In the early 1960s, researchers in Cebu, the Philippines, began tracking a group of pregnant women through childbirth and into the early lives of their children. Over the following two decades, the Cebu Longitudinal Health and Nutrition Survey collected dietary recall data, anthropometric measurements, household asset information, and, in time, cognitive assessments for the index children as they moved into adolescence. What the dataset eventually demonstrated was at once obvious and profound: low-birthweight infants who also experienced linear growth faltering in the first two years of life carried measurably reduced cognitive capacity, lower school attainment, and reduced adult economic productivity. These were associations that no cross-sectional survey could have confirmed, because they depended on the temporal sequence - the exposure had to precede the outcome by years, sometimes decades. The Cebu study, later incorporated into the COHORTS consortium, became one of the most-cited bodies of evidence for early-life nutrition investment ( Victora et al., 2008 ).
That trajectory - patient, longitudinal, methodologically demanding, and ultimately revelatory - captures what cohort studies do and why they occupy a distinctive position in the epidemiological toolkit.
The Cohort Study Definition
A cohort study is an observational study design in which a defined group of individuals - the cohort - is identified on the basis of exposure status (or other shared characteristic) and followed forward in time to observe the development of outcomes. The cohort is assembled when participants are free of the outcome of interest, and the key exposure is measured at or near the start of follow-up. The central question cohort studies answer is: does exposure X increase the rate at which outcome Y occurs?
The cohort study definition distinguishes this design from two superficially similar alternatives. Unlike the cross-sectional study, which measures exposure and outcome simultaneously, the cohort design preserves temporal ordering - a prerequisite for causal inference. Unlike the case-control study, which begins by identifying outcomes and looks back, the cohort study looks forward, allowing direct calculation of incidence and relative risk in exposed and unexposed groups.
Prospective and Retrospective Cohorts
The prospective cohort - sometimes called a concurrent cohort - recruits participants and begins exposure measurement before any outcome has occurred. Follow-up proceeds in real time. The COHORTS consortium exemplifies this design at an ambitious scale: five birth cohort studies across Brazil, Guatemala, India, the Philippines, and South Africa were each started prospectively, with data collection extending 20 years or more before the adult outcomes of interest could be assessed.
Prospective designs offer control over data quality. The investigator defines exposure measurements in advance, selects appropriate instruments, trains field workers consistently, and can collect biosamples that would be unavailable in a retrospective study. The trade-off is time and expense: prospective cohorts studying outcomes with long latency require sustained funding, stable research infrastructure, and mechanisms to minimise attrition over years or decades.
The retrospective cohort reconstructs past exposures from existing records - employment files, medical charts, birth registers, or administrative databases - and follows subjects forward (often to the present). The design is far more efficient when suitable records exist. Occupational epidemiology has relied heavily on retrospective cohorts to study associations between workplace exposures and cancer or respiratory disease. In SSA contexts, retrospective cohorts are constrained by the quality of historical data; incomplete birth registration, missing household records, and under-resourced health information systems limit what can be reconstructed reliably.
A hybrid design - the ambidirectional cohort - collects both historical (retrospective) and prospective data. This is increasingly used in research settings with existing health and demographic surveillance infrastructure, where baseline demographic data have been collected continuously for years and can anchor a prospective component.
Exposure Measurement in Cohort Studies
The internal validity of a cohort study depends critically on accurate exposure measurement. Misclassification - assigning participants to the wrong exposure category - attenuates associations toward the null when it is non-differential (i.e., unrelated to outcome status), and can bias associations in any direction when it is differential.
For nutritional cohort studies, the challenges are severe. Dietary intake varies day to day, seasonally, and over the life course, meaning that a single 24-hour dietary recall at baseline may poorly represent habitual intake over the follow-up period. The STROBE statement - the reporting guideline for observational epidemiology developed by Vandenbroucke and colleagues - requires explicit description of how exposures were measured and validated, acknowledgement of measurement error, and sensitivity analyses where appropriate ( Vandenbroucke et al., 2007 ).
Biomarker-based measurement reduces but does not eliminate this problem. Serum ferritin, for example, reflects iron stores at the time of measurement but is confounded by acute-phase response. Plasma zinc concentrations are affected by the time of day, recent food intake, and concurrent infection. Retinol-binding protein, used as a proxy for vitamin A status, is depressed during the acute-phase response regardless of true vitamin A reserves. Researchers increasingly advocate for composite biomarker panels with concurrent acute-phase protein measurement as a partial solution ( Victora et al., 2010 ).
Follow-up, Attrition, and Bias
Long-term cohort studies face attrition - participants who are lost to follow-up before outcomes are observed. In settings with high internal migration, cross-border movement, and limited administrative identification systems, attrition can be substantial. The Navrongo Health Research Centre in northern Ghana, one of Africa’s longest-running HDSS sites, has developed household visitation protocols and community engagement mechanisms specifically designed to minimise loss to follow-up, but even with these investments attrition over 20-year periods is non-trivial.
Attrition becomes biasing when it is not random with respect to exposure and outcome - that is, when the characteristics of those lost to follow-up differ systematically from those who remain. If, for example, the most severely malnourished children in a birth cohort are disproportionately likely to die before the age at which cognitive assessment is conducted, the observed association between early malnutrition and cognitive outcomes will be attenuated relative to the true effect. This is the “healthy survivor” bias, and it is particularly relevant in high-mortality SSA settings where child mortality rates mean that cohorts narrow substantially over time.
Sensitivity analyses - such as worst-case scenario analysis assuming that all lost participants had the worst outcome, or multiple imputation of missing outcome data under plausible assumptions - are standard approaches to assessing the potential impact of attrition. The STROBE reporting guidelines require transparent accounting of retention, with numbers followed up at each stage presented alongside reasons for loss ( Vandenbroucke et al., 2007 ).
Confounding in Cohort Research
Even in the best-designed prospective cohort, exposure groups are not randomised - participants select (or are selected into) exposures by mechanisms that often reflect the same structural forces that determine outcomes. Socioeconomic position, maternal education, household food security, access to clean water, and proximity to health services are simultaneously correlated with dietary exposures, nutritional status, and disease outcomes. Any uncontrolled confounder can generate or distort an association.
The analytic response is multivariable adjustment, typically through regression models - logistic regression for binary outcomes, Cox proportional hazards for time-to-event data, linear regression for continuous outcomes. Adjusted models can control only for measured confounders, and measurement is always imperfect. Residual confounding - the fraction attributable to unmeasured or poorly measured confounders - is the dominant threat to validity in observational cohort research.
Propensity score methods, instrumental variable analysis, and directed acyclic graphs (DAGs) are increasingly used to sharpen causal inference in cohort data, but each makes assumptions that must be stated and evaluated. The STROBE statement recommends explicit description of which potential confounders were controlled for and the rationale for their inclusion.
The Nested Case-Control Design
When outcomes are rare and biomarker analysis is expensive, researchers sometimes embed a case-control study within an established cohort. In a nested case-control design, cases who develop the outcome during follow-up are matched to controls who were at risk at the time the case occurred but did not develop the outcome. Stored biological samples, collected at cohort entry, are retrieved and analysed only for cases and their matched controls rather than for the entire cohort.
This design is efficient: it reduces laboratory costs by a factor of several while preserving the temporal advantages of the cohort (exposure was measured before outcome). Because controls are drawn from the same cohort, selection bias from outside the study population is avoided. Incubation periods, rare cancers, and uncommon nutritional deficiency-disease pairings are well suited to this approach ( Grimes & Schulz, 2002 ).
Landmark Cohort Studies in Global Nutrition Research
The COHORTS consortium deserves extended attention. The five cohort studies - in Cali, Colombia; Pelotas, Brazil; New Delhi, India; Cebu, Philippines; and Johannesburg/Soweto, South Africa - enrolled a total of more than 7,000 participants at or near birth and followed them into adulthood. Pooled analyses demonstrated that children who were stunted at age two had, on average, lower adult height, lower lean mass, less schooling, lower adult earnings, and higher risk of overweight - a finding that shifted global nutrition policy decisively toward the first 1,000 days as the primary intervention window ( Victora et al., 2008 ).
The Demographic Surveillance Sites of the INDEPTH Network represent a different kind of cohort: population-based longitudinal surveillance rather than a hypothesis-driven birth cohort. INDEPTH sites in Ghana (Navrongo, Kintampo), Kenya (Kilifi, Kisumu), Tanzania (Ifakara, Mwanza), and more than 40 others across Africa and Asia continuously record vital events - births, deaths, migrations - in defined populations. These data permit the calculation of age-specific mortality rates, cause-specific mortality fractions, and longitudinal nutritional outcomes that no cross-sectional survey could generate. The comparative analyses published by Streatfield and colleagues represent some of the most extensive mortality data available for SSA populations outside of vital registration ( Streatfield et al., 2014 ).
The Pelotas Birth Cohort (1982, 1993, 2004 cohorts) in southern Brazil has provided unusually rich multigenerational data, with participants from the 1982 cohort now enrolling their own children. The longitudinal depth of this data has allowed analysis of how early nutritional exposures interact with later-life economic circumstances and dietary transitions. Although the Brazilian context is not directly transferable to SSA, the methodological innovations - including the use of life-course analytical frameworks - have influenced African cohort research ( Victora et al., 2010 ).
In SSA specifically, the Kintampo Health Research Centre in central Ghana and the KEMRI/Wellcome Trust Research Programme in Kilifi, Kenya, have generated prospective data on child mortality, malaria, nutritional status, and vaccination outcomes that have directly shaped national and international policy. These sites are formally affiliated with the INDEPTH Network and exemplify how sustained investment in research infrastructure pays epidemiological dividends over decades ( Sankoh & Byass, 2012 ).
The role of HDSS sites as platforms for embedding cohort sub-studies is examined further in Implementing HDSS in Sub-Saharan Africa .
Strengths of the Cohort Design
The cohort study has several genuine methodological advantages. Temporality is established directly: exposure precedes outcome by design. Multiple outcomes can be studied in relation to a single exposure - a birth cohort enrolled to study undernutrition can simultaneously generate data on infectious disease incidence, cognitive development, and later non-communicable disease risk. Incidence rates can be calculated directly, and absolute risk differences - which are policy-relevant in ways that relative risks are not - can be estimated. The design is generally free of recall bias because exposures are measured prospectively rather than reconstructed after outcomes are known.
For rare exposures, the cohort design is more efficient than the case-control. A study of nutritional exposures in pregnancy, for example, can enrol pregnant women and follow them and their offspring - even if the outcome of interest, such as neural tube defects or severe stunting, occurs in only a small proportion.
Limitations of the Cohort Design
Cohort studies are expensive, slow, and logistically demanding. For outcomes with long latency - cardiovascular disease, many cancers, cognitive decline - the time from cohort assembly to publishable results may span decades, outlasting research funding cycles and investigator careers. Cohort studies are inefficient for rare outcomes, since a very large population must be enrolled and followed to observe enough events for stable estimates; nested case-control designs partially address this.
Loss to follow-up, as discussed above, is both a practical and an analytic challenge. Confounding is controlled by measurement and adjustment but never eliminated entirely. And the cohort design, like all observational designs, cannot establish causality with the certainty of a well-executed randomised trial - residual confounding and selection forces are always present. The 2013 Lancet series acknowledged these limitations explicitly while arguing that the convergence of evidence from multiple cohort studies across diverse settings makes the early-life nutrition hypothesis as robust as observational science allows ( Black et al., 2013 ).
For a broader account of observational study designs and the measures cohort studies generate, see Epidemiology: Definition, Core Methods, and Applications .
Limitations of This Article
This account focuses on cohort studies as used in nutrition and global health research, particularly in SSA contexts. The methodological literature on cohort designs in pharmacoepidemiology, occupational health, and cancer epidemiology is substantially more developed in some respects - notably in causal inference methods and administrative data linkage - and is not fully covered here. The examples cited reflect the published literature available through 2024; ongoing cohorts will generate additional findings. No meta-analysis or systematic review of the cited studies was conducted; the examples are illustrative rather than exhaustive.
Frequently Asked Questions
What is the cohort study definition in simple terms? A cohort study follows a group of people - who share a common characteristic or exposure - over time to see who develops a particular outcome. By measuring exposure at the outset and observing what happens subsequently, cohort studies preserve temporal ordering and allow researchers to calculate the rate at which outcomes develop in exposed versus unexposed groups. The design is distinct from cross-sectional studies (which measure everything at once) and from case-control studies (which start from outcomes and look back).
What is the difference between a prospective and retrospective cohort study? A prospective cohort enrolls participants and begins data collection before any outcomes have occurred, following them forward in real time. A retrospective cohort uses pre-existing records to reconstruct past exposures and then follows participants (often to the present). Prospective designs offer better control over data quality; retrospective designs are more efficient but depend on the completeness and accuracy of historical records.
What is attrition bias and why does it matter? Attrition bias occurs when participants who are lost to follow-up differ systematically from those who remain, in ways related to both exposure and outcome. If, for example, the sickest participants are most likely to drop out or die before outcomes are assessed, the observed association between an exposure and the outcome will be distorted. Reporting guidelines such as STROBE require transparent accounting of loss to follow-up and, where possible, sensitivity analyses exploring the impact of attrition on conclusions.
What made the COHORTS consortium methodologically important? The COHORTS consortium was notable for combining data from five distinct birth cohort studies across four continents, allowing pooled analyses with sufficient statistical power to detect modest associations between early-life growth and adult outcomes. The harmonisation of measurement methods across sites - made possible by shared protocols for anthropometry, dietary assessment, and socioeconomic indicators - meant that pooled estimates were stronger than any single-country study could provide. The consortium’s findings directly shaped the global policy consensus around the first 1,000 days of life as the critical window for nutrition intervention.