---
name: longitudinal-and-multi-study-design
description: >
  Designs multi-phase research programmes where each study genuinely depends on
  the last, with contingency planning for when study one fails, and longitudinal
  designs that can survive attrition, measurement change and cohort confounding.
  Use for "multi-study PhD design", "sequential mixed methods", "study two
  depends on study one", "panel versus cohort design", "attrition bias",
  "measurement invariance over time", "age period cohort", "publication-based
  thesis coherence", "three papers do not make a thesis", "will this fit in my
  PhD timeline".
category: 15 Academic University Research
ref: "15.17"
tier: 3
inherits: [K2, K3, K4, K5]
---

# Longitudinal and Multi-Study Design

## 1. One-line description

A method for designing research programmes that run over time or across multiple studies, where the sequencing logic is real rather than decorative, the failure of any one component has been planned for, and the design can survive the specific threats that only appear when data is collected more than once.

## 2. What this skill is used for

**The research problem it solves.** Multi-study doctorates fail in two characteristic ways. The first is the programme where three studies follow one another without depending on one another: study two would have been designed identically whatever study one found, which means the sequence bought nothing and the thesis is three studies stapled together. The second is the programme where the dependency is real and unhedged: study two cannot be designed until study one returns a usable result, study one returns nothing usable in month eighteen, and there is no time left. Longitudinal designs add their own failures, all of which are invisible in a single wave: differential attrition that quietly reshapes the sample, an instrument that measures something slightly different at wave three from wave one so that observed change is measurement change, and the impossibility of separating age, period and cohort effects with any single design. This skill builds the sequencing logic, forces the contingency plan, and installs the checks that make a change claim over time defensible.

**Where it sits in the research lifecycle.** At programme design, before any data collection, which is the only point at which most of these problems are cheap to fix. Also at the midpoint, when study one has returned something other than expected and the programme has to be re-planned, and at write-up for publication-based theses where coherence has to be constructed across papers.

**Typical use cases.**
- Designing a doctoral programme of two to four linked studies with a defensible sequencing argument.
- Planning what study two becomes under each plausible outcome of study one, including the null.
- Choosing between panel, cohort and repeated cross-section for a question about change.
- Designing a longitudinal study that can survive attrition and demonstrate measurement invariance.
- Specifying the integration point in a sequential mixed-methods design.
- Constructing the overarching argument that makes a publication-based thesis a thesis.
- Assessing honestly whether a proposed programme fits in the time and funding available.

**Who uses it.** Doctoral candidates designing multi-phase or longitudinal work; supervisors testing whether a proposed programme is feasible; researchers planning panel studies, cohort studies or sequential mixed-methods programmes.

## 3. When to use it

- Your thesis will contain more than one study and you need the sequence to be an argument rather than a chronology.
- Someone has asked why study two follows study one, and the honest answer is "it just does".
- You are designing a study that measures the same people or units more than once.
- You need to make a claim about change over time and want it to survive scrutiny.
- Your first study has returned an unexpected or null result and the programme has to be re-planned.
- You are writing a publication-based thesis and the papers do not obviously add up to one contribution.
- You are being asked whether the proposed programme fits in the remaining time, and you need a defensible answer rather than optimism.
- You have a sequential mixed-methods design and cannot say precisely where the strands meet.

## 4. When NOT to use it

- **The question can be answered by one study.** A multi-study design is expensive in time and risk and should be adopted because the question requires it, not because a doctorate is expected to look substantial. If one well-designed study answers the question, that is the design, and the extra studies would be padding an examiner will identify. Use **01.04 Research Method Selection**.
- **The single design has not been justified yet.** Each component study still needs its own comparative defence. That is **15.16 Doctoral Methodology Justification and Rigour**, and this skill assumes it, covering only the logic that connects the studies and the threats that arise from repetition over time.
- **The programme is already underway and cannot change.** Where studies are complete and the sequence was weak, the task is to construct the strongest honest account of what the programme did establish, not to retrofit a dependency that did not exist. Retrofitting is detectable, because the design of study two will not reflect study one's findings.
- **The time available cannot accommodate the design and there is no reduction that works.** The correct output is that finding, delivered as early as possible, not an optimistic schedule. Overrun in doctoral programmes is caused more often by a design that never fitted than by any other single factor. `RESEARCHER DECISION REQUIRED` (K5 §2.7).
- **The claim is causal and the longitudinal design does not license it.** Repeated measurement establishes temporal order, which is necessary for a causal claim and not sufficient. Time-varying confounding, reverse causation within intervals shorter than the measurement gap, and selection into exposure all survive a panel design (K4 §3.2). Where causal inference is the requirement, the design question is about identification strategy, not about frequency of measurement.
- **The analysis of longitudinal data is the task.** Growth models, fixed and random effects specifications, survival analysis and handling of missing data over waves are analytical decisions. See **05.01 Statistical Analysis Planning** and **06.02 Advanced Statistical Modelling**. This skill covers the design that makes those analyses possible.
- **The thesis format decision is institutional and already made.** Where the institution mandates a monograph or a publication-based format, the choice is not open, and this skill addresses the coherence work each format requires rather than the choice between them.
- **Academic integrity.** Per §12.1: this skill does not write programme rationales, linking chapters or design justifications for submission as the candidate's own unaided output. It interrogates the sequencing logic, exposes unhedged dependencies and returns what the coherence argument would require. Where an institution prohibits AI assistance for a task, this skill must not be used for it.

## 5. Required inputs

**Required.**
- **The overarching research question the programme as a whole answers.** Not the questions of the individual studies. Without it there is no basis for sequencing and no basis for coherence, and this is the input most often missing in troubled multi-study designs.
- **The time and funding actually available**, including fixed deadlines, funded period, and any period already consumed. Feasibility is not an optional consideration in a programme design; it is a constraint that shapes the design.
- **The access position for each intended data source**, with honest status: secured, likely, requested, or hoped for. A programme built on unsecured access is a programme with an undeclared single point of failure.

**Optional, and what each one adds.**
- **Results from any completed study.** Converts the design of the next study from hypothetical into determined, and lets the dependency be demonstrated rather than asserted.
- **Ethics approval status and timelines for each phase.** Approval lead times are a common and underestimated source of slippage, and a design requiring three separate approvals has three separate delays.
- **Attrition estimates from comparable studies.** Lets the longitudinal sample be sized for the wave that matters rather than for wave one.
- **The instruments intended for repeated administration.** Lets measurement invariance be planned rather than discovered as a problem at analysis.
- **The institution's rules for a publication-based thesis.** These vary considerably: number of papers, required publication status, whether co-authored work counts and how contribution is evidenced, and what the linking material must contain.
- **Co-author and collaborator commitments.** Determine what parts of the programme depend on other people's timelines, which is a dependency candidates routinely fail to plan for.

## 6. Questions to ask before starting

1. **What does study two need from study one that it could not get otherwise?** The test of a real sequence. Default: if the answer is "context" or "background", the dependency is decorative and the studies can run in parallel, which is usually better for the timeline.
2. **What are the plausible outcomes of study one, including the null and the ambiguous?** Determines whether a contingency plan is possible. Default: enumerate at least three outcomes and design study two under each.
3. **How long does each phase actually take, including approval, recruitment, data collection, analysis and writing?** Default: estimate each separately and add the approval and recruitment periods that candidates habitually omit, then compare with the calendar.
4. **What is the single point of failure?** One site, one gatekeeper, one dataset, one collaborator, one instrument licence. Default: identify it and either remove it or plan around it explicitly.
5. **Is the question about change within units, or difference between units over time?** Determines panel versus repeated cross-section, and the two answer different questions. Default: derive it from the research question, since candidates frequently choose the design first.
6. **What would make the same instrument mean something different at a later wave?** Ageing of respondents, changed context, changed salience of the topic, a public event. Default: assume invariance must be tested rather than assumed.
7. **If this is a publication-based thesis, what is the claim that only the whole set supports?** Default: if there is no such claim, the papers do not yet make a thesis and the linking work is the priority.

## 7. Step-by-step methodology

**Step 1. State the programme-level question and check each study earns its place.**
Write the overarching question first, then for each proposed study write what it contributes to answering it and what would be lost if it were deleted. A study whose deletion loses nothing is padding. A study whose deletion loses "breadth" is usually padding too, since breadth is what a programme sacrifices for depth, and an examiner asking why there are three studies wants to hear a dependency, not a survey of the territory. Then check the converse: is there a study missing without which the programme's claim does not hold? *Correct result: a programme question, and per study a one-sentence statement of the unique contribution plus what its deletion would cost.*

**Step 2. Classify the sequencing logic, and test whether it is real.**
Sequences come in a small number of genuine forms. Study one identifies the constructs or population study two measures, so study two's instrument cannot exist without it. Study one establishes an effect and study two explains its mechanism. Study one identifies an anomaly and study two tests competing explanations. Study one develops and study two validates. Study one is broad and study two is a theoretically sampled deep case selected on its results. In every real form, study two's design is a function of study one's output. The test is brutal: write down what study two would look like under two different outcomes of study one. If it is the same design either way, the sequence is a chronology, not a logic, and should be stated as such rather than dressed as dependency. `RESEARCHER REVIEW RECOMMENDED` where a sequence is decorative, since running the studies in parallel is often faster and equally defensible (K5 §2.7). *Correct result: a named sequencing form per link, and the demonstration that study two's design varies with study one's outcome.*

**Step 3. Build the contingency plan for every dependency.**
For each link, enumerate the plausible outcomes of the upstream study: the expected result, the opposite, a null, an ambiguous or noisy result, and failure to collect adequate data at all. For each, write what study two becomes. Three properties make a contingency plan real. It is written before study one runs, so it cannot be a rationalisation. It includes a version of study two that works under the null, the outcome most likely to strand a candidate and more common than optimism allows. And it names a decision date, chosen so there is still time to execute the branch. Add the failure branch explicitly: if study one produces nothing usable, what study two can still be run, and does the thesis still have a contribution? A programme with no answer to that carries an unmanaged risk of catastrophic overrun. *Correct result: a branch table per dependency, with a decision date, and a viable path under the null and under the failure case.*

**Step 4. Choose the longitudinal design from the question, and know what each supports.**
Three basic forms answering different questions. A panel follows the same units over time and is the only design supporting within-unit change, individual trajectories and the temporal ordering of individual events; it costs attrition, conditioning where repeated measurement changes the respondent, and expense. A cohort follows a group defined by a shared starting characteristic, supporting claims about that group's development but unable to separate the cohort's experience from the period it lived through. A repeated cross-section draws a fresh sample each time, supports population-level change with no attrition or conditioning, and says nothing about individual change, so a stable figure across waves is entirely compatible with large offsetting individual movements. Sequential and accelerated designs, following several cohorts across overlapping age ranges, buy some separation of age from cohort at the cost of complexity. Choose from what the question requires, and state explicitly what the chosen design cannot show. *Correct result: a named design, a statement of what it supports, and an explicit statement of the claim it cannot support.*

**Step 5. Plan the wave structure on the phenomenon's timescale, not on convenience.**
Spacing is a substantive decision. Waves too close together capture noise and risk conditioning; too far apart and the process happens invisibly between them, so an effect that rises and returns to baseline within the interval appears as no effect at all. Set the interval from what is known or theorised about how long the process takes, and state the assumption. Decide the number of waves from what the analysis requires: two support change, three are the minimum for any claim about trajectory or non-linearity, more for turning points. Then work backwards to the calendar: the last wave must complete with enough time left for analysis and writing, the constraint that most often forces a redesign, and far better met now than at month thirty. *Correct result: an interval justified by the phenomenon's timescale, a wave count justified by the analysis, and a schedule that completes with writing time intact.*

**Step 6. Plan for attrition, which is never random.**
Estimate attrition per wave from comparable studies, then design against it. Size the sample for the wave the main analysis needs, not for wave one, since a design powered at baseline is under-powered where it matters. Build retention in: multiple contact routes collected at baseline, planned contact between waves, ethically cleared incentives, and a tracing procedure. Then plan the analysis of attrition itself, because the question is not how many were lost but whether loss is related to the variables of interest. At each wave compare leavers and stayers on baseline characteristics including the outcome, report the comparison, and state the direction of bias it implies. Where attrition is related to the outcome, the missingness is not ignorable and the analysis must address it rather than default to listwise deletion, which silently assumes exactly what the attrition analysis has just disproved. Report attrition as a result with its own table, not a sentence in limitations. *Correct result: a retention plan, a per-wave attrition analysis specification, and a stated approach to non-ignorable missingness.*

**Step 7. Establish measurement invariance, without which observed change may be measurement change.**
The same items administered at two points do not automatically measure the same construct. Meaning drifts: a term acquires new connotations, a behaviour becomes more or less socially acceptable, an event changes what a question is heard to ask, and respondents change how they use a scale as they age or gain experience. Where the design permits, test invariance formally in ascending levels: the same items loading on the same factors, then the same loadings, then the same intercepts. Only the strongest level supports comparing means across time, which is precisely what most longitudinal claims do. Where formal testing is impossible, the equivalent discipline is qualitative: keep wording identical, treat any changed item as a new item, watch for context effects from question order and from events between waves, and note where a construct's local meaning may have shifted. Never change an instrument between waves for improvement without carrying the original items alongside; the improved measure will not be comparable, and the comparison is the entire point of the design. *Correct result: an invariance testing plan, or a documented qualitative equivalent, and a rule that no instrument change occurs without overlap.*

**Step 8. Handle age, period and cohort confounding honestly.**
The three are linearly dependent: age plus birth year equals the year of measurement, so no design and no statistical technique separates all three without an assumption imposed from outside. This is not a limitation to be overcome by a better model; it is an identification problem. The honest options are to constrain one effect on substantive grounds and state the constraint, to use a design that provides leverage such as multiple cohorts observed at the same ages, or to report the confounding and interpret with it visible. What must never happen is a change over time attributed to ageing when the sample also lived through a period effect, or a difference between cohorts attributed to generational character when the cohorts were observed at different ages. Name which of the three your claim rests on and what you have assumed away. *Correct result: an explicit statement of which effect is identified, which is assumed, and what the assumption costs.*

**Step 9. Specify the integration point in a mixed-methods sequence.**
A sequential mixed-methods design is defined by what crosses between strands, and the commonest weakness is two strands reported side by side with a paragraph asserting they converge. Specify three things. What crosses: constructs to be measured, items to be written, a sampling frame for the qualitative phase drawn from quantitative results, an explanation to be tested, or anomalies to be investigated. When it crosses, as a named decision point in the schedule. And how the two are brought together in analysis and write-up: a joint display setting quantitative results against qualitative findings case by case or theme by theme, and an explicit treatment of divergence. Divergence is the most valuable output of a mixed design and it is routinely smoothed away (K4 §4.1); plan in advance what you will do when the strands disagree, because deciding afterwards invites the decision that makes the story cleaner. *Correct result: a specified integration point with content, timing and method, and a pre-committed approach to divergence.*

**Step 10. Assess feasibility against the calendar, in reverse.**
Work backwards from the submission date. Reserve the writing period, which is longer than candidates estimate and is what gets consumed when anything upstream slips. Then place analysis, final data collection, and each earlier phase, including ethics approval per phase, recruitment periods, and the Step 3 decision dates. Add the dependencies on other people: co-authors, gatekeepers, site approvals, supervisor availability, and any period when the field is inaccessible. Then compare with the time available. If it does not fit, the reduction options in order of preference are usually: fewer waves, fewer studies, smaller sample where power permits, a narrower population, or a component moved to future work. Reducing the writing period is not on the list. Deliver the honest assessment even when unwelcome, because a programme that does not fit is the largest single source of doctoral overrun and the cost of discovering it in year three is measured in years. `RESEARCHER DECISION REQUIRED` on any programme that does not fit, since the trade-off between scope and completion belongs to the candidate and supervisor (K5 §2.7). *Correct result: a reverse schedule with named dates, an explicit fit-or-not verdict, and a ranked reduction plan.*

**Step 11. For a publication-based thesis, build the coherence the papers do not supply.**
Three publishable papers do not automatically make a thesis, and this is the format's specific failure. Papers are written for different journals, audiences and lengths, define constructs slightly differently, position against different literatures, may use inconsistent terminology, and each is written to stand alone, which is the opposite of what a thesis requires. The coherence work is the linking material and it has real content. State the overarching contribution no single paper makes, which is the thesis's actual claim. Reconcile construct definitions across papers explicitly, and where a definition changed, say so and why rather than leaving an examiner to find it. Make the sequencing visible: what each paper established that the next depended on. Explain overlaps and repeated material. Document the candidate's own contribution to each co-authored paper specifically, since examiners test this and vague statements read badly. And write an integrated discussion doing the work no individual paper's discussion could, because each was constrained to its own scope. *Correct result: an overarching contribution statement, a construct reconciliation table across papers, linking sections that establish dependency, and an authorship contribution statement per paper.*

**Step 12. Write the programme-level threats and their treatment.**
Some threats belong to the programme rather than to any study: differential attrition across the whole panel, drift in the research question between phases, findings from study one shaping how study two's data is interpreted rather than only how it is collected, and the accumulation of researcher expectation across phases. Name them and say what is done about each: pre-specifying study two's analysis before study one's results are known where the design permits, keeping the coding of later phases blind to earlier results where feasible, and recording the decision points with dates so the sequence of reasoning is inspectable (K2 §5). *Correct result: a programme-level threat register distinct from the per-study one.*

## 8. Analytical framework

A multi-study programme is defended on a dependency chain:

    Programme question → Study 1 output → What Study 2 needs from it →
    Study 2 design determined by that output → Study 2 output →
    Programme-level claim that neither study makes alone

The chain is tested at the third link. If what study two needs from study one is only context, there is no chain. It is tested again at the last link: if the programme-level claim is just the two studies' claims listed together, the programme has produced two studies rather than one contribution.

Longitudinal designs are evaluated on a separate structure, and every claim about change must pass all four:

    Same units observed?  → Same construct measured?  →
    Attrition unrelated to the outcome?  → Age, period and cohort separated?

A "no" at the first means no within-unit change claim. A "no" at the second means observed change may be measurement change, and this is the failure most often undetected. A "no" at the third means the sample at the later wave is not the sample at the earlier one. A "no" at the fourth means the change is real but its attribution is not. Each failure has a specific correction, and none is fixed by a larger sample.

## 9. Output format

**1. Programme question and study map.**

| Study | Question it answers | Contribution to programme question | Cost if deleted |
|---|---|---|---|

**2. Sequencing logic.** Per link: the named form, what crosses, and the demonstration that study two's design varies with study one's outcome.

**3. Contingency table.**

| Upstream outcome | Probability judgement | What study two becomes | Decision date | Still viable? |
|---|---|---|---|---|

**4. Longitudinal design specification.** Type, units, wave count and spacing with justification, what it supports and what it cannot support.

**5. Attrition plan.** Expected rates, retention measures, per-wave attrition analysis specification, missingness approach.

**6. Measurement invariance plan.** Instruments, invariance levels to be tested or the qualitative equivalent, and the rule on instrument change.

**7. Age, period and cohort statement.** What is identified, what is assumed, and the cost of the assumption.

**8. Integration specification** for mixed designs: content, timing, method, and the pre-committed approach to divergence.

**9. Reverse schedule and feasibility verdict**, with named dates and a ranked reduction plan.

**10. Publication-based coherence pack** where applicable: overarching contribution, construct reconciliation table, linking structure, authorship contribution statements.

**11. Programme-level threat register.**

**When the evidence is thin, the format must not force fabrication (K4 §1).** Attrition rates from comparable studies are cited or marked as unavailable; they are never estimated to fill the table. Where the timeline does not fit, the output says so rather than compressing an estimate. Where a dependency has no viable contingency, that is reported as an unmanaged risk, which is a finding the supervisor needs, not a gap to be smoothed.

## 10. Quality checks

Run before a programme is committed to. These sit on top of K4 §8.

1. Does each study make a contribution the programme would lose without it?
2. For every dependency, does study two's design demonstrably differ under different outcomes of study one?
3. Is there a written contingency branch for the null and for outright failure of each upstream study?
4. Does every branch have a decision date with enough time after it to execute?
5. Is the longitudinal design type derived from the question, and is what it cannot show stated?
6. Is the wave interval justified by the phenomenon's timescale rather than by convenience?
7. Is the sample sized for the wave the main analysis needs, after expected attrition?
8. Is attrition analysed against the outcome at each wave, and is the missingness assumption stated and defended?
9. Is measurement invariance tested, or the qualitative equivalent documented, before any change over time is claimed?
10. Has any instrument changed between waves without overlap, and if so is the comparison abandoned?
11. Is the age, period and cohort identification problem addressed explicitly rather than modelled away?
12. Does the mixed-methods design name what crosses between strands, when, and how divergence will be handled?
13. Does the reverse schedule preserve the writing period intact?
14. For a publication-based thesis, is there a claim that only the whole set supports, and are construct definitions reconciled across papers?
15. Is the candidate's own contribution to each co-authored paper stated specifically?

## 11. Common failure modes

| Failure | How to recognise it | How to prevent it |
|---|---|---|
| **Chronology posing as logic** | Study two would be identical whatever study one found | Test the design against two outcomes (Step 2) |
| **Unhedged dependency** | No plan for study one returning nothing | Write branches including the null (Step 3) |
| **The optimistic schedule** | No ethics lead time, no recruitment period, writing at the end | Schedule in reverse from submission (Step 10) |
| **Wrong longitudinal form** | Individual change claimed from repeated cross-sections | Choose the design from the question (Step 4) |
| **Convenient wave spacing** | Waves at semester boundaries with no substantive rationale | Justify the interval from the phenomenon (Step 5) |
| **Attrition counted, not analysed** | A retention percentage with no leaver-stayer comparison | Compare on baseline outcome at each wave (Step 6) |
| **Improved instrument mid-study** | A better measure introduced at wave two | Never change without carrying the original (Step 7) |
| **Measurement change read as real change** | A mean shift with no invariance evidence | Test invariance before comparing means (Step 7) |
| **Cohort effect called ageing** | Development inferred from a single cohort observed once per age | State what is identified and assumed (Step 8) |
| **Integration by assertion** | Strands reported separately, convergence claimed in a paragraph | Specify the integration point in advance (Step 9) |
| **Divergence smoothed** | Qualitative findings that contradict the survey do not appear | Pre-commit to reporting divergence (Step 9, K4 §4.1) |
| **Three papers, no thesis** | Linking chapters that summarise rather than connect | Build the overarching claim (Step 11) |
| **Drifting constructs across papers** | The same term defined differently in papers two and three | Reconcile definitions explicitly (Step 11) |
| **Contamination across phases** | Later data interpreted through earlier expectations | Pre-specify and blind where feasible (Step 12) |

## 12. AI guardrails

Skill-specific. The universal prohibitions in K4 apply in full and are not repeated.

1. **Academic integrity, and it binds this whole skill.** These skills assist a researcher's thinking, structure and rigour. They do not produce work to be submitted as the student's own unaided output. The user must comply with their institution's AI use policy and its declaration requirements, which vary by institution and by assessment. Where an institution prohibits AI assistance for a task, this skill must not be used for it. The skill never writes a passage for submission as though the student wrote it; it interrogates, structures, critiques and teaches. Operationally here: linking chapters in a publication-based thesis are examined precisely because they are where the candidate's own synthetic thinking is visible, and having them written elsewhere defeats their purpose. The skill returns the coherence problems; the candidate resolves them in their own words.

2. **Never estimate attrition, recruitment yield, effect sizes or timelines from model knowledge and present them as planning figures.** These come from comparable studies with citations, from the candidate's own pilot, or from the institution. An invented attrition rate produces a sample size that is precisely wrong (K4 §2.1).

3. **Never construct a dependency that did not exist.** Where studies were run in parallel or where study two was designed before study one reported, say so. A retrofitted sequencing argument is detectable from the study two protocol dates and it damages the whole programme's credibility.

4. **Never present a schedule as feasible without the arithmetic.** Feasibility claims are checkable. Where the phases and the calendar have not been laid out, the honest statement is that feasibility has not been assessed.

5. **Never let a longitudinal design license a causal claim on its own.** Temporal order is necessary and not sufficient. Time-varying confounding and selection into exposure survive repeated measurement (K4 §3.2), and the design section must say what identification strategy, if any, addresses them.

6. **Never assume measurement invariance.** Where it has not been tested and cannot be, any claim about change over time is capped at moderate confidence and must carry the caveat in the same passage (K3 §4.2).

7. **Never report a mixed-methods integration that did not happen.** Two strands presented sequentially with a paragraph asserting convergence is not integration, and describing it as such misrepresents the design.

8. **Never resolve divergence between strands or between waves by selecting the more coherent result.** Report both, and treat the divergence as a finding.

9. **Never soften an infeasible programme.** If the design does not fit the time available, that is the output, delivered plainly and early. Producing an encouraging plan that cannot be executed is the most consequential failure available in this skill, because its cost is measured in years of the candidate's life.

## 13. Best-practice principles

- **A sequence is real only if the downstream design is a function of the upstream result.** Everything else is a chronology, and there is no shame in a parallel programme honestly described.
- **Plan for the null.** It is the most likely single outcome of a well-designed study of an uncertain effect, and it is the outcome for which candidates have no plan.
- **Set decision dates, not just milestones.** A milestone marks progress; a decision date forces the branch to be chosen while there is still time to execute it.
- **Attrition is a finding about your sample, not a nuisance in your denominator.** Who leaves tells you something, and it usually tells you which way your remaining estimates are biased.
- **Never improve an instrument mid-study without overlap.** The improvement costs you the comparison, which was the reason for measuring twice.
- **Wave spacing encodes a theory about how long the process takes.** Make that theory explicit, because if the interval is wrong the study can return a null for a real effect.
- **Repeated cross-sections and panels answer different questions and are routinely confused.** A flat population trend is entirely compatible with enormous individual movement in both directions.
- **Reserve the writing period first and defend it.** Everything upstream expands into it, and a thesis written in a compressed period is a thesis with the coherence problems in **15.18 Thesis Architecture and Chapter Coherence**.
- **Every additional approval is an additional delay of unknown length.** A three-phase programme with three separate ethics submissions carries three schedule risks, and they do not overlap.
- **In a publication-based thesis the linking material is the thesis.** The papers are evidence; the argument that they add to one contribution is what is being examined.
- **Divergence between strands or waves is the most valuable output you will get.** It is also the one most likely to be quietly dropped, so decide how you will report it before you see it.
- **The most common cause of doctoral overrun is a design that never fitted.** Every month spent making the design fit before it starts is worth several months later.

## 14. Worked example

Generic fictional scenario, academic, environmental social science.

**INPUT**

A doctoral candidate has proposed three studies on household adoption of water-saving behaviour in a drought-affected region. Study one: a survey of 800 households. Study two: interviews with 30 households. Study three: a two-wave panel measuring behaviour before and after a tariff change. She has 30 months remaining and access secured only for study one.

**PROCESS**

*Step 1 and 2.* The programme question is drafted: what sustains water-saving behaviour after the immediate crisis that prompted it has passed? Study one contributes prevalence and correlates. Study two is proposed as "deeper understanding", which fails the earning test. Interrogating the sequence reveals that study two's design would be identical whatever study one found, because the interview guide was written from the literature. The fix is to make the dependency real: study two samples theoretically from study one's results, specifically the households whose stated commitment and reported behaviour diverge most, which study one can identify and nothing else can.

*Step 3.* Contingency branches are written. If study one finds the expected commitment-behaviour gap, study two investigates it. If the gap does not appear, study two instead samples the households whose behaviour persisted longest and investigates maintenance, using the same guide structure with different sampling. If study one's response rate collapses, study two recruits through a community organisation with a different, weaker sampling logic that is documented as a fallback. Decision date is set at month 11.

*Step 4 and 5.* Study three is examined. The intended claim is about individual behaviour change around the tariff event, so a panel is required and a repeated cross-section will not do. Two waves support change but not trajectory, which is accepted since the question is about the event. Spacing is set from the tariff implementation date rather than from convenience, with the pre-wave close enough to the change to avoid intervening events and the post-wave far enough after to be past the initial reaction, which is theorised at four months.

*The judgement call.* Step 10, the reverse schedule. Working back from submission with six months reserved for writing, study three's second wave must complete by month 22. The tariff change is scheduled for month 16, the post-wave at month 20, which fits, but only if study three's ethics approval is obtained by month 12 and recruitment for the panel runs concurrently with study two's fieldwork. It is tight and it depends on a policy date outside her control. The temptation is to proceed and hope the tariff date holds. The decision taken is to build an explicit alternative: if the tariff change is delayed past month 18, study three becomes a single-wave study of the anticipatory period, with a different and smaller claim, and this is written into the proposal now rather than negotiated in a panic later.

*Step 6 and 7.* Panel attrition is estimated from two comparable regional studies with citations, at around 30 per cent, so wave one is sized at 400 to retain the 280 the analysis needs. Instrument invariance is planned: the behaviour items are identical across waves, and a new item that would have been useful at wave two is added alongside rather than replacing an existing one, so the comparison survives.

**OUTPUT**

A three-study programme with one genuine dependency and one honest parallel component, a contingency table with three branches and a decision date, a panel design sized after attrition with an invariance plan, an explicit fallback for the external policy risk, and a reverse schedule that preserves the writing period. The programme is judged feasible with one named external risk that is now planned for rather than hoped away.

`RESEARCHER DECISION REQUIRED` on whether to accept the dependency on an externally scheduled tariff change at all, or to redesign study three around a within-household comparison that carries no external date risk (K5 §2.7).

## 15. Advanced usage

**Designed replication within a programme.** Where study one produces a striking result, a direct or conceptual replication as study two is a strong and under-used design: it strengthens the contribution, addresses the field's reproducibility concerns, and provides a defensible study two that does not depend on study one being interesting. Pre-register the replication, and specify in advance what would count as a successful replication, since a post-hoc judgement about whether an effect "replicated" is unfalsifiable.

**Cohort-sequential and accelerated designs.** Following several cohorts across overlapping age ranges buys partial separation of age from cohort in a fraction of the calendar time a single cohort would need. The cost is complexity in analysis and a set of assumptions about convergence between cohorts that must be tested rather than assumed.

**Measurement burst designs.** Clusters of intensive measurement (daily or weekly) separated by long intervals capture both short-term process and long-term change. They are powerful for questions where the mechanism operates on a much shorter timescale than the outcome, and they place a high burden on participants that must be planned for in retention.

**Programme-level pre-registration.** Registering the whole programme, including the contingency branches and their decision rules, is unusual and strong: it demonstrates that the branch taken was pre-specified rather than chosen after seeing the results, which is otherwise very hard to establish. See **15.16 Doctoral Methodology Justification and Rigour**.

**Recovering a programme mid-flight.** Where study one has already returned something unexpected, run Steps 1 to 3 in reverse: identify what the completed study actually established, what programme question that could support, and what study two would now have to be. This frequently produces a better programme than the original, and it must be reported as a change with reasons rather than presented as the original plan.

## 16. Skill chain

**Recommended previous skills:**
- **15.13 Doctoral Research Positioning and Originality.** Hands over the programme-level contribution the sequence must deliver.
- **15.15 Theory Building and Conceptual Contribution.** Hands over the propositions that determine what each study must test and in what order.
- **15.16 Doctoral Methodology Justification and Rigour.** Justifies each component design; this skill connects them.
- **01.04 Research Method Selection.** Hands over the method choice per component.

**Recommended next skills:**
- **15.09 Ethics Clearance Application.** Takes each phase into approval, on the timeline this design sets.
- **05.01 Statistical Analysis Planning** and **06.02 Advanced Statistical Modelling.** Take the longitudinal design into the analysis it permits.
- **15.18 Thesis Architecture and Chapter Coherence.** Takes the programme structure into the thesis structure, particularly for publication-based formats.
- **15.19 Examiner-Proofing and Defence Strategy.** Prepares the sequencing argument and the change claims for challenge.

**Runs well alongside:**
- **04.04 Longitudinal and Tracking Study Management**, for the operational running of multi-wave fieldwork.
- **15.24 Journal Article Development**, where each study becomes a paper.
- **13.01 Research Quality Review**, which audits the programme design independently.

---
A Yazi Supplied Skill and resource.