ACADEMIC WRITING GUIDE

How to Conduct and Report Survival Analysis in a Dissertation (With Worked Examples)

By Acadelyra Editorial TeamPublished

Survival analysis helps you answer a question ordinary averages often miss: how long does it take for a defined event to occur when some participants leave observation before that event is observed? In a dissertation, the event might be student withdrawal, employee resignation, equipment failure or recovery. “Survival” means remaining free of your chosen event; it does not necessarily mean remaining alive.

A defensible analysis connects the research question, time origin, event definition, follow-up rules, estimator and interpretation. This guide develops a hypothetical student-withdrawal example from hand-calculated Kaplan–Meier estimates to adjusted Cox regression, then explains what belongs in your methodology and results chapters. Every dataset and numerical finding below is synthetic and illustrates a method, not evidence about real students or support programmes.

When does a dissertation need survival analysis?

Use time-to-event methods when both whether an event occurs and when it occurs matter, particularly when follow-up is incomplete or differs between participants. If two students remain enrolled when observation ends, neither has an observed withdrawal time. Excluding them wastes information; assigning the study-end date as their withdrawal time creates events that did not happen.

A binary model can answer whether withdrawal occurred by a specified horizon when that status is adequately observed, but it does not automatically solve unequal follow-up. A model of withdrawal time among withdrawers alone answers a different question and selects participants using their outcome. Start by writing the outcome and target comparison in your dissertation data analysis plan.

For the example, the question is: “From initial enrolment, how does the time to first formal withdrawal differ between students with standard support and students offered enhanced support?” Withdrawal is undesirable, so higher survival means longer retention. If the event were graduation, higher event-free survival would instead mean remaining ungraduated. Choose language that matches the event.

Define the clock, event and censoring before fitting a model

Choose a time origin that matches eligibility and exposure

The time origin is the point from which elapsed time is measured. We use initial enrolment, record time in months and classify support at enrolment. Do not start one group's clock at enrolment and another's at their first support visit. The latter group would have to remain enrolled long enough to attend, introducing a selection problem.

Define eligibility, baseline assessment and when participants become observable together. If students enter the records after enrolment but the scientific clock still starts at enrolment, consider delayed entry, also called left truncation: they join risk sets only after their entry time. If withdrawal is known only to have occurred between assessments, that is interval censoring rather than an exact event date. These structures require suitable methods; they are not repaired by rounding dates or using the ordinary right-censored example unchanged. The official Surv documentation distinguishes these data representations.

Record what is observed, including why follow-up stops

For right-censored data, record the earlier of the event time and the censoring time, together with an event indicator: event = 1 for an observed withdrawal and event = 0 for right censoring. A censored student was known to be withdrawal-free up to the relevant observation time; their later withdrawal time is unknown. Censoring does not establish that withdrawal never occurred.

Administrative censoring occurs because a planned observation limit is reached. Loss to follow-up occurs when reliable observation ends earlier. Both can be coded zero in the analysis indicator, but preserve their reasons separately. A fixed closing date is often easier to justify than unexplained disappearance, yet administrative censoring still needs consideration when enrolment periods, policy changes or recruitment timing relate to withdrawal.

The usual Kaplan–Meier analysis requires censoring to be noninformative for the survival comparison being estimated, within groups where appropriate. Cox regression ordinarily needs conditional independent censoring given the model covariates. These are substantive assumptions: students leaving the tracking system because they are about to withdraw could violate them. Clark and colleagues' introductory tutorial explains why incomplete event times need special treatment.

Survival probability, hazard and cumulative hazard are different

Three quantities that answer different time-to-event questions
QuantityMeaningAppropriate interpretation
Survival probability, S(t)Probability that event time T exceeds time tProportion expected to remain withdrawal-free beyond a stated month
Hazard, h(t)Instantaneous event rate conditional on being event-free just before tSpeed at which withdrawal occurs among students still at risk; units are inverse time
Cumulative hazard, H(t)Accumulated hazard from the origin to tIntegrated event rate, not a probability or a count of withdrawals

For continuous event times, H(t) = integral of h(u) from 0 to t and S(t) = exp(-H(t)). The event probability by time t is 1 - S(t) in the single-event setting used here. A hazard can exceed one in the chosen time units; a probability cannot. Avoid calling a hazard “the probability of withdrawing this month” without an explicitly discrete-time model.

These relationships describe population quantities. In finite samples, the Nelson–Aalen cumulative-hazard estimate, which sums event-count/risk-set fractions, does not generally give exactly the Kaplan–Meier estimate when exponentiated. The survfit.formula documentation separates the product-limit and cumulative-hazard estimation options.

Worked example 1: build the Kaplan–Meier curve by hand

A transparent hypothetical dataset

Suppose all students are enrolled and observable from the origin. Group A receives standard support; Group B is offered enhanced support at baseline. Times are recorded exactly in this teaching dataset. Earlier censoring represents loss to follow-up; censoring at the final observation month is administrative. No competing event is included.

Hypothetical student records used for the hand calculations
StudentBaseline support groupObserved monthsEvent indicatorObservation ends because
A1A: standard21Withdrawal
A2A: standard41Withdrawal
A3A: standard41Withdrawal
A4A: standard71Withdrawal
A5A: standard40Loss to follow-up
A6A: standard100Administrative limit
B1B: enhanced31Withdrawal
B2B: enhanced61Withdrawal
B3B: enhanced81Withdrawal
B4B: enhanced50Loss to follow-up
B5B: enhanced80Loss to follow-up
B6B: enhanced100Administrative limit

There are 12 students, 7 observed withdrawals and 5 censored observations. These counts describe observed records; they do not reveal how many censored students would eventually withdraw.

Calculate risk sets before each event

At an event time, the risk set contains students still observed and not previously withdrawn immediately before that time. Let n_j be its size and d_j the number of withdrawals at that time. The Kaplan–Meier estimate multiplies the conditional survival factors:

S_hat(t) = product over event times t_j <= t of (1 - d_j/n_j)

Start at survival probability one. Censoring changes later risk sets but does not create a downward step. When events and censoring have the same recorded time, this example uses the standard convention that those censored then remain in the risk set for the events at that time; censoring removes them afterward. If actual within-day ordering is known, preserve it rather than manufacture a tie.

Pooled hypothetical risk sets and Kaplan–Meier estimates after each recorded time
MonthAt risk immediately beforeWithdrawals at timeCensored at timeSurvival factorEstimated S(t)
2121011/120.9167
3111010/110.8333
410218/100.6667
570110.6667
66105/60.5556
75104/50.4444
84113/40.3333
1020210.3333

At month 4, two students withdraw and one is censored. All ten are at risk immediately before the time. Therefore, S_hat(4) = (11/12) × (10/11) × (8/10) = 0.6667, rounded. The subsequent risk set contains 10 - 2 - 1 = 7 students. Removing the censored student before calculating the withdrawal factor would use the wrong denominator.

The month-5 censoring leaves survival unchanged, but only six students remain for the month-6 event. Keep full precision through the calculation; round only the displayed estimates. The final flat portion means there were no further observed withdrawals while those last students were followed. It does not justify projecting the curve beyond observed support.

Interpret survival, uncertainty and the median

The pooled estimate at month 6 is 0.5556: estimated withdrawal-free survival beyond six months. Using Greenwood variance with a log–log transformation, its pointwise 95% confidence interval is 0.2311–0.7899. This interval is wide because the dataset is tiny. It is not a simultaneous confidence band for the entire curve or an interval containing an individual student's withdrawal time.

For reproducibility, Greenwood's accumulated term at month 6 is 1/(12×11) + 1/(11×10) + 2/(10×8) + 1/(6×5) = 0.0750. If g = log(-log(S_hat)), its approximate standard error is sqrt(0.0750)/abs(log(S_hat)); transform g ± 1.959964 × SE(g) back using exp(-exp(g)), reversing the endpoints. That produces the interval above. Near survival zero or one, boundary handling needs care; use documented software rather than forcing this expression through an undefined logarithm.

The estimated median is the first time the curve reaches or falls below 0.5. Here it is 7 months: survival is above one-half at month 6 and below it at month 7. It is not the median of the observed times after censoring times are treated as events. If a curve never reaches one-half, report “median not reached” or “not estimable within follow-up.” Median confidence limits require inversion of survival confidence limits; none are claimed for this small illustration.

What should the plot show?

A dissertation figure should show step curves, labelled time units, the event-free outcome, group names, censor marks, confidence intervals and a numbers-at-risk panel. State whether risk counts refer to immediately before each displayed time. Do not use a smooth line that implies unobserved changes between event times. A long tail supported by few students deserves visible caution.

In this example, at month 4 the group risk counts are A = 5 and B = 5 immediately before the events. At month 6 they are A = 2 and B = 4. These counts are neither cumulative enrolments nor the number who will eventually withdraw. Explain caption and panel conventions using the same care as other research tables and figures.

Worked example 2: compare the support groups

Group-specific Kaplan–Meier curves use each group's own risk sets. They are unadjusted descriptions; different baseline characteristics may contribute to the observed separation.

Hypothetical unadjusted group summaries from the small dataset
GroupStudentsObserved withdrawalsS(6)Pointwise 95% CI for S(6)Estimated median months
A: standard640.50000.1109–0.80374
B: enhanced630.62500.1419–0.89318

For Group A, survival falls to one-half at the tied month-4 withdrawals and stays there until month 7. For Group B, the month-5 censoring reduces the risk set to four before the month-6 withdrawal, producing S_B(6) = (5/6) × (3/4) = 0.6250. The apparent difference is too uncertain to support a substantive programme claim.

Calculate and interpret the log-rank test

The ordinary log-rank test compares observed events with those expected from the group risk-set proportions under the null of equal survival curves, assuming appropriate independent censoring. At each event time, E_Aj = d_j × n_Aj/n_j. Its variance contribution is n_Aj × n_Bj × d_j × (n_j-d_j) / (n_j² × (n_j-1)), which accounts for tied events.

Hypothetical log-rank contributions using the pooled event times
MonthA at riskB at riskA withdrawalsB withdrawalsExpected A withdrawalsVariance contribution
266100.50000.2500
356010.45450.2479
455201.00000.4444
624010.33330.2222
723100.40000.2400
813010.25000.1875

Across event times, A has 4 observed withdrawals, 2.9379 expected withdrawals and variance 1.5921. Using unrounded contributions gives chi-square = (O_A - E_A)² / V = 0.7086, with 1 degree of freedom and approximate p = 0.3999. R's survdiff documentation identifies rho = 0 as the ordinary log-rank test.

This p-value does not show that the programmes are equivalent. The test supplies no adjusted association or measure of effect magnitude, and its chi-square approximation is fragile in such a small sample. Proportional hazards are not required for the null hypothesis of identical survival, but the ordinary test is most sensitive to proportional-hazard alternatives. Crossing curves can make opposite differences cancel. Choose an alternative test or estimand for scientific reasons in advance, rather than trying several tests and selecting the smallest p-value.

Worked example 3: fit and interpret a Cox model

Move to an explicitly larger hypothetical cohort

The hand dataset is useful for understanding denominators but unsuitable for a credible multivariable dissertation model. We therefore generate a separate hypothetical cohort of 240 students with the same withdrawal question, baseline support definition and enrolment origin. It is not an expansion that recovers additional observations from the first dataset.

The generator assigns enhanced support with probability 0.5 and a baseline high-support-need indicator with probability 0.4, independently. Conditional event times follow an exponential distribution with monthly hazard 0.075 × exp(-0.5 × group + 0.7 × severity), where standard support and lower baseline need are coded zero. Independent potential loss occurs with probability 0.18, with its time uniform between months 2 and 12; other students have no potential early loss. Everyone has an administrative limit of 12 months. Only the earliest event, loss or administrative time is observed.

This fixed synthetic realisation contains 132 standard-support and 108 enhanced-support students, 140 withdrawals, 23 losses to follow-up and 77 administrative censorings. No covariate is missing. The mechanism deliberately satisfies conditional proportional hazards and independent censoring; it does not establish these conditions for empirical data. The self-contained R generator below defines the precise reproducible sequence.

Specify the adjustment set and the estimand

The Cox proportional hazards model is h(t given X) = h_0(t) × exp(beta_group × group + beta_severity × severity). The baseline hazard is left unspecified. A coefficient describes a log hazard ratio; exponentiating it gives a hazard ratio conditional on the other included covariates. The official coxph documentation describes model fitting and Efron, Breslow and exact tie-handling options.

Here the adjusted comparison is enhanced versus standard support among students with the same baseline-need category. We fit group alone for the unadjusted comparison and add the prespecified baseline indicator for the adjusted comparison. Both use the same students. In a real study, choose confounders using subject knowledge and the temporal relationships between variables; automatic p-value screening can omit important confounders. Adjusting for a mediator or a consequence of withdrawal risk may change or bias the question.

Executed synthetic Cox estimates with model-based Wald confidence intervals
Model and predictorLog HR coefficientStandard errorHazard ratio95% CI for HRTwo-sided Wald p-value
Group only: enhanced versus standard-0.2798090.1721580.75590.5394–1.05930.1041
Adjusted: enhanced versus standard-0.3382340.1738130.71300.5072–1.00240.0517
Adjusted: high versus lower baseline need0.4379590.1720331.54951.1060–2.17090.0109

These values were computed in Python by maximizing the Cox partial likelihood and independently checked using a separate CSV-based likelihood calculation with numerical derivatives. They are not copied or invented R output. Efron tie handling was specified; the generated event times are distinct, so alternative ordinary tie methods coincide for this cohort. Do not round the stored analysis times before fitting.

For the adjusted support coefficient, exponentiating the unrounded estimate gives HR = 0.7130. Its model-based 95% Wald interval is exp(beta ± 1.959964 × SE) = 0.5072–1.0024. Holding baseline need constant, the fitted withdrawal hazard for enhanced support is lower by approximately 28.70%, calculated as 100 × (1 - HR). The interval includes one; these synthetic results remain compatible with essentially no hazard difference as well as a substantial lower hazard. Avoid turning rounding or a threshold near the p-value into a categorical finding.

A hazard ratio is not a risk ratio

The support HR compares instantaneous withdrawal rates among students still event-free at each time, subject to the model. It does not mean that withdrawal probability by month 12 is lower by 28.70%, that retention probability is 0.7130, or that students stay enrolled 28.70% longer. Those statements concern different quantities.

Absolute survival prediction also needs an estimated baseline survival function and a stated covariate profile or standardization population. Under proportional hazards, a profile's survival is S_0(t)^exp(X beta); a hazard ratio alone is insufficient. Hernán's discussion of hazard-ratio interpretation also cautions that comparing selected survivors complicates causal claims.

Adjustment can change a hazard ratio even without confounding because the measure is noncollapsible, as Dumas and Stensrud explain. The difference between our unadjusted and adjusted estimates therefore does not prove that baseline need was a confounder. The programme labels in this synthetic example do not justify a real-world recommendation.

Check proportional hazards and other model assumptions

The proportional hazards assumption means that a fitted covariate's hazard ratio is constant over follow-up, conditional on the model. It does not mean that hazards themselves are constant. A Cox baseline hazard may rise or fall while the group hazard ratio stays constant; our exponential generator uses a constant baseline only for simplicity.

Use scaled Schoenfeld residual plots to examine time patterns, together with term-specific and global tests such as cox.zph. A roughly horizontal estimated coefficient curve supports the working assumption; systematic trends raise concern. A large p-value does not prove proportionality, especially with few events. Inspect uncertainty and influential observations, and do not interpret crossing raw curves as a complete adjusted diagnostic.

Also assess the functional form of continuous covariates, sparse categories, influential records and dependence between observations. Consider splines for scientifically plausible nonlinear associations. Students clustered in institutions may need an appropriate clustered variance or multilevel survival model. A robust standard error addresses a variance issue; it does not repair a misspecified mean structure, informative censoring or an incorrect time origin.

The R code below includes diagnostic commands for readers to execute. No Schoenfeld test statistics, residual-plot findings or empirical assumption checks are claimed for this article: those R commands were not run here. Knowing the synthetic generating mechanism is different from diagnosing the fitted realisation.

Time-varying covariates are different from time-varying effects

A time-varying covariate changes value during follow-up, such as a student's current financial hardship. A time-varying effect means a covariate's association changes with elapsed time, even if its value was fixed at enrolment. The official time-dependent survival vignette explains both structures and the danger of using future information.

If enhanced support starts after enrolment, do not classify every eventual recipient as exposed from the origin. Students must remain enrolled long enough to receive it. Use correctly ordered start–stop records with exposure updated when it actually begins, or a justified landmark design. Covariate values must be available just before the event comparison. Time-dependent confounding may require methods beyond simply adding an updated covariate to Cox regression.

For nonproportional hazards, consider prespecified time interactions, a stratified Cox model for a nuisance covariate, flexible parametric survival models or an alternative estimand. Stratification allows separate baseline hazards, but does not estimate a hazard ratio for the stratifying variable. An accelerated failure-time model can estimate time ratios when its distributional assumptions are appropriate.

Restricted mean survival time is the area under the survival curve up to a chosen horizon. Its difference compares average event-free time accumulated within that horizon and need not assume proportional hazards. Specify a horizon with adequate follow-up support and report uncertainty; it is not an unrestricted lifetime mean. Royston and Parmar discuss this alternative. No restricted-mean estimate is calculated in our examples.

Missing data, competing risks and informative censoring

Separate missing covariates from incomplete event times

Right censoring records partial outcome information under a defined observation rule. A missing baseline predictor is a different problem; so is an unknown event date or an incorrectly recorded withdrawal status. Audit the dataset and report the amount and reasons for missingness before selecting a method. Do not automatically code an absent event record as event = 0 at the planned study end.

Complete-case analysis changes the analysed sample and can introduce bias. Where justified, multiple imputation should accommodate the survival outcome and analysis structure, including nonlinear terms or interactions. White and Royston explain the role of the event indicator and a cumulative-hazard estimate when imputing Cox-model covariates; routine imputation that ignores survival information can weaken associations. State the assumptions and conduct appropriate sensitivity analyses. Our broader missing-data guide explains the starting distinctions, but a survival dataset needs a compatible imputation specification.

Choose the estimand when competing events exist

A competing event makes the event of interest impossible to experience first. For withdrawal, a terminal programme completion or death may be competing events, depending on the precise study definition. Distinguish those records from students whose status simply becomes unknown.

If the target is actual probability of withdrawal in the presence of competing events, estimate a cumulative incidence function, commonly using Aalen–Johansen. Treating competing events as ordinary censoring and reporting 1 - Kaplan–Meier generally overstates that probability. Cause-specific Cox models address the instantaneous rate of a particular event among those still free of all events; removing competing events from that risk set is appropriate for that estimand. Fine–Gray models address a subdistribution hazard connected to cumulative incidence. Neither hazard ratio is automatically a risk ratio. Austin, Lee and Fine explain the distinction. Choose the question before choosing the model.

Investigate whether observation ending carries information

Compare loss patterns, timing and recorded reasons across groups. If dropout relates to unobserved withdrawal propensity, ordinary estimates may be biased. Depending on the available information, inverse-probability-of-censoring weighting or sensitivity analyses can examine the problem, but they require their own model, positivity and identification assumptions. A low proportion lost does not by itself establish negligible bias. Document what remains unknowable, rather than saying the software “controlled for censoring.”

Plan sample size around events and the intended analysis

The information for a Cox coefficient depends strongly on observed events and the distribution of predictors across risk sets. Total enrolment alone is insufficient. Plan recruitment, accrual duration, follow-up, expected event incidence, censoring and the precision or power needed for the primary estimand.

For a prespecified group comparison, use a survival-specific power calculation compatible with the allocation, target effect and expected survival pattern, then translate required events into participants under defensible follow-up assumptions. If nonproportional hazards are anticipated, use a design method that matches the chosen comparison rather than treating one constant hazard ratio as sufficient.

Prediction-model development is a different planning problem from testing a single exposure association. Riley and colleagues derive context-dependent requirements involving participants, events, candidate predictor parameters, anticipated model fit and overfitting. A universal events-per-variable rule cannot certify adequacy. Count parameters, including category contrasts, nonlinear terms and interactions, rather than just variable names. Our power-analysis guide explains planning logic; its simple examples should not be substituted for survival-specific calculations.

Our extended cohort has 140 events and two fitted predictor parameters. Those numbers describe the demonstration, not a sample-size recommendation or a claim that every model with that event count is reliable.

Reproducible R syntax for the hypothetical examples

The following code uses the established survival package. It creates the small dataset, draws Kaplan–Meier curves with a separate risk-count table, reproduces the extended generator and specifies the Cox models. It does not download data. R was unavailable in the validation environment, so this R code was reviewed but not executed. The displayed numerical results were independently computed and checked in Python. Run the code and inspect diagnostics before adapting it to a dissertation; record the actual R and package versions used.

library(survival)

# All observations are hypothetical. event: 1 = withdrawal; 0 = censored.
small <- data.frame(
  id = c(paste0("A", 1:6), paste0("B", 1:6)),
  group = factor(rep(c("A", "B"), each = 6)),
  time = c(2, 4, 4, 7, 4, 10, 3, 6, 8, 5, 8, 10),
  event = c(1, 1, 1, 1, 0, 0, 1, 1, 1, 0, 0, 0)
)
pooled <- survfit(Surv(time, event) ~ 1, data = small,
                  conf.type = "log-log")
km <- survfit(Surv(time, event) ~ group, data = small,
              conf.type = "log-log")
summary(pooled, censored = TRUE)
summary(km, times = 6)
print(km) # Includes median estimates and their software-derived limits.
plot(km, col = c("navy", "darkorange"), lty = 1:2,
     conf.int = TRUE, mark.time = TRUE,
     xlab = "Months since enrolment", ylab = "Withdrawal-free probability")
legend("bottomleft", legend = c("A: standard", "B: enhanced"),
       col = c("navy", "darkorange"), lty = 1:2, bty = "n")
# Keep this risk table with the figure; counts are immediately before time t.
ticks <- c(0, 2, 4, 6, 8, 10)
risk <- sapply(ticks, function(t) {
  tapply(small$time >= t, small$group, sum)
})
colnames(risk) <- ticks
print(risk)
survdiff(Surv(time, event) ~ group, data = small, rho = 0)

# Deterministic cross-language uniform generator; no set.seed dependency.
state <- 80024
u <- function() {
  state <<- (16807 * state) %% 2147483647
  state / 2147483647
}
n <- 240
large <- data.frame(id = seq_len(n), group = 0L, severity = 0L,
                    time = 0, event = 0L, reason = "")
for (i in seq_len(n)) {
  group <- as.integer(u() < 0.5)
  severity <- as.integer(u() < 0.4)
  event_time <- -log(u()) / (0.075 * exp(-0.5 * group + 0.7 * severity))
  loss <- if (u() < 0.18) 2 + 10 * u() else 99
  time <- min(event_time, loss, 12)
  event <- as.integer(event_time <= min(loss, 12))
  reason <- if (event == 1) "withdrawal" else if (loss < 12) "loss" else "administrative"
  large[i, ] <- list(i, group, severity, time, event, reason)
}
table(large$group)
table(large$reason)
unadjusted <- coxph(Surv(time, event) ~ group, data = large,
                    ties = "efron", x = TRUE)
adjusted <- coxph(Surv(time, event) ~ group + severity, data = large,
                  ties = "efron", x = TRUE)
summary(unadjusted)
summary(adjusted)
exp(cbind(HR = coef(adjusted), confint(adjusted)))

# Run and interpret these diagnostics; no diagnostic results are claimed here.
ph <- cox.zph(adjusted)
print(ph)
plot(ph)
# Profile-specific predictions, distinct from unadjusted Kaplan-Meier curves.
profiles <- data.frame(group = c(0, 1), severity = c(0, 0))
predicted <- survfit(adjusted, newdata = profiles)
plot(predicted, col = c("navy", "darkorange"), lty = 1:2,
     xlab = "Months since enrolment", ylab = "Predicted withdrawal-free probability")
packageVersion("survival")
sessionInfo()

The curve command produces a figure when executed; the code does not claim to have generated or inspected it here. The support groups in large are numeric indicators with standard support as the reference. For an empirical dataset, inspect coding explicitly, validate dates and check which records the model actually includes.

How to report survival analysis in a dissertation

Methodology paragraph you can adapt

The following paragraph describes the executed hypothetical demonstration, not an empirical study:

Time was measured in months from initial enrolment to first formal withdrawal. Students without observed withdrawal were right-censored at their last reliable observation or the administrative observation limit. We retained censoring reasons separately. Kaplan–Meier estimates described withdrawal-free survival, with pointwise log–log confidence intervals based on Greenwood variance. The small illustrative support-group comparison used the ordinary log-rank test with a tied-event variance correction. In a separate synthetic cohort, Cox regression estimated the association of baseline enhanced support with withdrawal hazard, first unadjusted and then adjusted for the prespecified baseline high-need indicator. Model-based Wald confidence intervals were calculated on the log-hazard scale and exponentiated. The partial-likelihood analysis was executed in Python and independently reproduced; accompanying R code was not executed. Proportional-hazards diagnostics are specified for readers to run, so no empirical diagnostic result is asserted.

For your dissertation, replace the demonstration details with your actual eligibility rules, dates, competing-event handling, missing-data strategy, covariate rationale, software versions, diagnostics and sensitivity analyses. Distinguish a planned method from an analysis you actually completed. The STROBE checklist is useful for cohort reporting, including participant flow, follow-up, missingness, events and adjusted estimates.

Results paragraph grounded in the calculations

In the hypothetical hand-calculation cohort, 7 withdrawals were observed among 12 students. Estimated pooled withdrawal-free survival at month 6 was 0.5556 (pointwise 95% CI 0.2311–0.7899), and the estimated median was 7 months. Group-specific month-6 estimates were 0.5000 for standard support and 0.6250 for enhanced support. The approximate log-rank comparison gave chi-square(1) = 0.7086, p = 0.3999; this small example provides little precision for evaluating a group difference. In the separate 240-student synthetic cohort, 140 withdrawals occurred. The adjusted hazard ratio for enhanced versus standard support was 0.7130 (95% CI 0.5072–1.0024; Wald p = 0.0517), conditional on baseline need. The interval includes one, and the estimate does not describe an absolute reduction in withdrawal probability. These demonstration findings support no conclusion about real programmes.

In an empirical results chapter, add the analysed sample flow, reasons for exclusion, event and censoring counts by group, follow-up summary, risk-set panel and diagnostic findings. If reporting median follow-up, identify the estimator; the median of observed event-or-censoring times is not automatically a suitable follow-up measure. Explain any survival horizon chosen for interpretation. Use our statistical-reporting guide to keep estimates, precision and substantive interpretation together.

Common mistakes that weaken survival-analysis reporting

  • Treating every student without an event record as followed to study end. Check the last reliable event-free observation and retain the reason observation stopped.

  • Deleting censored students. They contribute valid risk-set information while observed, under the relevant censoring assumptions.

  • Removing same-time censorings before events without justification. State the ordering convention and preserve finer timing when available.

  • Calling the observed event fraction a survival estimate at every horizon. Unequal follow-up and censoring change the calculation.

  • Reporting a median that the curve never reaches. Report its non-estimability and a supported time-specific estimate instead.

  • Interpreting an HR as a risk ratio or a time ratio. Name the estimand and the reference group.

  • Claiming assumptions passed from one nonsignificant test. Combine plots, tests, design knowledge and sensitivity checks.

  • Presenting unadjusted curve separation as a causal programme effect. Address selection, confounding, survivor conditioning and the study design.

  • Using future support attendance as a baseline exposure. Align exposure status with the time it becomes known.

  • Confusing competing events with ordinary missing follow-up. Choose cumulative incidence or a hazard estimand deliberately.

Frequently asked questions

Can I use survival analysis when nobody dies?

Yes. The event can be withdrawal, resignation, recovery, equipment failure or another clearly defined occurrence. Explain what remaining event-free means in your study and whether the event is desirable or undesirable.

Is a censored observation the same as a missing observation?

No. Right censoring preserves information that the participant remained event-free up to an observation limit, while the later event time is unknown. A missing baseline variable or an unknown last-observed date requires separate handling.

Should I report Kaplan–Meier, log-rank and Cox results together?

Often they complement one another: Kaplan–Meier describes unadjusted survival, log-rank tests an unadjusted group comparison, and Cox regression estimates conditional hazard associations. Use only analyses that address the question, and do not present their estimates as interchangeable.

What if my survival curves cross?

Check the data and the adjusted model, then consider whether changing effects are plausible. A single hazard ratio and an ordinary log-rank p-value can conceal important timing. A prespecified time-dependent effect or a supported restricted-mean comparison may be more informative.

What if median survival is not reached?

State that clearly. Report an estimate at a meaningful time with adequate follow-up, its confidence interval and the number at risk. Do not extrapolate a median from a flat tail unless a justified model and its assumptions are explicitly reported.

Does an adjusted hazard ratio establish causation?

No. Adjustment addresses measured covariates within a chosen model. Selection, measurement error, unmeasured confounding, informative censoring and conditioning on survival can still affect interpretation. Match causal language to the design and identification assumptions.

Sources and further reading

Need help making your analysis understandable and defensible?

Acadelyra Services can support analysis planning, interpretation and clear methodological reporting. Bring your research question, data dictionary, actual output and institutional requirements so the guidance fits your study. You remain responsible for the analysis decisions, authorship and work you submit; support should strengthen your understanding and follow your institution's academic-integrity rules.