[ Supplement An Overview of Study Design and Statistical Considerations ]

Cohort Studies
Design, Analysis, and Reporting
Xiaofeng Wang, PhD; and Michael W. Kattan, PhD

Cohort studies are types of observational studies in which a cohort, or a group of individuals
sharing some characteristic, are followed up over time, and outcomes are measured at one or
more time points. Cohort studies can be classified as prospective or retrospective studies, and
they have several advantages and disadvantages. This article reviews the essential charac-
teristics of cohort studies and includes recommendations on the design, statistical analysis, and
reporting of cohort studies in respiratory and critical care medicine. Tools are provided for re-
searchers and reviewers. CHEST 2020; 158(1S):S72-S78

KEY WORDS: bias; cohort studies; confounding; prospective; retrospective

General Overview of Cohort Study outcome is reasonable. Cohort studies are

Design the design of choice for determining the
The term “cohort” in modern epidemiology incidence and natural history of a condition.
refers to “a group of people with defined Due to their longitudinal design feature, one
characteristics who are followed up to can look at disease progression and natural
determine the incidence of, or mortality history.3 Cohort studies allow us to calculate
from, some specific disease, all causes of the incidence rate, cumulative incidence,
death, or some other outcome.”1 A cohort relative risk, and hazard ratio. Causality
study observes people as two or more groups, cannot be established definitively through a
from exposure to outcome.2 A key feature of cohort study.2 Nevertheless, cohort studies
the cohort study design is that subjects are are useful to provide evidence that suggests
followed up over time. It begins with subjects causality and information regarding the
who are exposed and not exposed to a factor strength of the association between the risk
and then evaluates the subsequent factors and the outcome.
occurrence of an outcome. Unlike cross- Description of Subtypes of Cohort
sectional studies, which are often used to Studies
determine prevalence, cohort studies are used
Cohort studies can be either prospective or
to study incidence, causes, and prognosis.
retrospective. The type of cohort study is
In clinical research, cohort studies are determined by the outcome status. If the
appropriate when there is evidence to outcome has not occurred at the start of the
suggest an association between an exposure study, then it is a prospective study; if the
and an outcome, and the time interval outcome has already occurred, then it is a
between exposure and the development of retrospective study.4 Figure 1 presents a

ABBREVIATION: CAP = community-acquired pneumonia 9500 Euclid Ave/JJN3-01, Cleveland, OH 44195; e-mail: wangx6@ccf.
AFFILIATIONS: From the Department of Quantitative Health Sciences, org
Lerner Research Institute, Cleveland Clinic, Cleveland, OH. Copyright Ó 2020 American College of Chest Physicians. Published by
CORRESPONDENCE TO: Xiaofeng Wang, PhD, Department of Quan- Elsevier Inc. All rights reserved.
titative Health Sciences, Lerner Research Institute, Cleveland Clinic, DOI: https://doi.org/10.1016/j.chest.2020.03.014

S72 Supplement [ 158#1S CHEST JULY 2020 ]

 Past Present Future

Start Study
Prospective Cohort Study

Retrospective Cohort Study (”Look Ahead”)

(”Look back”)

Study Population

Exposed Group Unexposed Group

Event No event Event No event

Group Group Group Group

Figure 1 – Graphical representation of the timeline in a prospective vs a retrospective cohort study design.

graphical representation of the designs of prospective cohort design was proposed by Prentice8 as a cost-
and retrospective cohort studies. The distinguishing effective alternative to the nested case-control design. In a
feature of a prospective cohort study is that at the time case-cohort design, a subcohort is randomly drawn from
that the investigators begin enrolling subjects, none of the full cohort, and the case-cohort sample consists of the
the subjects has developed the outcome of interest. In subcohort plus those subjects from the entire cohort
contrast, a retrospective study is conceived after subjects whose outcome occurred during the study period.
have already developed the outcome. The investigators Figure 2 illustrates the subject selection process of a case-
jump back in time to identify a cohort of subjects at a cohort sample. The case-cohort study design is efficient
point in time when they did not have the outcome. A when only a very small fraction of the full cohort develops
prospective cohort study design is ranked higher in the the outcome in the given study time frame and the
hierarchy of evidence than a retrospective design exposure measurement of interest is expensive to obtain.9
because the outcome, predictor, and confounding
variables can be better measured and controlled.5
Information gained from a retrospective study can be Use Cases of Cohort Studies
helpful in planning a future prospective study.6 Example 1
Nijkeuter et al10 conducted a prospective cohort study to
A study combining two study designs, the case-cohort understand the natural course of hemodynamically
design, is a combination of a case-control and cohort stable pulmonary embolism (PE). The study aimed to
design that can be either prospective or retrospective. The evaluate the incidence of recurrent VTE, hemorrhagic
case-cohort design can be viewed as a variant of the
nested case-control design.7 In a nested case-control
Cohort
study, one starts with identifying cases that have already
occurred (retrospective) or as they occur (prospective) in Case
subjects
a defined cohort. A specific number of control subjects
are then selected from among those in the cohort.
Limitations in this type of design include: (1) inefficiency Random Subcohort
due to the need to align each selected case subject to its Cases in subcohort
matched control subject; and (2) when there is more than
one outcome considered, strict implementation of the
Figure 2 – An illustration for subject selection in a case-cohort study.
design requires the selection of a new set of control The case-cohort sample consists of the subcohort members as well as all
subjects for each distinct disease outcome. The case- the case subjects who are outside of the subcohort.

chestjournal.org S73
complications, and mortality in patients with PE, and to with a single exposure or multiple exposures in a single
identify risk factors and the time course of these events. study. Even the combined effect of multiple exposures
Between November 2002 and September 2004, a total of on the outcome can be determined. Cohort study
3,503 patients with clinically suspected PE were designs also allow for the study of rare exposures.
screened, and PE was diagnosed in 674 patients. Three- Investigators can specifically select subjects exposed to a
month follow-up was completed in 673 of the 674 certain factor. Furthermore, cohort studies often have
patients with PE. The authors found that recurrent VTE broader inclusion and fewer exclusion criteria compared
occurred in a small percentage of patients treated for an with randomized controlled trials. The investigators may
acute PE, and the majority of recurrent VTEs were fatal. obtain large samples and reach greater power in
Immobilization, hospitalization, age, COPD, and statistical analysis relative to a randomized controlled
malignancies were risk factors for recurrent VTE, trial. For these reasons, results from cohort studies may
bleeding, and mortality. be more generalizable in clinical practice. Finally, the
longitudinal nature of cohort studies means that changes
Example 2
in levels of exposure over time, and changes in outcome,
Short et al11 performed a retrospective cohort study to can be measured to provide insight into the dynamic
examine the effect of b-blockers in the management of relation between exposure and outcome.
COPD. They searched a disease-specific database of
Prospective and retrospective studies have different
patients with COPD and linked to the Scottish
strengths and weaknesses. Prospective cohort studies are
morbidity records of acute hospital admissions, the
conducted from the present time to the future, and thus
Tayside community pharmacy prescription records, and
they have an advantage of being accurate regarding the
the General Register Office for Scotland death registry. A
information collected about exposures, end points, and
total of 5,977 patients aged > 50 years with a diagnosis
confounders. The disadvantage could be the long period
of COPD were identified and divided into two groups
of follow-up while waiting for events to occur, leading to
according to b-blocker use. The study found that
vulnerability to a high rate of loss to follow-up.
b-blockers might reduce mortality and COPD
exacerbations when added to established inhaled Retrospective studies rely on data collected in the past to
stepwise therapy for COPD, independently of overt identify both exposures and outcomes. These studies use
cardiovascular disease and cardiac drugs, and with no data that have already been collected, such as would be
adverse effects on pulmonary function. obtained using a database extracted from electronic
medical records. Thus, cohort studies are often time-
Example 3
efficient and cost-effective. However, many retrospective
Skull et al12 described the epidemiology of community- cohort studies use data that were collected in the past for
acquired pneumonia (CAP) in elderly Australian another objective. Hence, the investigators lack control
subjects. Using a case-cohort design, cases with CAP over the collection of data. The measurement of
were identified as in-patients aged $ 65 years with variables might be inaccurate or inconsistent, which
International Statistical Classification of Diseases and results in a source of information bias. Table 1
Related Health Problems, Tenth Revision, Australian summarizes the advantages and disadvantages of cohort
Modification codes J10 to J18 admitted over 2 years to studies.
two tertiary hospitals. The cohort sample was randomly
selected from all hospital discharges and frequency-
matched to case subjects according to month. A total of Study Subject Considerations
4,772 inpatients were studied. The strongest predictors There are several considerations related to the subjects
of CAP were previous pneumonia, history of other of a cohort study. These include selection of an
respiratory disease, and aspiration. ICU admission, renal appropriate sample of the population of interest, the
disease, and increasing age were the strongest predictors sampling method that will be used, access to
of mortality, whereas influenza vaccination conferred longitudinal data for the subjects chosen, and the sample
protection. size required to properly power the study. The criteria
for inclusion and exclusion should be determined at the
Benefits and Downside of Cohort Studies study design stage. The study subjects selected should be
A major advantage of the cohort study design is the appropriate for the study question and should be
ability to study multiple outcomes that can be associated generalizable to the population of interest. Avoiding bias

S74 Supplement [ 158#1S CHEST JULY 2020 ]

TABLE 1 ] Advantages and Disadvantages of Cohort Studies
Advantages
Can investigate multiple outcomes that may be associated with multiple exposures
Able to study the change in exposure and outcome over time
Good for examining rare exposures
Can measure incidence of outcome
May be able to infer causality
Prospective Study Retrospective Study
Able to control design, sampling, data collection, and follow-up methods Time-efficient and inexpensive
Can measure all variables of interest Easy to obtain large sample
Disadvantages
Susceptible to loss to follow-up compared with cross-sectional studies
Confounding variables are the major problem in analyzing the data compared with RCTs
Prospective Study Retrospective Study
May be expensive to conduct Less control over variables
Time-consuming Susceptible to information bias and recall bias

RCTs ¼ randomized controlled trials.

in subject selection, ensuring generalizability of the developed to help readers to perform the sample size
results, and determining the feasibility of performing an estimation for cohort studies. It can be found at http://
adequately powered study are crucial elements of the riskcalc.org:3838/samplesize/.
study design.
Sample size determination for cohort studies has been Statistical Considerations
widely discussed in the literature.13-15 A comparison of Investigators often use cohorts to assess the association
incidence rates is usually the major aim of a cohort between multiple exposures and multiple outcomes over
study. Assume that and p1 are the incidence rates of the time and to build prognostic/prediction models. The
end point of interest in the exposed and unexposed modeling and analysis strategy could be sophisticated in
samples. The sample size is typically calculated based on cohort studies. Here we emphasize a few important
the following statistical hypothesis: aspects of statistical analysis.

H0 : p1 ¼ p2 vs H1 : p1 sp2 Bias
Bias may be defined as any systematic error in a clinical
The sample size formula can be found in Fleiss et al.16
study that results in an incorrect estimate of the true
For paired cohort studies or case-cohort designs, the
effect of an exposure on the outcome. A major source of
formulae can be found in Kasiulevicius et al17 and Cai
potential bias in cohort studies is due to loss to follow-
and Zeng.9 When the outcome of interest in a cohort
up. This occurs due to dropouts or death, which often
study is continuous (although it is less common), we
occurs in studies with long follow-up durations. A
would like to compare the means of two cohorts. The
general rule of thumb requires that the loss to follow-up
formula based on the minimum detectable difference
rate does not exceed 20% of the sample.20 It is
can be found in Woodward.18
recommended that investigators examine any systematic
It is also important to consider subject loss to follow-up differences related to the outcome and/or exposures
in designing a cohort study. Any sample size calculated between those who completed the study and those who
should be inflated to account for the expected dropouts. were lost to follow-up. Methods of minimizing loss to
For instance, if the dropout rate is expected to be 10%, follow-up in a prospective cohort study have been
the estimated sample size would be N multiplied by 1/(1- comprehensively discussed by Hulley et al.21 We
0.1). A general discussion about sample size suggest that the investigators report median follow-up
determination is presented in the article by Wang and for patients without the event or the number followed
Ji19 included in this supplemental issue of CHEST. As up without an event at a given follow-up time.
part of that article, an online calculator has been For example, consider the case of a cohort of 1,000

chestjournal.org S75
patients with COPD treated in 1970 and followed up Many statistical methods can be applied to control for
until 2010. The median follow-up for all patients confounding factors, both at the design stage and in the
might be far less than the median follow-up for data analysis. The aim of controlling for confounding is
patients who survived. The latter statistic may provide to make the groups as similar as possible with respect to
a more accurate impression of how long the cohort the confounders. At the design stage, restriction is a
had been followed up. Now assume that in 2009, a common method for controlling confounders. The
second cohort of 2,000 patients were added to the investigators first identify potential confounding factors
study. The median follow-up for survivors will now be based on previous studies or the knowledge that
around 1 year, which is again misleading. An alternative confounding is biologically plausible. The investigators
would be to report a statistic such as “312 patients have then limit participation in the study to individuals who
been followed up without a death event for at least are similar with respect to those confounders. For
35 years.” example, a lung cancer study restricted to smokers will
eliminate any confounding effect of smoking. A
There are many other types of bias in clinical studies.
drawback of this method is that it may be difficult to
Examples include allocation bias, prevalence-incidence
generalize the findings to the rest of the population.
bias, recall bias, and detection bias. In the accompanying
cross-sectional study article included in this At the analysis stage, stratification is one of the popular
supplemental issue of CHEST, Wang and Cheng22 controlling methods. Stratification allows the association
provide a detailed discussion regarding common types between exposure and outcome to be examined within
of biases and their definitions in clinical studies. different strata of the confounding variables. For
example, a study is conducted to examine the association
Confounding between lung cancer and exposure to asbestos. To
control for smoking, the study population could be
Confounding often occurs in cohort studies. For a
stratified according to smoking status. The association
variable to be a confounder, it should meet three
between exposure to asbestos and cancer can then be
conditions: (1) be associated with the exposure being
assessed separately within each stratum. An issue with
investigated; (2) be associated with the outcome being
stratifying is that strata with more individuals will tend
investigated; and (3) not be in the causal pathway
to have a more precise estimate of the association (with a
between exposure and outcome. Confounding could
smaller SE) than strata with fewer individuals. For this
result in a distortion of the effects; it may lead to
reason, the Cochran-Mantel-Haenszel method is often
overestimation or underestimation of an effect, or even
used in stratification analysis. It allows calculating an
reverse the direction of an effect. For example, a study
overall and adjusted effect estimate of a given exposure
found that alcohol consumption was associated with
for a specific outcome by combining (pooling with
lung cancer. A person who drinks alcohol is more likely
weight) stratum-specific relative risks or OR.23
to smoke, and smoking is a risk factor for lung cancer.
Controlling for the potential confounding effect of Multivariable regression analysis is a model-based
smoking may show that there is no association between method to control for confounding. One builds a
alcohol consumption and lung cancer. Figure 3 shows multivariable regression model for the outcome and
the relation among the exposure, confounder, and exposure as well as other confounding variables. Based
outcome in this example. on the regression equation, the effect of the variable of
interest can be examined with confounding variables
held constant statistically. Multivariable regression has
Confounder the advantage in that it can control simultaneously for
Smoking more confounding variables than can stratification. It
has the disadvantage in that this model may not fit the
data well. The investigators have to be careful to use
accepted variable selection procedures.

Exposure The propensity score method is also popular for

Outcome
Alcohol
Lung cancer controlling confounding.24 The propensity score is the
consumption
probability of treatment/exposure assignment
Figure 3 – An example of illustrating the relation among exposure, conditioned on observed baseline characteristics. It
confounder, and outcome. allows investigators to mimic some of the characteristics

S76 Supplement [ 158#1S CHEST JULY 2020 ]

of a randomized controlled trial in a cohort study. In include some form of internal validation, such as cross-
practice, propensity score analysis involves a two-step validation or bootstrapping, particularly in the situation
procedure. The first stage of the analysis is to estimate that has no additional external validation performed.
the propensity score. We consider the exposure variable Controlling for confounding when building a prediction
as the response variable and build a regression model model is less common than that when modeling to
with the variables that influence exposure group assess for associations/causality. The article by Kattan
membership, such as sex and age. This model is used to and Gerds28 included in this supplemental issue of
give each subject a propensity score that measures the CHEST offers guidance on building prediction models.
propensity (probability) to be exposed given the
subject’s characteristics. In the second stage, the Reporting Considerations
outcomes of interest are compared between exposed and We suggest that investigators report their cohort studies
unexposed following adjustment for propensity scores. following the Strengthening the Reporting of
There are different approaches on using propensity Observational Studies in Epidemiology (STROBE)
scoring but they all yield similar results: matching on the statement, which contains a checklist of 22 items that are
propensity score, stratification on the propensity score, considered essential for reporting of observational
inverse probability of treatment weighting using the studies.29 If multivariable prognostic prediction models
propensity score, and covariate adjustment using the are developed in a cohort study to be used in predicting
propensity score.25 Although propensity score methods future outcomes in individuals at risk, we recommend
are powerful, they involve sophisticated statistical that investigators consult the Transparent Reporting of a
techniques. A deep understanding of the methodology is multivariable prediction model for Individual Prognosis
necessary when implementing the specific analysis. or Diagnosis (TRIPOD) statement.30,31
There are other methods of controlling for confounding
such as instrumental variable analysis and regression
discontinuity design; details are provided in Merrill1 and Short List of Questions to Guide the Reviewer
Rothman et al.2 The application of directed acyclic When reviewing a cohort study, consider commenting
graphs in observational studies assessing associations is on the following:
described in a separate article included in this
supplemental issue of CHEST (Etminan et al26). 1. The study cohort. Was the study cohort well
described? Was the method for selection of cohort
Model Building members and the inclusion/exclusion criteria
Model building is often crucial in cohort studies. appropriate? Were there potential biases intro-
Investigators may need to build explanatory models or duced by the methods chosen? Was the sample
predictive models. In explanatory modeling, one is size adequate for the primary study question? How
interested in identifying variables that have a was subject dropout, death, and missing data
scientifically meaningful and statistically significant handled?
relation with an outcome. In predictive modeling, the 2. The exposures and outcomes. Were they clearly
goal is to predict the probability of or the risk for the defined? Are there concerns about the accuracy of
presence (diagnosis) or future occurrence (prognosis) of their measurement? Were there potential biases
an outcome for an individual. When building a model introduced by the definitions and measurements?
(explanatory or predictive), the variables selected for 3. Potential confounders. Were potential con-
inclusion should be based on the critical consideration of founders identified based on prior knowledge?
relevant literature or knowledge of medical experts. Use Were they properly controlled for in the study
of stepwise selection should be restricted to a limited design and/or analysis? Are causal directed acyclic
number of circumstances, such as during the initial graphs included or required?
stages of developing a model, or if there is poor 4. The interpretation of the strength of the asso-
knowledge of what variables might be predictive.27 ciation(s) identified. Were the measures used to
Modern shrinkage or penalization procedures such as describe the association between the exposure and
LASSO/least absolute shrinkage and selection operator, the outcome clearly described and appropriate?
elastic net, and their variants are recommended for the Was the interpretation of the association(s) iden-
study of rare events or when there are a large number of tified appropriate?
predictors. If predictive models are built, one should

chestjournal.org S77
Acknowledgments 17. Kasiulevicius V, Sapoka V, Filipavici
ut_e R. Sample size
calculation in epidemiological studies. Gerontologija. 2006;7(4):
Financial/nonfinancial disclosures: None declared. 225-231.
18. Woodward M. Formulae for sample size, power and minimum
References detectable relative risk in medical studies. J R Stat Soc Ser D Stat.
1. Merrill RM. Introduction to Epidemiology. 7th ed. Burlington, MA: 1992;41(2):185-196.
Jones & Bartlett Publishers; 2015. 19. Wang X, Ji X. Sample size estimation in clinical research: from
2. Rothman KJ, Greenland S, Lash TL. Modern Epidemiology. 3rd ed. randomized controlled trials to observational studies. Chest.
Philadelphia, PA: Wolters Kluwer Health/Lippincott Williams & 2020;158(suppl 1):S12-S20.
Wilkins; 2012. 20. Conato J, Shah N, Horwitz RI. Randomized, controlled trials,
3. Carlson MDA, Morrison RS. Study design, precision, and validity in observational studies, and the hierarchy of research design. N Engl J
observational studies. J Palliative Med. 2009;12(1):77-82. Med. 2000;342:1887-1892.

4. Euser AM, Zoccali C, Jager KJ, Dekker FW. Cohort studies: 21. Hulley SB, Cummings SR, Browner WS, Grady DG, Newman TB.
prospective versus retrospective. Nephron Clin Prac. 2009;113(3): Designing Clinical Research. 3rd ed. Philadelphia, PA: Lippincott
c214-c217. Williams & Wilkins; 2007.
5. Vandenbroucke JP. Observational research, randomised trials, and 22. Wang X, Cheng Z. Cross-sectional studies: strengths, weaknesses,
two views of medical science. PLoS Med. 2008;5(3):e67. and recommendations. Chest. 2020;158(suppl 1):S65-S71.
6. Hess DR. Retrospective studies and chart reviews. Respir Care. 23. Sullivan LM. Essentials of Biostatistics in Public Health. Burlington,
2004;49(10):1171-1174. MA: Jones & Bartlett Learning; 2017.
7. Ernster VL. Nested case-control studies. Prev Med. 1994;23(5):587- 24. Rosenbaum PR, Rubin DB. The central role of the propensity
590. score in observational studies for causal effects. Biometrika. 1983;70(1):
41-55.
8. Prentice RL. A case-cohort design for epidemiologic cohort studies
and disease prevention trials. Biometrika. 1986;73(1):1-11. 25. Austin PC. An introduction to propensity score methods for
reducing the effects of confounding in observational studies.
9. Cai J, Zeng D. Sample size/power calculation for case–cohort studies. Multivariate Behavio Res. 2011;46(3):399-424.
Biometrics. 2004;60(4):1015-1024.
26. Etminan M, Collins GS, Mansournia MA. Using causal diagrams to
10. Nijkeuter M, Söhne M, Tick LW, et al. The natural course of improve the design and interpretation of medical research. Chest.
hemodynamically stable pulmonary embolism: clinical outcome and 2020;158(suppl 1):S21-S28.
risk factors in a large prospective cohort study. Chest. 2007;131(2):
517-523. 27. Kattan MW, Vickers AJ. Statistical analysis and reporting guidelines
for Chest. Chest. 2020;158(suppl 1):S3-S11.
11. Short PM, Lipworth SIW, Elder DHJ, Schembri S, Lipworth BJ.
Effect of b blockers in treatment of chronic obstructive pulmonary 28. Kattan MW, Gerds TA. A framework for the evaluation of statistical
disease: a retrospective cohort study. BMJ. 2011;342:d2549. prediction models. Chest. 2020;158(suppl 1):S29-S38.
12. Skull SA, Andrews RM, Byrnes GB, et al. Hospitalized community- 29. von Elm E, Altman DG, Egger M, et al. The Strengthening the
acquired pneumonia in the elderly: an Australian case-cohort study. Reporting of Observational Studies in Epidemiology (STROBE)
Epidemiol Infect. 2009;137(2):194-202. statement: guidelines for reporting observational studies. Int J Surg.
2014;12(12):1495-1499.
13. Altman DG. Practical Statistics for Medical Research. Boca Raton,
FL: CRC Press; 1990. 30. Moons KG, Altman DG, Reitsma JB, et al. Transparent Reporting of
a multivariable prediction model for Individual Prognosis or
14. Everitt BS. Statistical Methods for Medical Investigations. London, Diagnosis (TRIPOD): explanation and elaboration. Ann Intern Med.
England: Edward Arnold; 1994. 2015;162(1):W1-W73.
15. Kahn HA, Sempos CT. Statistical Methods in Epidemiology. New 31. Collins GS, Reitsma JB, Altman DG, Moons KG.
York, NY: Oxford University Press; 1989. Transparent Reporting of a multivariable prediction
16. Fleiss JL, Tytun A, Ury HK. A simple approximation for calculating model for Individual Prognosis or Diagnosis
sample sizes for comparing independent proportions. Biometrics. (TRIPOD): the TRIPOD statement. Ann Intern Med.
1980;36(2):343-346. 2015;162(1):55-63.

S78 Supplement [ 158#1S CHEST JULY 2020 ]