Japanese Journal of Clinical Oncology

INTRODUCTION

Molecular genetic technology has become available for examining joint effects of genetic traits and environmental exposures (gene-environment interactions) in the etiology of diseases. Traditional study designs, i.e. cohort study and case-control study, can be and indeed have been applied to detect interactions, but usually their power is not sufficient for detecting interactions (1). A new design, case-only study, has therefore been introduced to improve the power to identify interactions, with the assumption of independence between exposure and genotype (2-4). The logic is clear, but there is still skepticism among clinical researchers and even epidemiologists. This paper demonstrates advantages of case-only study over case-control study, using hypothetical and real reported data. It is hoped that the results will be convincing for clinical researchers, given the advantages of an inductive approach based on examples over a deductive strategy based on mathematics.

MATERIALS AND METHODS

Notations used for definitions of odds ratio (OR) and gene-environment interaction are shown in Table 1. Notations a₁, b₁, c₁ and d₁ are, respectively, numbers for cases without the genotype under study who were not exposed to the environmental factor under study, exposed cases without the genotype, unexposed cases with the genotype and exposed cases with the genotype. Notations a₂, b₂, c₂ and d₂ are numbers for corresponding controls. The relevant ORs are defined as shown in Table 1. Based on a multiplicative model, interaction is defined here as the ratio of the OR for the exposed divided by the OR for the unexposed or the OR of those with the genotype divided by the OR of those without, equal to (a₁d₁/b₁c₁)/(a₂d₂/b₂c₂), i.e. the OR among cases divided by the OR among controls. The latter OR is equal to 1 when exposure and genotype are independent among controls. In that case, the value for the interaction is equal to a₁d₁/b₁c₁. This means that we can obtain an estimate for interaction solely from cases. This is the logic underlying case-only studies (2-4).

Hypothetical data were generated as follows: prevalence of the exposure under study in the population (e) was set at 0.1 or 0.3 and the prevalence of the genotype under study in the population (g) at 0.1, 0.3 or 0.5. The OR for exposure (ORe) was 1 or 2, that for genotype (ORg) was 1 and that for interaction (ORi) was 2. Several sets of subjects were selected for demonstrating ORs and 95% confidence intervals, in such a way that point estimates became close to the above-adopted ORs.

Table 1. Notations used for definitions of odds ratios (OR) and an interaction

Exposure	Genotype	Cases	Controls
No	No	a1	a2
Yes	No	b1	b2
No	Yes	c1	c2
Yes	Yes	d1	d2
OR for exposure:    Among those without genotype a2b1/a1b2     Among those with genotype c2d1/c1d2 OR for genotype:     Among the unexposed a2c1/a1c2     Among the exposed b2d1/b1d2 Interaction:     (c2d1/c1d2)/(a2b1/a1b2)    = (b2d1/b1d2)/(a2c1/a1c2)     = (a1d1/b1c1)/(a2d2/b2c2)

Cancer case-control studies of genetic polymorphism and lifestyle were searched by MEDLINE and four papers that were found to list the numbers of subjects according to polymorphism and smoking habit (5-8) were used as examples. The polymorphism was classified into two categories, when more than two categories were used in the papers. Age and other confounding factors were not adjusted for, because the data were not available. This did not detract from our analysis, because the purpose was not to provide a real estimate of interaction of each study, but to compare estimates between two different study designs.

An unconditional logistic model was used for estimating ORs for exposure, genotype and gene-environment interaction. For case-control studies, two models were adopted, a two-parameter model of exposure and genotype and a three-parameter model of exposure, genotype and interaction. The ORs estimated using the two-parameter case-control design were termed marginal. The analysis was conducted with the SAS Logistic Procedure (9).

RESULTS

DISCUSSION

Epidemiological studies have identified many risk factors for cancers from records and/or questionnaires (10) and in several cases, including smoking, some dietary components, hazardous agents, reproductive history and family history, the link could be appreciated only on the basis of a document-based approach. The next step is to identify those individuals sensitive to environmental factors, i.e. to examine gene-environment interactions. Molecular genetic technology has opened the door for researchers to a new field of research into interactions between genetic traits and exposure, constantly expanding with increase in genes and their polymorphisms identified. New study designs are clearly necessary to cope with this situation.

It is not always necessary to examine the contribution of well established risk factors such as smoking when an interaction is the main focus of study, since this may not help with the elucidation of biological mechanisms and cancer prevention. When associations with exposure or genotype are to be measured separately in the same dataset studied for interaction, an alternative incomplete-data case-control design is applicable, where data on exposure only or genotype only are measured among controls (11).

As shown in this paper, case-only studies provide narrower confidence intervals than their case-control counterparts. The standard errors of log(ORi) in our hypothetical case-only study with 200 cases were 0.52-0.70 times those of a case-control study with the same number of cases and controls. Thus, the case-control study requires twice the number of subjects to obtain half to two thirds of the precision in the estimate. Comparisons between the case-only study with 200 cases and case-control study with 200 cases and 400 controls showed that threefold subjects of the case-control study gave a lower precision in the estimate. Similar findings have been reported with regard to sample size calculation for interaction detection; a case-only study requires fewer cases than the number of cases in the case-control studies with two controls per case (1).

A case-only study provides valid estimates only on the assumption that genotype and exposure are independently distributed in the study population. If they are associated, the odds ratio for interaction derived from a case-only study would be biased depending on the strength of the association. However, the association, e.g. between ALDH2 genotype and alcohol drinking, seems to be the exception rather than the rule. In four studies used as examples in this paper, no significant association was observed between the polymorphism and smoking habit. If an association was found, it would be of major interest in terms of biological mechanism and disease prevention.

In case-control studies concerning genetic traits and diseases, there is uncorrectable confounding due to `population stratification' or `population admixture' (4,12). Some subgroups may have different patterns of genotype frequency, whose members have a high risk condition due to lifestyle and/or other genes. When a genotype frequent in the subgroup is studied in the population including the subgroup, a spurious association may be observed between the genotype and disease in case-control studies. This is analogous to case-control studies without adjustment for age and gender. It is easy to imagine that confounding will occur even if the controls are sampled randomly from persons without the disease under study.

In conclusion, the present investigation demonstrated that case-only study is powerful with regard to the detection of interactions. The approach will become more important as genes and polymorphisms are identified at exponentially increasing speed and it appears particularly suitable for screening purposes. Understanding its advantages and disadvantages is now essential for clinical researchers, epidemiologists and all those health professionals engaged in disease prevention.