Prospective cohort studies are inherently superior to retrospective cohort studies

Cohort Study

A cohort study is an observational study design in which groups of subjects are identified based on their exposure to a particular risk factor and then compared to a group that have not been exposed to that same factor. This study design can be conducted from either a forward-looking (i.e., prospective) or backward-looking (i.e., retrospective) viewpoint. Prospective cohort studies are planned in advance and carried out over a period of time to assess outcome incidence among exposure groups. In comparison, retrospective cohort studies look at data that already exist and attempt to identify risk factors for outcomes that have already occurred. Retrospective analyses of existing surgical databases, in particular, constitute one of the most common applications of this study design.10 For both types of cohort studies, a higher incidence of outcomes in the exposed group suggests an association between that factor and the outcome. However, because prospective studies collect information about exposures and outcomes purposefully and systematically, they provide stronger evidence for causation.

There are several strengths and limitations of cohort studies. The advantages of cohort studies are that they are easier and less expensive to conduct than RCTs, the incidence (or rate) of exposure and outcomes can be estimated, and subjects in cohorts can be matched to limit the influence of confounding variables. In addition, enrollment criteria and outcome measures can be standardized, and—unlike case-control studies—multiple simultaneous outcome assessment is possible. However, cohort studies also have several disadvantages. Because there is no randomization, one cannot account for unmeasured imbalances in patient characteristics. Blinding or masking is difficult (or impossible retrospectively), and outcomes of interest can take a long time to occur, requiring many years of information about exposures. For retrospective studies in particular, treatment selection bias and confounding variables may lead to unmeasured differences in exposure groups over time that cannot be controlled for with statistical analysis. Moreover, interpretations can be limited because of missing data that is impossible for researchers go back in time to collect.

In order to assess the evidence quality derived from different types of cohort studies, several simplified checklist methods have been developed. The Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement is one of these checklists that provides an organized system for assessing observational study methodology and results that is similar to the CONSORT system used for RCTs.11 STROBE provides guidance on how to report critical components of research studies from all types of observational study designs, including cohort, case-control, and cross-sectional study designs.

As part of a series of articles on research methods published in a special issue ofJAMA Surgery in 2018, another checklist was proposed for evaluating the quality of evidence derived from databases commonly used in surgical outcomes research.10 This checklist consists of 10 items for review that can be used when evaluating the findings of retrospective cohort studies, or considering questions that might be reasonably addressed using these databases. In the same issue ofJAMA Surgery, in-depth information was provided for 13 of the most popular surgical databases, including the National Inpatient Sample; Surveillance, Epidemiology, and End Results (SEER) Program12; Medicare Claims13; National Surgical Quality Improvement Program (NSQIP)14; Society of Vascular Surgery Vascular Quality Initiative (SVS-VQI)15; and the Society of Thoracic Surgeons (STS) National Database.16

Clinical Data

Cohort studies might also utilize data that were collected for routine clinical use and/or patient monitoring. Studies have shown that using automated health records collected for clinical purposes to conduct cohort studies can be an effective method of data collection (Hashikata et al., 2011).

In the United Kingdom, one of the largest sources of electronic health records used for research is the CPRD, which has been used in conducting several retrospective cohort studies. One example is a study that aimed to confirm whether corticosteroid treatment was associated with an increased fracture risk (van Staa et al., 2001). In fact, literature reviews have found that, among epidemiological studies, the most frequently used data source was the CPRD, formally known as General Practice Research Database, which accounts for about 20% of the studies using large existing data. Of these, 40% used a cohort design (Perrio et al., 2007).

Kaiser Permanente (founded in 1945) is one of the largest health plans in the US, serving approximately 11.7 million members (Kaiser Permanente, 2018). In addition to providing care, its Division of Research conducts and publishes health services and outcomes research (including pharmacoepidemiologic cohort studies). Using electronic health records obtained from Kaiser Permanente, investigators were able to determine that there was no increase in the risk of recurrent breast cancer in subjects taking paroxetine concurrently with tamoxifen, despite widespread controversy about this drug combination (Haque et al., 2016).

Studies using clinical data might also aim to answer questions about the effectiveness of emerging therapies for which guidelines have not yet been established. The Clinical Data Analysis and Reporting System (CDARS) of the Hong Kong Hospital Authority (the only publicly funded health care provider in Hong Kong) provides electronic health records for research. When investigators were concerned about the growing trend of using gastro-protective agents to prevent gastrointestinal bleeds (GIB) in patients using the anticoagulant dabigatran, they were able to construct a cohort study using data from CDARS to examine this relationship. With dabigatran dispensing records identifying exposure status and diagnosis records for GIB to determine outcomes, the investigators used Poisson regression controlling for baseline comorbidities and medication use to demonstrate that that patients who used gastro-protective agents with dabigatran had a 48% decrease in risk of GIB compared to those who used dabigatran alone (Chan et al., 2015).

Single-Cohort Studies

Although the body of evidence regarding PAE is growing, the literature is primarily based on small cohort series concentrated in a handful of centers. It is important to note that there have only been two randomized trials published to date, both comparing PAE to TURP.Pisco et al. (2011) were first to report on a small series of patients who underwent PAE prospectively. They noted statistically significant improvements in AUA-SS (6.5 points,P = .005) at a mean follow-up of 7.9 months. Peak urinary flow only increased by 3.85 mL/sec (P = .015), and PSA and prostate volume decreased by 26% and 27%, respectively. The procedure was unable to be completed in 6.7% of patients, and 28.6% of patients met criteria for clinical failure. There was one major complication, an ischemic area of the bladder wall that required surgical intervention. Mean procedure time was 85 minutes (range 25 to 135 minutes), with patients undergoing amean fluoroscopy exposure time of 35 minutes (range 15 to 45 minutes) (Pisco et al., 2011). Later studies have reported a mean fluoroscopy time ranging from 18 to 85.9 minutes, which in part, may be caused by the learning curve associated with the procedure (Kuang et al., 2017).

Antunes et al. (2013) included urodynamic data in their cohort of 11 patients who presented with urinary retention, with a prostate volume between 30 and 90 g. In this study, 10 of the 11 patients were catheter-free at a minimum follow-up of 12 months (mean follow up 22.3 months), and mean AUA-SS score at 12 months was 2.8 (range 1 to 7). UDS performed before and after treatment showed a decrease in detrusor pressure at Qmax from a mean of 85.7 to 51.5 cm H2O. All patients had a BOOI greater than 40 before PAE. Post-PAE, only 30% of patients were clearly unobstructed (BOOI <20), 40% in the equivocal range (BOOI 20 to 40), and 30% demonstrated persistent obstruction (BOOI >40). Almost all patients reported mild, transitory pelvic pain, with three patients having minor rectal bleeding. One patient had an a focal area of hypoperfusion of the bladder wall on MRI at 1 month, which normalized on MRI at 90 days and 2 years.

Bilhim et al. (2013) investigated the effect of unilateral versus bilateral PAE in a single-center cohort study of 122 consecutive patients. A comparison of 103 patients who underwent bilateral PAE and 19 who underwent unilateral PAE revealed no statistically significant differences in change in prostate volume, PSA, IPSS, QoL, Qmax, PVR, or IIEF. A significant proportion of patients in both groups required medical therapy for persistent LUTS after PAE, 52% and 77.8% of patients in the bilateral and unilateral PAE groups, respectively. In addition, 20% of patients who underwent bilateral PAE went on to require prostatic surgery, compared with 11.1% of patients in the unilateral PAE group.

Observational studies

Cohort study & case control study

In observational studies, the exposure and outcome are observed over a certain period of time and the association is described (Fig. 6.3 and 6.4).

No interventions are carried out by the examiner.

Although observational studies are more prone to confounders and bias, compared to interventional studies, they are a very useful tool in epidemiology since practical or ethical matters can preclude the design of an RCT with certain research questions.

Advantage

Less expensive and easier to carry out than intervention studies.

Exposure and outcome can be linked in relation to time.

Ethically less limitations compared to interventional studies.

Disadvantage

The investigational exposure can be influenced by unknown factors.

Observational studies are prone to recall bias.

Observational studies are prone to Neyman bias.

In observational studies, confounders are unevenly distributed.

A long observational period is required for diseases that appear late after exposure. e.g., cancer.

Prospective studies

Data is collected prospectively after the study is planned. (e.g., questionnaire)

Prospective cohort study or randomized controlled studies are prospective studies.

Advantage

Data can be collected specifically focused and targeted to the research question.

Data quality can be improved with good study planning.

Disadvantage

Ethically challenging research data cannot be collected.

It is time consuming as data collection requires a certain collection and follow-up time.

Retrospective studies

Data is collected and stored during before the study was planned without a specific research question (e.g., electronic medical record).

Advantage

In retrospective studies, research questions can be answered for which a prospective study does not exist or cannot be carried out for ethical reasons.

It is relatively inexpensive compared to a prospective study, since the data pool is already available.

Disadvantage

In retrospective studies, certain factors (bias) cannot be controlled, or only with difficulty.

Since the data collection took place before the study planning, the quality of the available data is not always good.

Data quality cannot be improved since the dataset already exists.

Additional data cannot be collected.

Cohort study (Fig. 6.5)

A cohort study can be performed prospectively or retrospectively.

It examines the association between a given exposure (risk factor) and the outcome (disease).

Advantage

Ethically safe.

A cohort study is generally easier to perform and less expensive than intervention study.

Disadvantage

Large numbers of cases or long follow-up times are necessary for rare diseases.

There must be an exposure for the disease.

The exposure may be influenced by a hidden variable (confounder).

A randomization is not possible.

No “blinding” available.

Observational studies are prone to selection bias and reporting bias.

Example

Smoking causes lung cancer. 1000 people are examined, including 500 smokers (Fig. 6.6).

Figure 6.6. Example of a cohort study.

Cohort study

A cohort study recruits its participants from the general population, and those recruited must be without the disease of interest. Using the example given inTable 123.4, all participants would need to be SCC-free at the onset of the study. They are then categorized as having or not having high-dose UV exposure and are followed over time. At the end of the study, one can calculate the rate of cutaneous SCC among those with high-dose UV exposure (a/a+b) and compare it to the rate of SCC among those without high-dose UV exposure (c/c+d). Therefore, a cohort study compares the incidence of disease between those exposed and not exposed to particular risk factor(s). This is usually expressed in terms of a RR (see definition above). Similar to an OR, a RR of 1 represents a null association, whereas an RR statistically significantly greater or less than 1 indicates a positive or negative association, respectively. A cohort study can be done prospectively, allowing for more control over measuring outcomes (preferred particularly for fatal diseases), or retrospectively, which is less expensive and requires less time.

One of the advantages of a cohort study is that it can determine theincidence of a disease, which is the number of new cases divided by the total number of people at risk in a particular population (e.g. participants in the study) over a specified time period (e.g. length of study)1. It also gives information on the natural history of a disease, since it captures data from the time of disease onset. Another advantage is that a cohort study establishes temporality, as the exposure precedes the outcome, and therefore allows for causal inference between the exposure and disease. The disadvantages are that a cohort study often requires a large sample size and is less feasible for rare outcomes.

A recent cohort study by Zhang et al.26 looked at the association between tanning bed use and development of skin cancers. Their study population included >70 000 female nurses who were followed between 1989 and 2009. After adjusting for possible confounders (e.g. age, skin type, sun exposure, UV index at place of residence), they found that for an incremental increase of 4 sessions/year in mean use of tanning beds prior to age 35 years, the RR of developing BCC was 1.15 (95% CI, 1.11 to 1.19). Thus the use of tanning beds a mean of 4 or 8 times per year resulted in a ~15% or ~30% higher risk of BCC, respectively, compared to not using tanning beds. This demonstrates a “dose response” effect, i.e. women who used tanning beds more often had a higher skin cancer risk. Furthermore, the authors noted that women exposed to tanning beds in their teenage or college years were at higher risk than those exposed later in life.

Effect modification orinteraction occurs when the association between the exposure and disease varies across levels of a variable (e.g. age;Fig. 123.4). Therefore, the exposure and effect modifier have an interdependent role in their effect on the disease. When this happens, the association between exposure and disease should be displayed at various levels or strata of the effect modifier1,2.

Cohort Study

A cohort study involves enrolling/identifying a group of people defined by a specific characteristic who are followed over time. Cohort studies generally compare exposed patients to unexposed patients, but they can also compare one exposure to another. An association is identified when the outcome occurs more frequently in the exposed than the unexposed group. Comparators can be persons not exposed to the drug, or taking a drug from the same therapeutic class. Yet, due to the nonrandom assignment of groups, increased attention must be given to the selection of appropriate comparators.

Routinely collected pharmaceutical data may be used to identify a cohort of persons all exposed to the same drug, who may then be observed for years to examine the long-term side effects of that prescription drug use. If these data include all drug dispensings, researchers use other prescription drug use as a proxy for comorbidities to calculate the risk of a specific outcome occurring based on the presence of disease.

2.2.1 Cohort studies

In a cohort study a well-defined group of subjects (e.g., workers) at different levels of exposure is followed-up for a specified amount of time that could be short (e.g., hours, days, weeks, or months) or long (e.g., years or decades), and occurrence of selected diseases is recorded and compared across exposure groups. Put it simply the cohort study answers the question: “What are the effects of this agent?”

The key features are the following: (1) We start with defining exposure (usually only one, or at most a few); for this reason occupational cohort studies are usually performed in occupational settings (a factory, an agricultural site), or based on data regarding workers (e.g., members of an agricultural association, licensed pesticide workers). (2) We investigate disease occurrence across exposure groups during the follow-up time, so a synonym for these studies is follow-up studies. (3) When follow-up period is long, we have to deal with time-varying variables in an appropriate way (sometimes this applies also to studies with short follow-up period). (4) Many diseases can be investigated in a single study.

According to classical definitions, cohort studies may be prospective (carried on from present to future) or retrospective (performed using data collected in the past, like the majority of occupational studies). However, since the study direction is inherently from exposure to disease, some prefer to avoid the term retrospective and instead use the term historical cohort studies.

53.4.2 Retrospective cohort

The retrospective cohort study looks back on existing data when outcomes have already occurred. These data often come from large, computer databases. However, data may also come from paper charts or medical records. In these studies, cohorts are assembled the same way as in prospective cohort studies, namely, on the basis of exposure to certain drugs of interest. The major advantage of retrospective cohort studies is lower costs. These are much less expensive to conduct than either clinical trials or prospective studies. The major disadvantage is that there are many forms of bias found in retrospective studies. Conceptually, retrospective cohort studies are conducted the same way as prospective cohort studies.

Given the importance of defining drug exposure in the field of pharmacoepidemiology, it is worth discussing this issue in some detail. A cohort may be defined by determining the date a drug was first prescribed (i.e., index date). Although the index date may differ for each patient, it acts as the anchor for two key viewpoints: (1) for looking forward in time for the occurrence of the outcomes of interest [e.g., myocardial infarction (MI) and renal insufficiency] and (2) for looking backward in time for the baseline factors that may affect the outcome independent of the drug exposure. The typical claims-based pharmacoepidemiologic study might define the index date based on the first pharmacy-prescription claim for the drug that comes after an observable period with no claims for a particular drug. However, defining the index date as the true time of drug initiation can require additional considerations. For example, when a researcher examines a public insurance claims database, the researcher has to make sure that subjects are not exposed to medication from other sources outside of what can be tracked in that database (Moga et al., 2013, 2017). Accounting for potential use of illicit drugs, which are not typically found in medical records, can also present challenges when studying drugs with abuse potential. Drug exposure should be verified, sometimes using physiologic measurements, to prevent misclassification of cohort membership. A common problem in pharmacoepidemiologic studies is that patients could get a medication from a pharmacy different than the pharmacy submitting the claim, physicians may provide the drug as a sample from their office, or the drug may be available as an over-the-counter (OTC) drug and, as such, can be directly available to the patient with no official record into a database. If a drug is available from multiple sources, the investigator should demonstrate that all sources were included in the cohort formed for study.

Interestingly, even pharmacists themselves are considered the exposure in some retrospective cohort studies. Consider that Gatwood et al. (2018) work that showed that veterans with diabetes (cohort) adhered to their medications (outcome) more successfully when working with a clinical pharmacist (exposure) than not. Differences between prospective and retrospective cohort studies can be technical and confusing even for pharmacoepidemiologists as discussed by Klebanoff and Snowden (2018). Rather than focus on the formal nomenclature for these studies, it is more important to think critically about defining study parameters such as the recognizing exposure groups whether they be drugs or health-care services as provided in examples previously.

Tumorigenicity

A retrospective cohort study in 5052 elderly subjects, of whom 451 were taking verapamil, diltiazem, or nifedipine, showed that these drugs were associated with a cancer risk of 1.72 (95% CI = 1.27, 2.34), and there was a significant dose–response relation (8). A small risk of cancer (RR = 1.27; 95% CI = 0.98, 1.63) with calcium channel blockers was reported in a nested case-control retrospective study involving 446 cases of cancers in hypertensive patients (128). However, the authors concluded that this finding may have been spurious, as there was no relation between the cancer risk and the duration of drug use. Another study did not show any excess cancer risk with short-acting nifedipine after myocardial infarction in patients followed up for 10 years, although there were only 22 cancer deaths in 2607 patients (129). Neither did the much larger Bezafibrate Infarction Prevention (BIP) Study, which reported cancer incidence data in 11 575 patients followed for a mean period of 5.2 years, with 246 incident cancer cases, 129 among users (2.3%) and 117 (2.1%) among non-users of calcium channel blockers (130). Others also failed to find a positive link between calcium channel blockers and cancer (14,131). However, elderly women taking estrogens and short-acting calcium channel blockers had a significantly increased risk of breast carcinoma (hazard ratio = 8.48; 95% CI = 2.99, 24) (132). This controversy can perhaps only be resolved by prospective studies with longer follow-up periods (133), although ideal studies are unlikely ever to be conducted.

8.34.7.7.2 Cohort studies

Data obtained from cohort studies have the greatest power to determine a true relationship between an exposure and disease outcome because one starts with a healthy cohort, obtains biomarker samples, and then follows the cohort until significant numbers of cases are obtained. A nested study within the cohort can then be designed to match cases and controls. An advantage of this method is that the controls are truly representative of the source population since both were recruited at the same time and with the same health status. A major disadvantage, however, is the time needed in follow-up (often years) to accrue the cases. This disadvantage can be overcome in part by enrolling large numbers of people (often tens of thousands) to ensure case accrual at a reasonable rate. To date, two major cohort studies with aflatoxin biomarkers have demonstrated the role of this carcinogen in the etiology of HCC. The first study, comprising > 18,000 people in Shanghai, examined the interaction of HBV and aflatoxin biomarkers as independent and interactive risk factors for HCC. The nested case–control data revealed a statistically significant increase in the RR of 3.4 for those HCC cases where urinary aflatoxin biomarkers were detected. For HBsAg-positive people RR = 7, but for individuals with both urinary aflatoxins and positive HBsAg status, RR = 59 (Qian et al., 1994; Ross et al., 1992). These results strongly support a causal relationship between the presence of the chemical and viral-specific biomarkers and the risk of HCC. Subsequent cohort studies in Taiwan have confirmed the results from the Shanghai investigation. Wang et al. (1996c) examined HCC cases and controls nested within a cohort and found that in HBV-infected people there was an adjusted odds ratio of 2.8 for detectable aflatoxin–albumin adducts compared with nondetectable aflatoxin–albumin adducts and 5.5 for high levels compared with low levels of aflatoxin metabolites in urine. In a follow-up study, there was a dose–response relationship between urinary AFM1 levels and risk of HCC in chronic HBV carriers. Similar to the Shanghai data, the HCC risk associated with AFB1 exposure was more striking among the HBV carriers with detectable AFB1–N7-Gua in urine.

Therefore, these cohort data from two different populations demonstrate the power of validated aflatoxin biomarkers to define a heretofore unrecognized chemical–viral interaction in the induction of human HCC (Harris, 1994). These findings have significant public health implications. First, vaccination to prevent HBV infection would substantially reduce a major factor in HCC. Unfortunately, in most parts of the world HBV infection is acquired before age 3 years; therefore, complete worldwide elimination of HBV infection by vaccination will require much of the next century to accomplish. Second, elimination or reduction of aflatoxin exposure should also reduce the risk of HCC. This goal can be attained by using available technologies and the dose–response data from epidemiological studies indicate that, similar to smoking cessation, the reduction in aflatoxin exposure during an individual’s lifetime should reduce risk of HCC. Taken together, these cohort studies provide the final data sets in the validation scheme to establish the use of some of the aflatoxin biomarkers as validated risk markers.