دانلود رایگان ترجمه مقاله بررسی داده های از دست رفته در مطالعات اپیدمیولوژی مولکولی (نشریه aacrjournals 2011)

رسیدگی به داده های از دست رفته در مطالعات اپیدمیولوژی مولکولی

The Handling of Missing Data in Molecular Epidemiology Studies

Abstract

Molecular epidemiology studies face a missing data problem, as biospecimen or imaging data are often collected on only a proportion of subjects eligible for study. We investigated all molecular epidemiology studies published as Research Articles, Short Communications, or Null Results in Brief in Cancer Epidemiology, Biomarkers & Prevention from January 1, 2009, to March 31, 2010, to characterize the extent that missing data were present and to elucidate how the issue was addressed. Of 278 molecular epidemiology studies assessed, most (95%) had missing data on a key variable (66%) and/or used availability of data (often, but not always the biomarker data) as inclusion criterion for study entry (45%). Despite this, only 10% compared subjects included in the analysis with those excluded from the analysis and 88% with missing data conducted a complete-case analysis, a method known to yield biased and inefficient estimates when the data are not missing completely at random. Our findings provide evidence that missing data methods are underutilized in molecular epidemiology studies, which may deleteriously affect the interpretation of results. We provide practical guidelines for the analysis and interpretation of molecular epidemiology studies with missing data.

1 Introduction

With the advent of new technology to measure biomarkers, studies in molecular epidemiology have become increasingly more common. As a result, many epidemiology studies now collect biospecimens such as blood, buccal, urine, or tissue samples to evaluate biomarkers that may provide insight into the underlying pathogenesis of disease or that may be predictive of prognosis. Imaging studies, such as mammography, positron emission tomography, and functional MRI, are also used to measure relevant biomarkers of disease.

Generally, biospecimens and image-based data are available only for a subset of the subjects in the study, posing a missing data problem. Occasionally, even when samples are available, measurements may be subject to censoring (i.e., partially missing) due to the detection limit of an assay. Missing data methods, however, are not typically being employed. In a 1995 study, Greenland and Finkle (1) discussed the underutilization of missing data methods in epidemiology studies due to their inaccessibility and complexity. Although missing data methods such as imputation are more readily available at present, a recent study by Klebanoff and Cole in 2008 (2) found that less than 2% of articles published in epidemiology journals make use of imputation-based methods. Instead, a common approach is to conduct a complete-case (CC) analysis (1, 2): exclusion of subjects missing data on at least one variable considered in the analysis. Our study characterizes the prevalence of missing data specifically in molecular epidemiology studies and provides an in-depth description of how the issue is addressed.

There are a variety of reasons biomarker data may be missing in molecular epidemiology studies, some of which may be related to the actual values of the biomarkers themselves and/or other variables; these underlying reasons matter. Specifically, CC approaches are statistically valid, that is, they provide unbiased point estimates and CIs that achieve nominal coverage (3), only when data are missing completely at random (MCAR), that is, when missingness is unrelated to observed or unobserved data yielding a study sample that is representative of the larger cohort (3, 4). For example, consider a batch of randomly selected samples for which measurements are not observed because of an instrumentation malfunction, as occurred in the study by Clendenen and colleagues (5); it is reasonable to assume that these data are MCAR. In this case, a CC analysis should not yield biased estimates, although the estimates may suffer from efficiency loss. If missingness is related only to observed variables, the data are considered missing at random (MAR). An example of this may be given by Mavaddat and colleagues (6), who examined the role of common single-nucleotide polymorphisms (SNP) in subtypes of breast cancer. These authors found that those eligible for study without samples for genotyping were more likely to have advanced stage breast cancer (III/IV). In this case, the data may be MAR if, conditional on stage, the probability of missing SNP information is not related to the unobserved SNP values. If, however, the reason for missing data is related to the unobserved values, the data are not missing at random (NMAR). For example, suppose tumor size is measured less frequently on smaller tumors, as in the study described by Gilcrease and colleagues (7), these data would be considered NMAR. CC analyses conducted on data that are not MCAR (i.e., MAR or NMAR) can lead to biased and inefficient estimates.

Often one can infer whether missingness is related to observed variables, as Mavaddat and colleagues (6) conducted in their analysis comparing those included in the analysis with those excluded from the analysis, which may suggest MCAR is not a reasonable assumption for the variable in question. Distinguishing between NMAR and MAR patterns, however, is not feasible without making unjustifiable assumptions, as it is impossible to examine the nature of missingness for data that do not exist. Thus, one may rely on assumptions based on biological, clinical, and epidemiologic understandings.

There are theoretically sound methods for analyzing data that are either MAR or NMAR. For MAR data, likelihood-based methods and standard multiple imputation (MI) are examples of statistically valid approaches. Furthermore, MI is particularly simple to implement and readily available (4). Analogous methods (likelihood-based and MI-based) exist for NMAR data, although they are not as easily accessible and are more complex to implement (4, 8–14). The increase in complexity is due to the need to model the missing data distribution (or missing data mechanism), whereas assuming the data are MAR generally allows one to ignore this aspect. The goals of this article are to characterize the extent that missing data are present in molecular epidemiology studies, to elucidate how the issue is being addressed, and to discuss MI as a possible, practical solution.