Ezekiel Yeo

Writing

Guerrilla research in community care: building evidence from everyday work

How Community teams can turn fast, local inquiry into findings taken seriously.

Guerrilla research in community care: building evidence from everyday work

Some of the most useful questions in community care never become research. A nurse notices that night shifts keep going sideways for the same handful of patients, and wonders why. A medical social worker runs family meetings every week and suspects that a small change to how they open might lower the temperature in the room. A GP hears steady grumbling about a new discharge workflow and would like to know whether it actually reduces unplanned readmissions. These questions matter to the people asking them, often urgently, but they rarely attract grants, and the people asking them rarely have protected time to pursue them.

I have spent nearly nine years at a home hospice, working in the space between frontline care and research, and I have come to believe these questions can be answered well. Not perfectly, and not always to the standard of a funded trial, but well enough to change practice and well enough for others to learn from. What it takes is less a budget than a set of habits, and a willingness to be honest about what our evidence can and cannot tell us.

I have been calling this guerrilla research. In this piece I want to set out what I mean by it, describe the habits that I think make it trustworthy, and walk through two projects from HCA Hospice where I saw it work: a call centre that grew out of nurses' ringing phones, and a protocol for bringing dying patients home from hospital.

What I mean by guerrilla research

I started using the term before I knew it already existed. Only later, looking it up out of curiosity, did I find that UX designers use it for "user research in the wild": short, informal conversations with real people in the places they already are, rather than in a lab or through a long survey (Interaction Design Foundation). That description fits community care rather well. Our services happen in living rooms, clinic corridors and phone calls, and much of what we need to know is already there, in records we keep and conversations we have anyway.

Health care has older relatives of the idea, and I owe a lot to them. Quality improvement teams test small changes in quick Plan-Do-Study-Act cycles, guided by the Model for Improvement (IHI). Participatory action research brings the people affected into the inquiry as partners rather than subjects. What I am describing sits somewhere between the two. It borrows the pace of quality improvement and the respect for local knowledge that participatory work insists on, and then asks one further question: what would it take for someone outside our team to trust this finding?

It is worth being clear about what it is not. It is not a way around ethics review, and a quick count or a handful of conversations does not prove anything. What they offer is direction, and whether direction is enough depends on what you plan to do with it. Rewording a brochure needs very little evidence; changing who answers the phone when a dying patient's family calls deserves a great deal more.

So the principle I keep returning to is proportionate rigour: let the care you put into the evidence match the weight of the decision it will support, and be plain about where your evidence stands.

A few habits that make quick work trustworthy

None of these habits need money. Most of them need a little discipline at the start of a project, which is usually the moment when it feels least necessary.

The first is to spend more time on the question than feels comfortable. I find it helps to write three sentences before anything else: who is affected, what we hope will change, and how we would know. Then choose one outcome to treat as the main one. It sounds small, but it keeps a project from drifting, and it makes every later decision easier to explain.

The second is to write the plan down before collecting any data. A registration on the Open Science Framework is ideal, but a dated document on a shared drive does most of the work. A paragraph on who you will sample, what you will ask and how you will analyse it is enough. The point is not to lock yourself in. It is to leave a record of what you intended, so that neither you nor your readers have to wonder whether the analysis was shaped by the results.

The third is to measure things the same way each time. Ask the same two or three questions, and keep a short codebook so that five colleagues do not invent five names for the same idea. At MediHELP, our call centre, every incoming call is tagged with who called and one of about ten reasons, such as a request for patient information or a question about symptom control. That small and slightly tedious habit is what later let us see patterns across thousands of calls.

The fourth is to keep a trail. Raw data, cleaned data and results live in separate places, and the cleaning steps are written down. Most readers will never check, but the ones who do care far less about which software you used than about whether a number can be traced back to where it came from.

The fifth is to look at the same question from more than one side. Interviews tend to tell you why something happens, counts tell you how much, and documents tell you what actually happened, as opposed to what people remember. When two of these agree, I start to trust a finding. When they disagree, that is usually where the interesting work begins.

When it comes to writing up, I try to report how large a change was and how uncertain we are about it, rather than leaning on a p-value. I also try to name the one limitation that I think could genuinely change the conclusion.

Thinking in rungs

It helps me to think of evidence as a ladder. Each rung answers a slightly different question, and perhaps the most important habit of all is to claim only what the rung you are standing on can support. The table below is how I map it, with an example from our own work where I have one.

Rung Design What you can claim What you cannot claim Example from my work
1. Count and describe Coded routine records, descriptive trends What happens, how often, when and to whom That anything caused it MediHELP call coding: 5,640 calls, Jan–Sep 2025
2. Ask the people involved Exit questions, short surveys, field notes, interviews How the service is experienced, and where it hurts That outcomes improved MediHELP peer review: five satellite teams and the MediHELP team rated each other
3. Before and after Pre-post comparison, run chart, interrupted time series That a change coincided with your intervention That nothing else caused it, such as season, staffing or policy MediHELP call diversion tracked month by month
4. Compare with a group that did not change Stepped rollout, matching, difference-in-differences A credible estimate of the effect That unmeasured differences between groups do not matter Not yet done; the 71 discharges outside the compassionate discharge protocol are a ready comparison group
5. Explain how and for whom Realist evaluation, mixed methods Which mechanisms work, in which contexts, and what to keep when transferring How big the effect is Compassionate discharge realist evaluation: 51 caregivers, 114 clinicians
6. Transfer and accumulate Same outcomes across sites, shared packages, adoption elsewhere That it holds beyond your own team That it suits every setting Compassionate discharge materials in the Singapore Hospice Council e-library

Rungs four and five are not really above and below each other; they answer different questions. One asks how much something changed, and the other asks why, and for whom. Our compassionate discharge study, described below, reached rung five without passing through rung four, so it can tell a careful story about mechanisms but cannot say how large the effect was. The ladder lines up quite closely with Nesta's Standards of Evidence, which move from showing that change happened (level 2), to showing causality against a comparison group (level 3), to explaining why and how (level 4), and finally to replication and scale (level 5) (Nesta).

One caution about rung three, because it is where it is easiest to overreach. A before-and-after comparison with only a few time points on either side can look convincing on a chart. With so few points, I would treat a jump as something to keep watching rather than a result, partly because neighbouring months are rarely independent of one another.

MediHELP

The problem was obvious to anyone who had spent a day shadowing one of our home hospice nurses. Their phones rang constantly. Our 30 frontline nurses could receive up to 135 calls a day between them, often in the middle of a nursing procedure, or while sitting with a family at a difficult moment. Each call was reasonable on its own, but together they were pulling attention away from the patient in the room.

MediHELP was our response. During office hours, from 08:30 to 17:30, calls to home care nurses are diverted to a small central team, which tries to answer within three rings. That team can see the electronic medical records, arrange a quick video consult or an urgent home visit if needed, and then hand each case back to the patient's own nurse.

Getting nurses to actually divert their calls turned out to be a small research problem of its own. A quality improvement project mapped the reasons on a fishbone diagram and tracked the diversion rate using phone-call notes that nurses were already writing in the records. [Add the baseline and final diversion rates from the paper.] Between April and October 2024, 58% to 69% of calls each month went to MediHELP, even as total calls grew from 931 to 1,190 a month. Calls from colleagues trying to track down a patient's primary nurse fell to somewhere between none and three a day.

Coding every call by who called and why is what made the service legible to us. Of the 5,640 daytime calls we coded between January and September 2025, 40.8% were requests for patient information, which are best answered by someone with the medical records open on a screen rather than a nurse in a patient's living room. Caregivers made 54% of the calls and hospital staff 33%, and 70% lasted under five minutes. The calls clustered between 9 and 11 in the morning and again around 4 in the afternoon. Those turned out to be exactly the times of the nurses' morning roll call and their end-of-day documentation, the moments the phones had always seemed to interrupt most.

We also simply asked people. In October and November 2024, the five satellite teams (one doctor and 31 nurses) and the MediHELP team (one doctor and five nurses) rated one another on three five-point questions about communication, problem-solving and how easy it was to work together. MediHELP scored between 4.1 and 4.3. The written comments were more useful than the scores. Nurses said it was a relief not to answer the phone during visits, but some caregivers still wanted to hear from their own nurse, and patients with complex psychosocial needs were better kept with the primary team. Those comments shaped a twelve-point guideline that tried to draw clearer lines between MediHELP and the primary teams, which was the main gap our reviewers had noticed.

Because we kept the coding scheme mostly stable, a slower trend became visible too. Hospital staff made 26% of calls between August and November 2024, and 33% between January and September 2025, mostly about patient information and medication. That suggests the next useful change is not more call-centre staff but a better way for hospitals and the hospice to share information.

Looking back, there are things I would do differently. We never counted the internal calls before MediHELP started, so "far fewer than before" is a claim nobody can check. We changed the coding between years: the 2024 scheme had a code for returned missed calls, which made up 16.8% of calls, and the 2025 scheme dropped it, so the two years are not quite comparable. The 2025 poster describes a one-year period when the data cover nine months. And we measured the flow of calls more carefully than anything families would actually feel, such as how long they waited for a callback, or whether they ended up in an emergency department.

Compassionate discharge: going home to die

The second project climbed further up the ladder than most of our work, and it was for families at one of the most fragile moments they will ever go through.

Some patients, as they near the end, ask to go home. When they are already in hospital and death is likely within days, that means a terminal or compassionate discharge, often arranged within hours. In 2020 our hospice received 259 such referrals, roughly one every working day and almost a tenth of all our referrals. Each one involved a hospital team, our team and a family, all trying to coordinate something urgent and emotionally enormous. There were no shared criteria, handovers varied from one hospital to the next, and families often arrived home not quite ready for what was coming.

We began, as guerrilla research usually does, by counting. A retrospective audit of our own terminal discharges set out what was going wrong in practice (Chong, Hii & Yeo, 2023). From there, a small core team, which I was part of, spent six months in stakeholder meetings, consensus discussions and prototyping workshops with hospital and hospice clinicians. What came out of that was the Compassionate Discharge (ComD) protocol: shared eligibility criteria, a standard handover checklist, orientation videos and an information pack for clinicians, and a factsheet and guide for caregivers.

To evaluate it, we chose a realist evaluation. Rather than asking simply whether an intervention works, a realist evaluation asks how it works, for whom, and in what circumstances, which is exactly what another hospital would want to know before adopting it. Of the 162 terminal discharges that came to us in 2023, 65 followed the new protocol. We heard from 51 of those families, from 66 hospital staff who referred them and from 48 of our own clinicians (Chong et al., 2026).

We tried to keep the burden on families light. After getting home, caregivers could send short reflections to our research assistant by WhatsApp, and 46 of them did. At least three months after the death, 45 completed a survey. Clinicians answered a short electronic questionnaire within two weeks of each discharge. The analysis was done in Excel and SPSS, and the study had no external funding.

Some of what we found was reassuring. Only 2 of the 51 patients (3.9%) went back into hospital. Among clinicians, 83.1% of hospital staff and 66.0% of hospice staff felt that discharges under the new pathway went better than before. Caregivers' preparedness averaged 2.44 on a scale from 0 to 4, a moderate level, with a lot of variation between families.

The most interesting finding, though, overturned our own starting assumption. We had built the protocol on the idea that better information exchange would leave families better prepared. On five-point scales, caregivers rated the support they received from the hospice (3.20) above the information (2.78) and training (2.69) they received from the hospital. Read alongside the interviews, the picture was that success depended less on passing information along than on helping everyone live with uncertainty, and that a reliable safety net mattered more to families than training did. Hospital palliative teams, involved in 46 of the 51 discharges, also mattered a great deal, often acting as the bridge between institutions. For me, this is the real value of a realist evaluation: it can tell you that your theory was wrong in a way you can use.

The materials are now hosted in the Singapore Hospice Council's e-library, including the handover checklist, the clinicians' information pack, and caregiver guides in English, Chinese, Malay and Tamil (Singapore Hospice Council). In April 2026 the Health Minister noted that a compassionate discharge procedure is now in place in all public hospitals (Ministry of Health). I would not claim that our study caused that, but it is the wider system our work now sits within.

There is a rung we did not reach. Without a matched comparison group, we can describe mechanisms carefully but cannot say how large the effect was. The 71 discharges that reached our hospice without using the protocol are an obvious place to look for that comparison next, alongside the question of cost-effectiveness that the paper leaves open.

Calculating cost

Sooner or later, someone will ask what a change costs. I think community teams should be ready with an honest rough answer, rather than waiting for an economist to arrive. The simplest version sets the new way of working beside usual practice and asks what each costs, what each achieves, and what each extra unit of benefit costs.

\text{Cost per extra outcome} = \frac{C_{\text{new}} - C_{\text{usual}}}{E_{\text{new}} - E_{\text{usual}}}

Here C is cost and E is whatever outcome you care about, such as completed referrals or avoided readmissions. A small sheet with these figures in it is often enough to start a sensible conversation.

The arithmetic is easy; the judgement around it is where things tend to go wrong, and a few traps are worth naming. A cost avoided is not the same as cash saved. An hour of nursing time freed from the phone is not refunded; it is spent on something else, and it is worth saying what. It matters whose costs you are counting, because the organisation, the health system and the family will each see a different picture. The comparison should be against what would have happened otherwise, not against nothing. The intervention has costs of its own: MediHELP runs on a doctor and five nurses, and their time belongs in the sum. And time itself can be measured directly. A time-motion study, like the one we carried out with our home hospice teams, gives you the minutes that later turn into costs (Yeo, Chua & Chong, 2026).

There is a point where a rough sheet is no longer enough: when the figure will decide whether something gets funded, when several options are being compared, or when costs and benefits stretch over years. That is the moment to bring in a health economist, and the conversation is likely to go better if the rough sheet is already done.

Ethics, and the people behind the data

Working quickly does not mean being casual with people's lives or their data, and in hospice care that is never far from mind. The line I watch most closely is the one between service evaluation and research. Measuring your own service against its own standards is usually service evaluation or quality improvement. Collecting information beyond routine care, randomising people, or setting out to publish findings meant to apply elsewhere usually makes it research, and research needs review. When in doubt, talk to your institutional review board early. Our compassionate discharge study went through the SingHealth Centralised Institutional Review Board (reference 2021/2728).

Timing matters as much as permission. In that study, caregivers were asked for written consent at the first home visit, only after the most urgent clinical needs had been dealt with, and the main survey waited until at least three months after the death. When approaching anyone, it helps to say plainly who you are, why you are asking, and that saying no will not change anyone's care (Fox, 2019).

Small datasets deserve particular care. A hospice dataset of a few dozen patients can identify people even without names. I would share instruments, code and aggregate results openly, and keep record-level data behind an agreement; the compassionate discharge paper, for instance, offers summary and aggregate data on request. All of this sits within data protection law, which in Singapore includes the Personal Data Protection Act, as well as your own organisation's policies.

From a single project to a line of evidence

A single project, however good, tends to be forgotten once the people who did it move on. What lasts, I think, is a series of projects that measure things in compatible ways and leave behind enough for others to pick up.

Reusing the same measures is the quiet foundation of this. It was only because MediHELP kept its call codes largely stable across two years that we noticed hospital calls creeping up, and where a code changed, the comparison simply broke.

It also helps to think of every project as a small, complete bundle rather than a report: the protocol, the instruments, the codebook, the cleaning steps, the code, the figures and a one-page summary. A colleague in another organisation, or an academic who wants to extend the work, can pick up a bundle like that without a long series of explanatory calls.

I have come to think of sharing in three layers. There is the complete bundle, kept somewhere with a stable link. There is something written for practitioners, such as a poster, a practice note or a toolkit. And there is something written for researchers, such as a paper, a preprint or a short methods letter. Looking back, both of our projects followed roughly that path. MediHELP led to a quality improvement paper in Palliative Medicine Reports (2025) and a conference poster describing 5,640 calls. Compassionate discharge led to an audit letter in the Annals, Academy of Medicine, Singapore (2023), a case report (2024), the realist evaluation in Palliative Medicine (2026) and materials in a shared national e-library.

None of this needs much infrastructure. I picture a shared drive with a folder of templates, a few short guides that people actually read, and a monthly call where someone kindly says, "try this first, before you run the numbers again". One person keeps the protocol template tidy, another holds open office hours, and a third collects good examples from teams down the road rather than from a distant journal. If a sector can manage one shared drive and one recurring call, it already has the bones of a research community.

If I were helping a team start from nothing, this is what I would put in their first folder:

  • a three-sentence theory of change, with one main outcome
  • a one-page protocol template covering who you will sample, what you will ask and how you will analyse it
  • a codebook listing each code, what it means and an example
  • a consent script written in plain language
  • a folder structure separating raw data, cleaned data and results, with a log of the cleaning steps
  • a simple two-row cost sheet
  • a one-page summary template: the question, the setting, the design, the main result and one honest limitation

Closing thoughts

Neither of these projects began as a research programme. One began with phones that would not stop ringing, and the other with families arriving home not quite ready for a death they had chosen. In both cases the questions were already there in the everyday work, waiting for someone to take them seriously.

That, in the end, is what I mean by guerrilla research: keep the closeness and the pace of everyday inquiry, and add just enough care and structure that what you learn locally can be trusted, and used, by others. If more of us share the same basic tools, use the same tags and leave behind complete little bundles, community care can build a body of evidence of its own, one that funders and academics can hear clearly. Perhaps more importantly, the next person with a good question will not have to start alone.

References

My own work

Other sources