As our attention is increasingly focused on the role of diet in chronic, degenerative diseases, epidemiology is becoming central to the field of nutritional sciences. Through epidemiology, diet-disease relationships first observed or hypothesized in the laboratory can be examined at the level of free-living populations and clinically-defined subgroups. Conversely, observed associations between dietary patterns and disease incidence rates across populations can give rise to biological hypotheses of disease causation to be more fully explored in laboratory or clinical settings. As such, nutritional epidemiology is tightly interwoven with other branches of nutrition research. Its growing prominence highlights the need for nutritional scientists to hone their abilities to critically appraise and accurately interpret epidemiological research.

The purpose of this paper is to examine key issues in the interpretation of findings from epidemiologic studies undertaken to investigate relationships between dietary factors and cardiovascular disease and cancers  the chronic diseases of greatest concern currently. A brief review of epidemiologic study designs and issues specific to their application to nutrition questions is presented. The estimation of risk associated with dietary factors is then examined in more depth, and particular issues which arise in the interpretation of epidemiological research findings in nutrition are discussed.

BASIC RESEARCH DESIGNS IN EPIDEMIOLOGY

Experimental epidemiology

Table 1. Calculation of relative risk and odds ratio1 [View Table]

Application of experimental epidemiology to nutrition questions is limited by the difficulty in operating controlled interventions with dietary variables and by the high costs associated with population-based intervention trials aimed at modifying risk of chronic diseases. It is difficult, although not impossible (Burr et al. 1989, Henderson et al. 1990), to effect the sizable and sustained changes in the food intake behaviour of a free-living adult sample which are required for a controlled intervention study. Attempts to study the effects of more moderate changes have sometimes been compromised by unexpected behavioural changes among members of the control group, owing in part to growing public awareness of diet-disease relationships (Burr et al. 1989, Multiple Risk Factor Intervention Trial Research Group, 1982). It is perhaps more logistically feasible to apply experimental study designs to contrast the effects of pharmacologic doses of specific nutrient or food components since placebos can be manufactured for these treatments and exposures can be well controlled. Such studies may lack generalizability and applicability to free-living populations, however, insofar as their relationships to dietary intake patterns are not readily apparent.

Descriptive and analytic epidemiology

The application of case-control study designs to questions of interest to nutritionists is limited by the particular nature of diet-disease relationships. The role of diet in the etiology of chronic diseases is generally believed to be most important before disease onset or in the early stages of the disease  likely before it is clinically observable. The insight to be gained from a cross-sectional comparison of dietary exposures between cases and controls is limited by the possibility that individuals' present dietary patterns are not representative of their patterns during the time period when diet was most important in the disease process. Retrospective case-control studies attempt to overcome this limitation by measuring past diet using food frequency or diet history methods. The accuracy with which individuals can recall past intake patterns is questionable, however (Friedenreich et al. 1992). The presence of disease may also color recollections of past dietary practices by the cases in case-control studies, leading to serious biases in the study results (Coughlin 1990). The refinement of dietary assessment methods to minimize measurement errors associated with the recall of past dietary patterns is an area of ongoing research.

Because dietary patterns are assessed prior to disease onset, cohort study designs can provide important insight into causal relationships between dietary exposures and disease outcomes. However, determination of the relevant time period of dietary assessment in cohort studies of cancer and cardiovascular disease is complicated by the need for a clear conceptualization of the disease process and the role of dietary factors within this process (Sempos et al. 1993). As well, such studies can be very expensive, requiring large sample sizes and lengthy follow-up to study the occurrence of disease states which are relatively rare and/or have lengthy time periods between exposure to causal factors and detection of the disease. Both conditions hold for most cancers. Although cancers are a major cause of death, the malignancy rates for most sites are relatively low, and the time lag between initiation and promotion phases and clinical diagnosis of the disease may be several years (Hebert and Miller, 1988).

THE ESTIMATION OF RISK ASSOCIATED WITH DIETARY FACTORS

Error in the measurement of dietary factors

Error in the measurement of dietary intake is a serious problem in epidemiologic studies because it can hamper the identification of associations between dietary factors and disease occurrence (Beaton et al. 1979 and 1997, Freudenheim and Marshall 1988, Freudenheim et al. 1989, Liu 1988, Liu et al. 1978). Measurement errors can be divided into two general classes: random error and bias or systematic error (Beaton 1994). Error is considered to be random if it occurs in no predictable direction; i.e., intake is as likely to be overestimated as underestimated. This kind of error may be a function of day-to-day fluctuations in intake, noise in the estimation of true intake, or errors in the analysis of food composition. Random error in the classification of subjects according to their usual intakes will bias risk estimates towards the null and thus lessen the likelihood of detecting a significant association between diet and disease.

Bias refers to the systematic under- or over-estimating of intake in an individual or group. It may be a function of the reporting process or a result of errors in the food composition databases used to code intake data (Beaton 1994). One potential source of bias in reported food intakes is social desirability; those affected report intake practices consistent with popular perceptions of health (Hebert et al. 1995, Worsley et al. 1984). A related issue is the apparent tendency of overweight individuals to systematically underreport their food intakes (Bingham et al. 1995, Black et al. 1993, Heitmann and Lissner 1995). As noted earlier, bias in the reporting of dietary practices by individuals who have been diagnosed with cancer or cardiovascular disease is of particular concern in case-control studies (Coughlin 1990). Bias in dietary data can inflate or deflate the relative risk or odds ratio, depending on the direction of the bias and how widespread the problem is within the data. The problem is compounded when the source of the bias in dietary intake data is also related to the disease outcome variable (e.g., in a situation where fat intake is underreported by overweight subjects and being overweight is thought to be a risk factor for the disease).

Although the refinement of dietary assessment methods to reduce measurement errors is ongoing, it is unlikely that measurement error will ever be entirely eliminated from dietary assessment (Beaton 1994, Kushi 1994). However, our understanding of the sources and nature of errors in the measurement of usual intake and the impact of these errors on statistical analyses of diet-disease associations is becoming increasingly sophisticated. Large dietary studies now often include substudies designed to obtain dietary intake estimates among a representative subsample of the population using more accurate dietary assessment methods (methods which are not logistically feasible to employ with the full sample). For example, if dietary intake is being assessed by means of single 24-hour dietary recalls (e.g., in large, cross-sectional population surveys), then multiple 24-hour recalls may be collected from a representative subsample of the study population. The substudy yields an estimate of day-to-day variation in intake (a primary source of measurement error in the survey) which can be used to statistically adjust estimates of associations derived from the main survey (e.g., Liu 1994). The use of substudies for calibration purposes is an important development in dietary assessment methodology, but the absence of a "gold standard" for dietary assessment continues to plague the field. The effectiveness of statistical procedures to minimize error is limited by the fact that estimates of dietary intake derived from substudies also contain measurement errors (Wacholder et al. 1993).

Insufficient variation in intake levels across the study population

In summary, the strength of an observed association between a dietary pattern and health or disease outcome is influenced by study design and measurement issues. The presence of random error in the measurement of intake, for example, is particularly problematic when there is little range in intakes within a study population. Whereas the true relative risk associated with the upper vs. lower ends of a narrow range of intakes is likely small, the observed effect will be even smaller given the attenuation associated with random error in intake measurements (Freudenheim and Marshall 1988). The absence of an observed relationship between diet and disease therefore cannot be construed as evidence of the absence of a relationship until methodologic effects have been ruled out.

THE INTERPRETATION OF EPIDEMIOLOGICAL FINDINGS

Intercorrelation of dietary factors

Analytic and interpretational problems also arise in attempts to separate generic from specific effects of complex food components (e.g., the overall effect of carbohydrate vs. the effects of simple and complex carbohydrates, or the effect of fruits and vegetables vs the effects of specific antioxidants found in some fruits and vegetables) (Ames et al. 1995, Wacholder et al. 1994), and to differentiate the effect of total energy intake from specific macronutrient effects on disease risk (Sempos et al. 1993). A number of analytic approaches have been proposed to control for the effect of energy while examining the effect of a specific macronutrient (Beaton et al. 1997, Howe et al. 1986, Pike et al. 1989). However, comparisons of the results of these approaches suggest that they are not simply alternative approaches to analysis; they represent very different assumptions about the biological relationships being examined (Beaton et al. 1997, Kipnis et al. 1993, Kushi et al. 1992).

In summary, the intercorrelation of dietary factors limits the possibilities for identification of etiologic relationships between dietary factors and disease risks through observational studies. This problem highlights the need for good experimental methods to tease apart the biologic effects of specific, intercorrelated food constituents.

Correlation between nutrition variables and other risk factors

DRAWING CONCLUSIONS ABOUT DIET-DISEASE RELATIONSHIPS

In utilizing epidemiologic research findings, it is important to recognize that measures of association or relative risk are not biological constants. Just as the results of experimental studies are influenced by issues of design and measurement, so too are the results of epidemiologic studies. The ability to detect and accurately characterize the disease risk associated with a particular dietary factor is dependent on the sample size, variation in dietary patterns within the study population, and measurement of and statistical adjustment for confounding factors. It is also dependent on the magnitude and nature of measurement errors present in the estimation of dietary variables as well as other variables considered. Identifying the precise role(s) that dietary factors play in chronic diseases is further complicated by the intercorrelation of dietary components and by the correlation of dietary patterns with other behavioural and environmental factors which may also impart or exacerbate risk of disease.

In nutritional epidemiology, the determination of causal relationships between dietary factors and disease outcomes is generally a matter of inference, requiring the meticulous assembly and thoughtful review of evidence from a wide variety of sources. A number of criteria have been identified for the inference of causality from epidemiologic data (Rothman, 1986; Susser, 1991). To be considered causal, an association between a specific dietary factor and disease occurrence should be observed consistently across a number of population-based studies, conducted by different research groups in different settings. Conclusions cannot generally be drawn from a single study. Different study designs will have different abilities to contribute to our understanding of causality (Susser, 1991). Even when a hypothesized association is subject to rigorous testing through randomized controlled trials, consistent findings from more than one trial are generally required to draw definitive conclusions. Only through the recent culmination of evidence from several clinical trials (Greenberg et al. 1990 and 1994, Hennekens et al. 1996, Omenn et al. 1996, The Alpha-Tocopherol, 1994), for example, has the hypothesis that beta carotene supplementation lowers risk of cancer and cardiovascular disease been finally dismissed (Greenberg and Sporn, 1996).

The interweaving of epidemiologic evidence with insights gained from physiology and molecular biology is particularly important in the elucidation of relationships between dietary factors and chronic diseases (Greenberg and Sporn 1996, Smith and Shipley 1991). In order to infer causality, the association should be biologically plausible and should be supported by experimental research into the disease process. It is also important that exposure to the risk factor clearly precedes the onset or progression of disease. In addition, the argument for causality is strengthened when the associated risk is sizable, the disease outcome is specific to the dietary factor and a biologic gradient (dose-response relationship) is observable.

Despite the difficulties inherent in the determination of causal relationships in nutritional epidemiology, it is not uncommon to find published estimates of the proportion of disease risk within a population which can be attributed to a particular dietary factor or pattern (e.g., Doll and Peto 1981, Hankin et al. 1992, Kune et al. 1992, McGinnis and Foege 1993, Riboli 1992, Willett and Ascherio 1994). Perhaps the most widely cited of these is Doll and Peto's (1981) estimate that 35% of all cancer deaths could be attributed to diet. Most commonly, current estimates are in the form of population attributable risks (PAR), a statistic which combines information about the disease relative risk associated with the factor of interest with an estimate of the prevalence of exposure to that factor within the general population. The term "attributable" conveys the impression of a causal relationship. However, it cannot be assumed that the evidence for a causal relationship has been evaluated and a definitive conclusion reached when a PAR statistic is presented (Brooker et al. 1996, MacMahon and Trichopoulos, 1996). Interpretation of the PAR becomes even more complex when this statistic is applied to one risk factor for a disease of multifactorial etiology (Leviton 1995). The PAR estimate is expressed as a percentage, giving the illusion that the risk attributable to a particular dietary practice is being appraised out of a total of 100%. However, this is not true when the disease in question is multifactorial. It has been shown that the sum of the PAR's for all the known risk factors of a single disease of multifactorial etiology can easily exceed 100% (Brooker et al. 1996, Rothman 1976). Thus, even if the PAR for a particular dietary pattern in relation to colon cancer is greater than 50%, diet may still not be the most important risk factor for that disease! Although PAR estimates appear to offer an easily interpretable means to appraise the public health importance of dietary patterns in relation to chronic diseases, their simplicity is seductive  particularly for non-epidemiologists unaware of the interpretational issues. More cautious interpretation and application of attributable risk estimates is warranted.

Lastly, it must be recognized that the identification of diet-disease associations is only one part of a much bigger puzzle. Dietary patterns exist within a social and material context, one shaped by a variety of sociocultural, political, economic, and geographic factors. If the study of diet-disease associations occurs in abstraction, without examination of the social and material environment within which particular dietary practices arise (either because these factors are not considered relevant or because they are viewed as confounders to be statistically factored out of the final analysis), then the public health utility of study findings will be severely, and unnecessarily limited. Many authors have called for the broadening of traditional models of health and disease to include social and material influences (Krieger 1994, Loomis and Wing 1990, Pearce 1996, Ruzek 1993). Such developments would be important in enhancing our understanding of dietary and other risk factors in relation to major chronic diseases.