods that return an uncertainty.

Parameter Selection & Tuning

Which error variance to use. The kriging variance describes prediction error at a point; it does not include the error from assigning a residential value to a person who spends eight hours elsewhere. A defensible total combines the surface’s cross-validated error with an allowance for the residential proxy, and the allowance should be stated rather than omitted because it is hard to estimate.

Number of imputation draws. Fifty is generally sufficient for a stable pooled standard error; the between-draw variance is estimated from that many values, so twenty is too few when the exposure uncertainty is large.

Whether to correct at all. A correction that multiplies the estimate by two also multiplies any residual confounding by two. Where the study’s main threat is confounding rather than measurement error, reporting the naive estimate with the reliability ratio stated may be more useful than a corrected number that invites over-interpretation.

Clustering for spatially correlated error. Where many subjects share a surface neighbourhood, standard errors should be clustered on an area identifier or computed with a spatial correction, following the diagnostics in Diagnosing Spatial Autocorrelation in Regression Residuals.

Non-linear models. The attenuation results above are exact for linear models and approximate for logistic and Poisson ones. For a rare outcome the approximation is usually adequate; for a strong effect it is not, and simulation is the reliable route.

Edge Cases & Failure Modes

No variance available. Land-use regression, inverse distance weighting and proximity metrics supply no per-location uncertainty. The options are to derive one from held-out validation error, treat the error as a fixed fraction with a stated sensitivity range, or acknowledge that no correction is possible.

Error correlated with the exposure itself. Surfaces are typically least accurate where concentrations are highest, because those places are near sources with steep gradients. That makes the error heteroscedastic, and a single reliability ratio understates the problem at the top of the exposure range where the dose-response matters most.

Correcting the wrong direction. Applying a classical-error correction to a Berkson-dominated assignment inflates an estimate that was already unbiased. Diagnose the structure before correcting: if everyone in an area shares a value, the dominant structure is Berkson.

Validation sets that are not representative. A subset with personal monitoring is usually volunteers, who differ. Use it to characterise the error structure, not to calibrate a correction factor applied to everyone, unless the representativeness can be argued.

Reporting only the corrected number. The naive estimate, the reliability ratio and the corrected estimate belong together. A corrected figure alone hides a large and assumption-laden adjustment.

Report All Three, Not Just the One You Prefer Three effect estimates with confidence intervals from the same data. The naive estimate using the assigned exposure directly is the smallest, with the narrowest interval. The regression-calibrated estimate is larger, scaled up by the inverse reliability ratio, with an interval scaled by the same factor. The multiple imputation estimate is similar in magnitude but has a visibly wider interval because the between-draw variance from the exposure uncertainty has entered the standard error. A note states that the widest interval is the one that has accounted for the most, not the weakest result. The correction moves the estimate and the interval, and both matter no effect naive reliability 0.58 regression calibration scaled by 1/0.58 multiple imputation 50 draws The widest interval has accounted for the most, not concluded the least

The reliability ratio converts all of this into one number that bounds how far a naive estimate could be from the truth:

How Far a Naive Estimate Can Fall Under classical error, the naive slope is the true slope multiplied by the reliability ratio, so the relationship is a straight diagonal line. At a reliability of ninety per cent the naive estimate retains ninety per cent of the true effect. At fifty per cent it retains half. At thirty per cent it retains under a third, enough to turn a meaningful effect into an apparent null. Three typical exposure pipelines are marked along the line to show where they usually fall. Reliability is the share of the effect a naive estimate keeps dense network, rooftop geocodes typical state kriging sparse network, street-level geocodes 0 0.5 1.0 reliability ratio of the assigned exposure share of true effect kept

Compliance & Audit Controls

Record the error variance source and how it was derived, whether the dominant structure was judged classical, Berkson or mixed and on what evidence, the reliability ratio, the correction method with its parameters and seed, the naive and corrected estimates with intervals, and the clustering used for spatially correlated error.

Two reporting practices make the analysis defensible. State the assumed error structure explicitly in the methods, because a reader cannot evaluate a correction without it. And publish the naive estimate alongside the corrected one, so the size of the adjustment is visible. A study reporting only a corrected effect has combined a measurement and a modelling assumption into one number without showing the split.

Production Implementation Checklist

Frequently Asked Questions

How do I know whether my error is classical or Berkson? Look at how the value was assigned. If each person got the surface value at their own location, the error is mostly classical — the assigned value scatters around their truth. If everyone in an area got the area’s value, it is mostly Berkson — their truths scatter around the shared assignment. Buffer means sit between the two, and the buffer radius relative to the exposure’s spatial scale determines where.

Is it always right to correct for attenuation? No. Correction scales up the estimate and everything correlated with it, including uncontrolled confounding. Where confounding is the dominant threat, reporting the naive estimate with an explicit reliability ratio is often more informative than a corrected number that reads as more certain than it is.

What if my exposure model gives no uncertainty at all? Derive one from held-out validation error — the cross-validated root mean squared error squared is a usable estimate of the error variance, though it describes an average rather than a per-location value. Failing that, run the analysis across a stated range of assumed reliabilities and report how the conclusion changes.

Does measurement error explain a null result? It can, and the reliability ratio bounds how much. If reliability is 0.9, attenuation cannot turn a large effect into a null one; if it is 0.3, it easily can. Compute the number before attributing a null to anything.

Why does the corrected interval get wider under imputation but not under regression calibration? Because they account for different things. The first-order calibration scales both estimate and standard error by the same factor and does not add the uncertainty of the exposure itself. Imputation samples the exposure repeatedly, so the variation between draws — the genuine uncertainty about what each person was exposed to — enters the pooled standard error.