Psychology dissertation survival analysis examines both whether an event occurs and how long it takes to occur. It can answer questions about relapse, symptom remission, treatment discontinuation, return to work, participant dropout, or another clearly defined transition. This guide explains how to define time and events, handle censoring, choose Kaplan-Meier or regression methods, check assumptions, and report results without overstating what the design can support.
The word “survival” does not restrict the method to mortality. In psychology, the event might be the first panic attack after treatment, disengagement from a digital intervention, readmission to care, or sustained recovery. The essential outcome has two parts: an event indicator and a measured duration from a defensible starting point.
When survival analysis fits a psychology dissertation
Ordinary linear regression ignores that some participants have not experienced the event by the end of follow-up. Logistic regression retains whether an event occurred but discards its timing. Survival methods use the available follow-up from every participant while explicitly representing incomplete event times through censoring.
The Columbia University time-to-event overview emphasises four foundations: define the target event, time origin, time scale, and the ways participants exit follow-up. If any of these is ambiguous, the analysis cannot rescue the research question.
| Psychology question | Event | Time origin and scale | Possible censoring |
|---|---|---|---|
| How long until anxiety symptoms return after therapy? | Predefined relapse threshold | Treatment completion; days or weeks | No relapse by final assessment |
| What predicts disengagement from an app? | Account becomes inactive under a stated rule | Registration; days | Still active when observation ends |
| Does a support programme accelerate return to work? | Verified return to paid employment | Programme entry; weeks | Not returned or lost to follow-up |
| When do students withdraw from a study? | Formal withdrawal or confirmed dropout | Consent; days | Completes study without withdrawal |
Define the event before viewing outcomes
A useful event is observable, meaningful, and applied consistently. “Recovery” is too vague unless tied to a validated score, clinical judgement, functional criterion, or composite rule. State whether the event must occur once, persist for a minimum duration, or be confirmed at a later assessment.
Changing an event definition after inspecting results can introduce outcome-selection bias. Preregister the primary definition and label alternatives as sensitivity analyses. The psychology dissertation preregistration guide explains how to separate planned and exploratory decisions.
Choose the time origin and time scale
Time zero must represent a comparable starting condition. Treatment allocation, discharge, diagnosis, study enrolment, or first app use may be appropriate. Using enrolment for some participants and treatment completion for others creates incomparable risk periods.
Decide whether time is measured in days, weeks, sessions, or another meaningful unit. Record event and censoring dates precisely enough for that scale. If participants enter the risk set after the formal study origin, consider delayed entry or left truncation rather than pretending they were observed earlier.
Understand censoring and follow-up
Right censoring occurs when the event time is only known to be later than the last observed time. A participant may complete follow-up without relapse, remain engaged when the dataset closes, or be lost before an event is observed. Their data still contribute information up to their censoring time.
Standard methods usually assume censoring is non-informative conditional on variables in the model. That means censored participants have comparable future event prospects to similar participants who remain observed. The classic Survival Analysis Part I tutorial warns that withdrawal caused by worsening condition or treatment toxicity can make censoring informative.
Do not code loss to follow-up as “no event”
Coding every lost participant as event-free until the planned study end invents observation time. Instead, censor at the last time their status was known. Compare follow-up completeness across groups, summarise reasons for loss, and assess whether predictors of censoring overlap with predictors of the event.
Missing covariates and censored outcomes are different problems. Survival methods address incomplete event times, not missing baseline scores or unrecorded confounders. Plan covariate handling with the psychology dissertation missing data guide.
Recognise other observation patterns
Left censoring means the event occurred before it was first observed, while interval censoring means it occurred between two assessments. Recurrent events arise when the same participant can experience repeated panic attacks, relapses, or service contacts. A simple first-event Cox model may be unsuitable for these designs.
Competing events prevent or alter observation of the target event. For example, moving to another service may preclude dropout from the original clinic. Treating a competing event as ordinary censoring can distort cumulative incidence. The official R survival vignette on competing risks illustrates cumulative-incidence and multi-state approaches.
Prepare the survival dataset
At minimum, one row per participant needs a start time, stop or event time, and event indicator. Add group membership and pre-specified predictors. Recurrent events, time-varying covariates, or delayed entry usually require start-stop data with multiple rows per participant.
Audit impossible sequences, negative durations, events recorded after censoring, duplicate identifiers, inconsistent date formats, and events before eligibility. Keep a reproducible derivation log so every duration and status can be traced to source records. Link this audit to the broader psychology dissertation data analysis workflow.
| Variable | Example coding | Quality check |
|---|---|---|
| Participant ID | Unique pseudonymous code | No duplicate person unless long format is intended |
| Time origin | Therapy completion date | Same conceptual origin for every participant |
| Follow-up time | Days from origin to event or censoring | Positive and consistent with source dates |
| Event status | 1 = relapse, 0 = right-censored | Matches the registered event definition |
| Predictors | Baseline severity, treatment group, age | Coding, reference levels, missingness, plausible ranges |
| Competing event | Transfer to another service | Not silently combined with ordinary censoring |
Describe time-to-event outcomes
Kaplan-Meier curves
The Kaplan-Meier estimator describes the probability of remaining event-free beyond each time. The curve changes when events occur and accounts for the changing number at risk. Display censoring marks and a number-at-risk table so readers can see when estimates depend on few remaining participants.
Report event counts, censoring counts, follow-up duration, and survival probabilities at substantively meaningful times with confidence intervals. The median event-free time is the point at which the estimated survival probability reaches 0.5. If the curve never falls to 0.5, the median is not estimable and should not be fabricated.
Log-rank tests
A log-rank test compares entire survival curves under assumptions that include independent observations and compatible censoring. It provides a global group comparison but does not estimate an adjusted causal effect. Curves can cross, hazards can vary over time, and a single p-value can hide important timing differences.
Show the curves and examine whether the comparison matches the question. If early differences matter more than late differences, pre-specify an appropriate summary such as survival at a fixed time or restricted mean survival time. Avoid selecting a time point after seeing where curves separate most strongly.
Use Cox proportional hazards regression carefully
The Cox model relates predictors to the instantaneous event rate among participants who remain at risk. Exponentiated coefficients are hazard ratios. A hazard ratio above one indicates a higher hazard for the coded comparison, while a value below one indicates a lower hazard, conditional on included covariates.
A hazard is not a probability, and a hazard ratio is not a risk ratio. It also does not state how many weeks earlier an event occurs. Interpret the direction, reference group, confidence interval, follow-up period, and model assumptions. The peer-reviewed Survival Analysis Part II tutorial explains the Cox model and alternatives such as accelerated failure-time models.
Select covariates from the research question
Do not screen predictors solely by univariable p-values. Choose covariates from theory, design, prior evidence, and a stated causal or prognostic aim. Adjustment for a mediator or collider can introduce bias, while omitting a common cause can leave confounding. Keep the number and complexity of parameters compatible with the observed number of events.
Survival study information depends heavily on event count, not just total sample size. Plan recruitment, follow-up duration, expected censoring, and expected event frequency together. The psychology dissertation power analysis guide provides general planning principles, but specialised survival calculations or simulations should reflect the planned model.
Check proportional hazards
The proportional-hazards assumption means a covariate’s hazard ratio is constant over time. Assess it with subject-matter reasoning, survival plots, log-minus-log plots when appropriate, scaled Schoenfeld residuals, and tests such as cox.zph. The official R survival package supports Cox models, proportionality checks, Kaplan-Meier estimation, Aalen-Johansen curves, and parametric accelerated failure-time models.
Do not treat a non-significant diagnostic test as proof of proportionality, especially with few events. Inspect plots and effect patterns. If proportionality is implausible, consider stratification, a time-by-predictor interaction, piecewise effects, restricted mean survival time, or an alternative model. Select a response that preserves the research estimand.
Consider restricted mean survival time
Restricted mean survival time is the average event-free time up to a pre-specified horizon. A difference between groups can be expressed directly in days or weeks, which is often easier to interpret than a hazard ratio. It does not require proportional hazards, although adequate follow-up and a defensible horizon remain essential.
Choose the horizon from the protocol, clinical relevance, or reliable common follow-up, not from the point that produces the largest difference. Report the estimated restricted means, their difference, confidence interval, and the time horizon. Explain how many participants remained under observation near that boundary.
Check functional form and influence
The Cox model often assumes a linear relationship between a continuous predictor and log hazard. Examine residual plots or fit pre-specified flexible terms such as restricted cubic splines. Arbitrary categorisation wastes information and creates cut-point dependence. Investigate influential observations, sparse categories, interactions, and clustering.
Assess overall model adequacy and calibration if prediction is the goal. Prediction requires more than statistically significant coefficients. Internal validation through resampling can quantify optimism, and an independent sample is needed to assess transportability.
Choose alternatives when Cox regression is unsuitable
An accelerated failure-time model describes how predictors multiply expected event times under a specified distribution. Flexible parametric models can estimate smooth survival functions and time-varying effects. Discrete-time survival models suit data observed only in intervals such as weekly assessments or academic terms.
Competing-risks methods estimate cause-specific hazards or cumulative incidence, depending on the question. Recurrent-event and multi-state models address repeated or sequential transitions. Mixed-effects or frailty models can represent clustering. These extensions require adequate events and careful supervision; complexity should serve the estimand rather than impress the reader.
Repeated follow-up may connect naturally to a psychology dissertation longitudinal study, but the outcome structure determines the model. If event dates come from an existing cohort or administrative source, document provenance, eligibility, linkage quality, and measurement changes using the psychology dissertation secondary data analysis guide.
| Method | Best suited to | Key caution |
|---|---|---|
| Kaplan-Meier plus log-rank | Unadjusted event-free experience and group comparison | Weak summary when curves cross or risk sets become sparse |
| Cox proportional hazards | Adjusted relative hazard with plausible proportionality | Hazard ratio is not a probability or time ratio |
| Accelerated failure time | Direct modelling of relative event time under a distribution | Distributional fit must be assessed |
| Discrete-time survival | Events known only within scheduled intervals | Interval definition and link function affect interpretation |
| Competing-risks model | Different events prevent or change the target event | Cause-specific hazard and cumulative incidence answer different questions |
| Recurrent-event or multi-state model | Repeated events or several transitions | Within-person dependence and state definitions matter |
Psychology-specific worked example
Consider a study of 240 adults completing cognitive behavioural therapy for panic disorder. The event is the first confirmed recurrence meeting a pre-specified symptom threshold. Time zero is the final treatment session, follow-up is measured in days, and participants without recurrence are censored at their last completed assessment.
The primary comparison is between standard aftercare and a digital maintenance programme. Baseline severity and previous episode count are pre-specified covariates. The student plots Kaplan-Meier curves with risk tables, estimates event-free probabilities at 90 and 180 days, and fits a Cox model. The proportionality assessment suggests the programme effect weakens over time, so a time-varying coefficient is reported instead of one constant hazard ratio.
Several participants transfer to specialist care before recurrence can be observed. The student treats transfer as a competing event in a sensitivity analysis and reports cumulative incidence. Loss to follow-up is described by group and baseline severity. Conclusions are framed as model-adjusted associations unless the design and assumptions support a causal interpretation.
Software and reproducibility
R’s survival package is a mature option for Kaplan-Meier estimation, Cox regression, AFT models, diagnostics, multi-state models, and time-dependent covariates. The package’s official vignette index links methods, validation, competing-risks, and time-dependent-covariate documentation. Stata, SAS, and SPSS also offer survival procedures, but capabilities differ.
Record software and package versions, time and event coding, reference categories, tie handling, model formulas, diagnostic commands, seeds, and any data transformations. Export readable tables and figures rather than screenshots of software output. Retain code and a data dictionary in line with ethical and governance constraints.
Report a psychology dissertation survival analysis
Report the study population, time origin, event definition, censoring rules, follow-up distribution, event and censoring counts, losses, competing events, and missing covariates. For Kaplan-Meier results, include curves, risk tables, meaningful time-specific estimates, confidence intervals, and medians only when estimable.
For regression, report the model formula, covariate rationale, reference groups, hazard ratios with confidence intervals, functional-form checks, proportionality assessment, influential observations, and sensitivity analyses. Explain any response to assumption violations. Avoid describing a hazard ratio as a percentage change in probability.
The STROBE statement provides a reporting checklist for observational cohort, case-control, and cross-sectional research. Use the cohort items where relevant, together with survival-specific details. Reporting guidance improves transparency but does not certify design quality.
Common mistakes to avoid
- Using logistic regression for a time-to-event question: it discards timing and handles unequal follow-up poorly.
- Leaving time zero vague: incomparable origins create biased risk sets.
- Coding lost participants as event-free: censor at the last verified observation instead.
- Hiding the number at risk: tail estimates can look precise when few participants remain.
- Calling a hazard ratio a risk ratio: the measures have different meanings.
- Ignoring proportionality: one constant effect may conceal changing hazards.
- Treating competing events as ordinary censoring: cumulative incidence may be distorted.
- Choosing covariates by p-value alone: align adjustment with theory and the intended estimand.
Frequently asked questions
Does survival analysis require death as the event?
No. Any clearly defined event with meaningful timing can be analysed, including relapse, remission, dropout, readmission, disengagement, or return to work.
What is the difference between censoring and missing data?
A right-censored participant contributes verified event-free time up to a known point. Missing covariates or unknown follow-up status require separate investigation and handling.
Can I use Kaplan-Meier curves with more than two groups?
Yes, but the plot and global comparison can become difficult to interpret. Pre-specify meaningful contrasts, show risk tables, and avoid unsupported pairwise testing.
What does a hazard ratio of 0.70 mean?
Under a proportional-hazards model, the estimated instantaneous event hazard is 30% lower for the coded group at each time, conditional on the model. It does not mean 30% fewer participants will experience the event.
What if survival curves cross?
Crossing may indicate a changing effect or different timing patterns. Investigate proportionality and consider time-varying effects, restricted mean survival time, or another pre-specified summary.
How many participants are needed?
The number of observed events, censoring rate, follow-up duration, effect size, group allocation, and model complexity all matter. Use a survival-specific calculation or simulation rather than a fixed participant rule.
Can survival analysis establish causation?
Not by itself. Causal interpretation depends on design, exchangeability, measurement, follow-up, censoring, competing events, and model assumptions.
Conclusion
A defensible psychology dissertation survival analysis begins with an explicit event, comparable time origin, appropriate scale, and honest censoring rules. Describe event-free experience before modelling, choose a method that matches the estimand, check proportionality and functional form, address competing events, and report uncertainty with the number at risk.
If you need methodological support, choose guidance that helps you understand and defend every analytic choice. Ethical dissertation assistance should strengthen design, data preparation, code, interpretation, and reporting while preserving your authorship and complying with institutional academic-integrity requirements.
