OCR Text |
Show Editorial Review of Cohort Studies: A Companion Article to Malmqvist et al "Progression Over 5 Years of Prelaminar Hyperreﬂective Lines to Optic Disc Drusen in the Copenhagen Child Cohort 2000 Eye Study" Victoria L. Tseng, MD, PhD, Melinda Y. Chang, MD A cohort study is an observational study design that follows a group of individuals (the "cohort") over time to identify risk factors for an outcome of interest (usually a disease) (1). Some individuals in the cohort are exposed to the potential risk factor(s) being studied, whereas others are not exposed. The proportions of subjects in the exposed vs nonexposed groups who develop the outcome of interest are used to estimate the incidence of the outcome based on the exposure status. A cohort study can be closed or open. A closed cohort is one with ﬁxed membership, meaning that no one can be added to the cohort after follow-up begins, and the size of the population at risk decreases as people develop the outcome of interest. An open cohort can take additional members after follow-up has started, and the population at risk can change over the course of the study (2). Because cohort studies are nonrandomized, results may be affected by confounders-that is, factors that inﬂuence both the exposure and outcome. Investigators must therefore correct for potential confounders in statistical analyses. Cohort studies are used to calculate the relative risk (RR), which is a measure of the strength of association between an exposure and an outcome (see "statistical analysis for cohort studies"). A RR greater than 1 indicates exposed individuals are at a greater risk of the outcome than nonexposed individuals, whereas a RR less than 1 signiﬁes that people who are exposed are protected against the outcome. A RR equal to 1 means that exposed and nonexposed individuals have the same risk of developing the outcome-that is, the exposure does not affect the outcome. Cohort studies are an efﬁcient method to assess the inﬂuence of rare exposures because the investigator can enrich the cohort with subjects who are exposed. They are generally less efﬁcient to study rare outcomes because a large cohort will be required, which is expensive and time consuming. To reduce expense and effort for analyzing a rare outcome within a cohort study, a nested case control study within a cohort study may be performed by analyzing only a subset of the cohort (3), as in the accompanying article by Malmqvist et al "Progression Over 5 Years of Prelaminar Hyperreﬂective Lines to Optic Disc Drusen in the Copenhagen Child Cohort 2000 Eye Study" (4). In this article, we will review types of cohort studies, relevant statistics, and advantages and disadvantages of this statistical design. COHORT STUDY DESIGNS Table 1 provides a summary of 4 different cohort study designs that will be described below. Department of Ophthalmology (VLT), Stein Eye Institute at the University of California, Los Angeles, California Heed Ophthalmic Fellowship (VLT), San Francisco, California; The Vision Center at Children's Hospital of Los Angeles (MYC), Los Angeles, California; and Department of Ophthalmology (MYC), Roski Eye Center, University of Southern California, Los Angeles, California. Supported by an Unrestricted grant to the Department of Ophthalmology from Research to Prevent Blindness, New York, NY. The authors report no conﬂicts of interest. Address correspondence to Melinda Y. Chang, MD, Vision Center, Children's Hospital of Los Angeles, 4650 Sunset Blvd, Mailstop #88, Los Angeles, CA 90027; E-mail: Melinda.y.wu@gmail.com 286 Tseng and Chang: J Neuro-Ophthalmol 2020; 40: 286-291 Copyright © North American Neuro-Ophthalmology Society. Unauthorized reproduction of this article is prohibited. Editorial TABLE 1. Summary of 4 different cohort study designs Exposure Prospective cohort study Retrospective cohort study Risk factor present before the outcome develops Risk factor present before the outcome develops Nested case-control study Risk factor present before the outcome develops Case-cohort study Risk factor present before the outcome develops Controls Statistics Reported All subjects who develop the outcome during follow-up All subjects who develop the outcome during follow-up All noncases Relative risk All noncases Relative risk All subjects who develop the outcome during follow-up Noncases that are matched to cases Odds ratio All subjects who develop the outcome during follow-up Noncases from within a randomly selected subcohort Weighted Cox proportional hazards regression model Outcome Cases Disease of interest and diagnosed after the study begins Disease of interest and diagnosed before the study begins Disease of interest and diagnosed before or after the study begins Disease of interest and diagnosed before or after the study begins Prospective Cohort Design In a prospective cohort study, the investigator chooses the cohort at the beginning of the study and follows the subjects longitudinally. The exposure status is assessed at the start of the study, and cases are identiﬁed as subjects who develop the outcome of interest. The outcome may be assessed at the end of the study or periodically during the study period. Data on potential confounders may also be collected during the study period. A neuro-ophthalmic example of a prospective cohort study is the pediatric optic neuritis prospective outcomes study, in which children with optic neuritis were enrolled at diagnosis and will be followed prospectively for 2 years (5). One of the outcomes that will be examined is development of multiple sclerosis. Data on exposures, such as treatment with steroids, were collected. The investigators will calculate the RR of multiple sclerosis in children treated with steroids, to determine if this exposure affects the outcome in children with optic neuritis. Retrospective Cohort Design In a retrospective cohort study, the investigator reviews records of a cohort of subjects and determines the exposure status at an earlier point in time. The outcome of interest is evaluated at a later time point, and the RR is calculated. For example, Lam et al performed a retrospective cohort study to determine whether cataract surgery in patients with unilateral nonarteritic anterior ischemic optic neuropathy (NAION) increased the risk of fellow eye NAION (6). They performed a retrospective chart review of the medical records of all patients diagnosed with unilateral NAION at their institution over a 15-year period. The exposure was cataract surgery, and the outcome was fellow eye NAION. Statistical analysis calculated the RR of fellow eye Tseng and Chang: J Neuro-Ophthalmol 2020; 40: 286-291 NAION in patients who underwent cataract surgery compared with those who did not. Because statistical analysis in a retrospective cohort study is performed post-hoc (i.e., the research question is not posed until after the data are collected), the data collection may be incomplete, inaccurate, or inconsistent. In addition, data on potential confounders may not be available. Therefore, compared with prospective cohort studies, retrospective cohort studies are more susceptible to bias from confounding and certain types of information bias, such as misclassiﬁcation (see section below on "bias in cohort studies"). Nested Case-Control Design A nested case-control design allows an investigator to perform additional analyses on a subset of the full cohort, reducing time and expense (7). This design may be especially useful for examining rare outcomes from within a larger cohort study. Cases who developed the outcome of interest are matched to those who did not (controls) within the cohort. The exposure status of the cases and controls is determined, and this is used to calculate the odds ratio (OR) for developing the outcome. Similar to RR, the OR calculates the strength of association between an exposure and outcome (see "statistical analysis for cohort studies" for calculation). However, ORs compare the frequency of the exposure based on the outcome status, whereas RRs compare the frequency of the outcome based on the exposure status (3). ORs do not take into account the absolute risk of developing an outcome and are therefore more appropriate than risk ratios in case-control studies, where 287 Copyright © North American Neuro-Ophthalmology Society. Unauthorized reproduction of this article is prohibited. Editorial absolute risk is unknown. ORs approximate RRs when outcomes are rare. A nested case-control design is useful when the exposure status would be expensive or time consuming to obtain for the entire cohort. For example, in the accompanying article by Malmqvist et al (4), a nested case-control design was used to assess the scleral diameter in children with optic disc drusen (ODD) and controls, which enabled the authors to measure the scleral diameter in only 115 controls instead of 708 controls in the full cohort. Another advantage in using the nested case-control design instead of analyzing the full cohort is the potential to avoid sparse data bias when outcomes are rare (8) because cases are enriched in the nested case-control analysis. Case-Cohort Design A case-cohort study is similar to a nested case-control study in that a subset of the full cohort is chosen for analysis of a particular exposure. In a case-cohort study, a group of subjects from the full cohort is randomly selected to be included in a subcohort, regardless of the ﬁnal outcome. All of the cases in the full cohort who develop the outcome of interest, in addition to the controls in the subcohort, are evaluated for the exposure. Statistical analysis is more complicated because it must take into account the over representation of cases in the sample (a weighted Cox proportional hazards regression model, which is outside the scope of this article, is used) (9). The advantage of the case-cohort design over a nested case-control study is that the subcohort may be selected before cases develop the outcome of interest, and the subcohort may be used to study multiple outcomes because matching is not required, and the case status is irrelevant. As an example, Schrijvers et al performed a casecontrol study nested within the prospective populationbased Rotterdam cohort study (10). The investigators assessed whether plasma clusterin was associated with development of Alzheimer disease (AD). At the beginning of the study, 3,709 participants were at risk of AD. Fasting blood samples were drawn from all subjects. A random subcohort of 952 participants was identiﬁed. At follow-up, 52 participants in the subcohort developed AD, in addition to 156 in the rest of the cohort. Plasma clusterin levels were assessed in the blood samples of all 208 cases (156 + 52) and 900 controls (952-52) in the subcohort. Statistical analysis showed that plasma clusterin was not related to the risk of developing AD during follow-up. STATISTICAL ANALYSIS FOR COHORT STUDIES Cohort studies can be used to estimate risks and incidence rates. Risks are measured with individuals or other count 288 data (e.g., eye) as the unit in the numerator and denominator. The denominator represents the number at risk at baseline, and the numerator represents the number who develops the outcome. Risk can be calculated for groups with and without the exposure of interest, to determine the ratio of risk in the exposed vs the unexposed. This is called the risk ratio or RR. When calculating risk, measures of association may be calculated using Table 2. 1. Risk in exposed = R1 = A1/N1 Risk in unexposed = R0 = A0/N0 2. Risk ratio or RR = R1/R0 3. Absolute risk difference (attributable risk [AR]) = R1 2 R0 4. Number needed to harm (or treat) = 1/AR 5. OR = (A1/B1)/(A0/B0). Note: the OR would also be the unit of measure for a nested case-control study conducted within a cohort study. In the accompanying article by Malmqvist et al (4), the unit of measure for calculations is the eye. The authors reported that 21 eyes had ODD at ﬁnal follow-up. Of these 21 eyes, 16 eyes had ODD at baseline (age 11- 12 years). They performed a subgroup analysis of new ODD developing over 5 years of follow-up in these 16 eyes, which represents the total number at risk at baseline. Eight eyes developed new ODD at follow-up (age 17 years) at the location of prelaminar hyper-reﬂective lines seen on baseline optical coherence tomography. These lines were distant from the site of baseline ODD. Seven eyes developed new ODD during followup that were not at the location of previous prelaminar hyper-reﬂective lines. Presumably, the last eye did not develop new ODD. If we assume that distant prelaminar hyperreﬂective lines were not present in this eye at baseline, then the risk of new ODD development based on the presence of distant prelaminar hyper-reﬂective lines in eyes with ODD at baseline may be calculated as shown in Table 3: 1. Risk in exposed = R1 = 8/8 = 1.0 Risk in unexposed = R0 = 7/8 = 0.875 2. Risk ratio or RR = R1/R0 = 1.0/0.875 = 1.14 TABLE 2. Measures of association for calculating the risk Outcome No Outcome Exposed A1 B1 Not exposed A0 B0 Total # cases Total # non-cases Total # exposed = N1 Total # not exposed = N0 Total Tseng and Chang: J Neuro-Ophthalmol 2020; 40: 286-291 Copyright © North American Neuro-Ophthalmology Society. Unauthorized reproduction of this article is prohibited. Editorial TABLE 3. Calculation example Exposed = distant prelaminar hyper-reﬂective lines Not exposed = No distant prelaminar hyper-reﬂective lines Outcome = New ODD No Outcome = No New ODD A1 = 8 B1 = 0 Total # exposed (N1) = 8 A0 = 7 B0 = 1 Total # not exposed (N0) = 8 Total # cases = 15 Total # noncases = 1 Total = 16 3. Absolute risk difference (AR) = R12R0 = 1.020.875 = 0.125 4. Number needed to harm (or treat) = 1/AR = 1/0.125 = 8 5. OR = (A1/B1)/(A0/B0) = (8/0)/(7/1) (unusual situation because all of the exposed developed the outcome). Another unit of measure used in many cohort studies is the incidence rate and the corresponding rate ratio, which is the incidence rate in exposed vs unexposed. Similar to risk, incidence rates have a count unit in the numerator (e.g., persons, eyes). However, the denominator is calculated in person-time, which is the count unit multiplied by the amount of time each person or eye contributed to the study. Compared with the risk, the incidence rate is a more ﬂexible unit of measure that can be used in more situations. Speciﬁcally, when many subjects are lost in a study because of (1) developing the outcome, (2) death from a competing risk, or (3) loss to follow-up, measuring risk may be less accurate and incidence rates should be considered, because each person or eye contributes signiﬁcantly different amounts of time to the study and thus has differing lengths of time where they could develop the outcome of interest (11). In the accompanying article by Malmqvist et al (4), if several children in the study did not present for follow-up at age 17 years, then use of an incidence ratio would be appropriate. When calculating the incidence rate, measures of association may be calculated using the Table 4: 1. Incidence Incidence 2. Incidence 3. Incidence rate rate rate rate in exposed = I1 = A1/T1 in unexposed = I0 = A0/T0 ratio (IRR) = I1/I0 difference (IRD) = I12I0 ADVANTAGES OF COHORT STUDIES Cohort studies are most similar to standard experimental strategies, with the exception that exposures are observed rather than assigned (2,11). As such, cohort studies are secondary to randomized controlled trials in validity ranking. Cohort studies allow the investigation of multiple exposures over time without assigning individuals to randomized treatments. Because cohort studies are conducted over time, temporal associations can be established between exposures and outcomes, and disease rates and time to event can be calculated. Cohort studies can be conducted Tseng and Chang: J Neuro-Ophthalmol 2020; 40: 286-291 retrospectively from existing databases, which can improve study efﬁciency and reduce expense compared with randomized trials or prospective studies. BIAS IN COHORT STUDIES Selection Bias Selection bias occurs when the association between an exposure and outcome is not the same for those who participate and do not participate in a study (2,11). Selection bias can arise from self-selection into a study, which can happen both from the participant or the investigator. In a prospective cohort study, a participant may volunteer for a study because of special concern for the exposure or outcome under investigation and as such, this participant may have reasons for self-referral that are associated with the outcome under study. Conversely, investigators may choose to compare outcomes from a speciﬁc group in a study with outcomes in the general population; for example, comparing the mortality rate in a group of active working individuals with the mortality rate in the general population. Although the mortality rate in the working population may be lower, this may be due to factors that allow this population to keep working rather than due to the exposure under study. This is referred to as the "healthy worker effect" and has traditionally been classiﬁed as a form of selection bias, although it may actually represent confounding. In both prospective and retrospective cohort studies, selection bias can occur when there is loss to follow-up within the cohort, that is, differential regarding both disease and/or exposure. In a hypothetical example from the accompanying article by Malmqvist et al (4), parents of children with hyper-reﬂective prelaminar lines may be more likely to bring their children for follow-up examination at TABLE 4. Calculating measures of association Exposed Not exposed Cases Person-Time (Unit = Time21) A1 A0 T1 T0 289 Copyright © North American Neuro-Ophthalmology Society. Unauthorized reproduction of this article is prohibited. Editorial the age of 17 years and parents of children without these ﬁndings may be more likely not to present for follow-up. This would lead to selective dropout of children without the exposure. This differential dropout related to the exposure would lead to selection bias within the study. Selection bias is not easily addressed by common statistical methods, and thus must be taken carefully into account when the study is designed and participants are recruited. Confounding Confounding is a central source of bias for all epidemiologic studies. It refers to a confusion of effects, where the effect of the exposure is mistaken for the effect of another variable. From the deﬁnition by Rothman and colleagues (2,11), a confounding factor (1) must be an extraneous risk factor for the disease, (2) must be associated with the exposure under study in the population at risk from which the cases are derived, and (3) must not be affected by the exposure or the disease and cannot be an intermediate step in the causal path between the exposure and the disease. In both prospective and retrospective cohort studies, confounding can be addressed at multiple stages. At the design stage, the study can be restricted to certain groups (e.g., if sex is a potential confounder, the study can be conducted only in women or only in men), or subjects can be matched on potential confounding factors. In the analysis stage, confounding can be addressed with multivariable analyses, stratiﬁed analyses, or more advanced statistical methods such as propensity scores or marginal structural models. In the accompanying article by Malmqvist et al (4), all study covariates could potentially be considered as confounders, and the authors could consider multivariable analyses to assess these factors. Information Bias Bias may occur from measuring the variables used in a study (2,11). For example, if administrative billing codes are used to assess the presence of a disease in a retrospective study, there may be people who have the disease who do not have the code present, or people without the disease who are erroneously coded as having the disease. This type of bias is often referred to as misclassiﬁcation. Misclassiﬁcation can be differential, meaning it differs according to values of other study variables, or nondifferential, meaning the misclassiﬁcation is unrelated to other study variables. A common type of differential misclassiﬁcation is recall bias, where participants are asked about exposure information after the outcome is already known. Conversely, if a study is conducted using administrative billing codes that were collected for clinical purposes without knowledge of future studies that would use those codes, any coding misclassiﬁcation would more likely be nondifferential. As an example, Chang and Keltner performed a retrospective cohort study of risk factors for fellow eye 290 involvement in patients with unilateral NAION seen at a single institution over a 10-year period (12). They identiﬁed patients with unilateral NAION by searching the electronic medical record for patients with International Classiﬁcation Disease (ICD) codes for ischemic optic neuropathy. The charts were then reviewed to determine whether criteria for NAION diagnosis were met. One of the risk factors identiﬁed for fellow eye NAION was noncompliance with continuous positive airway pressure (CPAP) in patients with moderate-to-severe obstructive sleep apnea. The study was subject to differential misclassiﬁcation by recall bias because patients who developed NAION in the fellow eye may have been more likely to recall noncompliance with CPAP. If they had performed the study using ICD codes only, there would also have been signiﬁcant nondifferential misclassiﬁcation bias because the authors found that only 43% of charts coded as "ischemic optic neuropathy" met the criteria for NAION. In general, nondifferential misclassiﬁcation tends to lead to more predictable patterns of bias toward the null value, whereas the pattern of bias with differential misclassiﬁcation is less predictable. A complicating layer in cohort studies is the factor of time because participants are being followed over extended periods, and the status of some exposures such as exercise or smoking could change over time. These factors must also be taken into account during design and analysis. In the accompanying article by Malmqvist et al (4), the criteria for both the exposure (hyper-reﬂective prelaminar lines) and outcome (ODD) were both clearly described and graded in a standardized fashion, which makes the possibility of misclassiﬁcation low. CONCLUSION Cohort studies are effective for performing research on populations over time in a real-world setting. When performing a cohort study, investigators may choose from a variety of prospective and retrospective study designs. The most commonly used statistical measure in a cohort study is the RR, but other measures such as the incidence ratio may be calculated. Because of the observational nature of cohort studies, biases must be considered in study design and analysis. As demonstrated in the article by Malmqvist et al (4), cohort studies can provide useful information on a multitude of exposures for chronic diseases. REFERENCES 1. Miettinen O. Design options in epidemiologic research. An update. Scand J Work Environ Health. 1982;8(suppl 1):7-14. 2. Rothman KJ. Epidemiology: An Introduction. New York, NY: Oxford University Press, 2012. 3. Hamedani AG, Pineles SL, Moss HE. The case-control study in neuro-ophthalmology. J Neuroophthalmol. 2020;40:144-147. 4. Malmqvist L, Li XQ, Hansen MH, Thomsen AK, Skovgaard AM, Olsen EM, Larsen M, Munch IC, Hamann S. Progression over 5 Years of prelaminar hyperreﬂective lines to optic disc drusen in the Tseng and Chang: J Neuro-Ophthalmol 2020; 40: 286-291 Copyright © North American Neuro-Ophthalmology Society. Unauthorized reproduction of this article is prohibited. Editorial Copenhagen Child cohort 2000 eye study. J Neuroophthalmol. 2020. doi: 10.1097/WNO.0000000000000911. 5. Pineles SL, Liu GT, Waldman AT, Lazar E, Kupersmith MJ, Repka MX, Pediatric eye disease investigator G, the neuroophthalmology research disease investigator C. Pediatric optic neuritis prospective outcomes study. J Neuroophthalmol. 2016;36:115-117. 6. Lam BL, Jabaly-Habib H, Al-Sheikh N, Pezda M, Guirgis MF, Feuer WJ, McCulley TJ. Risk of non-arteritic anterior ischaemic optic neuropathy (NAION) after cataract extraction in the fellow eye of patients with prior unilateral NAION. Br J Ophthalmol. 2007;91:585-587. 7. Ernster VL. Nested case-control studies. Prev Med. 1994;23:587-590. Tseng and Chang: J Neuro-Ophthalmol 2020; 40: 286-291 8. Greenland S, Mansournia MA, Altman DG. Sparse data bias: a problem hiding in plain sight. BMJ. 2016;352:i1981. 9. Sharp SJ, Poulaliou M, Thompson SG, White IR, Wood AM. A review of published analyses of case-cohort studies and recommendations for future reporting. PLoS One. 2014;9:e101176. 10. Schrijvers EM, Koudstaal PJ, Hofman A, Breteler MM. Plasma clusterin and the risk of Alzheimer disease. JAMA. 2011;305:1322-1326. 11. Rothman KJ, Greenland S, Lash TL. Modern Epidemiology, 3rd edition. Philadelphia, PA: Lippincott Williams & Wilkins, 2008. 12. Chang MY, Keltner JL. Risk factors for fellow eye involvement in nonarteritic anterior ischemic optic neuropathy. J Neuroophthalmol. 2019;39:147-152. 291 Copyright © North American Neuro-Ophthalmology Society. Unauthorized reproduction of this article is prohibited. |