Contents
On this page
- Outcome type as the determinant of the statistical pathway
- Analysis of a binary outcome: relative risk and logistic regression
- Analysis of a time-to-event outcome: Kaplan-Meier estimation and Cox regression
- Sample size, power and planning for attrition
- Testing whether an association differs across subgroups
- Continuous outcomes, confounding and common pitfalls
Academy · Research Data, Analytics & AI · Cohort Study Statistical Analysis
Possible Data and Statistical Analyses for a Research Question Based on a Prospective Cohort Study: Matching the Model to the Outcome Type
The research question of a prospective cohort study determines its statistical analysis to a greater extent than is often anticipated: whether the outcome is binary, continuous or time-to-event changes the entire analytic pathway, from relative risk to Cox regression, before any model is fitted. This article sets out how each outcome type maps to an appropriate analysis.
A prospective cohort study follows a defined group of people forward in time, measuring exposures at baseline and recording which participants do and do not develop an outcome of interest. The appropriate statistical analysis depends almost entirely on how the outcome is measured: whether it is binary (developed the condition or not), continuous (a measured value such as blood pressure), or time-to-event (not only whether, but when, an event occurred). Selecting the analysis before data collection, rather than afterwards, is what enables the eventual results of a cohort study to answer the question it was designed to address.
Learning objectives
- How the outcome type of a cohort study determines the applicable statistical analysis
- The difference between relative risk and the hazard ratio, and the circumstances in which each is the appropriate measure
- Why time-to-event data require Kaplan-Meier estimation and Cox regression rather than simple proportions
Common misconceptions
- “Relative risk and the hazard ratio are two names for the same quantity, and either may be reported for any cohort outcome.” Relative risk compares the cumulative proportion of each group experiencing an outcome by a fixed point, whereas the hazard ratio, produced by Cox proportional hazards regression, compares the instantaneous rate of the event over the full follow-up period and accommodates participants who leave the study or are observed for different lengths of time, which a simple relative risk calculation does not.
Outcome type as the determinant of the statistical pathway
Before any statistical test can be chosen, the research question must specify exactly what the outcome is and how it will be measured, because the analysis differs according to whether the outcome is binary, continuous or time-to-event. Song and Chung’s widely used methods primer on cohort and case-control designs frames this as a design decision made at the outset of a prospective cohort study, not an afterthought left to the analysis stage, because the sampling and follow-up plan must support whichever analysis the outcome type will eventually require1. A question asking “does Exposure X increase the risk of Outcome Y”, with Y measured as present or absent, calls for a binary-outcome analysis; the same question with Y measured as “time until Outcome Y occurs” calls for a time-to-event analysis instead, and the two are not interchangeable after the fact.
Analysis of a binary outcome: relative risk and logistic regression
When the outcome is binary and measured at a fixed follow-up point, the most direct measure of association is relative risk, the ratio of the cumulative incidence in the exposed group to the cumulative incidence in the unexposed group, which cohort studies can calculate directly because they track exposure forward to outcome, unlike case-control designs1. When the analysis needs to adjust for multiple covariates simultaneously, logistic regression is the standard extension, producing an adjusted odds ratio rather than a crude relative risk; the two measures converge when the outcome is rare but diverge as outcome prevalence rises, which is why the STROBE reporting guidelines specifically call for authors to state which measure was used and to justify it in the context of the outcome's frequency2.
Analysis of a time-to-event outcome: Kaplan-Meier estimation and Cox regression
When the research question is about not just whether but when an event occurs, and when participants are followed for different lengths of time or leave the study before the follow-up period ends, a time-to-event, or survival, analysis is the appropriate choice rather than a simple proportion. The Kaplan-Meier method estimates the probability of remaining event-free over time directly from the observed data, including participants who are lost to follow-up partway through, by treating their last known event-free time as a censored observation rather than discarding it3. To estimate the effect of an exposure on the rate of an event while adjusting for other covariates, Cox’s 1972 proportional hazards model remains the standard approach; it produces a hazard ratio, the relative instantaneous rate of the event between exposure groups, without requiring the underlying distribution of event times to be specified in advance, which accounts for its wide adoption across very different types of outcome4.
Sample size, power and planning for attrition
Because the statistical power of a prospective cohort study depends directly on the number of outcome events observed, not simply on the number of participants enrolled, sample-size planning must begin from an estimate of the expected event rate in the exposed and unexposed groups, together with the smallest effect size that the study would consider clinically or scientifically meaningful. For a rare outcome, this often requires a much larger cohort than intuition would suggest, since a study of a thousand participants with a two percent annual event rate will accumulate far fewer events over a given follow-up period than the same thousand participants followed for an outcome occurring in twenty percent of the exposed group. Attrition compounds this problem in time-to-event designs, because participants who are lost to follow-up stop contributing additional at-risk time to the analysis, which reduces the power of the study below what the enrolled sample size would suggest; building an explicit attrition allowance into the initial sample-size calculation, rather than treating loss to follow-up as an unplanned shortfall discovered during the study, is standard practice for this reason6.
Testing whether an association differs across subgroups
When the research question of a cohort study asks whether an exposure–outcome association differs across a subgroup, such as age band, sex or baseline severity, it concerns a moderator, and the correct analysis adds a formal interaction term between the exposure and the candidate moderator to the regression model, rather than presenting the association separately within each subgroup and comparing the two estimates informally. Baron and Kenny’s foundational distinction between moderation and mediation applies here: a cohort study testing whether baseline body mass index (BMI) changes the strength of the relationship between physical activity and cardiovascular risk is testing a moderator and requires an interaction term in either a logistic or a Cox model, whereas a cohort study testing whether the effect of physical activity on cardiovascular risk operates partly through reduced blood pressure is testing a mediator and requires a different mediation analysis with an estimated indirect effect. Subgroup analyses that are conducted and reported without this formal interaction test, simply because the data were available by subgroup, are a well-documented source of spurious, non-replicable findings in cohort research; for this reason, a subgroup or moderator analysis should be pre-specified in the study protocol rather than added on an exploratory basis once the primary results are known.
Continuous outcomes, confounding and common pitfalls
When the outcome is a continuous measurement, such as a change in blood pressure or a biomarker level, linear regression is the standard choice, with the exposure and any covariates entered directly as predictors of the continuous outcome, producing a coefficient interpretable as the average change in the outcome per unit change in exposure. Across all three outcome types, confounding remains the central threat to validity in an observational cohort design, since exposure was not randomly assigned; adjustment through regression, matching or stratification should be planned according to which variables are plausible common causes of both exposure and outcome, not simply which variables are available in the dataset1. A second common error is to treat a non-significant p-value as proof of no effect; Sterne and Davey Smith’s widely cited critique of significance testing argues that a p-value alone, without the accompanying effect size and confidence interval, systematically misleads readers about what the result of a cohort study shows, particularly when the sample size was too small to detect a real but modest effect5.
Summary: selection of the model according to the research question rather than the dataset
The principle on which this analysis plan rests is that the outcome type named in the research question, decided before data collection, should determine the statistical model, rather than the model being chosen afterwards according to which analysis the available dataset makes convenient. A cohort dataset that recorded both a binary indicator and a precise event date for the same outcome can technically support either a logistic regression or a Cox model, but the research question, specifically whether it concerned occurrence by a fixed point or timing across the whole follow-up period, should determine which is reported as the primary analysis; the other, if conducted at all, should be clearly labelled as a secondary or sensitivity analysis rather than presented as an equally valid alternative primary result.
Key points
Key points
The statistical analysis of a prospective cohort study follows directly from its outcome type: binary outcomes require relative risk or logistic regression, time-to-event outcomes require Kaplan-Meier estimation and Cox proportional hazards regression, and continuous outcomes require linear regression, with adjustment for confounding and reporting of effect sizes required in all three cases.
Practical application
Practical application
- The outcome type (binary, continuous or time-to-event) should be decided before the data collection plan is finalised. This single decision determines the entire downstream analytic pathway and cannot readily be changed once data collection has ended.
- For time-to-event outcomes, censoring should be planned for from the outset. Exact dates of loss to follow-up and study withdrawal should be recorded so that Kaplan-Meier and Cox models can treat these participants correctly rather than discarding their partial follow-up data.
- Effect sizes and confidence intervals should be reported alongside any p-value. A p-value alone does not convey the magnitude or precision of an association, and reporting statistical significance alone risks misrepresenting the findings of a study.
Frequently asked questions
Relative risk is appropriate when every participant is followed for the same fixed length of time and the outcome is simply present or absent by that point; once follow-up times vary or participants are lost before the end of the study, a hazard ratio from a Cox model is the more appropriate and more statistically efficient measure.
It can, provided that every participant has complete, equal-length follow-up and the exact timing of the event is not the focus of the research question; if timing matters, or follow-up lengths differ meaningfully across participants, Cox regression is more appropriate because it explicitly accounts for both.
When an outcome is common rather than rare, the odds ratio overstates the size of the association relative to the relative risk, which can mislead readers about the real-world magnitude of an effect if the two measures are not clearly distinguished in the reporting.
A commonly cited rule of thumb is roughly 10 outcome events per covariate included in a regression model, though this varies by context; a model with too many covariates relative to the number of observed outcome events risks producing unstable, unreliable estimates.
No. If participants who are lost to follow-up are systematically different from those who remain, for example because they were sicker or healthier, treating the loss as random censoring in a Kaplan-Meier or Cox analysis can bias the result, and sensitivity analyses exploring different assumptions about why participants left are advisable.
References
- Song JW, Chung KC. Observational Studies: Cohort and Case-Control Studies. Plastic and Reconstructive Surgery. 2010;126(6):2234-2242. read 10 October 2026. View on PMC
- von Elm E, Altman DG, Egger M, Pocock SJ, Gøtzsche PC, Vandenbroucke JP; STROBE Initiative. The Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement: guidelines for reporting observational studies. Lancet. 2007;370(9596):1453-1457. read 10 October 2026. View on The Lancet
- Clark TG, Bradburn MJ, Love SB, Altman DG. Survival analysis part I: basic concepts and first analyses. British Journal of Cancer. 2003;89(2):232-238. read 10 October 2026. View via DOI
- Cox DR. Regression models and life-tables. Journal of the Royal Statistical Society Series B. 1972;34(2):187-220. read 10 October 2026. View via DOI
- Sterne JAC, Davey Smith G. Sifting the evidence-what's wrong with significance tests? BMJ. 2001;322(7280):226-231. read 10 October 2026. View on BMJ
- Farrugia P, Petrisor BA, Farrokhyar F, Bhandari M. Practical tips for surgical research: Research questions, hypotheses and objectives. Canadian Journal of Surgery. 2010;53(4):278-281. read 10 October 2026. View on PMC
- Baron RM, Kenny DA. The moderator-mediator variable distinction in social psychological research: conceptual, strategic, and statistical considerations. Journal of Personality and Social Psychology. 1986;51(6):1173-1182. read 10 October 2026. View via DOI
- Munn Z, Moola S, Riitano D, Lisy K. The development of a critical appraisal tool for use in systematic reviews addressing questions of prevalence. International Journal of Health Policy and Management. 2014;3(3):123-128. read 10 October 2026. View on IJHPM
- Page MJ, McKenzie JE, Bossuyt PM, Boutron I, Hoffmann TC, Mulrow CD, et al. The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ. 2021;372:n71. read 10 October 2026. View on BMJ
- Richardson WS, Wilson MC, Nishikawa J, Hayward RS. The well-built clinical question: a key to evidence-based decisions. ACP Journal Club. 1995;123(3):A12-A13. read 10 October 2026. View via DOI
Every source above is a peer-reviewed article, a publisher’s record of one, or a standard issued by the body that owns it. Each carries the date it was last read. Where a standard moves, this article is revised rather than left up with a note.
How to cite this article
Pick your referencing style and copy it. Your university or target journal may vary the punctuation — check its guide if it has one.
Pubrica Academy. (2026, October 10). Possible Data and Statistical Analyses for a Research Question Based on a Prospective Cohort Study: Matching the Model to the Outcome Type. Pubrica. https://pubrica.com/academy/statistical/hypothetical-research-question-based-on-prospective-cohort-study/
Pubrica Academy. “Possible Data and Statistical Analyses for a Research Question Based on a Prospective Cohort Study: Matching the Model to the Outcome Type.” Pubrica, 10 Oct. 2026, pubrica.com/academy/statistical/hypothetical-research-question-based-on-prospective-cohort-study/.
Pubrica Academy (2026) Possible Data and Statistical Analyses for a Research Question Based on a Prospective Cohort Study: Matching the Model to the Outcome Type. Pubrica. Available at: https://pubrica.com/academy/statistical/hypothetical-research-question-based-on-prospective-cohort-study/ (Accessed: 15 October 2026).
Pubrica Academy. “Possible Data and Statistical Analyses for a Research Question Based on a Prospective Cohort Study: Matching the Model to the Outcome Type.” Pubrica. October 10, 2026. https://pubrica.com/academy/statistical/hypothetical-research-question-based-on-prospective-cohort-study/
Pubrica Academy, ‘Possible Data and Statistical Analyses for a Research Question Based on a Prospective Cohort Study: Matching the Model to the Outcome Type’ (Pubrica, 10 October 2026) <https://pubrica.com/academy/statistical/hypothetical-research-question-based-on-prospective-cohort-study/> accessed 15 October 2026.
Pubrica Academy. Possible Data and Statistical Analyses for a Research Question Based on a Prospective Cohort Study: Matching the Model to the Outcome Type. [Internet]. Pubrica; 2026 Oct 10 [cited 2026 Oct 15]. Available from: https://pubrica.com/academy/statistical/hypothetical-research-question-based-on-prospective-cohort-study/
Pubrica Academy, “Possible Data and Statistical Analyses for a Research Question Based on a Prospective Cohort Study: Matching the Model to the Outcome Type,” Pubrica, Oct. 10, 2026. [Online]. Available: https://pubrica.com/academy/statistical/hypothetical-research-question-based-on-prospective-cohort-study/
@misc{/la_statistical_analysis_prospective_cohort_study_2026,
author = {{Pubrica Academy}},
title = {Possible Data and Statistical Analyses for a Research Question Based on a Prospective Cohort Study: Matching the Model to the Outcome Type},
howpublished = {Pubrica},
year = {2026},
month = {oct},
url = {https://pubrica.com/academy/statistical/hypothetical-research-question-based-on-prospective-cohort-study/},
note = {Accessed: <span data-cite-acc="iso">2026-10-15</span>}
}This article is written to be learned from. Cite it if you are referring to what it says; if you are using the evidence behind it, cite the sources listed above instead.
Cohort Study Analysis-Plan Worksheet
This worksheet should be used before data collection to align the analysis plan with the outcome type.
[ ] Defined the outcome as binary, continuous, or time-to-event [ ] If binary: planned for relative risk and/or logistic regression [ ] If time-to-event: planned to record exact event/censoring dates for Kaplan-Meier and Cox regression [ ] If continuous: planned for linear regression [ ] Identified plausible confounders to adjust for, based on subject-matter reasoning, not just data availability [ ] Planned to report effect sizes and confidence intervals alongside any p-value [ ] Planned a sensitivity analysis for participants lost to follow-up
Pubrica AcademyWritten and maintained by Pubrica’s subject matter experts. Articles are revised when the underlying standards change, and each article states the date on which it was last reviewed.
How Pubrica can help
Pubrica’s biostatistics team aligns the analysis plan of a prospective cohort study with its outcome type before data collection begins, and builds the Kaplan-Meier, Cox, logistic or linear models that the research question requires.