Aliza K. Fink

Semi-automated Sensitivity Analysis to Assess Systematic Errors in Observational Data

Background Published epidemiologic research usually provides a quantitative assessment of random error for effect estimates, but no quantitative assessment of systematic error. Sensitivity analysis can provide such an assessment. Methods We describe a method to reconstruct epidemiologic data, accounting for biases, and to display the results of repeated reconstructions as an assessment of error. We illustrate with a study of the effect of less-than-definitive therapy on breast cancer mortality. Results We developed SAS code to reconstruct the data that would have been observed had a set of systematic errors been absent, and to convey the results. After 4,000 reconstructions of the example data, we obtained a median estimate of relative hazard equal to 1.5 with a 95% simulation interval of 0.8-2.8. The relative hazard obtained by conventional analysis equaled 2.0, with a 95% confidence interval of 1.2-3.4. Conclusions Our method of sensitivity analysis can be used to quantify the systematic error for an estimate of effect and to describe that error in figures, tables, or text. In the example, the sources of error biased the conventional relative hazard away from the null, and that error was not accurately communicated by the conventional confidence interval.

A null association between smoking during pregnancy and breast cancer using Massachusetts registry data (United States)

Objective: An earlier investigation reported a five-fold increase in breast cancer risk among women who smoked during pregnancy. Using a similar design, we re-examined this hypothesis. Methods: The source population comprised Massachusetts residents who gave birth between 1987 and 1999 with a birth record in the Massachusetts Vital Statistics Registry. Cases were diagnosed with breast cancer between 1988 and 2000 at ages 25–55 with a record in the Massachusetts Cancer Registry. Three controls were matched to each case on maternal age, year of giving birth, and birth facility. Information on smoking, the matched factors, and potential confounders were collected from the birth certificate. The data were analyzed using conditional logistic regression. Results: After adjusting for potential confounders, women who smoked during pregnancy did not have an increased risk of breast cancer compared to women who did not smoke during pregnancy (relative risk = 1.0, 95% Confidence interval CI = 0.81–1.2). We observed no dose response relation between number of cigarettes smoked per day during pregnancy and breast cancer risk. There was no evidence that our results were biased by misclassification from women inaccurately reporting their smoking status. Conclusion: In contrast to the previous study, we did not observe an increased risk of breast cancer in women who smoked during pregnancy.

The Relationship Among Physicians’ Specialty, Perceptions of the Risks and Benefits of Adjuvant Tamoxifen Therapy, and Its Recommendation in Older Patients With Breast Cancer

Objectives:The objectives of this study were to determine whether tamoxifen recommendation differs by physician specialty, to determine whether perception affects tamoxifen recommendation, and to investigate the association between the physicians specialty and the perception of risks and benefits of tamoxifen. Methods:We enrolled a cohort of geographically diverse women aged 65 and older with stage I through IIIa breast cancer in a prospective cohort study. We recruited their surgeons and, when applicable, their medical oncologists to provide patient-specific information about their perceptions of the risks and benefits of tamoxifen and whether they recommended tamoxifen. Each physician also completed a questionnaire regarding his or her demographic and practice characteristics. Patient data were collected through medical record review and a patient interview 3 months after definitive breast cancer surgery. Results:We collected physician treatment recommendation forms for 585 women. Oncologists were 2.5 times more likely to recommend tamoxifen, compared with surgeons, after adjusting for patient and tumor characteristics (95% confidence interval, 1.5–4.2). For both specialties, their perceptions of the risks and benefits of tamoxifen were strong predictors of tamoxifen recommendation. However, there were differences in perception by physician specialty. Distant metastases and tolerance of tamoxifen side effects were more important to oncologists, whereas local recurrence and risk of cataracts were more important to surgeons. Conclusion:Physicians’ perceptions of the risks and benefits of tamoxifen therapy for older women are important in their decision-making process.

Null association between pregnancy termination and breast cancer in a registry-based study of parous women

Studies suggesting a positive association between pregnancy termination and breast cancer risk have often been of retrospective case‐control design, so subject to selection and recall biases. We undertook a registry‐based analysis with minimal selection bias and prospective record‐based ascertainment of terminations. The source population comprised Massachusetts women with a record of giving birth between 1987 and 1999 in the Massachusetts Vital Statistics Registry. Primary breast cancer cases were 25–55 years old at diagnosis between 1988 and 2000 and had a record of the diagnosis in the Massachusetts Cancer Registry. We matched 3 controls to each case on maternal age, year of giving birth and birth facility. Information on terminations (induced and spontaneous) before the birth of record, the matched factors and potential confounders were collected from the birth certificate. After adjustment for the matched factors, age, parity and maternal and paternal education, the odds ratio associating any termination history with breast cancer risk equaled 0.91 (95% CI = 0.79–1.05). The marginally protective adjusted odds ratio largely derived from a protective effect among women with parity equaled to 1 (OR for any termination = 0.68; 95% CI = 0.45–1.03), suggesting a protective effect of terminated pregnancy among women with one live birth.

Predictors and outcomes of surgeons' referral of older breast cancer patients to medical oncologists.

Older women are less likely than younger women to receive definitive care for a new diagnosis of breast cancer, but the reasons are not well understood. Although coordination of referral among specialists is an important component of quality of care, it has not been studied as a factor that contributes to observed age‐related variations in breast cancer care.

Probabilistic Bias Analysis

To this point we have considered situations in which the bias parameters for a bias analysis are known with certainty or modeled as if they are known with certainty (i.e., simple bias analysis, see Chaps. 4– 6). We have also considered models of the impact of combinations of bias parameters on an observed estimate of association (i.e., multidimensional bias analysis, see Chap. 7). Simple bias analysis is an improvement over conventional analyses, which implicitly assume that all the bias parameters are fixed at values that confer no bias. However, the usefulness of simple bias analysis is limited by its assumption that the bias parameters are known without error, a situation that is rarely, if ever, a reality. Multidimensional bias analysis improves on simple bias analysis by examining the impact of more than one set of bias parameters, but even this approach only examines the bias conferred by a limited set of bias parameters. For any analysis, many other possible combinations of bias parameters are plausible, and a multidimensional analysis will not describe the impact of these possibilities. More important, multidimensional analysis gives no sense of which corrected estimate of association is the most likely under the assumed bias model, which can make interpretation of the results challenging.

A Guide to Implementing Quantitative Bias Analysis

Estimates of association from nonrandomized epidemiologic studies are susceptible to two types of error: random error and systematic error. Random error, or sampling error, is often called chance, and decreases toward zero as the sample size increases and the data are more efficiently distributed in the categories of the adjustment variables. The amount of random error in an estimate of association is measured by its precision. Systematic error, often called bias, does not decrease toward zero as the sample size increases or with more efficient distributions in the categories of the analytic variables. The amount of systematic error in an estimate of association is measured by its validity.

Unmeasured and Unknown Confounders

Confounding occurs when the effect of the exposure of interest mixes with the effects of other variables that are causes of the exposure or that share common causal ancestors with the exposure (Kleinbaum et al.,1982). Understanding and adjusting for confounding in epidemiologic research is central to addressing whether an observed association is indeed causal. It is imperative to control confounding because it can make an association appear greater or smaller than it truly is, and can even reverse the apparent direction of an association. Confounding can also make a null effect (i.e., no causal relation between the exposure and the disease) appear either causal or preventive. For a covariate variable to induce confounding, there must be a relation between both the exposure and the covariate in the source population and between the covariate and the disease (among those unexposed). In addition, the covariate must not be affected by the exposure. In nonrandomized epidemiologic studies, confounding can be controlled by design or in the analysis, but only for known and measured confounders.

Data Sources for Bias Analysis

All bias analyses modify a conventional estimate of association to account for bias introduced by systematic error. These quantitative modifications revise the conventional estimate of association (e.g., a risk difference or a rate ratio) with equations that adjust it for the estimated impact of the systematic error. These equations have parameters, called bias parameters, that ultimately determine the direction and magnitude of the adjustment. For example: The proportions of all eligible subjects who participate in a study, simultaneously stratified into subgroups of persons with and without the disease outcome and within categories of the exposure variable of interest, are bias parameters. These parameters determine the direction and magnitude of selection bias. The sensitivity and specificity of exposure classification, within subgroups of persons with and without the disease outcome of interest, are bias parameters that affect the direction and magnitude of bias introduced by exposure misclassification. The strength of association between an unmeasured confounder and the exposure of interest and between the unmeasured confounder and the disease outcome of interest are bias parameters that affect the direction and magnitude of bias introduced by an unmeasured confounder.

Multidimensional Bias Analysis

The preceding three chapters have described the techniques for conducting simple bias analysis to assess errors caused by selection bias, residual confounding, or misclassification. However, simple bias analysis implies that the researcher has one and only one estimate to assign to each of the values for the error model’s bias parameters. In many situations, that is not the case. There are many bias parameters for which validation data do not exist, so the values assigned to the bias parameter are educated guesses. In this situation, the analyst is better served by making more than one educated guess for each value and then combining values in different sets. In other situations, multiple different measures of the bias parameter may exist, and there may be no basis for the analyst to select just one as the best estimate of the truth from among those available. For example, when both internal and external validation studies have been conducted, or there were multiple external estimates each in populations slightly different to the one under study, the analyst has no basis to select one value for the bias parameter over another. Frequently, internal estimates are more useful than external estimates because they derive from the same source population as yielded the study’s estimate of association. If there is the possibility of selection bias into the internal validation study, however, then it is possible that the subjects included in the validation study do not provide a good estimate of the bias parameter in the remainder of the study population. In this situation, the analyst may want to use values informed by all of the available validation studies as independent estimates.