Testing departure from additivity in Tukey's model using shrinkage: application to a longitudinal setting

While there has been extensive research developing gene–environment interaction (GEI) methods in case‐control studies, little attention has been given to sparse and efficient modeling of GEI in longitudinal studies. In a two‐way table for GEI with rows and columns as categorical variables, a conventional saturated interaction model involves estimation of a specific parameter for each cell, with constraints ensuring identifiability. The estimates are unbiased but are potentially inefficient because the number of parameters to be estimated can grow quickly with increasing categories of row/column factors. On the other hand, Tukey's one‐degree‐of‐freedom model for non‐additivity treats the interaction term as a scaled product of row and column main effects. Because of the parsimonious form of interaction, the interaction estimate leads to enhanced efficiency, and the corresponding test could lead to increased power. Unfortunately, Tukey's model gives biased estimates and low power if the model is misspecified. When screening multiple GEIs where each genetic and environmental marker may exhibit a distinct interaction pattern, a robust estimator for interaction is important for GEI detection. We propose a shrinkage estimator for interaction effects that combines estimates from both Tukey's and saturated interaction models and use the corresponding Wald test for testing interaction in a longitudinal setting. The proposed estimator is robust to misspecification of interaction structure. We illustrate the proposed methods using two longitudinal studies—the Normative Aging Study and the Multi‐ethnic Study of Atherosclerosis. Copyright © 2014 John Wiley & Sons, Ltd.     

Mixed effect models of longitudinal Alzheimer's disease data: a cautionary note

Longitudinal studies of cognitive function in Alzheimer's disease (AD) patients are powerful tools to better understand the biology and natural history of the disease, but the attributes of the studies that make them valuable also pose special challenges to analysts. A fundamental problem is the accurate measure of time at which cognitive decline begins. Investigators typically use the date of AD diagnosis or the date of enrollment in an AD study. If the rate of cognitive decline is non-linear, variables associated with the time of diagnosis or enrollment might artificially be associated with the rate of decline. Unlike the mixed effects models typically used to analyse cognitive decline, summary measure analyses do not directly compare the rate of decline with time since decline began, and, therefore, are less sensitive to biased measures of time of decline. We simulated trajectories of cognitive decline using the multivariate normal random effect model and tested the ability of the two analytic techniques to discriminate between true and spurious associations. Our analyses suggest summary measure models are less likely to detect spurious associations generated by biased measures of time at which decline begins, and more likely to detect true associations concealed by biased time measurement.     

Hospital visitors as controls in case-control studies.

OBJECTIVE: Selecting controls is one of the most difficult tasks in the design of case-control studies. Hospital controls may be inadequate and random controls drawn from the base population may be unavailable. The aim was to assess the use of hospital visitors as controls in a case-control study on the association of organochlorinated compounds and other risk factors for breast cancer conducted in the main hospital of the "Instituto Nacional de Cancer" - INCA (National Cancer Institute) in Rio de Janeiro (Brazil). METHODS: The study included 177 incident cases and 377 controls recruited among female visitors. Three different models of control group composition were compared: Model 1, with all selected visitors; Model 2, excluding women visiting relatives with breast cancer; and Model 3, excluding all women visiting relatives with any type of cancer. Odds ratios (OR) and 95% confidence intervals were calculated to test the associations. RESULTS: Age-adjusted OR for breast cancer associated with risk factors other than family history of cancer, except smoking and breast size, were similar in the three models. Regarding family history of all cancers, except for breast cancer, there was a decreased risk in Models 1 and 2, while in Model 3 there was an increased risk, but not statistically significant. Family history of breast cancer was a risk factor in Models 2 and 3, but no association was found in Model 1. In multivariate analysis a significant risk of breast cancer was found when there was a family history of breast cancer in Models 2 and 3 but not in Model 1. CONCLUSIONS: These results indicate that while investigating risk factors unrelated to family history of cancer, the use of hospital visitors as controls may be a valid and feasible alternative.     

Defining exposure in case-control studies: a new approach

Many epidemiologists conducting case-control studies choose to dichotomize their exposure data to make the analysis of the data easier and its presentation more straightforward. The choice of a specific rule for dichotomization can have a large effect on the outcome measure, the odds ratio, although this effect is rarely studied. The authors present a graphic approach for exploring this effect, the quantile-quantile (Q-Q) plot. By examining a Q-Q plot, an investigator simultaneously gains information about the distribution of exposures among cases, the distribution of exposures among controls, and odds ratios at all possible cutpoints and their standard errors. In addition, by finding the slope at each point along the Q-Q curve, it is possible to estimate the rate ratios for each possible value of exposure. The authors view the use of the Q-Q plot as an exploratory tool. It enables the investigator to become more familiar with the data and check for irregularities such as outliers, nonlinearities, or nonmonotonic dose-response curves, and idiosyncratic variations of the odds ratio. The authors present an example evaluating the risk of low birth weight as a function of mother's age for Boston births in 1984.     

Practical and analytical aspects of using friend controls in case-control studies: experience from a case-control study of childhood cancer

We report empirical data on the use of friend controls, specifically response rates, case-control concordance and analytical approaches. The data derive from a North American multi-institutional study of childhood cancer that was conducted in 2002-07 and that focused on paternal exposures. Case parents nominated friends as potential controls; up to three controls participated per case. For 137 (69%) of the 199 case families, at least one control parent participated. Of 374 potential controls contacted, 247 (66%) participated. Case fathers with controls were markedly more likely to be non-Hispanic White, college graduates and non-smokers compared with case fathers without controls. Odds ratios adjusted for demographic characteristics were generally similar but occasionally differed between analyses that included only members of matched sets and those that included all participants, i.e., controls and cases with and without controls. For demographic characteristics, simulations demonstrated that the observed concordance of cases and controls within matched sets exceeded that expected under random ascertainment, indicating probable overmatching. However, the observed concordance of smoking and other exposures was similar to the expectation under random ascertainment, suggesting little overmatching on exposures. Although not ideal, friend controls were convenient, had a reasonably high response rate and provided controls closely matched on race/ethnicity, education and age.     

Cancer case-control studies with other cancers as controls

Theoretical considerations concerning the use of other cancer patients as controls in cancer case-control studies are reviewed. Selection bias may be a problem in that some other cancers may be caused by the exposure under study biasing the odds ratio towards unity. Such bias is noted to be greatest with low prevalence exposures associated with high attributable risks for other cancers. However, it may be possible to identify selection bias with other cancer controls using census or other general population data. In addition, using other cancers as controls has important advantages with regard to recall and interviewer bias, which may be of unknown magnitude and direction when using general population controls. A further disadvantage of general population controls is that separate selection of decreased controls should usually be made for deceased cases, whereas a mixture of live and deceased controls can be expected when selecting other cancer patients as controls. Since there are also logistical and cost advantages in using other cancer patients as controls, this study design is likely to be used increasingly in the future, particularly in cancer registry settings.     

Comparison of clinical indicators for performance measurement of health care quality: a cautionary note

The use of clinical performance data is increasing rapidly. Yet, substantial variation exists across indicators designed to measure the same clinical event. We compared indicators from several indicator measurement systems to determine the consistency of results. Five measurement systems with well-defined indicators were selected. They were applied to 24 hospitals. Indicators for mortality from coronary artery bypass graft surgery and mortality in the perioperative period were chosen from these measurement systems. Analyses results and concludes that it is faulty to assume that clinical indicators derived from different measurement systems will give the same rank order. Widespread demand for external release of outcome data from hospitals must be balanced by an educational effort about the factors that influence and potentially confound reported rates.     

Smoking as a confounder in case-control studies of occupational bladder cancer in women.

BACKGROUND: In studies in men, risk estimates on occupation and bladder cancer are distorted by about 10% when not adjusting for smoking. We examined the degree to which occupational risk estimates for bladder cancer in women are confounded by smoking, and the degree of residual confounding by inadequate control of this effect. METHODS: Primary data of 11 case-control studies on occupation and bladder cancer from Denmark, France, Germany, Greece, Italy, and Spain were pooled. Information for smoking and lifetime occupational history for 700 female cases and 2,425 female controls ages 30-79 was abstracted and recoded. Logistic regression was used to calculate odds ratios (OR) by occupation, applying five models which differed in their degree of adjustment for smoking. RESULTS: In major occupational groups, risk estimates were distorted by less than 10% when not adjusting for smoking. A statistically significant excess risk for bladder cancer was found in 13 specific occupations and industries. In most occupations, adjustment for smoking led the ORs towards the null value, but all statistically significant associations were maintained after adjustment. In three occupations (lathe operators, field crop workers, and wood manufacturers), a statistically significant excess risk was masked when not adjusting for smoking. In six occupations, estimates were distorted by more than 10% (-22% up to +40%). In occupations where smoking acted as a positive confounder, the proportion of confounding removed using a dichotomous smoking variable (ever/never) was around 60%. In one occupation (buyers), controlling for smoking status (ever, never) led to over-adjustment, because the percentage of smokers was high but the quantity smoked was low.     

Selection bias in case-control studies using relatives as the controls

Investigators have suggested using relatives of cases as the control group when studying complex diseases thought to have a major genetic component. However, there is a concern about possible bias and we developed a model to examine the possibility of bias in the selection of relatives as the control group. Assuming the exposure-specific risks of disease remain constant over time, the results indicate that even when there is a correlation in the exposure status among relatives, selection of controls from relatives of cases does not, of itself, introduce bias in the estimate of effect.     

Bias in case-control studies

Case-control studies are largely used to explore differences between groups of individuals. They can identify potential risk factors associated with disease, or they can investigate patient behaviour, such as why some people do not attend for services. As such, case-control studies are often used to generate or test hypotheses about causal factors. Nonetheless, bias is always a danger in case-control studies, arising especially from the way in which study samples are selected or from the collection of retrospective data. Confounding also remains a problem. This short paper explores ways in which such flaws can be uncovered in published studies.     

Using collectively-derived standards to evaluate individual performance: a cautionary note on clinical practice guidelines.

The purpose of this work is to demonstrate the problem of evaluating an individual physician's performance relative to practice guidelines which have typically been derived from group consensus or some measure of central tendency. It is argued that when evaluated against a set of criteria derived at the macro-level, an individual physician's performance may justifiably vary due to the patient characteristics or the evolving process of care. It is also argued that it is not necessarily true that costs are reduced when practice variation is reduced. The results indicate that there are cost reduction in areas not targeted by the guidelines, suggesting a possible 'spillover effect' due to the increased vigilance in monitoring provider performance. The results also provide some evidence of increased costs following a reduction in variation. Caution should be exercised when evaluating individual physician performance relative to guidelines established at the aggregate level. Acceptable individual physician performance should be judged within the upper and lower boundaries of the implicit distribution of physicians' performances from which the established guidelines generated.     

Logistic regression in case-control studies: The effect of using independent as dependent variables

In case-control studies, cases are sampled separately from controls. In such studies the primary analysis concerns the estimation of the effect of covariables on being a case or a control. To explore causal pathways, further secondary analysis could concern the relationships among the covariables. In this paper the validity of such secondary analysis is addressed. In particular, the use of multiple logistic regression in case-control studies where the dependent variable is not the case/control indicator is explored. It is shown that only under very restrictive conditions will sample regression coefficients correctly estimate their true value. In many situations, it may be valid to regress one covariable on others in the control group, but not in the case group or the combined sample. This principle is illustrated by a study of sexually transmitted disease in Kenya.     

The use of linear instrumental variables methods in health services research and health economics: a cautionary note

OBJECTIVE: To investigate potential bias in the use of the conventional linear instrumental variables (IV) method for the estimation of causal effects in inherently nonlinear regression settings. DATA SOURCES: Smoking Supplement to the 1979 National Health Interview Survey, National Longitudinal Alcohol Epidemiologic Survey, and simulated data. STUDY DESIGN: Potential bias from the use of the linear IV method in nonlinear models is assessed via simulation studies and real world data analyses in two commonly encountered regression setting: (1) models with a nonnegative outcome (e.g., a count) and a continuous endogenous regressor; and (2) models with a binary outcome and a binary endogenous regressor. PRINCIPAL FINDINGS: The simulation analyses show that substantial bias in the estimation of causal effects can result from applying the conventional IV method in inherently nonlinear regression settings. Moreover, the bias is not attenuated as the sample size increases. This point is further illustrated in the survey data analyses in which IV-based estimates of the relevant causal effects diverge substantially from those obtained with appropriate nonlinear estimation methods. CONCLUSIONS: We offer this research as a cautionary note to those who would opt for the use of linear specifications in inherently nonlinear settings involving endogeneity.     

Interaction between genetic and environmental risk factors for Alzheimer's disease: A reanalysis of case-control studies

To study the interaction among genetic and environmental risk factors, a reanalysis of case-control studies of Alzheimer's disease (AD) was conducted based on the original data of all studies carried out to January 1, 1990. Seven studies were included in the present analysis, comprising a total of 814 AD patients and 894 control subjects. When comparing those with a positive and negative family history of dementia, similar odds ratio were found for late maternal age [1.7; 95% confidence interval (0.6–4.8) vs. 2.0 (1.1–3.5)], head trauma [1.7 (0.7–4.2) vs. 1.9 (1.1–3.2)], and history of depression [2.0 (0.2–19.8) vs. 2.1 (0.8–1.7)]. This suggests a model in which these risk factors increase the risk for AD independent of family history of dementia. Among those with a positive family history of dementia, the odds ratios for family history of Down's syndrome [4.2 (0.9–20.0))] and of Parkinson's disease [3.3 (0.4–28.2)] tended to be higher than among those with a negative family history of dementia [2.6 (0.8–8.5) and 2.4 (0.8–7.0), respectively]. However, for both disorders the difference in odds ratio was not statistically significant. For history of cigarette smoking, there was no association to AD for those with no first degree relatives with dementia and an inverse relation with AD for those with a positive family history. Although in all analyses, family history of dementia remained significantly associated with AD in the absence of other factors, the odds ratio associated with family history of dementia tended to be lower for those with a positive smoking history, particularly for those with two or more affected relatives. These findings suggest that smoking may interact specifically with a genetically determined process. © 1994 Wiley-Liss, Inc.