The three common observational designs are: cross-sectional, case-control, and cohort study designs. Observational studies share the common characteristics such as: no control on patients, interested in the effect of an exposure on an outcome, and susceptible to potential confounders. In this section, we will briefly illustrate these designs and the statistical analysis for each design.
A cross-sectional study is an observational study that involve data collected from a specific population at a specific point in time (4). If the investigator is interested in the association between an exposure and certain outcome, then the exposure and the outcome variable are collected from the same time point. Ideally, the cross-sectional study should perform random sampling from a population. In this way, the sample will be a representative subset of the whole population. This is essential if the study objective is to evaluate the prevalence (event rate) of certain disease or event. If the purpose is to evaluate the association between an exposure and a disease (event), most clinical studies choose the convenient sample due to practical reasons. In this case, caution should be taken when interpreting the disease rate or prevalence.
A distinct feature of the cross-sectional study is that investigator don't control the proportion of subjects in the exposure or the proportion of subjects in the disease group when perform the sampling. This is the feature separating the cross-sectional studies from the other two types of observational studies. Obviously, the cross-sectional studies often cost less compared to other types of studies. If the sample is representative, it can be used to estimate the disease prevalence. Furthermore, since neither the disease proportion, nor the exposure proportion is controlled, cross-sectional study can be used to study the association between several exposure factors and multiple outcomes simultaneously.
Clearly, there are limitations in cross-sectional studies due to the design features. First, the exposure and the outcome variables are collected at the same time point. It doesn't provide any causal relationship as it takes time for the exposure variable to affect the outcome variable. In a study conducted in a tertiary hospital to evaluate whether the anesthesiologists actually prepared for loss of airway or respiratory function and its association with regional/general anesthesia (5). Among the 87 procedures, 29.9% didn't have verification of the availability of material for orotracheal intubation. The procedures with regional anesthesia have a significantly higher rate of not having verification than the procedures with general anesthesia. This association between the type of anesthesia and verification rate should be further study to established causal relationship. Second limitation is that when the exposure or outcome events are very rare, this study design may collect either only a very small number of patients in exposure or small number of patients with an outcome event. The statistical test performance and the estimator's accuracy are compromised when the number of patient in either of these two categories are small. The features and limitations of cross-sectional study design are listed in Table along with the other two study design introduced below. Regardless these limitations, there are cross-sectional studies published in high profile journal such as the New England Journal of Medicine (6) and the Lancet (7, 8).
A case-control study is an observational study conducted to study the association between exposures and an outcome (case/event) in which the cases and controls are selected first and the exposure information were collected retrospectively. In a case-control study, investigators usually define the case for identification. The control is then defined according to the research questions to be answered. The number of cases and controls are pre-specified in the designing stage. Therefore, the proportion of cases in the sample is determined by design and not a representation of prevalence of case in the general population. For each identified case and control, their previous exposure status is inquired by interview or through searching databases. For example, in a study conducted in Minnesota, USA, to test whether the exposure to procedure requiring general anesthesia during adulthood is associated with dementia (9). Eight hundred dementia patients were identified through Mayo Clinic Alzheimer's Disease Registry and 877 matched controls free of dementia were sampled from the general pool of Olmsted County residents. Medical records were reviewed to determine exposures to procedures requiring anesthesia after age 45 years and before the index year. Six hundred and fifteen dementia case (70%) underwent procedures requiring general anesthesia. There is not significant statistical association between dementia and experience of general anesthesia (odds ratio, 0.89; 95% CI, 0.73-1.10; P=0.27).
Case-control studies are retrospective and are valid tools for evaluating association. Compared to the cohort studies illustrated in next subsection, it usually enjoys the features of lower cost and less time-consuming as there is no prospective follow up period. It is also important to note that for the diseases with long latency, conducting a clinical trial or cohort study will take a long follow up period to observe enough number of disease events for valid statistical analysis. Case-control studies are much more efficient in this case. Similarly, if the disease is a rare disease (e.g. intraoperative awareness), it is more efficient to conduct a case-control study than other study designs.
There are limitations in case-control studies. Due to its retrospective feature, case-control studies are susceptible to bias including recall bias and selection bias (10). These biases should be carefully evaluated and reduced through better information collection tools and statistical analysis. Another difficulty is the selection of controls. The cases are often defined more clearly. However, the control selection can vary much among different investigators which may result in different results (11). Thirdly, since a case-control study pre-specifies the proportion of cases and controls in the sample, the disease prevalence or incidence in the general population is not estimable through the case-control design without making strong assumptions. Case-control studies are also less efficient when the exposure is rare as it takes a larger sample size to collect enough patients with exposure (12). More studies employing case-control published can be seen in the Lancet (13, 14), the New England Journal of Medicine (15, 16), Journal of American Medical Association (17), and Anesthesiology (18).
In a cohort study, the investigator selects a group of non-diseased people and follows them over time to determine if they develop a disease/outcome. The cohort is defined based on the exposures to be investigated (19). For a prospective cohort study, the patients are recruited after the study design and events are prospectively observed. A retrospective cohort study defines the cohort based on the exposure from a historical time point and uses the historical data (e.g. medical record) to identify the event occurrence for each patient. Although constructed retrospectively, it possesses the feature of specifying the exposure group and target enrollment number in each exposure group which is different from cross-sectional and case-control studies.
In some settings, it is difficult to perform a randomized study as the treatment assignment is controversial. In this setting, cohort study design will provide a direction and magnitude of causal relationship (19). For example, in a study conducted to evaluate whether epidural anesthesia in labor is associated with chronic low back pain, a random assignment of epidural is not ethical. Therefore, a prospective cohort with one hundred sixty-four women had received epidural analgesia and 165 without analgesia for labor and delivery were recruited. They were followed for 12 months to quantify their back pain (20). Since the epidural and non-epidural groups were pre-specified, the low back pain rate can be calculated for each group which is, 9.9% and 13.8%, respectively. But the difference of rates is not significant statistically. Compared to other observational study designs, cohort study is not subject to recall bias. Due to the longitudinal feature of the cohort design, one can account for change within and between subjects and improve the efficiency through repeated measurements. Prospective cohort studies are considered to yield most reliable results in observational studies.
While providing more reliable results, cohort studies also have limitations in practice. First, like other longitudinal studies, attrition (lost to follow-up of study subjects) is a major concern that often biases the analysis results. The cost of a prospective cohort study is usually higher than other type of observational studies. Over the follow up period, the study condition might change between the baseline and primary end point. The analysis technique is more sophisticate and complex which often requires professional statistical support. Last, the cohort study is less efficient when studying the rare disease and the disease with long latency. More cohort studies can be found in the New England Journal Medicine (21, 22), the Lancet (23-25), and Anesthesiology (26).