---
name: qualitative-and-quantitative-integration
description: >
  Brings a qualitative and a quantitative evidence stream together properly,
  including when they disagree. Use for "integrate the qual and quant", "the
  survey and the interviews say different things", "mixed methods analysis",
  "triangulate these findings", "why does the qual contradict the data", "build
  a joint display", "combine the depth interviews with the survey".
category: 07 Qualitative Analysis
ref: 07.06
tier: 1
inherits: [K2, K3, K4, K5]
---

# Qualitative and Quantitative Integration

## 1. One-line description
Joins a qualitative and a quantitative stream into a single account by deciding which stream leads on which question, testing convergence and divergence explicitly, and investigating disagreement rather than averaging it away.

## 2. What this skill is used for

**The research problem it solves.** Most mixed-method studies are not integrated. They are two studies reported in sequence, with a sentence at the join claiming they support each other, which nobody has checked. Two opposite failures follow. In the first, the qualitative material is demoted to decoration: the survey supplies every finding and the interviews supply pull-quotes underneath them, which wastes the entire qualitative investment and makes it impossible for the qual to contradict anything. In the second, the qualitative material is quantified: percentages appear from twenty interviews, and a purposive sample of twenty is reported as though it estimated a population. Both failures come from the same root, which is treating the two streams as the same kind of evidence with different sample sizes. They are different kinds of evidence answering different questions, and the value of running both is not that they agree. **It is that where they disagree, something is true that neither stream could have shown alone.** This skill makes disagreement visible and then does something disciplined with it.

**Where it sits.** Synthesis. It takes analysed qualitative themes and analysed quantitative results and hands one integrated evidence position to insight development and reporting.

**Typical use cases.**
- A survey with a qualitative phase before it, after it, or alongside it.
- Depth interviews that need testing against behavioural or transactional data.
- Programme or service evaluation where outcome measures and participant accounts must be reconciled.
- A tracker whose numbers moved and whose qualitative phase explains, or fails to explain, why.
- Two commissioned studies on the same question, run separately, now needing one answer.
- A situation where the qualitative and quantitative findings point different ways and someone must decide what to tell the client.

**Who uses it.** Research directors owning a mixed-method deliverable; evaluators; insight leads reconciling survey data with customer research; academics writing an integration chapter; anyone holding two evidence streams and one narrative.

## 3. When to use it

- You hold analysed findings from both streams, and a single account has to be produced.
- The two streams point in different directions and the difference needs explaining rather than smoothing.
- The study design was mixed-method by intent, and the integration is a promised deliverable.
- A qualitative phase was run to explain a quantitative result, or a quantitative phase to test a qualitative one.
- Confidence in a conclusion needs to be raised or lowered on the basis of what more than one method shows.
- A stakeholder is quoting the stream that suits them and someone needs to establish what both together support.
- The deliverable will be scrutinised on method, and the integration logic has to be visible rather than asserted.

## 4. When NOT to use it

- **Either stream has not been analysed yet.** Integration works on findings, not on raw material. Interviews go through **07.01 Thematic Analysis** or **07.03 Interview and Transcript Analysis** first; survey data goes through **05.01 Descriptive Analysis** and, where comparisons are claimed, **05.02 Statistical Testing**. Integrating raw qual with analysed quant guarantees the qual is subordinated, because it arrives without structure of its own.
- **The two streams are not about the same thing.** Before integrating, check that the constructs, populations and periods overlap enough to be comparable. Two studies of adjacent constructs on different populations cannot be integrated; they can be reported side by side with the difference stated. Forcing them produces a conclusion neither supports.
- **One stream is being used to authorise a conclusion the other refuses.** Where a stakeholder wants the survey to override inconvenient interviews, or the interviews to override an inconvenient survey, that is selection, not integration, per K4 §4.2. Say so, and produce the honest position, which is that the streams disagree.
- **The qualitative material is only there to be quoted.** If no qualitative finding is permitted to change a quantitative conclusion, there is no integration to do. Report the quantitative study and label the verbatims as illustration, per **07.04**.
- **A population estimate is wanted from the qualitative stream.** Qualitative samples are recruited for range and depth, not incidence. A proportion from them is not an estimate of anything, per K4 §3.3, and it will be quoted as one.
- **The streams come from populations that cannot be reconciled.** A customer survey and interviews with lapsed customers are about different people, and apparent divergence between them is a definition, not a finding. Establish the population overlap first, or integrate nothing.
- **Multi-source compilation is what is actually needed.** Where five or more heterogeneous inputs (desk research, client data, two surveys, a workshop) must be assembled, the task is **10.03 Multi-Source Research Synthesis** or **12.04 Research Evidence Integration**, not two-stream integration.
- **Causal claims are the deliverable.** Integration raises or lowers confidence in an association. It does not license causal language, whichever way the two streams point, per K4 §3.2. A mechanism described in interviews and a correlation in a survey are still not a causal design.
- **The streams are separated by a long period and the question is change.** Where a qualitative phase ran two years after the survey it is being compared with, the difference between them is at least partly time, and the first analysis is a trend question rather than an integration one. Use **05.05 Trend and Tracker Analysis** to establish what moved before asking why the streams differ.
- **One stream is known to be unsound.** Integration cannot repair a broken instrument, an unreliable code frame or an untested comparison. Combining a weak stream with a strong one produces a conclusion that inherits the weakness and conceals it behind the appearance of triangulation. Fix or discard the weak stream first, per K3 §3.6.

## 5. Required inputs

**Required.** Without these the skill cannot run. If absent, ask. If no answer is available and the work must proceed, state the assumption at the point where it bites, per K5 §5.

- **Analysed findings from both streams**, with their evidence attached: themes with prevalence, counts and denominators from the qualitative side; results with question wording, bases and test status from the quantitative side.
- **The sample definition and fieldwork period for each stream.** Almost every apparent divergence is partly explained here, and integration without them is guesswork.
- **The question wording or discussion guide for each.** What was asked determines what a result means, and the two streams frequently asked adjacent rather than identical things.
- **The research objectives**, so that stream leadership can be assigned per question rather than in general.

**Optional, and what each one adds.**

- **The design intent** (which phase was meant to inform which): tells you what the study licensed itself to conclude, which is not always what the client now wants concluded.
- **Behavioural, transactional or operational data**: adds a third stream that is frequently the arbiter when stated and reported evidence disagree, because it is not self-reported.
- **Mode and channel details for each stream**: makes social-desirability and mode-effect explanations testable rather than speculative.
- **Overlapping respondents** (the same people in both streams): converts group-level comparison into individual-level comparison, which is far stronger evidence and makes the interesting disagreements visible at the person level.
- **Prior waves of either stream**: lets you check whether an apparent divergence is a period effect.
- **The weighting scheme on the quantitative stream**: determines what population the numbers describe, which is what the qualitative stream is being compared against.

## 6. Questions to ask before starting

1. **Which stream leads on which research question?** They are not equal on everything, and treating them as equal is how integration becomes averaging. *Default if unanswered:* the quantitative stream leads on incidence, size and distribution; the qualitative stream leads on mechanism, meaning, language and why; and the assignment is written down per objective before any comparison is made.
2. **What was the design, and what does it license?** *Default:* infer it from the sequence and state it, along with the limitation it carries.
3. **Do the two streams cover the same population and period?** *Default:* map both, state the overlap, and treat any non-overlap as a standing candidate explanation for divergence.
4. **Are the constructs the same, or only similarly named?** Satisfaction with a brand and satisfaction with a recent service event share a word and are different things. *Default:* align constructs explicitly at step 4 and record every pairing that is approximate.
5. **Are there overlapping respondents?** *Default:* check, because individual-level comparison changes what integration can establish.
6. **What is the client expecting to be told?** Useful as information about the stakeholder, dangerous as a target, per K4 §4.2. *Default:* record it before integrating so that agreement with it is visible if it happens.
7. **If the streams disagree, who decides what is reported?** Divergence is a judgement point, not an arithmetic one. *Default:* produce the divergence analysis with the candidate explanations assessed, and mark it for researcher decision per K5 §3.1.

## 7. Step-by-step methodology

**1. Inventory what each stream can and cannot answer.** Before comparing anything, write, for each stream: the population it represents, the period, the mode, the base or participant count, the constructs it measured and how, and the claims it can and cannot support. *Correct result:* a two-column capability table. This is where you discover that the survey cannot speak to mechanism and the interviews cannot speak to size, which is the whole reason both exist. An integration that starts with the findings rather than the capabilities will compare things that were never comparable.

**2. Identify the design, and state what it licenses.** Four designs cover most work, and each permits a different kind of conclusion.

- **Sequential exploratory** (qualitative first, then quantitative). The qualitative phase generates the constructs, language and hypotheses; the quantitative phase measures them. *Licenses:* claims about the incidence of things the qualitative phase discovered. *Does not license:* treating the qualitative phase as validated because the survey was large. The survey can only measure what the qual thought to put in it, so anything the qual missed is invisible in both.
- **Sequential explanatory** (quantitative first, then qualitative). The survey establishes the pattern; the qualitative phase explains it. *Licenses:* mechanism claims about the pattern, at qualitative confidence. *Does not license:* treating the explanation as tested. The participants were often selected because they exemplified the pattern, so the qualitative phase cannot disconfirm it.
- **Concurrent triangulation** (both at once, independently). *Licenses:* the strongest convergence claim available, because neither stream shaped the other. *Does not license:* averaging. It also has the highest chance of genuine divergence, which is a feature.
- **Embedded** (one stream inside the other, subordinate by design: open ends inside a survey, a survey inside a qualitative programme). *Licenses:* enrichment of the dominant stream. *Does not license:* presenting the embedded stream as independent corroboration of the dominant one, because it came from the same respondents at the same moment.

*Correct result:* the design named in the output, with its licence and its limitation stated in one line each. Where the design was accidental rather than intended, say that.

**3. Assign stream leadership per question, in writing.** For each research objective, name the lead stream, the supporting stream, and what the supporting stream is allowed to do to the lead's answer. **This is the step that prevents both signature failures.** If the qualitative stream leads on mechanism, then a survey result cannot overrule a mechanism finding, only qualify it. If the quantitative stream leads on incidence, then no number from the interviews enters the report. *Correct result:* a leadership table covering every objective, with no objective marked "both equally", which is not a decision.

**4. Align the constructs and the units before comparing.** For each pairing you intend to make, check four things: are the two measuring the same construct, or adjacent ones; do the populations overlap; do the periods overlap; and are the units commensurable (a theme prevalence in participants against a percentage of respondents is not a like-for-like comparison and must never be presented as one). Record every approximate pairing as approximate. *Correct result:* a pairing list in which each entry is marked exact, approximate or not comparable. **Roughly half of the divergence in typical mixed-method work is resolved at this step**, before any investigation is needed, because the two streams were not measuring the same thing.

**5. Build the joint display.** The joint display is the working artefact of integration and the thing most often skipped. It is a matrix with one row per research question or finding, and columns for: the quantitative result with its base and test status; the qualitative finding with its prevalence and denominator; the relationship between them; and the integrated statement. Build it before writing anything, and build it for every finding, not only the ones that agree. *Correct result:* a complete display in which each row's relationship cell is populated from step 6's four categories. A display with empty rows is telling you where one stream is silent, which is itself a finding.

**6. Classify every pairing into one of four relationships.** Not two. The two-category habit (agree or disagree) hides the two most useful cases.

- **Convergence.** Both streams point the same way about the same thing.
- **Complementarity.** They address different facets and neither confirms nor contradicts the other. The survey gives the size, the interviews give the mechanism. **This is the most common relationship in good mixed-method work and it is routinely mislabelled as convergence**, which inflates confidence.
- **Divergence.** They point different ways about the same thing.
- **Silence.** One stream has nothing to say where the other does. Record it: a survey finding no qualitative participant ever raised, or a theme the survey never asked about, is a gap in the study and belongs in the output.

*Correct result:* every row classified, with complementarity distinguished from convergence.

**7. Handle convergence without over-claiming it.** Genuine convergence from independent streams raises confidence, per K3 §3.1 and §3.3, and it is one of the few legitimate ways to reach high confidence in applied research. Two tests before claiming it. First, **were the streams actually independent?** In sequential designs they are not: a survey built from qualitative findings agreeing with those findings is not corroboration, it is the design working. In embedded designs they are not, because the same respondents answered both at the same sitting. Second, **is this convergence or is it the same source twice?** Two measures of stated attitude from the same people are one piece of evidence, not two. *Correct result:* convergence claims that name the independence of the streams, and a confidence uplift only where independence holds.

**8. Investigate divergence with the standard explanations, one by one.** This is the analytical core of the skill. **Divergence is investigated, never averaged, never split, and never resolved by preferring the larger base.** Work through the candidate explanations in order, and record the assessment for each rather than stopping at the first plausible one.

- **Different populations.** Do the two samples actually contain the same kinds of people? Check recruitment criteria, not sample descriptions.
- **Different periods.** Did something happen between them? A three-month gap spans events.
- **Different constructs.** The most frequent cause. The survey asked about the brand and the interviews asked about the last interaction.
- **Stated versus observed.** One stream measures what people say, the other what they do or recall doing. Divergence here is expected and is usually the finding, not the problem.
- **Prompted versus volunteered.** A survey prompts every item; interviews largely do not. A topic ranked third when prompted and never raised unprompted is not a contradiction, it is a salience finding, per **07.02**.
- **Social desirability differing by mode.** Self-completion, telephone and face-to-face produce systematically different answers on sensitive topics, and an interview with a person in the room is the most susceptible of the three. The direction is predictable and should be stated.
- **Sample composition and weighting.** A weighted survey describes a population; a purposive qualitative sample describes a designed range. They can disagree simply because the qualitative sample deliberately over-represented an edge.
- **Analytical artefact.** Before concluding anything about the world, check the coding, the base, the question routing and the test. Some divergence is a mistake in one stream.

*Correct result:* a divergence record per case, with each explanation assessed and evidenced, a conclusion about which explanation the evidence supports, and a confidence level. Where no explanation is supported, the divergence is reported as unexplained, which is an honest and frequently important output. **The one thing never permitted is a split-the-difference statement**, which reports a number neither stream produced.

**9. Decide what the integrated position is, and keep both streams visible in it.** For each row, write a statement that names both streams and their relationship, rather than one that dissolves them. "Two in five report the problem, and the interviews show it is experienced as a failure of the organisation rather than of the product" keeps both. "Customers are frustrated by the product" has deleted the size and the mechanism at once. Where the streams diverge and the divergence is unexplained, the integrated statement says so and gives both figures with their sources. *Correct result:* integrated statements in which a reader can see which stream supports which part of the claim, per K2 §4.4.

**10. Set confidence per finding from the integration, not from the larger stream.** Convergence from independent streams supports high confidence. Complementarity supports the confidence each stream independently earned for its own part, and no more. Unexplained divergence caps confidence at moderate at best and frequently at low, per K3 §3.5, and the contradiction is reported either way. *Correct result:* a confidence column in the joint display, with the reason expressed in integration terms.

**11. Write the integrated narrative.** Structure the report by research question, not by method, because a chapter of survey results followed by a chapter of interview findings is two studies bound together. Under each question, state the answer, then the evidence from both streams with their sources, then the relationship. Where the streams diverge, the divergence gets its own passage rather than a footnote: it is usually the most valuable material in the study and the part a stakeholder most needs to understand. *Correct result:* a narrative in which no section belongs to one method, and in which the qualitative material appears as evidence rather than as illustration.

**12. State what could not be integrated.** Rows where one stream was silent; pairings marked not comparable; divergences left unexplained; and questions neither stream answered. *Correct result:* an explicit section, which is where the next study's brief comes from, per K3 §5.2.

## 8. Analytical framework

The chain the output is built on:

    Capability inventory → Design licence → Stream leadership → Construct alignment
        → Joint display → Relationship classification → Divergence investigation
        → Integrated statement (both streams visible) → Confidence

**Applying it.** The first four steps are all preparation, and they are where integration is won or lost. Every failure mode in Section 11 is a case of starting at the joint display without having done them.

**The four relationships, and what each does to confidence.**

| Relationship | Definition | Effect on confidence | The error to avoid |
|---|---|---|---|
| **Convergence** | Both streams, same thing, same direction | Raises it, if the streams were independent | Claiming it in a sequential design, where agreement was built in |
| **Complementarity** | Different facets, neither confirms nor contradicts | No uplift. Each part keeps its own confidence | Reporting it as convergence, which inflates confidence for free |
| **Divergence** | Same thing, different directions | Caps at moderate until explained | Averaging, splitting, or preferring the larger base |
| **Silence** | One stream has nothing to say | Neither raises nor lowers. Reported as a gap | Treating absence in one stream as support for the other |

**The divergence decision rule.** When the streams disagree, there are only four defensible outputs, and none of them is a compromise figure:

1. The divergence is explained, and the explanation is the finding.
2. One stream is shown to be measuring something the claim does not concern, and the other stands, with the reason stated.
3. One stream is shown to contain an analytical error, which is corrected and logged.
4. The divergence is unexplained and is reported as unexplained, with both figures, both sources, and reduced confidence.

**Assigning stream leadership.** The question type determines the authority, not the base size. Use this as the starting allocation and vary it only with a stated reason.

| Question type | Lead stream | What the supporting stream may do |
|---|---|---|
| How many, how often, how large, how it varies by segment | Quantitative | Qualitative may show the finding is a construct artefact, or that the measured thing is not what respondents understood |
| Why, how, in what sequence, under what conditions | Qualitative | Quantitative may show the mechanism is rare, or concentrated in a segment |
| What language people use, how they categorise the domain | Qualitative | Quantitative may test whether the vocabulary generalises |
| Which of several things matters most | Both, and they will disagree | Prompted ranking and volunteered salience are different measures. Report both, never merge |
| What changed since last time | Quantitative | Qualitative may explain the change or show the measure moved for a reason unrelated to the construct |
| What would happen if | Neither, on its own | Both are weak on the counterfactual. Say so rather than choosing the more confident stream |

**The joint display, in its minimum form.**

| Research question | Quantitative finding [source, base, test] | Qualitative finding [theme, n of N, denominator] | Relationship | Integrated statement | Confidence |
|---|---|---|---|---|---|

## 9. Output format

**A. Integration note (front matter)**
Design identified and what it licenses; capability inventory for both streams; stream leadership per objective; population, period and mode for each; construct alignment list with exact, approximate and not-comparable pairings; statement of AI involvement per K4 §7; consolidated review points per K5 §3.3.

**B. Joint display**
The full matrix, every research question, including rows where one stream is silent.

**C. Convergence register**
Each convergent finding, the independence assessment, and the resulting confidence uplift with its reasoning.

**D. Divergence register**
One entry per divergence: what each stream shows with its source; each of the eight standard explanations assessed with the evidence; the conclusion; the confidence; and, where unexplained, what would resolve it.

**E. Integrated findings by research question**
Each finding written with both streams visible, sourced per K2 §4.4 using the multi-source convention, and marked converging or diverging.

**F. Silence and gap register**
Findings in one stream that the other never addressed, and what that means for how the finding should be read.

**G. What could not be integrated**
Not-comparable pairings, unexplained divergences, and questions neither stream answered.

**When the evidence is thin.** No row is completed for symmetry. Where one stream is silent, the cell says silent, not "consistent". Where a pairing is approximate, the integrated statement carries that word. Where divergence could not be explained, the output says so rather than choosing the more convenient stream, and the confidence drops accordingly. Where the qualitative stream has nothing to add to a quantitative finding, the finding is reported as a quantitative finding without a decorative quote attached to it.

## 10. Quality checks

Run before anything is presented. Sits on top of K4 §8.

1. Is the integration design named, with what it licenses and what it does not?
2. Is stream leadership assigned per objective, with no objective marked as equally led?
3. Has construct alignment been checked, and is every approximate pairing marked approximate?
4. Do the populations and periods of the two streams overlap, and is any non-overlap stated?
5. Is every pairing classified into one of the four relationships, with complementarity distinguished from convergence?
6. Is any convergence claim made across streams that were not independent?
7. Has every divergence been run against all eight standard explanations, with each assessed rather than the first plausible one accepted?
8. Is there any averaged, split or compromise figure anywhere in the output?
9. Does any percentage appear that was derived from the qualitative stream?
10. Does every integrated statement name both streams, or state that one is silent?
11. Is the report structured by research question rather than by method?
12. Does any qualitative material appear purely as illustration under a quantitative finding, with no independent finding of its own?
13. Does confidence follow the integration relationship rather than the larger base?
14. Are unexplained divergences reported, with what would resolve them?
15. Is any causal claim present that neither stream's design licenses?

## 11. Common failure modes

| Failure | How to recognise it | How to prevent it |
|---|---|---|
| **Qualitative as decoration** (the commonest failure) | Every quote sits under a survey finding; no qualitative finding stands alone or contradicts anything | Assign stream leadership at step 3. If the qual cannot change a conclusion, it was not integrated |
| **Quantified qualitative** (the opposite failure) | "60 percent of interviewees said", on a base of 20 | No percentages from qualitative bases, per K4 §7. Report counts and denominators |
| **Averaging divergence** | A statement lands between two figures neither stream produced | The four defensible outputs in Section 8. Never a compromise number |
| **Complementarity claimed as convergence** | Confidence is high because "both streams support it", when they addressed different facets | Step 6's four categories. Complementarity gives no uplift |
| **Manufactured convergence in a sequential design** | The survey confirms the themes it was built from, reported as triangulation | Independence test at step 7. Agreement built in is not agreement found |
| **Construct mismatch read as a finding** | The streams "disagree" about satisfaction, and each measured a different satisfaction | Step 4, before any comparison |
| **The larger base wins** | The survey overrules the interviews on a mechanism question | Leadership per question. Size does not confer authority on questions the method cannot answer |
| **Two studies stapled together** | The report has a quant chapter and a qual chapter and a paragraph at the join | Structure by research question. No section belongs to one method |
| **Stated versus observed treated as error** | Divergence between what people say and what they do is reconciled away | This divergence is usually the finding. Report it as one |
| **Silence read as agreement** (the signature AI failure) | A survey finding no participant ever raised is described as supported by the qual | Classify silence explicitly. Absence in one stream supports nothing |
| **Mode effects ignored** | Sensitive topics diverge sharply and nobody mentions the interviewer in the room | Explanation six at step 8, assessed and stated, with the direction named |
| **Divergence buried** | The disagreement appears in an appendix, or as "some nuance" | Divergence gets its own passage, per K4 §4.1. It is usually the most valuable material |
| **Causal upgrade at the join** | A mechanism from interviews plus a correlation from the survey becomes "X drives Y" | Neither design licenses causality. Integration raises confidence in an association only |

## 12. AI guardrails

Skill-specific only. Universal prohibitions are inherited from K4 and are not repeated here.

1. **Never average, split or otherwise reconcile two diverging figures into a third.** The four defensible outputs are in Section 8, and none of them is a compromise number.
2. **Never report a percentage derived from a qualitative sample**, or present a theme prevalence and a survey percentage as like-for-like.
3. **Never claim convergence between streams that were not independent.** In sequential and embedded designs, agreement is partly built in and must be described as such.
4. **Never label complementarity as convergence**, and never take a confidence uplift from it.
5. **Never treat silence in one stream as support from it.** A survey finding no participant raised is a gap, not a corroboration.
6. **Never resolve a divergence by preferring the stream with the larger base.** Authority follows the question, not the sample size.
7. **Never stop the divergence investigation at the first plausible explanation.** All eight are assessed and the assessment is recorded.
8. **Never subordinate the qualitative stream to illustration** where the design gave it a question to lead on.
9. **Never upgrade to causal language at the integration step.** Two non-causal designs do not compose into a causal one, per K4 §3.2.
10. **Never present an integrated statement that does not name which stream supports which part of it**, per K2 §4.4.

## 13. Best-practice principles

1. **Disagreement is the return on running two methods.** If the streams always agreed, one of them was redundant. The divergence is what neither could have shown alone.
2. **Decide leadership before you look.** Assigning which stream answers which question after seeing the results is choosing the answer you prefer and calling it method.
3. **Most divergence is construct mismatch.** Check what each stream actually measured before theorising about the world.
4. **Stated and observed are supposed to differ.** Treating the gap as an error to be reconciled deletes one of the most reliable findings in applied research.
5. **Prompted and volunteered are different data.** A survey prompts everything; qualitative work mostly does not. Rank order from one and salience from the other are not comparable.
6. **A number cannot overrule a mechanism, and a mechanism cannot establish a size.** Each stream is authoritative in its own territory and a guest in the other's.
7. **Independence is what makes triangulation worth anything.** Where one phase built the other, agreement is architecture, not evidence.
8. **The joint display is the work.** If it exists only in the narrative, the integration has not been done, it has been asserted.
9. **Keep both streams in the sentence.** The moment one dissolves into the other, the reader loses the ability to see what supports what.
10. **Unexplained divergence is a legitimate output.** Reporting it costs credibility with nobody who matters and protects the client from acting on half the evidence.
11. **Structure by question, never by method.** Method-shaped reports guarantee the reader integrates the study themselves, badly.
12. **Write down what the client expects before you start.** Agreement with a pre-existing expectation is either reassuring or a warning, and you cannot tell which afterwards.

## 14. Worked example

*Fictional scenario, used for illustration only. All figures, participants and findings below are invented for the purpose of demonstrating method.*

**INPUT.** A university is investigating why first-year students withdraw. Two streams exist. A survey of 1,180 continuing and withdrawn students, fielded in June, weighted to the first-year population, including a battery of eleven possible reasons for considering withdrawal. And 24 depth interviews with students who withdrew, conducted in September and October, analysed per **07.01**. The two were commissioned separately and are now to be reported as one study.

**PROCESS.**

*Steps 1 to 2.* Capability inventory shows the survey can speak to incidence across the whole cohort and the interviews can speak to mechanism among leavers only. Design is identified as accidental concurrent, not designed triangulation, with the interviews running after the survey but not built from it. Licence stated: independent streams, so genuine convergence would carry weight; but different populations, so most pairings will need care.

*Steps 3 to 4.* Leadership assigned: the survey leads on how widespread each reason is; the interviews lead on how the decision was reached and what would have changed it. Construct alignment catches the central problem immediately. The survey asked all students to rate reasons for **considering** withdrawal. The interviews asked leavers about **actually withdrawing**. These are different constructs on different populations, and 6 of 11 planned pairings are marked approximate before anything is compared.

*Steps 5 to 6.* Joint display built across seven research questions. Two rows converge, three are complementary, one diverges sharply, and one is silence: the survey shows a substantial finding about timetable clashes that no interview participant raised at all.

*Step 8, and the judgement call.* The divergence. The survey ranks financial pressure first among reasons for considering withdrawal (58 percent rating it important or very important, n=1,180, weighted). In the interviews, finance is mentioned by 14 of 24 leavers but never as the reason they left; every one of the 14 describes it as background pressure that became decisive only after something else happened. The available shortcuts are both wrong: reporting "finance is the main driver" would follow the bigger base and delete the mechanism, and reporting "finance is not really the issue" would follow the qual and delete a real cohort-wide pressure. The eight explanations are worked through and recorded. Different populations: yes, considerers versus leavers, material. Different constructs: yes, considering versus doing, material. Different periods: June versus October, minor. Stated versus observed: not applicable, both are stated. Prompted versus volunteered: **material**, since the survey prompted finance from a list and the interviews did not. Social desirability by mode: assessed as plausible in the opposite direction, since finance is a more acceptable reason to give a stranger than academic struggle, which would inflate the survey figure. Sample composition: the qualitative sample deliberately over-sampled students from two faculties. Analytical artefact: coding and base checked, none found.

Conclusion reported: **the two streams are not in contradiction, they are answering different questions, and the integrated finding is that financial pressure is widespread as a background condition and rarely the proximate trigger.** Confidence moderate, because the construct mismatch means neither stream directly tests the integrated claim. Marked for researcher decision per K5 §3.1, because the university's funding case rests on which version is reported, and that is a judgement about consequence, not about data.

*Steps 9 to 12.* Integrated statements keep both streams. The silence row is reported: a cohort-wide timetable issue appearing in the survey and in no interview, most plausibly because the qualitative sample came from two faculties where it does not arise, which is stated as a limitation of the qualitative sample rather than as a doubt about the survey. Narrative structured by research question. "What could not be integrated" lists the three approximate pairings and one unexplained divergence about accommodation.

**OUTPUT.** A joint display across seven questions; a divergence register in which the headline disagreement is resolved as construct and population mismatch rather than as a contradiction; an integrated finding distinguishing background pressure from proximate trigger, at moderate confidence with the caveat attached; a silence row flagging a limitation in the qualitative sample; a decision point flagged for the research director because the reporting choice has funding consequences; and no percentage anywhere derived from the 24 interviews.

## 15. Advanced usage

**Individual-level integration.** Where the same people appear in both streams, integrate at the person level rather than the group level: match each participant's survey answers to their interview account. The most informative group is always the one whose two accounts disagree, and they are invisible in any group-level comparison. This is the strongest form of integration available in applied research and it requires planning at the design stage.

**Adding a third, non-self-reported stream.** Behavioural, transactional or administrative data is frequently the arbiter when a survey and interviews disagree, because it is not subject to recall or desirability. Treat it as a third stream with its own capability inventory: it is authoritative on what happened and silent on why, and it carries its own coverage problems.

**Integration in evaluation.** Where outcome measures and participant accounts diverge, resist the instinct to treat the measures as truth and the accounts as perception. A programme that improves a measure while participants describe it as unhelpful has produced a genuine finding about what the measure captures. Report both, and name the construct each one holds.

**Meta-inference across a programme.** Where several studies over time each integrated two streams, the integration itself becomes evidence: divergences that recur across studies are structural, and divergences that appear once are usually local. Hand to **14.04 Meta-Analysis Across Studies**.

**When the client will only accept one answer.** Where a decision genuinely requires a single position and the streams diverge, produce the position, name the stream it rests on, state what the other stream shows, and state what would resolve it. Per K4 §9, that is the strongest honest output, and it is more useful than a confident average because it tells the client what they are betting on.

**When the standard approach does not fit.** Where more than two heterogeneous inputs must be reconciled, move to **10.03 Multi-Source Research Synthesis**. Where the qualitative stream is itself two incompatible datasets, resolve that within **07.01** before integrating. Where the streams are separated by years, treat the comparison as a trend question first, per **05.05**, before treating it as integration.

## 16. Skill chain

**Recommended previous skills:**
- **07.01 Thematic Analysis.** Hands over themes with prevalence, denominators, counter-evidence and preserved tension, which is what makes the qualitative stream capable of leading rather than illustrating.
- **07.03 Interview and Transcript Analysis**, where the qualitative stream is a small set and case-level mechanism is what will be integrated.
- **05.01 Descriptive Analysis** and **05.02 Statistical Testing**, which hand over quantitative results with bases, wording and test status attached.
- **05.03 Cross-Tabulation**, where subgroup patterns are the thing the qualitative stream will explain.

**Recommended next skills:**
- **08.01 Finding to Insight Development.** Takes the integrated position, with divergence intact, and works from what the evidence shows to why it is happening.
- **12.04 Research Evidence Integration** and **12.03 Research Report Compilation**, which carry the joint display's source references into the deliverable per K2 §6.
- **11.01 Research Narrative Development**, where the integrated account has to be structured for an audience without collapsing back into two studies.

**Runs well alongside:**
- **13.03 AI Output Verification**, run against the joint display and every integrated statement.
- **13.04 Bias Detection**, particularly where a stakeholder expectation was recorded before integration.
- **K5**, at the review points this skill mandates: any divergence whose resolution changes what is reported, and any integrated finding with material consequence.

---
A Yazi Supplied Skill and resource.