Research design: turn a business decision into an answerable question

A research design connects a decision to an answerable question, suitable evidence and a credible way to interpret it. Start with what someone needs to decide. Then specify what you must learn, from whom or what, and what conclusion the chosen method can support.

The distinction between exploring a problem, describing a pattern and examining a cause is useful because those tasks require different evidence. Interviews can reveal how a delivery went wrong. A properly designed survey or record review can estimate how often a defined problem occurs. A defensible comparison can examine whether a proposed change caused an improvement.

This playbook uses a hypothetical garden centre considering a delivery checklist. The example and planning table are original teaching aids, not a report of a trial. The aim is a plan another colleague can challenge before anyone starts collecting data.

Name the decision and its owner

“Research our customers” is an activity label. “Decide whether to fund a rehearsal of a complete-order delivery service” identifies a commitment. The latter makes it possible to ask which uncertainty matters now and what research can reasonably resolve.

Our garden centre receives complaints about deliveries containing only part of an order. The manager wants a checklist. Before discussing a study, the team clarifies what remains open: revising the dispatch process, testing a checklist or changing the promise made to customers. It also identifies who can authorise each response.

GOV.UK’s research-planning guidance recommends connecting research questions, user groups and research activities. Apply that logic to the actual commitment. A broad discovery question may be appropriate before the team understands the problem; a narrow test may be appropriate once the mechanism is plausible.

The Double Diamond helps locate the work. It does not require every project to pass through the same research sequence. If the centre already has reliable evidence of a straightforward recording error, it need not delay a correction to run an expansive discovery programme.

Separate exploring, describing and explaining

An exploratory question asks what is happening and which explanations deserve attention. “How do orders become split between dispatch and delivery?” invites a reconstruction of a process. It can reveal variables or situations the team had not thought to measure.

A descriptive question asks about a defined population or pattern. “What proportion of completed delivery orders in the last month had a missing item reported within seven days?” needs an agreed denominator, a time window and a source that can identify the cases. It does not explain why the items were missing.

A causal question asks what difference a change makes. “Does the checklist reduce incomplete deliveries compared with the existing dispatch arrangement?” requires a credible account of what would have happened without the checklist. It also needs sufficiently clear outcomes and operating conditions.

GOV.UK’s guide to capturing research questions distinguishes what a team needs to learn from the questions it asks participants. “Why do deliveries fail?” is a research question. The participant prompt might instead ask a dispatcher to reconstruct the last order that was split.

Do not rank these question types as increasingly respectable research. A precise causal estimate of an irrelevant intervention is less useful than a careful exploratory account that changes the problem definition. Choose the contribution the decision needs.

Specify the people, cases and conditions

Define the unit you will study. It might be a customer, an order, a delivery trip or a dispatch shift. Different units produce different answers. A customer with three deliveries should not silently become three independent customers in the analysis.

For the garden centre, one order may leave on two vehicles because large pots and fragile plants need different handling. The team must decide whether an agreed split counts as incomplete delivery. Otherwise its headline measure can label an intended service arrangement as failure.

Specify dates, channels and exclusions. Does the study include click-and-collect orders, subcontracted routes and cancellations? Can the record system identify missing-item complaints received by telephone as well as online? Which cases never generate a completed order?

Use stakeholder mapping to identify dispatchers, drivers, reception colleagues and affected customer groups. Then write a sampling plan suited to the question. A list of organisational roles is not automatically a population frame or a research sample.

State access limitations early. If subcontractor records cannot be inspected, narrow the conclusion or arrange another evidence route. Do not leave that boundary until the final slide.

Choose evidence by what it can answer

Begin with desk research if current records or earlier studies might answer part of the question. Check how those records were produced before accepting their categories. A complaint log describes reported problems, not every delivery problem.

To reconstruct dispatch, observe agreed tasks and use customer interviews for recent accounts of receiving an incomplete order. Each reveals a different part of the sequence. Compare smooth cases as well as difficult ones so the team can examine meaningful differences.

For incidence, decide whether a suitable administrative dataset exists or a new survey is needed. AAPOR’s survey best-practice guidance asks whether a survey is the right method and whether existing evidence is sufficiently relevant and timely. Asking people is not the default answer to every information gap.

For comprehension of a proposed checklist, a task rehearsal or concept test may be sufficient. It can reveal unclear categories without estimating a real-world delivery effect. A successful rehearsal does not show that staff will use the checklist under a busy Saturday workload.

Keep a method’s useful contribution and its limit in the same sentence: “A dispatch walkthrough will reveal how these colleagues assemble these orders; it will not estimate all delivery failures.” That makes later reporting easier to keep honest.

Give causal claims a credible comparison

A before-and-after change can have several explanations. Demand may fall, weather may change, a different dispatcher may work the shift or suppliers may deliver more complete stock. A checklist introduced at the same time as these changes cannot automatically receive all the credit.

HM Treasury’s Magenta Book explains counterfactual reasoning and experimental and quasi-experimental approaches. Its central question is what would have happened without the intervention. Government policy evaluation is a different setting, but the identification problem also applies to this delivery question.

Random allocation can support a comparison where it is feasible and appropriate. It does not remove the need to define the intervention, measure consistently, consider interactions between groups and plan a suitable analysis. If dispatchers use the checklist on the comparison orders as well, the intended contrast may disappear.

The garden centre might first rehearse the checklist with ordinary orders to establish whether it is usable. A later effect study would need an appropriate unit of allocation: mixing arrangements within one dispatch process could cause confusion or contamination. A researcher with suitable statistical expertise should design that study before a broad causal claim is required.

Do not add a control group merely to make a plan sound scientific. The group must support a relevant comparison. If comparable data cannot be obtained, report what the study does show and narrow the causal conclusion.

Write the plan before recruitment

This compact record is an original planning template. Complete it with the people who will collect the data and those who will use the result. Leave questions visibly unresolved where the information is not yet available.

Planning element	Hypothetical garden-centre entry	Check before starting
Decision	Choose the next dispatch investigation	Who can authorise the work?
Research question	How do required items become separated?	Does this question precede a checklist decision?
Unit and scope	Delivery orders assembled at one site	How are intended split orders recorded?
Evidence route	Records plus contextual walkthroughs	Can telephone complaints be linked to orders?
Interpretation	Compare pathways and contrary cases	Which explanation would change the next action?
Review	Dispatch manager and research lead	Are limitations carried into the decision?

Add practical arrangements: recruitment, consent, accessibility, storage, analysis time and scheduling. The research plan should include the work after the last session. A calendar full of interviews with no analysis time is an incomplete commitment.

Use assumption mapping to identify which uncertainty could invalidate the proposal. The team may discover that stock is confirmed after delivery is promised. Testing checklist wording would then answer a secondary question while the main operational risk remains.

Define the decision rule without pretending certainty

Agree how different findings would affect the next step. A rule can be qualitative: if required items cannot be identified at dispatch, investigate the stock and order data before testing a checklist. If staff can identify them but consistently misunderstand one category, revise the wording and rehearse again.

For a quantitative effect study, the relevant difference, uncertainty and analysis should be planned in advance. Avoid a threshold selected after the team sees the result. Equally, do not invent “80% approval” when the actual question concerns a missing essential item or an operational burden.

Include conditions that would pause the work and what happens if the evidence is inconclusive. Missing data, a changed supplier process or an unavailable participant group can alter what the planned analysis means. An inconclusive study is not permission to describe the favoured option as proven.

Connect the next study to staged innovation funding. Fund the evidence needed for the next commitment and name what remains outside its scope. An exploratory study can justify another investigation without justifying a full implementation budget.

Analyse against the question and report the boundaries

Preserve the original question and record any revisions. Exploration legitimately changes as new circumstances emerge. The important distinction is between a documented change of focus and silently replacing the outcome because the original one was disappointing.

GOV.UK’s benefits guidance recommends baselines and revisiting benefit estimates as evidence develops. That supports planning what improvement would mean. A baseline alone does not establish that a later change caused the result.

At the review, separate findings, interpretation and the proposed action. The garden centre might find an unclear promise, a stock-data delay and occasional dispatch omissions. Those suggest different responses. Returning everything as “customers want better communication” would lose the operating distinctions the study was intended to reveal.

Use the findings during idea evaluation to compare proportionate next steps. If one assumption is ready for a bounded test, continue with experiment design. Carry the uncertainty into that plan instead of treating the earlier study as a blanket validation.

Evidence and limitations

This is an introductory planning method built from institutional research and evaluation guidance. It is not a complete statistical analysis plan, a sampling manual or a formal research protocol for every setting. The hypothetical garden centre supplies no evidence of effectiveness.

A research design can make a question and its limits explicit while still producing uncertain findings. Access, measurement, sample selection and interpretation all remain consequential. Combining methods helps only when each addresses a stated gap; several weak methods do not automatically cancel one another’s weaknesses.

Finish with a reviewable agreement: the decision, question, scope, method, analysis, owner and rule for the next action. If those cannot be connected, refine the design before spending people’s time on collection.