Table 1.

Major types and sources of dose estimation error in low-dose radioepidemiologic studies^*

Source	Description	Potential effect on dose-response	Common correction
All
Classical measurement error	Random (sampling) error in dose measurements, eg, the random error from an imprecise measurement device such as a film dosimeter	Loss of power. The effect depends on the error magnitude. A nondifferential classical error generally weakens the dose-response with linear risk coefficients biased towards zero (10, 11).	Dose-response models can be adjusted for random error using regression calibration or maximum likelihood methods (11).
Berkson measurement error	Error that occurs when the mean for a group is substituted for the individual dose within the group, eg, the use of a single factor to convert “recorded” external doses to organ doses results in Berkson error.	Loss of power. The effect depends on the error magnitude. Generally, Berkson error results in very little bias (12).	There is little need for adjustment.
Shared error	When there is error in a group mean assigned to all of the individuals in a group, or in a parameter used to calculate a quantity common to a group, this error is “shared” among those individuals.	Biased risk estimation (due to misspecification of group mean values and understatement of uncertainty due to correlated [nonindependent] dosimetry errors)	Dose error adjustments based on complex Monte Carlo simulation techniques are being applied in some analyses. These methods reduce bias and increase confidence interval width to reflect the correlated dose errors.
Differential and nondifferential error	Dose estimation error that is independent of case status (and other predictors) is said to be nondifferential. Sources of differential error include data collection bias from different recalls of cases compared to controls (ie, recall bias) or from selective data gathering by the exposure assessor (ie, observer or interviewer bias).	Nondifferential error commonly results in bias toward a null association; however, there are examples of nondifferential error in polytomous and continuous exposure measures that induce bias away from the null (13–15). Differential error can result in bias in either direction and can lead to spurious associations.	Collect exposure data prior to disease ascertainment or without prior knowledge of the hypothesized association. Keep exposure assessors blinded to case status.
Missing dose	Doses from occupational, environmental, and medical sources that were accrued by study subjects but were not accounted for in dose-response analyses, for example, historic practices of not measuring doses that were thought at the time to be trivial.	The bias can be in either direction, depending on the distribution of unmonitored dose among study participants.	Use evidence-based dose assignments to fill data gaps in exposure histories.
Environmental
Model validation	Indirectly obtained dose estimates should undergo some model validation. Model validation is the process of establishing the quality of dose estimates of being logically or factually sound (ie, the extent to which the estimate describes the true dose that is being measured).	The weight of evidence from a study relying on models is less if those models have not been validated.	Conduct intercomparisons of estimated from measurements and other dosimetry systems.
Group heterogeneity	Variance in true dose within the group of individuals assigned group-level estimates	Loss of power. The magnitude of variance depends on the resolution of the measurement and the homogeneity of the group (see Berkson error).	Identify similarly exposed groups.
Occupancy	Incomplete information on location and/or time when dose is estimated based on the product of the duration of time exposed and mean dose at a known location	Loss of power. Bias in either direction depending on the treatment of unmeasured exposure.
Medical
Missing data on patients	Lack of information on the physical characteristics of the patients (height and weight) with age often used as a surrogate	Unshared uncertainty	Impute height and weight based on growth curves could be performed. However, imputation strategy should avoid introducing biased values and thus systematic errors to the whole group.
Missing data on protocols	Lack of detailed information on specific protocol implemented for each individual examination. Typically, protocols are followed in ways that vary among individuals.	Shared uncertainties on imputed values with some uncertain individual variability	Impute values based on typical protocols implemented in the hospital are usually performed. However, imputation strategy should avoid introducing biased values and, thus, systematic errors to the whole group.
Uncertainty in the model and/or in phantom measurements	Uncertainties associated with on-phantom measurements, prediction equation parameters, and model specification	Shared uncertainty	Validate models with additional measurements.
Occupational
Unadjusted dose	Recorded doses that poorly estimate the absorbed dose to target tissues, which is the preferred quantity for use in dose-response analyses (16). This is an example of shared error.	For most exposure situations, using unadjusted doses in risk models will underestimate the dose-response association.	Adjust recorded dose to account for exposure geometry and incident radiation energies that differ from the calibration and dose quantitation protocols used (17–20).
Below detection limit (BDL) doses	Inaccurate estimates of dose resulting from treatment of exposures at levels below the minimum detection level of the instrument. The potential for this error is greatest in doses accrued prior to the 1960s, when dosimeters were least sensitive and weekly or biweekly monitoring was routine.	The bias can be in either direction, depending primarily on the exposure distribution, the BDL dose value assigned, and the variance in the measured exposure due to random measurement error (21).	Adjust recorded dose by substitution, multiple imputation, or other means to account for BDL doses (22–24).
Notional dose	Dose assigned to a worker’s dose record to account for exposures that were not quantified. These assignments were often based on a maximum allowable dose to prevent exceeding limits from subsequent exposure. The effects are likely minimized if assignments were realistic exposure scenarios or used data from measurements made in similar time and place.	The bias can be in either direction depending primarily on the distribution of notional dose among study participants.	Replace notional dose assignments with evidence-based dose estimates (25, 26).

^*Sources that are perceived to contribute substantially to dose estimation errors in study categories. Although listed for a single category, a source may contribute to errors in multiple categories.