eduKateSG Learning Node Series · 0089

How Evidence-Informed Decision Making Works | Combine Research, Local Data, Professional Judgement and Values Before You Act

A school can have more evidence than it knows what to do with.

There may be assessment dashboards, attendance records, student work, teacher observations, parent feedback, research summaries, inspection findings, intervention studies, budget constraints, curriculum requirements and years of professional experience. The hard part is not always finding another piece of information. The hard part is deciding what all of it should mean for the next action.

Evidence-informed decision making is the discipline of combining the best relevant research with local evidence, professional expertise, contextual constraints and the values attached to the decision. It is not a ritual in which somebody finds one study and says, “Research says we must do this.” It is a structured way of making judgement more inspectable, more defensible and easier to revise when reality disagrees.

Evidence should constrain a decision, improve a decision and sometimes overturn a decision. It should not pretend to make the decision by itself.

The 50-Second Read

• Evidence-informed decision making begins with a real decision, not with a favourite study or dashboard.
• Research evidence, local data, professional judgement, stakeholder knowledge and values answer different parts of the problem.
• Evidence quality and evidence relevance are separate questions.
• An average effect in research is not a guarantee for one school, learner or classroom.
• Descriptive data can reveal what is happening without proving why it happened.
• Good decisions make assumptions visible, especially assumptions about mechanism, capacity, cost and context.
• Triangulation is stronger when different evidence sources fail in different ways.
• High-stakes or hard-to-reverse decisions require a stronger evidence threshold than small, reversible experiments.
• Decision records preserve what was known, what was uncertain and why an option was chosen.
• A decision should include monitoring and stop rules, not merely an implementation date.
• Evidence use is a loop: decide, act, observe, update and decide again.

Canonical Owner Boundary

This Learning Node owns the decision layer that combines multiple forms of evidence into an educational choice. How Knowledge Brokering Works remains the owner for moving research across the boundary between evidence producers and users. How Student Work Analysis Works owns disciplined reading of learner artefacts. How Teacher Sensemaking Works owns how educators interpret reforms and messages through prior beliefs and local conditions. How Education Works | School Improvement owns the larger school-change architecture. This page owns the moment in which evidence streams, trade-offs and values become a decision.

1. Start With the Decision, Not the Evidence

“What does the research say?” sounds sensible, but it is incomplete until the decision is named. Are we choosing a reading programme? Deciding whether to reteach a unit? Reallocating intervention time? Changing homework? Selecting professional development? Each decision changes which evidence matters.

A useful opening sentence is: We need to decide whether to ___ for ___ because ___, by ___. That sentence forces the problem to acquire a population, an action, a reason and a deadline.

2. Evidence-Informed Is Not Evidence-Obedient

Education contains value judgements that research cannot settle. How much homework is acceptable for a ten-year-old? How should a school balance attainment, wellbeing, inclusion and workload? What level of uncertainty is tolerable before changing a timetable?

Research can estimate likely consequences under studied conditions. It cannot decide every priority. Evidence-informed practice therefore treats evidence as an input into judgement, not as a substitute for judgement.

3. Separate the Evidence Streams

Four streams commonly enter an educational decision. Research evidence tells us what has happened across studied settings. Local data tells us what appears to be happening here. Professional knowledge contributes practical understanding of learners, curriculum, staffing and feasibility. Values and stakeholder knowledge determine what outcomes matter and what costs are acceptable.

Problems begin when one stream impersonates all the others. A local test score is not a research synthesis. A research synthesis is not knowledge of one child. A teacher’s expertise is not a randomised trial. A parent’s concern is not noise merely because it is qualitative.

4. Ask What Kind of Question You Are Asking

Different questions require different evidence. “What is happening?” is descriptive. “Why is it happening?” is explanatory. “What will happen if we change X?” is causal. “Can we implement this?” is operational. “Should we do it?” combines evidence with values and trade-offs.

One of the fastest ways to misuse evidence is to answer a causal question with descriptive data. Attendance and attainment may move together without proving that one caused the other. Strong decisions keep the question type visible.

5. Search for the Best Available Research, Not the Most Convenient Citation

A persuasive article, conference slide or vendor case study can be useful, but it should not automatically become the evidence base. Look first for trustworthy syntheses, systematic reviews, practice guides and replications before building a large decision on one result.

The OECD’s 2025 work on education research use emphasises the infrastructure and capacity needed for evidence to inform policy and practice. The Education Endowment Foundation’s guidance on using its Toolkits likewise stresses that evidence summaries provide “best bets” rather than deterministic answers and must be combined with professional judgement.

6. Evidence Quality and Evidence Relevance Are Different

A beautifully designed study may involve a different age group, subject, staffing model or dosage from the decision in front of you. A less rigorous local study may be more contextually similar but weaker for causal inference.

Good decision-making does not collapse these into one score. Ask separately: How trustworthy is this evidence? and How directly does it speak to our situation?

7. Read Average Effects as Averages

An intervention with a positive average effect can contain settings with larger gains, smaller gains, no gain or harm. The average is useful because it summarises a distribution. It is dangerous when read as a promise.

Look for variation by age, prior attainment, implementation quality, subject, dosage and context. Then ask which parts of that variation resemble your own conditions.

8. Local Data Is Close to the Problem but Not Automatically True

Local data feels persuasive because it is ours. Yet local evidence can be noisy. Cohorts differ. Tests change. Missing data is rarely random. Teachers alter marking. Small groups swing sharply. A dramatic chart can be an artefact of scale.

Proximity increases relevance, not validity. Local data deserves the same disciplined questions as external research: what was measured, how, when, from whom and compared with what?

9. Professional Judgement Is More Than Opinion

Experienced educators carry compressed knowledge about misconceptions, pacing, classroom dynamics, curriculum dependencies and implementation friction. That expertise can identify conditions a study does not report.

But expertise can also contain habit, local folklore and confidence unsupported by outcomes. Strong professional judgement is therefore explicit about its reasoning and willing to be tested against evidence.

10. Values Enter Even When Nobody Names Them

Suppose two programmes produce similar attainment gains. One requires substantially more homework; the other costs more money. Research can estimate outcomes and workload. It cannot tell a community exactly how to value family time against budget.

If values remain hidden, they masquerade as facts. Name them.

11. Define the Outcome Before Comparing Options

“Better” is not an outcome. Better reading fluency? Better comprehension? Better examination performance? Lower teacher workload? Fewer absences? More equitable participation?

A decision can look evidence-informed while quietly changing the outcome halfway through. Define the primary outcome and important secondary outcomes before comparing alternatives.

12. Identify the Mechanism

Ask why an intervention is expected to work. Smaller groups may increase opportunities for feedback. Retrieval practice may strengthen access to knowledge. Coaching may change classroom behaviour through observation, modelling and rehearsal.

If the mechanism depends on conditions you do not have, the headline effect size may not travel. Mechanism is the bridge between external research and local implementation.

13. Make the Counterfactual Visible

The real question is rarely “Does this programme work?” It is “Does this programme outperform what we would otherwise do, for these learners, with these resources?”

Opportunity cost lives inside the counterfactual. An intervention that helps may still be a poor decision if a cheaper or more feasible alternative helps more.

14. Triangulate Evidence That Fails Differently

Triangulation is not collecting three versions of the same weak signal. A test score, teacher observation and student interview are useful together partly because their errors differ.

If all three point toward the same learning problem, confidence rises. If they disagree, the disagreement becomes information. Perhaps the test is narrow. Perhaps students can perform but cannot explain. Perhaps teachers overestimate independence.

15. Distinguish Signal From Measurement

Every measure is a representation. A score stands in for some capability. A survey stands in for attitudes. Attendance stands in for presence, not engagement. Behaviour referrals stand in for recorded incidents, not every behaviour that occurred.

Evidence-informed decisions keep a gap between the thing and its measure. Otherwise the organisation begins managing the number instead of the underlying reality.

16. Watch for Goodhart-Type Failure

When a measure becomes a target, behaviour can reorganise around the metric. A school may raise practice-test scores while narrowing the curriculum. An intervention may reduce recorded incidents because staff stop recording borderline cases.

This does not mean targets are useless. It means decisions need countermeasures: multiple indicators, audits, qualitative evidence and attention to unintended consequences.

17. Match Evidence Threshold to Decision Risk

A teacher trying a five-minute retrieval routine for one week does not need the same evidence threshold as a system purchasing a multi-year platform for every school.

The harder a decision is to reverse, the more expensive it is, the more learners it affects and the greater the potential harm, the stronger the case should be before committing.

18. Reversibility Is a Decision Variable

Small reversible moves can be used to learn. Pilot one timetable change. Test one feedback protocol. Trial one diagnostic routine. Observe before scaling.

This is not indecision. It is buying information cheaply before buying commitment expensively.

19. Define What Would Change Your Mind

Before implementation, write down disconfirming evidence. What result would make us stop? What would make us adapt? What would count as “not enough improvement”?

Predefining these thresholds protects against escalation of commitment, where an organisation keeps defending a decision because it has already invested in it.

20. Cost Is Part of the Evidence

Implementation consumes money, time, attention, training capacity and political capital. A programme with modest benefits and very high implementation burden may lose to a simpler option with slightly smaller average impact.

Cost-effectiveness is not anti-education. It is how finite educational resources become more learning rather than more activity.

21. Workload Is a Real Outcome

A new intervention can improve student outcomes while exhausting the staff required to run it. If workload makes fidelity collapse after one term, the intervention may not be sustainable.

Measure burden as well as benefit: preparation time, marking, meetings, data entry, training, supervision and opportunity cost.

22. Equity Changes the Question

An average improvement can conceal unequal effects. Who gains? Who loses? Who cannot access the intervention? Does the decision depend on devices, transport, parental time or prior knowledge that are unevenly distributed?

Disaggregate where sample size allows, but do not turn tiny subgroups into false precision. Equity requires both numbers and contextual knowledge.

23. Stakeholder Voice Is Evidence About Experience and Feasibility

Students, families and teachers can reveal friction that outcome data misses: confusing instructions, inaccessible materials, timetable clashes, anxiety, cultural mismatch or hidden support work at home.

Voice should not be treated as a referendum on causal effectiveness. It is evidence about experience, acceptability and implementation conditions.

24. Create a Decision Table

For competing options, build a table with criteria such as expected learning benefit, evidence strength, contextual fit, cost, workload, implementation capacity, equity, reversibility and risk.

The table does not make the decision. It prevents one attractive feature from swallowing the rest of the decision.

25. Record the Assumptions

“This will work if teachers receive coaching every fortnight.” “This will work if students actually complete the practice.” “This will work if the timetable protects the intervention.”

Assumptions are future failure points in disguise. Recording them turns implementation monitoring into a test of the decision model.

26. Write a Decision Record

• Decision: what was chosen?
• Problem: what need was being addressed?
• Options: what alternatives were considered?
• Evidence: which research and local sources mattered?
• Assumptions: what had to be true?
• Values: which outcomes or principles were prioritised?
• Risks: what could go wrong?
• Monitoring: what will be measured and when?
• Stop rule: what would trigger adaptation or abandonment?
• Review date: when will the decision be reconsidered?

27. Date the Evidence Base

Evidence ages. New reviews appear. Programmes change. Staff capacity changes. Technologies update. What was a reasonable choice in 2024 may need reconsideration in 2027.

Every consequential decision should have a “knowledge current as of” date and a review trigger.

28. Distinguish Adoption From Implementation

Choosing an evidence-informed intervention does not produce an evidence-informed outcome. Training, materials, leadership, data systems and ongoing support determine whether the intended mechanism reaches learners.

That is where How Implementation Support Systems Work takes ownership. The decision page hands the chosen option forward with its assumptions intact.

29. Cross-Domain Comparison: Clinical Decision-Making

Medicine often combines research evidence, clinical expertise and patient preferences. The analogy is useful because treatment evidence does not remove individual variation, feasibility or values.

Education should not imitate medicine mechanically. The transferable idea is epistemic: strong decisions integrate external evidence with local conditions and human judgement.

30. Cross-Domain Comparison: Aviation Safety

Aviation does not rely on one source when judging risk. Incident reports, maintenance data, flight data, engineering analysis and operating experience can all contribute. Each captures a different slice of reality.

The educational lesson is triangulation with traceability: do not wait for one perfect metric when several imperfect signals can be interpreted together.

31. Cross-Domain Comparison: Investment Decisions

Investment analysis separates expected return from risk, scenario and liquidity. A project with attractive upside may still be rejected if the downside is catastrophic or the commitment is difficult to reverse.

Education decisions also deserve scenario thinking. What happens if uptake is only half of plan? If staff turnover rises? If the intervention effect is smaller than the research average?

32. Example: Choosing a Reading Intervention

A school identifies a reading comprehension gap. Three programmes look plausible. Instead of asking which has the strongest marketing, the team defines the learner group, outcome, current provision and resource constraints. They examine independent evidence, compare programme dosage and staffing assumptions, inspect local reading data and speak with teachers about timetable feasibility.

One programme has stronger average effects but requires specialist staffing the school cannot sustain. Another has more modest effects but fits existing lessons and includes usable training. The decision is not “ignore the research.” It is “use the research together with feasibility so the chosen mechanism can actually operate.”

33. Example: A Falling Examination Score

A year group’s examination performance falls. The immediate temptation is to change the curriculum. Before doing so, the team checks item-level results, attendance, teacher changes, cohort prior attainment, marking consistency and student work.

The problem turns out to be concentrated in multi-step application questions rather than the whole curriculum. The decision narrows from “replace the programme” to “increase guided application practice and monitor transfer.” Better diagnosis produced a smaller, more defensible action.

34. Failure Mode: Research Says

A leader uses the phrase “research says” as a conversational full stop. Staff are shown one source. Questions about fit are treated as resistance.

Evidence has become authority theatre. Strong evidence use makes the chain inspectable and permits challenge.

35. Failure Mode: Dashboard Governance

The organisation makes decisions from whatever is easiest to count. Richer evidence disappears because it does not fit the dashboard.

A dashboard is a display layer, not reality. The decision process must be able to ask for evidence that the dashboard does not yet contain.

36. Failure Mode: Local Exceptionalism

“Our students are different, so outside research does not apply.” Sometimes contextual differences matter. Sometimes the phrase protects a familiar practice from scrutiny.

The correct response is not automatic transfer or automatic rejection. Specify which contextual feature is different and how it changes the mechanism.

37. Failure Mode: Evidence Accumulation Without a Decision

Teams can keep reading because reading feels safer than choosing. More documents accumulate while the decision deadline approaches.

Set a decision threshold: what information is essential, what uncertainty can remain, and what can be learned after a reversible first step?

38. A Practical Evidence-Informed Decision Stack

• Name: define the decision and deadline.
• Frame: specify population, outcome, options and constraints.
• Search: find the strongest relevant research.
• Appraise: separate quality from contextual relevance.
• Diagnose: examine local quantitative and qualitative evidence.
• Interpret: add professional knowledge without treating confidence as proof.
• Surface values: state what outcomes and trade-offs matter.
• Compare: make the counterfactual and alternatives explicit.
• Stress-test: ask what happens if assumptions fail.
• Decide: choose at the evidence threshold appropriate to the risk.
• Record: preserve sources, reasoning and uncertainty.
• Implement: carry mechanism and conditions forward.
• Monitor: collect evidence that can change the decision.
• Review: update rather than defend.

39. What Current Evidence-Use Guidance Adds

The OECD’s Everybody Cares About Using Education Research Sometimes (2025) documents how knowledge intermediaries across 34 countries support evidence engagement, collaboration and capacity. Its broader message matters here: evidence use is not merely access to studies; it depends on relationships, skills, organisational conditions and evaluative thinking.

The OECD’s earlier work on the changing landscape of research use in education distinguishes research evidence from wider evidence and emphasises that professional expertise remains part of evidence-informed judgement. The EEF’s current Toolkit guidance similarly warns against treating average findings as definitive instructions for one setting.

40. Missing-Node Scan: Questions Before a Consequential Decision

• What decision are we actually making?
• What would happen if we did nothing?
• Which outcome matters most?
• What is the strongest relevant research, and how current is it?
• How similar are the research conditions to ours?
• What does our local evidence show—and what can it not show?
• Which professional assumptions are carrying the decision?
• Whose values or experience have not been represented?
• What is the opportunity cost?
• Which option is easiest to reverse?
• What would change our mind?
• What unintended consequences should we monitor?
• When will the decision be reviewed?

41. The Return Path

The meeting begins with a sentence that sounds decisive: “The evidence says we should buy Programme A.”

Ten minutes later, the sentence has improved. Programme A has the strongest published evidence for a population similar to ours. The effect depends on a dosage we can only partly provide. Programme B has weaker evidence but fits the existing timetable. Our local data suggests the problem is concentrated in one subgroup. Teachers believe the key barrier is practice quality rather than curriculum materials. Families are concerned about the additional homework requirement. We can pilot A with one cohort, define the expected signal, and review before scaling.

The answer became less certain and more useful.

That is what evidence-informed decision making is for. Not to make human judgement disappear, but to make it better connected to reality.

A mature evidence system does not ask people to choose between data and judgement. It teaches them how to make judgement answer to evidence.