If you don't remember your password, you can reset it by entering your email address and clicking the Reset Password button. You will then receive an email that contains a secure link for resetting your password
If the address matches a valid account an email will be sent to __email__ with instructions for resetting your password
Address correspondence to: Barbara Bruemmer, PhD, RD, Graduate Program in Nutritional Sciences and Graduate Coordinated Program in Dietetics, University of Washington, 305 Raitt Hall, Box 353410, Seattle, WA 98195
The use of epidemiologic research designs and analytical methods is common in dietetics research. Food and nutrition professionals who seek to perform evidence-based practice or participate in research design, analysis, and communication need skills in the essentials of epidemiology. This is one of a series of monographs on research methodology that addresses these needs and supports the goals of the Board of Editors of the Journal of the American Dietetic Association to further enhance competency and skills. This monograph focuses on statistical approaches for univariate analyses used with the primary observational study designs associated with epidemiology. Tables illustrating the presentation and interpretation of these results are included.
This monograph continues the series on research design and analysis (
) as “the study of the distribution and determinants of health-related states or events in specified populations and the application of this study to control health problems.” Epidemiologic methods have been refined and used in genetic epidemiology (
). The target audience for this monograph is investigators who wish to strengthen skills or require review of the evaluation of evidenced-based practice. The business of research is dynamic, competitive, and costly. Efficiency of study resources should be targeted to the most powerful tools and current techniques. Only by ensuring the use of appropriate tools will the broader goal of health improvement be realized. Epidemiologic methods can support a wide range of research situations from descriptive studies to randomized controlled trials. Three key elements involved in epidemiologic research are common terms (Figure 1), a conceptual framework (Figure 2), and concepts of causal association. The elements of causal associations set forth by Hill (
) has further contributed to our understanding of these elements noting that the original tenets included “specificity,” which states that one cause leads only to a single effect; a concept which clearly has evolved.
Epidemiologic Study Designs
An investigator begins the research process with development of the hypothesis that includes a conceptual causal factor and an outcome of interest. Once that has been decided the appropriate study design is selected (
). A cross-sectional study does not allow examining a sequence of events but rather examines associations only at one point in time. Advantages of this approach include that it is typically less expensive and study participants are generally able to respond about current habits or status. However a major limitation is that, as in the example provided below, a cross-sectional design cannot provide insight into a temporal sequence leading to evidence on a causal factor. In other words, cross-sectional studies cannot provide answers to cause and effect.
Conversely, a second common way to pose a research question is to focus on a known outcome or endpoint, such as cardiovascular disease, nutrient deficiency, or breast cancer, and to explore the antecedents that may have contributed to the outcome. When an investigator is looking for an exposure to a known outcome, a case-control study may be appropriate (
). This type of retrospective study may be efficient in that cases and controls usually can be easily identified in a sample from their outcome status. Then questionnaires can be used to obtain information on subject characteristics and factors that may have influenced the etiology of the outcome at some specified time in the past. Food frequency questionnaires have often been used to assess diet in case-control studies where investigators are seeking information on past behaviors. A unique challenge in this study design is to appropriately target the time period in the past that represents the plausible biologic time period of exposure. In cancer studies for example, investigators must know if the hypothesized exposure is relevant in an early phase of cancer development (eg, initiation, or at a later phase, promotion). The study participant must be instructed to focus on a time frame for the usual intake relative to that period, such as 5 years before the interview or between 20 and 25 years of age. Clearly this is a cognitive challenge for individuals. It may also open the study to the phenomenon of recall bias where cases are motivated to search their memories in more detail because they are dealing with the psychological burden of disease. Because controls do not have the same motivation there may be bias in the responses from the two study groups. These tools were considered acceptable due to their relatively lower cost and previous reports of relative validity compared to food records. However, studies have recently documented that there are many limitations in assessing prior diet and the validity of food frequency questionnaires has recently been reevaluated using more objective biomarkers for comparison (
). The exposure may be any measureable event or physiological condition, such as weight gain, diet quality, or lack of policies that support breastfeeding, which may all lead to specific subsequent outcomes. For example, research questions such as: What are the consequences of weight gain?, Is a diet high in dark green leafy vegetables protective against hypertension?, and How long will a working woman breastfeed if she is not provided with a suitable place to pump breast milk? all apply to this situation. Each of these questions has identified a factor that can be temporally followed for an association with specified outcomes. The main advantage of the prospective cohort study is that multiple outcomes may be assessed over time. For example, many of the large national cohorts such as the Framingham Heart Study (
), and others have gathered baseline data that may be used to evaluate not only the initial hypothesized outcome but other health events or emerging health issues. The disadvantages are generally related to the cost of maintaining contact with a large number of study participants, the inefficiency of gathering information on individuals who may not contribute information on a rare outcome, and the limitations of our current diet assessment methods.
A variant of a cohort study is a nested case-control study that limits subject selection to members of a preexisting cohort. This is a very efficient design because exposure information, such as biological samples may be gathered at baseline but may only be evaluated for those subjects who are later selected to participate in the case-control study based on disease or disease-free status. This is generally only a small subset of the larger cohort so the cost of analysis of the exposure information is reduced. An example would be a large cohort study where blood samples were collected at baseline and subjects were followed forward for 10 years. Subjects who developed cancer during the 10 years could be compared to a group of disease-free controls sampled from the cohort. A comparison of baseline blood samples between these cases and the controls would be much more efficient than analyzing the blood samples of all members of the cohort.
Issues in Study Validity
The methods or design used in any study may influence the internal validity or external validity of the study. The internal validity of a study means that the study was able to address the stated hypothesis because the methods were free of bias. The outcome must be evaluated without consideration of the exposure classification which would compromise the internal validity by “stacking the deck” and making it less likely that the analysis of the data will provide insight into the true relationship of interest. Factors that may introduce bias include selection bias or differential recruitment of study participants into the case vs control groups. Nonresponse bias may occur if survey responders and nonresponders differ systematically and the rate of nonresponse differs for those who have the disease/condition being analyzed vs those who do not. Recall bias, as noted previously in regard to food frequency questionnaires, compromises internal validity. Bias due to misclassification occurs when exposure information has been measured or recorded with error.
It is important to distinguish between nondifferential misclassification, which indicates the error in exposure measurement occurs with the same pattern among those with and without the outcome, and differential misclassification where the error in exposure measurement differs between those with and without the outcome. Nondifferential misclassification happens often in nutrition research when subjects are unable to accurately report usual dietary intake. If both the cases and the controls similarly misjudge their intake the result is a diminished ability to find an association when one exists or an attenuation of the results. Differential misclassification is more serious because the bias may actually contribute to a false association. For example, in a study on total energy intake and blood glucose, if a site with a high proportion of subjects with the metabolic syndrome has a higher rate of error in dietary records than a site with a low proportion of subjects with the metabolic syndrome, the pattern of error may contribute to a finding of an association where one does not exist. This type of bias could even alter the direction of the finding suggesting a false association is protective or could be a risk factor.
The presence of one or more confounding factors is also a threat to internal validity. Confounding factors, confounders, covariates, or adjustment variables are all terms that refer to extraneous factors that may influence the association between the exposure and the outcome. If confounders are not identified or appropriately controlled for, the conclusions of the study may be flawed. A false association may be suggested between the exposure and the outcome that was due to the presence of the confounder. For example, if an investigator is interested in the association between body weight and mortality it would be important to identify subjects with cancer because this potential confounding factor is related to both body weight and mortality. Confounders may be identified through a number of methods such as a review of the literature, a review of pilot data, or a theoretical framework or conceptual model. Additional information on the potential for biased estimates due to inappropriate methods in the selection of confounding factors is provided by Harrell (
). In particular, the investigator will want to identify factors that are strongly associated with the outcome of interest. Another criterion for selection would be a factor that is associated with both the exposure and the outcome though the estimated strength may not be strong for either alone. The initial data analysis will provide the investigator with an opportunity to examine the relationship between the exposure and outcome with and without the confounder present. However, at the design stage, it is important to be inclusive and gather as much information on confounders as feasible. Typical confounders may be age, sex, education, income, and health behaviors such as tobacco use and exercise (
External validity refers to the extent to which the results from the study can be accurately generalized to the larger target population of interest. External validity may be compromised when study participants are not recruited from a representative sample of at-risk individuals. If the study participants are not representative of the target population the results will not be generalizable to that population. Data on response rates to questionnaires, participation rates, and loss to follow-up must be carefully gathered and analyzed. Many of the methods of analysis that are used in epidemiology may not be valid when studies are flawed by high rates of attrition or low response rates. The methods section should provide adequate information on how the exposure, outcome, and confounders were identified; how the information was captured; and how variables were categorized.
Preparation for Data Analysis
Beyond establishing the foundation for the research by stating a hypothesis, selecting the appropriate study design, and adhering to guidelines on methodology, the quality of an epidemiologic study also depends on having a well-considered plan for systematically gathering the data and subsequently effectively managing this information. Planning for analysis early in the execution of the project will assure a smooth transition from data gathering, through editing and into analysis and specific to randomized controlled trials is considered essential to drawing valid conclusions. Even simple decisions such as the use of a relational database in place of a spreadsheet may provide flexibility that speeds the data editing phase but also provides an increased level of security. A well-planned study design and implementation should also minimize missing data that can lead to bias.
Collecting as much raw data as possible will allow for flexibility in establishing categories during analysis. For example, if age is initially recorded as a continuous variable, it can be categorized as a dichotomous variable, such as <30 compared to ≥30 years of age, or in tertiles, quartiles, or other proportions during the analysis phase. Often dichotomous distributions provide limited insight into the trend across categories and therefore may not be suitable (
). Proportional distributions, such as percentiles, may be determined based on all subjects, or based on only the controls or only the cases. Other categories appropriate for age are biological ranges such as premenarche, adult, postmenopausal; or chronologic age such as decades, age 20 to 29 years, or age 30 to 39 years. For body mass index (BMI) common approaches may capture the Centers for Disease Control and Prevention guidelines of 18.5 to 24.9 as desirable; 25 to 29.9 as overweight, and ≥30 as obese (
). The investigator may determine that individuals below a certain BMI may not be appropriate as a reference group. The decision to drop this group or to establish a reference group in the mid-range of BMIs should be based on the hypothesis of interest and scientific evidence. The number of subjects in each category level may differ between the cases and the controls but highly uneven categories may limit the statistical power of the analysis.
The following is a brief description of the main univariate statistical tests that are used in epidemiologic studies. Multivariate methods will be described in a future monograph or the reader may wish to consult advanced texts (
). Each of these methods is a variation on a summary statement of the relationship between the exposure and the outcome. Outcome measures in epidemiologic studies are often presented as binary variables (ie, can potentially take on only two values) (
). This monograph presents the concepts of relative risks (RR) and odds ratios (OR) using binary examples. In particular, these measures tell the reader whether:
The exposure is positively associated with the likelihood of having the outcome (a RR or an OR >1),
The exposure is negatively associated with the likelihood of having an outcome (a RR or an OR <1), and
How much of a change relative to unity (where unity=1) was found.
For example, a point estimate of 1.4 strictly indicates that the risk of developing the outcome among those in the exposed group was 1.4 times the risk for those in the unexposed group. A summary statement may read, “There was a positive association between the exposed group compared to the unexposed group on the outcome of interest (OR 1.4).”
An investigator should pose the hypothesis with the potential direction of the association clearly identified. Then in the data gathering stage the information should be coded to mirror the question (Figure 4). A new investigator may lack an appreciation of how easily the results may be misinterpreted if the coding is not considered (
). When the results of a study are in conflict with the a priori hypothesis and/or the evidence in the literature, the first step should be a reconfirmation that the results are being interpreted correctly.
The RR estimate is used in prospective studies. It represents the magnitude of an association between an exposure and the likelihood of developing an outcome in the exposed group relative to the unexposed group (
). RR estimates are used in prospective studies under two options. The first option is a study that has counts of subjects according to their exposure and outcome status. The count of the outcome is generally for incident cases meaning the first time event during the follow-up period. The second option is a study that uses person time in the denominator. It is an estimate of the incidence of the outcome in the exposed group divided by that group's cumulative time of the study (eg, in person-days) compared to the incidence in the unexposed group during that group's cumulative time in the study. The RR as described above would be a crude estimate or univariate analysis because no confounders have been considered in the analysis. Prospective studies often employ more sophisticated models that allow for adjustment of confounders and inclusion of survival information (
A variant of the RR approach is used for cross-sectional and case-control studies where person time under a theoretical exposure period is not available. The OR is the odds of having the exposure among those with the outcome (cases) divided by the odds of having the exposure among those without the outcome. The OR is considered an estimate of the RR in case-control studies where the study design is built on the outcome rather than the exposure and it is a meaningful substitute for a RR. Additional information on the interpretation of binary outcomes from case-control studies has been offered by Kreamer (
) with suggestions on the translation of effect size to clinical practice.
The full description of the potential range for the estimate of the association is conveyed in the confidence interval. The confidence interval provides the information on the results of the statistical test that may either be statistically significant (otherwise known as not due to chance) or not statistically significant (due to chance). In other words, the data do not allow us to rule out the hypothesis that there is no true relationship between the exposure and the outcome. This interval represents the range of values that likely contain the true RR or OR value in the population as a whole and therefore should always be included when reporting a RR or OR. According to classical statistical theory the width of the interval is established by the α level, which is the probability of a type I error. A type I error occurs when the true value of the RR or OR falls outside of the estimated confidence interval. By convention the α level for many two-sided tests is .05 and the confidence interval is set at (1−α) or 95%. The second characteristic of a confidence interval for an OR or RR is the orientation of this range relative to unity or one. If the confidence interval includes one then the conclusion would be that the investigator cannot rule out the possibility that the relationship between the exposure and outcome estimated in the study is due to chance and thus the estimated relationship would be considered not statistically significant (Figure 5). For other types of statistical tests the relevant comparison would be whether the confidence interval includes zero. The last characteristic is the width of the range which is an indicator of precision or the likelihood of a type II error. There is more confidence that the point estimate is true when there are more subjects in an analysis. If a RR point estimate is 1.4 and the 95% confidence interval is 1.38 to 1.41, this very narrow range suggests a great deal of precision in placing the point estimate at 1.4. But if the confidence interval is 0.36 to 12.81 that point estimate of 1.4 looks quite imprecise and there is little assurance that the true association can be estimated at 1.4. A small sample size will often reduce the confidence and widen the interval. As the width of the confidence interval increases, the lack of precision will affect the ability to reject the null hypothesis because the range is more likely to include one. For information on alternate methods to the classical statistical approach, including a Bayesian framework for interval estimation and interpretation, see Iverson (
Following are three scenarios to model public health, clinical, and wellness hypotheses that may support further illustration of epidemiologic methods. (Note: all scenarios are developed with fictitious values for illustration purposes only.)
Cross-sectional Study: Scenario One
An investigator wishes to examine the association between the presence of a high-fructose corn syrup beverage in a household and the prevalence of obesity.
An example of a cross-sectional study addressing this hypothesis is provided in Figure 6. Hypothetical data examine 600 households where an audit was conducted for the presence of high-fructose corn syrup beverage in the home and reference subject's height and weight were measured. The generic two-by-two table for comparison of an exposure and an outcome is presented in Figure 7, panel A. An analysis of the hypothetical data is presented in Figure 7, panel B. An additional analysis illustrates the information that is obtained by examining the exposure in four levels compared to two levels (dichotomous) (Figure 7, panel C). As the results indicate, an association was demonstrated between the highest category and the reference group. However, if the analysis was limited to only a dichotomous variable, the shape of the association would not be visible at the lower two levels of intake.
The crude OR generated from a statistical program using a cross-tabular analysis by the Mantel-Haenszel method would yield an OR of 2.33 (95% confidence interval 1.64 to 3.29) for the hypothetical data. For this crude analysis without adjustment for confounding factors this estimated association is the same as that generated by the hand calculation of the OR shown in Figure 7, panel B. The summary statement of the results of this analysis would then be “There was a positive association between the presence of ≥7 high-fructose corn syrup beverage items in the home compared to the reference category of <7 on the prevalence of obesity, OR 1.90 (95% confidence interval 1.34 to 2.69).” A template for the presentation of results can be found in Figure 8.
Case-Control Study: Scenario 2
An investigator wishes to examine the association between the intake of fruit and vegetables and the incidence of hypertension.
The second example (Figure 9) uses the research question of scenario two to demonstrate a case-control design analyzed with an OR. Eligibility for the study is a first diagnosis or otherwise known as incident hypertension within the past 6 months compared to a control group with no current or previous diagnosis of hypertension. The results generated from a statistical program using a cross-tabular analysis would yield an OR of 0.48 (95% confidence interval 0.32 to 0.72), which again would be the same as the manual calculation because there were no confounding factors included in the analysis (see the Table). Following the template for the summary statement the results would be presented as, “There was a negative association between hypertension and the intake of fruits and vegetables at ≥5 servings per day compared to the reference category of <5 servings per day, OR 0.48 (95% confidence interval 0.32-0.72).” This can also worded as a protective effect of the exposure (fruits and vegetable consumption) on the outcome (hypertension).
TableCalculation of a crude odds ratio from a case-control study (from Scenario 2: The association between the intake of fruit and vegetables and the incidence of hypertension)
An investigator wishes to examine the association between obesity and the incidence of type 2 diabetes in subjects followed for 10 years.
Scenario three is a prospective cohort study of the association between obesity at baseline and the incidence of type 2 diabetes in subjects followed for 10 years comparing BMI ≥30 to the reference category of BMI <30 (Figure 10). A hypothetical database has been created with 600 subjects. Variables for the calculation of the RR include: the exposure (BMI in two categories <30 and ≥30), and the outcome—the incidence of type 2 diabetes (yes/no). There was a positive association between a BMI ≥30 compared to the reference group BMI <30 and the incidence of type 2 diabetes, RR 1.22 (95% confidence interval 1.10-1.36) (Figure 11).
This monograph has briefly presented essential principles, typical scenarios, hypothetical results, and areas for additional training in the use of epidemiologic techniques common to nutrition research. Appreciating the vitality of a well-defined hypothesis, selecting the optimal study design, appropriately gathering and analyzing the outcomes of interest, taking into account confounding variables, and communicating the oral and written findings of these scientific efforts is paramount to advancing the field and practice of nutrition. It is important to highlight that a thorough understanding of these principles before conducting rigorous nutrition-related research is fundamental. Readers of this monograph are encouraged to build relationships with a multidisciplinary team, including investigators in other areas of nutrition research, biostatisticians, and epidemiologists to strengthen skills in this area and contribute to the publication of quality nutrition research.
STATEMENT OF POTENTIAL CONFLICT OF INTEREST: No potential conflict of interest was reported by the authors.
In the October 2009 issue of the Journal, the Review article by Bruemmer and colleagues, “Publishing Nutrition Research: A Review of Epidemiologic Methods,” should have been identified as having been approved for Continuing Professional Education credit. The Continuing Professional Education Learning Need Codes for this article are 6000, 9000, 9010, and 9060. To take the Continuing Professional Education quiz for this article, log in to ADA's Online Business Center at www.eatright.org/obc , click the “Journal Article Quiz” button, click “Additional Journal CPE Articles,” and select this article's title from a list of available quizzes.