Power and sample size calculations for generalized regression models with covariate measurement error

Covariate measurement error is often a feature of scientific data used for regression modelling. The consequences of such errors include a loss of power of tests of significance for the regression parameters corresponding to the true covariates. Power and sample size calculations that ignore covariate measurement error tend to overestimate power and underestimate the actual sample size required to achieve a desired power. In this paper we derive a novel measurement error corrected power function for generalized linear models using a generalized score test based on quasi-likelihood methods. Our power function is flexible in that it is adaptable to designs with a discrete or continuous scalar covariate (exposure) that can be measured with or without error, allows for additional confounding variables and applies to a broad class of generalized regression and measurement error models. A program is described that provides sample size or power for a continuous exposure with a normal measurement error model and a single normal confounder variable in logistic regression. We demonstrate the improved properties of our power calculations with simulations and numerical studies. An example is given from an ongoing study of cancer and exposure to arsenic as measured by toenail concentrations and tap water samples.     

There is no impact of exposure measurement error on latency estimation in linear models

Identification of the latency period for the effect of a time-varying exposure is key when assessing many environmental, nutritional, and behavioral risk factors. A pre-specified exposure metric involving an unknown latency parameter is often used in the statistical model for the exposure-disease relationship. Likelihood-based methods have been developed to estimate this latency parameter for generalized linear models but do not exist for scenarios where the exposure is measured with error, as is usually the case. Here, we explore the performance of naive estimators for both the latency parameter and the regression coefficients, which ignore exposure measurement error, assuming a linear measurement error model. We prove that, in many scenarios under this general measurement error setting, the least squares estimator for the latency parameter remains consistent, while the regression coefficient estimates are inconsistent as has previously been found in standard measurement error models where the primary disease model does not involve a latency parameter. Conditions under which this result holds are generalized to a wide class of covariance structures and mean functions. The findings are illustrated in a study of body mass index in relation to physical activity in the Health Professionals Follow-Up Study.     

Performance of bias‐correction methods for exposure measurement error using repeated measurements with and without missing data

It is known that measurement error leads to bias in assessing exposure effects, which can however, be corrected if independent replicates are available. For expensive replicates, two‐stage (2S) studies that produce data ‘missing by design’, may be preferred over a single‐stage (1S) study, because in the second stage, measurement of replicates is restricted to a sample of first‐stage subjects. Motivated by an occupational study on the acute effect of carbon black exposure on respiratory morbidity, we compare the performance of several bias‐correction methods for both designs in a simulation study: an instrumental variable method (EVROS IV) based on grouping strategies, which had been recommended especially when measurement error is large, the regression calibration and the simulation extrapolation methods. For the 2S design, either the problem of ‘missing’ data was ignored or the ‘missing’ data were imputed using multiple imputations. Both in 1S and 2S designs, in the case of small or moderate measurement error, regression calibration was shown to be the preferred approach in terms of root mean square error. For 2S designs, regression calibration as implemented by Stata software is not recommended in contrast to our implementation of this method; the ‘problematic’ implementation of regression calibration although substantially improved with use of multiple imputations. The EVROS IV method, under a good/fairly good grouping, outperforms the regression calibration approach in both design scenarios when exposure mismeasurement is severe. Both in 1S and 2S designs with moderate or large measurement error, simulation extrapolation severely failed to correct for bias. Copyright © 2012 John Wiley & Sons, Ltd.     

Using measurement error models to assess effects of prenatal and postnatal methylmercury exposure in the Seychelles Child Development Study

Studies of the effects of environmental exposures on human health typically require estimation of both exposure and outcome. Standard methods for the assessment of the association between exposure and outcome include multiple linear regression analysis, which assumes that the outcome variable is observed with error, while the levels of exposure and other explanatory variables are measured with complete accuracy, so that there is no deviation of the measured from the actual value. The term measurement error in this discussion refers to the difference between the actual or true level and the value that is actually observed. In the investigations of the effects of prenatal methylmercury (MeHg) exposure from fish consumption on child development, the only way to obtain a true exposure level (producing the toxic effect) is to ascertain the concentration in fetal brain, which is not possible. As is often the case in studies of environmental exposures, the measured exposure level is a biomarker, such as the average maternal hair level during gestation. Measurement of hair mercury is widely used as a biological indicator for exposure to MeHg and is the only indicator that has been calibrated against the target tissue, the developing brain. Variability between the measured and the true values in explanatory variables in a multiple regression analysis can produce bias, leading to either over or underestimation of regression parameters (slopes). Fortunately, statistical methods known as measurement error models (MEM) are available to account for measurement errors in explanatory variables in multiple regression analysis, and these methods can provide an (either "unbiased" or "bias-corrected") estimate of the unknown outcome/exposure relationship. In this paper, we illustrate MEM analysis by reanalyzing data from the 5.5-year test battery in the Seychelles Child Development Study, a longitudinal study of prenatal exposure to MeHg from maternal consumption of a diet high in fish. The use of the MEM approach was made possible by the existence of independent, calibration data on the magnitude of the variability of the measurement error deviations for the biomarker of prenatal exposure used in this study, the maternal hair level. Our reanalysis indicated that adjustment for measurement errors in explanatory variables had no appreciable effect on the original results.     

The detection of gene-environment interaction for continuous traits: should we deal with measurement error by bigger studies or better measurement?

BACKGROUND: The search for biologically relevant gene-environment interactions has been facilitated by technological advances in genotyping. The design of studies to detect interactions on continuous traits such as blood pressure and insulin sensitivity is attracting increasing attention. We have previously described power calculations for such studies, and this paper describes the extension of those calculations to take account of measurement error. METHODS: The model considered in this paper is a simple linear regression relating a continuous outcome to a continuously distributed exposure variable in which the ratio of slopes for each genotype is considered as the interaction parameter. The classical measurement error model is used to describe the uncertainty in measurement in the outcome and the exposure. The sample size to detect differing magnitudes of interaction with varying frequencies of the minor allele are calculated for a given main effect observed with error both in the exposure and the outcome. The sample size to detect a given interaction for a given minor allele frequency is calculated for differing degrees of measurement error in the assessment of the exposure and the outcome. RESULTS: The required sample size is dependent upon the magnitude of the interaction, the allele frequency and the strength of the association in those with the common allele. As an example, we take the situation in which the effect size in those with the common allele was a quarter of a standard deviation change in the outcome for a standard deviation change in the exposure. If a minor allele with a frequency of 20% leads to a doubling of that effect size, then the sample size is highly dependent upon the precision with which the exposure and outcome are measured. rho(Tx) and rho(Ty) are the correlation between the measured exposure and outcome, respectively and the true value. If poor measures of the exposure and outcome are used, (e.g. rho(Tx) = 0.3, rho(Ty) = 0.4), then a study size of 150 989 people would be required to detect the interaction with 95% power at a significance level of 10(-4). Such an interaction could be detected in study samples of under 10 000 people if more precise measurements of exposure and outcome were made (e.g. rho(Tx) = 0.7, rho(Ty) = 0.7), and possibly in samples of under 5000 if the precision of estimation were enhanced by taking repeated measurements. CONCLUSIONS: The formulae for calculating the sample size required to study the interaction between a continuous exposure and a genetic factor on a continuous outcome variable in the face of measurement error will be of considerable utility in designing studies with appropriate power. These calculations suggest that smaller studies with repeated and more precise measurement of the exposure and outcome will be as powerful as studies even 20 times bigger, which necessarily employ less precise measures because of their size. Even though the cost of genotyping is falling, the magnitude of the effect of measurement error on the power to detect interaction on continuous traits suggests that investment in studies with better measurement may be a more appropriate strategy than attempting to deal with error by increasing sample sizes.     

SEGMENTED REGRESSION WITH ERRORS IN PREDICTORS: SEMI-PARAMETRIC AND PARAMETRIC METHODS

We consider the estimation of parameters in a particular segmented generalized linear model with additive measurement error in predictors, with a focus on linear and logistic regression. In epidemiologic studies segmented regression models often occur as threshold models, where it is assumed that the exposure has no influence on the response up to a possibly unknown threshold. Furthermore, in occupational and environmental studies the exposure typically cannot be measured exactly. Ignoring this measurement error leads to asymptotically biased estimators of the threshold. It is shown that this asymptotic bias is different from that observed for estimating standard generalized linear model parameters in the presence of measurement error, being both larger and in different directions than expected. In most cases considered the threshold is asymptotically underestimated. Two standard general methods for correcting for this bias are considered; regression calibration and simulation extrapolation (simex). In ordinary logistic and linear regression these procedures behave similarly, but in the threshold segmented regression model they operate quite differently. The regression calibration estimator usually has more bias but less variance than the simex estimator. Regression calibration and simex are typically thought of as functional methods, also known as semi-parametric methods, because they make no assumptions about the distribution of the unobservable covariate X. The contrasting structural, parametric maximum likelihood estimate assumes a parametric distributional form for X. In ordinary linear regression there is typically little difference between structural and functional methods. One of the major, surprising findings of our study is that in threshold regression, the functional and structural methods differ substantially in their performance. In one of our simulations, approximately consistent functional estimates can be as much as 25 times more variable than the maximum likelihood estimate for a properly specified parametric model. Structural (parametric) modelling ought not be a neglected tool in measurement error models. An example involving dust concentration and bronchitis in a mechanical engineering plant in Munich is used to illustrate the results. © 1997 by John Wiley & Sons, Ltd.     

A toolkit for measurement error correction,with a focus on nutritional epidemiology

Exposure measurement error is a problem in many epidemiological studies, including those using biomarkers and measures of dietary intake. Measurement error typically results in biased estimates of exposure‐disease associations, the severity and nature of the bias depending on the form of the error. To correct for the effects of measurement error, information additional to the main study data is required. Ideally, this is a validation sample in which the true exposure is observed. However, in many situations, it is not feasible to observe the true exposure, but there may be available one or more repeated exposure measurements, for example, blood pressure or dietary intake recorded at two time points. The aim of this paper is to provide a toolkit for measurement error correction using repeated measurements. We bring together methods covering classical measurement error and several departures from classical error: systematic, heteroscedastic and differential error. The correction methods considered are regression calibration, which is already widely used in the classical error setting, and moment reconstruction and multiple imputation, which are newer approaches with the ability to handle differential error. We emphasize practical application of the methods in nutritional epidemiology and other fields. We primarily consider continuous exposures in the exposure‐outcome model, but we also outline methods for use when continuous exposures are categorized. The methods are illustrated using the data from a study of the association between fibre intake and colorectal cancer, where fibre intake is measured using a diet diary and repeated measures are available for a subset. © 2014 The Authors. Statistics in Medicine Published by John Wiley & Sons, Ltd.     

On the potential of measurement error to induce differential bias on odds ratio estimates: an example from radon epidemiology

It is well established that odds ratios estimated by logistic regression are subject to bias if exposure is measured with error. The dependence of this bias on exposure parameter values, particularly for multiplicative measurement error, and its implications in epidemiology are not, however, as fully acknowledged. We have been motivated by a German West case-control study on lung cancer and residential radon, where restriction to a subgroup exhibiting larger mean and variance of exposure than the entire group has shown higher odds ratio estimates as compared to the full analysis. By means of correction formulae and simulations, we show that bias from additive classical type error depends on the exposure variance, not on the exposure mean, and that bias from multiplicative classical type error depends on the geometric standard deviation (in other words on the coefficient of variation of exposure), but not on the geometric mean of exposure. Bias from additive or multiplicative Berkson type error is independent of exposure distribution parameters. This indicates that there is a potential of differential bias between groups where these parameters vary. Such groups are commonly compared in epidemiology: for example when the results of subgroup analyses are contrasted or meta-analyses are performed. For the German West radon study, we show that the difference of measurement error bias between the subgroup and the entire group exhibits the same direction but not the same dimension as the observed results. Regarding meta-analysis of five European radon studies, we find that a study such as this German study will necessarily result in smaller odds ratio estimates than other studies due to the smaller exposure variance and coefficient of variation of exposure. Therefore, disregard of measurement error can not only lead to biased estimates, but also to inconsistent results and wrongly concluded effect differences between groups.     

Effect of measurement error on epidemiological studies of environmental and occupational exposures

Random error (misclassification) in exposure measurements usually biases a relative risk, regression coefficient, or other effect measure towards the null value (no association). The most important exception is Berkson type error, which causes little or no bias. Berkson type error arises, in particular, due to use of group average exposure in place of individual values. Random error in exposure measurements, Berkson or otherwise, reduces the power of a study, making it more likely that real associations are not detected. Random error in confounding variables compromises the control of their effect, leaving residual confounding. Random error in a variable that modifies the effect of exposure on health--for example, an indicator of susceptibility--tends to diminish the observed modification of effect, but error in the exposure can create a supurious appearance of modification. Methods are available to correct for bias (but not generally power loss) due to measurement error, if information on the magnitude and type of error is available. These methods can be complicated to use, however, and should be used cautiously as "correction" can magnify confounding if it is present.       

Correction of logistic regression relative risk estimates and confidence intervals for systematic within-person measurement error

Errors in the measurement of exposure that are independent of disease status tend to bias relative risk estimates and other measures of effect in epidemiologic studies toward the null value. Two methods are provided to correct relative risk estimates obtained from logistic regression models for measurement errors in continuous exposures within cohort studies that may be due to either random (unbiased) within-person variation or to systematic errors for individual subjects. These methods require a separate validation study to estimate the regression coefficient lambda relating the surrogate measure to true exposure. In the linear approximation method, the true logistic regression coefficient beta* is estimated by beta/lambda, where beta is the observed logistic regression coefficient based on the surrogate measure. In the likelihood approximation method, a second-order Taylor series expansion is used to approximate the logistic function, enabling closed-form likelihood estimation of beta*. Confidence intervals for the corrected relative risks are provided that include a component representing error in the estimation of lambda. Based on simulation studies, both methods perform well for true odds ratios up to 3.0; for higher odds ratios the likelihood approximation method was superior with respect to both bias and coverage probability. An example is provided based on data from a prospective study of dietary fat intake and risk of breast cancer and a validation study of the questionnaire used to assess dietary fat intake.     

A maximum likelihood latent variable regression model for multiple informants

Studies pertaining to childhood psychopathology often incorporate information from multiple sources (or informants). For example, measurement of some factor of particular interest might be collected from parents, teachers as well as the children being studied. We propose a latent variable modeling framework to incorporate multiple informant predictor data. Several related models are presented, and likelihood ratio tests are introduced to formally compare fit. The incorporation of partially observed subjects is addressed under a variety of missing data mechanisms. The methods are motivated by and applied to a study of the association of chronic exposure to violence on asthma in children.     

Regression calibration to correct correlated errors in outcome and exposure

Measurement error arises through a variety of mechanisms. A rich literature exists on the bias introduced by covariate measurement error and on methods of analysis to address this bias. By comparison, less attention has been given to errors in outcome assessment and nonclassical covariate measurement error. We consider an extension of the regression calibration method to settings with errors in a continuous outcome, where the errors may be correlated with prognostic covariates or with covariate measurement error. This method adjusts for the measurement error in the data and can be applied with either a validation subset, on which the true data are also observed (eg, a study audit), or a reliability subset, where a second observation of error prone measurements are available. For each case, we provide conditions under which the proposed method is identifiable and leads to consistent estimates of the regression parameter. When the second measurement on the reliability subset has no error or classical unbiased measurement error, the proposed method is consistent even when the primary outcome and exposures of interest are subject to both systematic and random error. We examine the performance of the method with simulations for a variety of measurement error scenarios and sizes of the reliability subset. We illustrate the method's application using data from the Women's Health Initiative Dietary Modification Trial.     

Correcting for bias in relative risk estimates due to exposure measurement error: a case study of occupational exposure to antineoplastics in pharmacists.

OBJECTIVES: This paper describes 2 statistical methods designed to correct for bias from exposure measurement error in point and interval estimates of relative risk. METHODS: The first method takes the usual point and interval estimates of the log relative risk obtained from logistic regression and corrects them for nondifferential measurement error using an exposure measurement error model estimated from validation data. The second, likelihood-based method fits an arbitrary measurement error model suitable for the data at hand and then derives the model for the outcome of interest. RESULTS: Data from Valanis and colleagues' study of the health effects of antineoplastics exposure among hospital pharmacists were used to estimate the prevalence ratio of fever in the previous 3 months from this exposure. For an interdecile increase in weekly number of drugs mixed, the prevalence ratio, adjusted for confounding, changed from 1.06 to 1.17 (95% confidence interval [CI] = 1.04, 1.26) after correction for exposure measurement error. CONCLUSIONS: Exposure measurement error is often an important source of bias in public health research. Methods are available to correct such biases.     

Estimating sample size for epidemiologic studies: the impact of ignoring exposure measurement uncertainty

Sample size requirements for epidemiologic studies are usually determined on the basis of the desired level of statistical power. Suppose, however, that one is planning a study in which the participants' true exposure levels are unobservable. Instead, the analysis will be based on an imprecise surrogate measure that differs from true exposure by some non-negligible amount of measurement error. Sample size estimates for tests of association between the surrogate exposure measure and the outcome of interest may be misleading if they are based solely on the anticipated characteristics of the distribution of surrogate measures in the study population. We examine the accuracy of sample size estimates for cohort studies in which a continuous surrogate exposure measure is subject to either classical or Berkson measurement error. In particular, we evaluate the consequences of not adjusting the sample size estimation procedure for tests based on imprecise exposure measurements to account for anticipated differences between the distributions of the true exposure and the surrogate measure in the study population. As expected, failure to adjust for classical measurement error can lead to underestimation of the required sample size. Disregard of Berkson measurement error, however, can result in sample size estimates that exceed the actual number of participants required for tests of association between the outcome and the surrogate exposure measure. We illustrate this Berkson error effect by estimating sample size for a hypothetical cohort study that examines an association between childhood exposure to radioiodine and the development of thyroid neoplasms. © 1998 John Wiley & Sons, Ltd.     

How independent are "independent" effects? Relative risk estimation when correlated exposures are measured imprecisely

A relative risk estimate which relates an exposure to risk of disease will tend to be estimated too close to unity if that exposure is subject to random measurement error or intra-subject variability. "Independent" relative risk estimates, for the effect of one exposure after adjusting for confounding exposures, may be biased in either direction, depending on the amount of measurement imprecision in the exposure of interest and in the confounders. We describe two methods which estimate the bias in multivariate relative risk estimates due to the effect of measurement imprecision in one or more of the exposure variables in the model. Results from the two methods are compared in an example involving HDL cholesterol, triglycerides and coronary heart disease. In this example, the degree of bias in relative risk estimates is shown to be highly dependent on the amount of measurement imprecision ascribed to the exposures. It is concluded that when two exposures are substantially correlated, and one or both is subject to sizeable measurement imprecision, a study in which exposures are measured only once will be inadequate for investigating the independent effect of the exposures. Where feasible, epidemiologists should seek study populations where the correlation between the exposures is smaller.     

Mediation analysis when a continuous mediator is measured with error and the outcome follows a generalized linear model

Mediation analysis is a popular approach to examine the extent to which the effect of an exposure on an outcome is through an intermediate variable (mediator) and the extent to which the effect is direct. When the mediator is mis‐measured, the validity of mediation analysis can be severely undermined. In this paper, we first study the bias of classical, non‐differential measurement error on a continuous mediator in the estimation of direct and indirect causal effects in generalized linear models when the outcome is either continuous or discrete and exposure–mediator interaction may be present. Our theoretical results as well as a numerical study demonstrate that in the presence of non‐linearities, the bias of naive estimators for direct and indirect effects that ignore measurement error can take unintuitive directions. We then develop methods to correct for measurement error. Three correction approaches using method of moments, regression calibration, and SIMEX are compared. We apply the proposed method to the Massachusetts General Hospital lung cancer study to evaluate the effect of genetic variants mediated through smoking on lung cancer risk. Copyright © 2014 John Wiley & Sons, Ltd.     

STRATOS guidance document on measurement error and misclassification of variables in observational epidemiology: Part 1—Basic theory and simple methods of adjustment

Measurement error and misclassification of variables frequently occur in epidemiology and involve variables important to public health. Their presence can impact strongly on results of statistical analyses involving such variables. However, investigators commonly fail to pay attention to biases resulting from such mismeasurement. We provide, in two parts, an overview of the types of error that occur, their impacts on analytic results, and statistical methods to mitigate the biases that they cause. In this first part, we review different types of measurement error and misclassification, emphasizing the classical, linear, and Berkson models, and on the concepts of nondifferential and differential error. We describe the impacts of these types of error in covariates and in outcome variables on various analyses, including estimation and testing in regression models and estimating distributions. We outline types of ancillary studies required to provide information about such errors and discuss the implications of covariate measurement error for study design. Methods for ascertaining sample size requirements are outlined, both for ancillary studies designed to provide information about measurement error and for main studies where the exposure of interest is measured with error. We describe two of the simpler methods, regression calibration and simulation extrapolation (SIMEX), that adjust for bias in regression coefficients caused by measurement error in continuous covariates, and illustrate their use through examples drawn from the Observing Protein and Energy (OPEN) dietary validation study. Finally, we review software available for implementing these methods. The second part of the article deals with more advanced topics.     

The performance of methods for correcting measurement error in case-control studies

BACKGROUND: It is generally agreed that adjustment for measurement error (when feasible) can substantially increase the validity of epidemiologic analyses. Although a broad variety of methods for measurement error correction has been developed, application in practice is rare. One reason may be that little is known about the robustness of these methods against violations of their restrictive assumptions. METHODS: We carried out a simulation study to assess the performance of two error correction methods (a regression calibration method and a semiparametric approach) as compared with standard analyses without measurement error correction in case-control studies with internal validation data. Performance was assessed over a wide range of model parameters including varying degrees of violations of assumptions. RESULTS: In nearly all the settings assessed, the semiparametric estimate performed better than all alternatives under investigation. The regression calibration method is sensitive to violations of the assumptions of nondifferential error and small error variance. CONCLUSIONS: The semiparametric method is a very robust method to correct for measurement error in case-control studies, but lack of functional software hinders widespread use. If the assumptions for the regression calibration method are fulfilled, application of this method, originally developed for cohort studies, in case-control studies may be a useful alternative that is easy to implement.     

Measurement errors in control risk regression: A comparison of correction techniques

Control risk regression is a diffuse approach for meta-analysis about the effectiveness of a treatment, relating the measure of risk with which the outcome occurs in the treated group to that in the control group. The severity of illness is a source of between-study heterogeneity that can be difficult to measure. It can be approximated by the rate of events in the control group. Since the estimate is a surrogate for the underlying risk, it is prone to measurement error. Correction methods are necessary to provide reliable inference. This article illustrates the extent of measurement error effects under different scenarios, including departures from the classical normality assumption for the control risk distribution. The performance of different measurement error corrections is examined. Attention will be paid to likelihood-based structural methods assuming a distribution for the control risk measure and to functional methods avoiding the assumption, namely, a simulation-based method and two score function methods. Advantages and limits of the approaches are evaluated through simulation. In case of large heterogeneity, structural approaches are preferable to score methods, while score methods perform better for small heterogeneity and small sample size. The simulation-based approach has a satisfactory behavior whichever the examined scenario, with no convergence issues. The methods are applied to a meta-analysis about the association between diabetes and risk of Parkinson disease. The study intends to make researchers aware of the measurement error problem occurring in control risk regression and lead them to the use of appropriate correction techniques to prevent fallacious conclusions.     

A conditional likelihood approach for regression analysis using biomarkers measured with batch‐specific error

Measurement error is common in epidemiological and biomedical studies. When biomarkers are measured in batches or groups, measurement error is potentially correlated within each batch or group. In regression analysis, most existing methods are not applicable in the presence of batch‐specific measurement error in predictors. We propose a robust conditional likelihood approach to account for batch‐specific error in predictors when batch effect is additive and the predominant source of error, which requires no assumptions on the distribution of measurement error. Although a regression model with batch as a categorical covariable yields the same parameter estimates as the proposed conditional likelihood approach for linear regression, this result does not hold in general for all generalized linear models, in particular, logistic regression. Our simulation studies show that the conditional likelihood approach achieves better finite sample performance than the regression calibration approach or a naive approach without adjustment for measurement error. In the case of logistic regression, our proposed approach is shown to also outperform the regression approach with batch as a categorical covariate. In addition, we also examine a ‘hybrid’ approach combining the conditional likelihood method and the regression calibration method, which is shown in simulations to achieve good performance in the presence of both batch‐specific and measurement‐specific errors. We illustrate our method by using data from a colorectal adenoma study. Copyright © 2012 John Wiley & Sons, Ltd.