Case-control studies are retrospective in design and have been referred to as "research in reverse"2; that is, the groups under investigation are defined by the presence or lack of an outcome of interest. When designing a case-control study, investigators first identify patients who have had the outcome of interest; these are the case patients. Investigators should provide explicit eligibility criteria for the case patients, such as age range and geographical location. A clear description of how the investigators established the presence or absence of the outcome being studied is also important (e.g., the clinical symptoms that were present and the diagnostic methods that were used). Selecting case patients from a discrete time period is preferable to selecting them from a broad time span due to the fact that diagnostic methods tend to change over time, possibly resulting in important differences between the case patients from different periods.
The investigators then identify control patients—persons who are reasonably similar to the case patients with respect to important determinants of outcome, such as age, sex, or comorbidity, but who have not experienced the target outcome. Control patients are selected independent of exposure status and represent the population at risk of experiencing the outcome of interest. For example, consider the following scenario: a researcher from a hospital with an outbreak of multidrug-resistant Acinetobacter baumannii wants to design a case-control study to test the hypothesis that pulsatile lavage wound treatment (the exposure) was the mode of transmission for the organism. The case patients would be inpatients who were infected with the organism (the outcome of interest); the control patients would be inpatients who were not infected but who had wound-care consultation. Investigators would then look back in time to assess the relative frequency of exposure to one or more variables in the case patients as well as the control patients, adjusting for differences in the known and suspected prognostic variables (Fig. 1). The collection of exposure data is often dependent on patient interviews. Interviews are typically conducted by study personnel who should be blinded, when possible, to the status of patients (i.e., whether the patient is a case patient or a control patient) so that information is not elicited differently depending on the status of the patient. If blinding of assessors is not possible, investigators should avoid revealing the study hypothesis. The case-control design permits the simultaneous exploration of multiple exposures that have a possible association with the outcome of interest. These data then allow for the calculation of a measure of association between the exposure(s) and the target outcome.
As with all study designs, case-control trials have benefits and limitations that impact the strength of the inferences that can be drawn from the results of the study (Table I). Due to their retrospective nature, case-control studies allow for the study of rare outcomes or outcomes that occur long after an exposure. The association between multiple exposures and an outcome of interest can be evaluated in a case-control study, and control patients can be matched with potentially confounding variables. Case-control studies are also relatively quick and inexpensive to conduct. As a general rule, however, cohort designs are more efficient when the prevalence of the outcome of interest is higher than the prevalence of exposure. For example, a cohort study would be the design of choice, as compared with a case-control study, to explore the association between pelvic irradiation (the exposure) and risk of pelvic fracture in older women (the outcome of interest). A cohort study can be prospective or retrospective and involves the exploration of a particular outcome in groups of subjects who are alike in most ways but who differ according to a certain characteristic or exposure.
Many case-control studies rely on patient interviews to elicit details on exposure and history and thus are susceptible to recall bias. Medical records that are used to validate exposure or outcome data may be lacking or incomplete. Selection of an appropriate control group can be challenging, and any control group is at risk for unequal distribution of prognostic variables as compared with a group of case patients, which can lead to spurious associations. For example, an influential case-control study published in the New England Journal of Medicine found evidence to suggest that higher coffee intake was strongly associated with the development of pancreatic cancer, leading investigators to conclude: "…coffee use might account for a substantial proportion of the cases of this disease in the United States."3 However, the investigators selected their control group from the practices of physicians who were caring for the case patients (patients with pancreatic cancer), and these control patients had a high rate of gastrointestinal disorders—some of which were exacerbated by coffee. A number of the control patients had learned to avoid coffee, whereas the patients with pancreatic cancer consumed coffee at a rate similar to population averages; as a result, a positive association was found between higher coffee intake and development of pancreatic cancer. A number of subsequent studies, with more appropriate control groups, refuted this association4,5.
Case-control studies can only explore one outcome of interest (such as a disease); therefore verification of exposure information (such as average coffee consumption) is often difficult. The case-control study design precludes calculation of the prevalence of an outcome of interest because investigators purposefully choose both the case patients and the control patients. Studies that make use of a case-control design should describe any association between the exposure and outcome event in terms of an odds ratio; however, if the event rate is low enough in both the exposed and unexposed groups (less than 5%), then the odds ratio can be considered to be an approximation of the risk ratio6. The same design features that make a case-control study ideal for exploring certain research questions also introduce susceptibility to bias, and interpretation of the results may therefore be challenging.
Example 1: Studying Outcomes that Occur Over Long Periods of Time
Some ecological studies have suggested that there is a positive association between the concentration of fluoride in drinking water and the incidence of hip fracture7,8. An ecological study is an observational study that is defined by the level at which data are analyzed, namely, at the population or group level rather than at the individual level. Hip fractures occur primarily in older adults, and the substantial length of time from when individuals first begin drinking fluoridated water to when they experience a hip fracture argues against the use of either a cohort study or a randomized controlled trial.
Hillier and colleagues used a case-control design to explore the association between exposure to fluoride through drinking water and the risk of hip fracture9. They found that, for hip fracture, the odds ratio associated with an average lifetime exposure to =0.9 ppm of fluoride was 1.0 (95% confidence interval, 0.7 to 1.5), which equates to no significant risk. The authors selected their case patients (adults with hip fracture) by identifying, through a search of hospital admission records, all patients fifty years of age or older who presented with newly diagnosed fractures of the femoral neck that were through or proximal to the lesser trochanter and not caused by cancer. The location of recruitment was any one of three hospitals that serviced the county of Cleveland in northeast England, and all reviewed patients presented during the same seventeen-month period. All potentially eligible patients were approached to participate in the study, and those who agreed were required to score six or higher on a Hodkinson abbreviated mental test to ensure they were capable of completing interview questionnaires.
The control group (adults without a hip fracture) was randomly selected from a list of all members of the study population (residents of Cleveland who were fifty years of age or older) who were registered with National Health Service general practitioners and who matched the case patients with regard to age and sex. Those who agreed to participate were also required to score six or higher on a Hodkinson abbreviated mental test. Both the case and control members were interviewed with use of structured questionnaires that asked about demographic variables and lifetime residential history as well as the following potential confounders that were adjusted for in the analysis: height, weight, physical activity, age at menopause, alcohol consumption, smoking history, recent medication (including corticosteroids), and dietary sources of calcium and fluoride.
Residential histories were reviewed by an investigator who was blinded to case or control status, and investigators used information from water suppliers to calculate the average fluoride concentrations that each individual had been exposed to throughout life. To explore the validity of their calculation, the investigators analyzed the fluoride concentration in the excised femoral heads of a subset of 105 patients who had undergone arthroplasty. Samples were analyzed without knowledge of fluoride exposure levels.
Nine hundred and fourteen case patients were identified during the study period, of which 514 (56%) were interviewed. The main reason given for case patients not completing an interview was a low mental test score (195 patients). There were 1196 control patients identified, of which 527 (44%) were interviewed. Reasons for exclusion of control patients were that their family doctor did not provide permission for them to be approached (108), they could not be reached (125), or they refused to participate (377). Data on lifetime fluoride exposure were at least 90% complete for 460 case patients and 423 control patients, and these data were used for the primary analysis.
This case-control study by Hillier et al. had a number of strengths. They clearly defined their case patients, the eligibility criteria, and how subjects were selected. All case patients had newly diagnosed fractures that presented in a relatively narrow time frame. Control patients were representative of the population at risk of becoming case patients, and selection was independent of the exposure being investigated. To provide greater assurance of valid questionnaire completion, all participants were screened for mental competence. The possibility of bias being introduced by information gatherers was reduced by blinding them to the outcome status. Furthermore, multiple known and suspected causes of increased fracture risk were accounted for in the analysis.
Limitations of their study included low participation rates in interviews (approximately 50%), which might have introduced a response bias if the association between fluoride and hip fracture in the individuals who were not interviewed was different from that in the individuals who were interviewed. The authors explored this possibility by performing an unadjusted analysis for fluoride exposure and risk of hip fracture among all participants, whether they were interviewed or not, the results of which were also nonsignificant. Despite some limitations, the overall study was rigorously conducted and the strength of inferences was therefore high.
Example 2: Studying Rare Outcomes
Leaving instruments or sponges inside a surgical patient is a rare event, and the factors that may influence the risk of this error are poorly understood. Gawande et al. recently undertook a case-control study to further investigate this issue10. Fifty-four case patients were identified (patients with retained instruments or sponges following surgery) through a review of all malpractice claims or incident reports filed between January 1, 1985 and January 1, 2001 with a single malpractice insurer who represented one-third of all physicians and twenty-two hospitals in the state of Massachusetts. Two hundred and thirty-five control patients were randomly selected from a group of patients who were identified from a search of hospital records as having undergone the same procedure during the same time period and, when feasible, at the same institution.
Through a review of the literature and interviews with surgeons, the following risk factors were identified and included in the analysis: age; sex; weight; height; the cavity of operation; start time; duration of the operation; the volume of blood loss; the volume of blood transfused; whether the operation was performed on an emergency basis; whether there was a change in or addition to the planned procedure; whether there was involvement of more than one surgical team, more than one procedure, or both; whether there was a complete count of sponges and instruments; whether the nursing personnel changed between counts; and whether the surgeon or another team member performed the closure. All subjective variables were clearly defined and all data were acquired from a review of hospital records. In an adjusted analysis, higher risk of retaining a foreign body was associated with emergency surgery (risk ratio, 8.8; 95% confidence interval, 2.4 to 31.9), unplanned change in the operation (risk ratio, 4.1; 95% confidence interval, 1.4 to 12.4), and higher body mass index (risk ratio for each one-unit increment, 1.1; 95% confidence interval, 1.0 to 1.2).
The investigators clearly defined the target outcome, the eligibility criteria, the method of selecting case patients, and the institutions from which the case patients were acquired. They did collect cases over a sixteen-year time span, but this is not likely to introduce bias because the methods of confirming retained instruments or sponges following surgery were unlikely to have changed during this time. Control patients were, for the most part, representative of the population at risk of becoming case patients, and 83% were matched not only for type of procedure and time period but also for the institution where the procedure occurred. The collection of exposure information may have been subject to bias because of the reliance on hospital records, and the investigators did not specify if the data collectors were blinded to case or control status of participants or to the overall hypothesis of the study. If they were not, data collectors may have been more vigilant in collecting exposure data when reviewing the records of case patients as compared with control patients.
The investigators elected to construct a multivariate regression model for their analysis, which yields adjusted odds ratios for all included independent variables. However, they reported these data as risk ratios, which is a reasonable assumption in that the outcome appears to be rare (<5%) in both the exposed and unexposed groups. Although there are some limitations to this study, the overall methodological conduct appears to have been well done, which strengthens the findings of the investigators.
Example 3: Case-Control Studies within Cohort Trials
The nested case-control or case-cohort design is an economic alternative to the standard cohort design. A case-control study is defined as nested when the case patients and control patients are acquired from a cohort study, either retrospectively or prospectively. The control patients are selected from a population at risk at the time of occurrence of each case that arises in a cohort, which allows for the confounding effect of time in the analysis. A cohort patient who is selected as a control patient at one time point can become a case patient at a later time point, and individuals can act as control patients for more than one case patient. Investigators attempt to match case patients and control patients with respect to known and suspected confounding variables.
When investigators have captured all cases of a disease of interest in a cohort and when control patients have been acquired from a random sample of the entire cohort, the resulting design is a case-cohort study. Unlike the nested case-control design, case patients in the case-cohort study are not matched to individuals in the comparison group on the basis of time or other variables. Advantages of the use of a case-cohort design (over the nested case-control design) are that the control subjects can be selected immediately from the cohort instead of being selected at the time that each case occurs.
With either a nested case-control or case-cohort design, because only a subset of the entire cohort is analyzed, there are substantial advantages in both cost and time. Wacholder and Boivin analyzed a large cohort study that was designed to explore the association between treatments for Hodgkin disease and second cancer occurrences and found that a case-cohort study would have saved 83% of the cost of gathering covariate information with only an 11% loss in statistical efficiency relative to the full cohort study11.
To explore exposures associated with increased risk of venous thromboembolism, Huerta et al. conducted a nested case-control analysis with use of a cohort of 1,856,206 individuals who were derived from the General Practice Research Database (1994 to 2000)12. Their 6,550 cases were patients who were reported to have had a deep vein thrombosis or pulmonary embolism and received anticoagulant therapy, independently of whether a diagnostic procedure was recorded in the database. Of the pool of eligible control patients, 10,000 individuals were randomly sampled and frequency-matched by age (within one year), sex, and calendar year to case patients with deep vein thrombosis and pulmonary embolism.
The General Practice Research Database contained detailed information on potential confounding variables that the authors used in their analyses; these variables were body mass index, cardiovascular disease, asthma, chronic obstructive pulmonary disease, gastrointestinal disease, cancer, renal disease, diabetes, anemia, osteoarthritis, rheumatoid arthritis, pregnancy, surgery, alcohol intake, smoking status, and prescription drug use. In their adjusted analysis, venous thromboembolism was strongly associated with fractures (odds ratio, 21.3; 95% confidence interval, 15.7 to 28.9) and surgery (odds ratio, 25.0; 95% confidence interval, 14.4 to 43.5). Within these respective groups, the strongest association was for hip fractures (odds ratio, 69.38; 95% confidence interval, 9.34 to 515.28) and musculoskeletal surgery (odds ratio, 21.31; 95% confidence interval, 15.70 to 28.91). In women, the risk of venous thromboembolism was 1.9 (95% confidence interval, 1.5 to 2.3) among those receiving hormone replacement therapy and 1.9 (95% confidence interval, 1.4 to 2.5) among those taking oral contraceptives.
By using a large prospective cohort, Huerta et al. benefited from a standardized collection of data, the lack of which is often a problem in case-control studies that are conducted outside of cohort trials. All data, including exposure information, were originally entered by the family physician of each patient and collected years before the current study was initiated and so would be unlikely to be biased by data collectors. In addition, a validation study allowed the investigators to confirm the diagnosis of venous thromboembolism in 94% of the patients, which provided substantial assurances against a misclassification bias. Analyses were adjusted with a considerable number of potential confounding variables, which increased the rigor of their results. This study has many strengths, and the use of a nested case-control design permitted exploration of important questions at a fraction of the cost and time that would have been expended to conduct such an analysis on all 1,856,206 members of the original cohort.
Example 4: When a Case-Control Study Is Not a Case-Control Study
The case-control study design involves methodological features that are unique, and interpretation of results can be challenging. Another source of difficulty is when studies are reported as a case-control design when, in fact, they are not. To explore the long-term outcomes after spinal arthrodesis in patients with adolescent idiopathic scoliosis, Danielsson and Nachemson conducted a study, which they identified as a case-control design13. The authors invited 156 consecutive patients who were treated surgically for adolescent idiopathic scoliosis between 1968 and 1977 to participate in their study. For their control group, they randomly selected 100 age and sex-matched individuals who had not undergone back surgery or been diagnosed with notable scoliosis. Eighty-nine percent (139 of 156) of the invited case patients participated and, along with the 100 control patients, underwent physical examination and plain radiographs and completed a series of questionnaires. The main finding of the authors was that, after a mean time of twenty-three years following spinal fusion surgery, case patients had minimal low back pain and no major differences in low back function and general quality of life when compared with control patients13.
The "case patients" in this study are patients who underwent spinal fusion for adolescent idiopathic scoliosis, but there is no defined exposure that was investigated for an association with this outcome. Therefore, despite being represented as a case-control design, this study is actually a prospective cohort trial with a comparison group that was examined in a cross-sectional manner.