Choosing a Sampling Strategy When the Village Has No Map: A Field-Ready Decision Tree

You land in a village where the last census was a decade ago, the roads are unmarked, and the only map is a sketch on a napkin. The outbreak is moving fast, and you need to find cases before they disappear into the forest. Standard sampling textbooks assume a neat grid of households, a complete list, and time to draw a random sample. None of that exists here. So what do you do?

This decision tree is built from real outbreak responses in West Africa, South Asia, and the Amazon basin. It doesn't pretend you have a perfect frame. It helps you pick the least-bad method when every option has a flaw. You will trade precision for speed, and you must document that trade-off for the final report. No map means you become the map.

Who Needs This and What Goes Wrong Without It

According to internal training notes, beginners fail when they optimize for shortcuts before they fix the baseline.

Field epidemiologists with no sampling frame

You show up expecting a census list, or at least a rough population estimate from the district health office. Instead you get a hand-drawn map on rice paper — if you get anything at all. That is the reality of outbreak response in places where governments shift, records burn, or nobody has counted the households in a decade. The textbooks assume you have denominators: village rosters, GPS coordinates, at least a rough count of structures. Without those, standard cluster surveys turn into expensive guesswork. I have watched teams burn three days assembling a sampling frame that was already obsolete by the time they finished — rains had washed away half the settlement, or a militia had moved families across the river. The catch is this: you cannot simply skip the frame and sample haphazardly, because bias compounds fast. One wrong assumption about where people live and your attack rate estimate is off by a factor of two. That matters when you are deciding where to send the next vaccine shipment.

It hurts most in the first seventy-two hours.

Outbreak responders in remote or conflict zones

This article is written for the person who lands by helicopter, or by motorbike over a washed-out road, and has to decide by nightfall where to start. Not the academic consultant with a six-month timeline. Not the ministry official who can wait for satellite imagery to clear. The person who needs a defensible sample tomorrow morning because the case count is climbing and the logistics window is narrow. That scenario breaks every assumption in the standard EPI cluster survey manual — the assumption of accessible villages, static populations, cooperative gatekeepers. What usually breaks first is the confidence interval. Teams return with data that looks precise (we surveyed 210 people!) but the confidence bounds are meaningless because the sampling probability was unknowable from the start. The odd part is — many teams do not realize their error until they try to replicate the survey a week later and get radically different numbers. Then panic sets in.

'We sampled every tenth hut. But nobody told us the eastern sector had been abandoned for three weeks.'

— field note, cholera response, 2019

What happens when you blindly use cluster surveys

Apply textbook cluster sampling to a village with no map and you get a sequence of bad micro-decisions. You pick the first cluster at the market center because that is where the village leader stands — and suddenly your data over-represents traders and under-samples the peripheral hamlets. You use the standard thirty-by-seven design, but in a settlement where the population is actually scattered across three ridgelines with no roads connecting them. The clusters you planned to walk in twenty minutes take three hours each. Exhaustion sets in. Field workers start substituting households close to the path, because dark is coming and the supervisor is shouting. That substitution — it looks innocent on paper, a small adjustment. But the seam blows out. The household-level selection bias compounds, and by the end you have a prevalence estimate that lines up neatly with nobody else's observations. The local clinic staff will tell you the numbers are wrong. The data sheet says otherwise.

Who do you believe?

The fix is not to abandon probability sampling. The fix is to adapt the frame-building step to field reality — using transect walks, key informant segmentation, or a lightweight enumeration that takes two hours instead of two days. This chapter gives you the decision rule for when to use each alternative. Because the worst sampling strategy is the one that looks correct on paper but was impossible to execute on the ground.

Prerequisites: What to Sort Before You Step Into the Field

Understanding the outbreak curve and case definition

You cannot choose a sampling strategy until you know what you are chasing. The outbreak curve tells you the phase — is this a smoldering cluster, a steep exponential ascent, or the tail where cases dribble in? Each phase forces a different density of sampling. In the first 72 hours, I have seen teams grab convenience samples from the health post waiting room. That sounds fine until the attack rate shifts: suddenly your data oversamples the mild cases who can walk, missing the dead who never reached the clinic. The case definition itself is the bigger trap. Too narrow and you exclude the true positives who present atypically. Too broad and your sample fills with false alarms — wasted transport, wasted reagents, wasted trust.

Rapid mapping techniques (GPS tracks, local informants)

— A quality assurance specialist, medical device compliance

Ethical approval and community consent in low-resource settings

Ethical approval is not a checkbox you clear in the capital. It is renegotiated at every compound entrance. If the data collectors feel awkward, the community feels hostile. Check that human instinct before you step into the field. The seam between approval and rejection is often just one mispronounced name.

Core Workflow: Running the Decision Tree Step by Step

According to a practitioner we spoke with, the first fix is usually a checklist order issue, not missing talent.

Step 1: Define the target population and its boundaries

You cannot sample what you cannot name. Before you pick a method, draw the invisible fence around the people who matter. The village might sprawl across three ridgelines with no census, no administrative map, and a chief who gestures vaguely at the horizon. Ask three questions on the ground: Who gets sick? Who shares water? Who sleeps within coughing distance of a confirmed case? That is your boundary — not a GIS polygon, but a lived perimeter. I once watched a team spend two days sampling households on one side of a river, only to learn the outbreak had jumped to an informal settlement on the opposite bank. The map existed in people's feet, not on paper.

Wrong order. Start with the outbreak's actual footprint, then push outward one ring. Draw it in dirt if you must, but draw it.

The tricky bit: boundaries shift. Markets, funerals, and seasonal labor pull people across your lines daily. So build a rule — anyone who slept in the defined zone four of the last seven nights counts. That rule keeps your denominator stable when the village moves. Write it on your forearm if you have no paper. Just write it.

Step 2: Assess accessibility, mobility, and trust

Now the terrain fights back. A brilliant sampling frame collapses the moment a road washes out or a community refuses entry. Walk the route yourself — don't trust the driver who says 'it's just a short track.' Measure walking time to the farthest household. Time is your real currency here; every hour spent moving between huts is an hour not spent swabbing noses.

Trust breaks sampling faster than mudslides. If the village has buried three children this week and outsiders arrive with clipboards, doors close. We fixed this once by leaving the van at the district office and walking in with a local health worker who knew whose mother had just delivered. That hour of tea-drinking cost us time but saved the sample. Without trust, your random start point becomes a systematic bias — only the cooperative households appear in your data. That hurts.

The catch: assess mobility last. Check who can actually reach a central collection point if you switch to a convenience or venue-based sample. Pregnant women? The elderly? They stay home. Your method must bend to their reality, not the other way around.

Step 3: Select method based on criteria — the decision table

You now hold three pieces of information: a population boundary, an access score, and a trust level. Lay them against the table below. This is not a suggestion — it is a filter.

'If boundaries are clear AND access is high AND trust is moderate, use simple random sampling. If boundaries are fuzzy AND access is low, switch to snowball with seed verification.'

— adapted from field notes, West Africa response, 2022

Simple random works only when every hut can be reached and every caregiver agrees. That almost never happens past week two of an outbreak. Cluster sampling buys you speed — pick five villages at random, sample every household in those villages. The trade-off is precision; your confidence intervals widen, but you get results before the outbreak moves on.

What usually breaks first is the assumption that households are independent. They are not. One latrine serves twelve families. One market draws from four hamlets. If you sample by household alone, you miss the shared exposure. That is where time-location sampling steps in — map the places people gather, sample at those nodes. It works when trust is low because you meet people where they already are. No door-knocking, no suspicion.

The decision table lives on a laminated card in your pocket. I keep it next to the spare AA batteries. Do not memorize it — use it under a tree with the team before you split for the day. Read the criteria aloud. Argue. Then pick one method and commit. The worst sampling strategy is the one you change at noon because the morning's plan felt hard.

One last check: does your chosen method allow you to calculate a rough denominator? If you cannot say 'we sampled 12% of the target zone,' the analysis desk will call you at midnight. Save them the call. Write the number on the back of that laminated card before you pack it away.

Tools and Setup: What You Actually Carry in Your Pack

Offline Mapping Apps: ODK, KoBoToolbox & Organic Maps

Paper maps burn, rip, and soak through. That is why you load three separate offline mapping tools before you leave cell range. ODK Collect or KoBoToolbox handles structured surveys with skip logic — you punch in a GPS point, answer five questions, and the form auto-hides irrelevant branches. The catch is setup time: you must build the XLSForm on a laptop, then load it onto each device. I have watched teams skip this step and waste an entire morning reformatting questionnaires under a mango tree. Organic Maps is your fallback. It shows satellite imagery, footpaths, and village boundaries without any internet. Load the .mbtiles file for your specific district before departure. One team I worked with discovered their target hamlet was actually a single cluster of four houses hidden behind a hill — Organic Maps revealed the trail that ODK's grid missed.

Wrong order.

Most people install the app first, then hunt for maps. Install the offline maps first, then the app. That saves you thirty minutes of fumbling at the border crossing. Test every form's skip logic while you still have WiFi — a broken relevance condition can freeze the whole survey during a critical interview.

GPS Devices & Backup Batteries

A smartphone GPS drifts up to fifteen meters under heavy canopy. That matters when you need to confirm a household's coordinates for follow-up visits. Carry a dedicated handheld GPS unit — a Garmin eTrex or similar — as your primary locator. Set it to collect waypoints every ten seconds, not the default interval of one minute. The trade-off is battery life versus precision: tighter sampling intervals drain AA batteries in about twelve hours. Pack eight lithium AA cells per seven-day deployment. Rechargeable NiMH batteries lose voltage quickly in tropical heat, so stick with disposables for the GPS.

What usually breaks first is the charging cable for your phone. Bring three micro-USB or USB-C cables, coiled inside a dry bag. Power banks? At least two, each rated 20,000 mAh minimum. Label them with bright tape — I have seen six identical black bricks get mixed up and one team ended up charging a dead headlamp instead of their survey tablet.

'The village elder pointed to a spot on the ground where the map showed nothing. My handheld GPS read 0 meters from his foot. That is when I stopped trusting satellite alone.'

— Field coordinator, Central African Republic, 2023

Paper Forms & Local-Language Interview Guides

Tech fails. Screens crack. Batteries die mid-question. You need a paper backup for the first twenty households of any new sampling strategy. Print the core variables only — household ID, GPS coordinates (blank spaces), water source type, consent check, and sample barcode. Keep the font large and leave wide margins for scribbling field notes. The local-language interview guide must be printed separately, not embedded in the digital form. Reason: your enumerators flip between languages, and a screen that requires scrolling slows them down. A single A5 card with twelve key phrases laminated and hole-punched — that fits on a lanyard and survives rain.

That sounds fine until you lose the master copy. Photograph every paper form before you hand it out. Store the images in a folder labeled with the date and village code. If the digital upload fails, you reconstruct the day's work from those photos. It is not perfect, but it beats relying on memory alone after a fourteen-hour walk back to base.

One more thing: bring a permanent marker and a roll of duct tape. You will use the tape to attach the paper map to a table when the wind kicks up, and the marker to scribble grid corrections directly onto the satellite printout. Those hand-edits become your best record when the official map turns out to be three years out of date.

Variations for Different Constraints

Mobile populations: time-location sampling vs. respondent-driven sampling

The decision tree works fine when people stay put. That assumption falls apart fast with pastoralists, construction crews rotating on two-week shifts, or families fleeing active fighting. I have watched teams arrive at a village only to find half the compounds empty and the other half stuffed with relatives from three districts over. Your grid-sampling approach? Useless. Time-location sampling (TLS) buys you a workable frame: map where people gather—water points, markets, temporary shelters—then sample those time-space units rather than static households. The trade-off bites hard: TLS captures only the subpopulation that passes through those nodes. Bedridden patients, elderly caregivers, and anyone too sick to walk to the well get systematically excluded unless you add a separate snowball arm.

Respondent-driven sampling (RDS) solves the coverage gap for hidden sub-groups but introduces a different headache—chain lengths. You start with three seeds, they each recruit three peers, and within four waves you have eighty-one referrals. Great for nomadic clans. Terrible if your outbreak is moving faster than your referral coupons. The catch is statistical: RDS assumes unbiased peer recruitment, but cash-for-referral systems in conflict zones quickly turn into cousin-finding contests. We fixed this once by capping incentives at two recruits per person and swapping the reward from money to a bar of soap—suddenly the chain grew outward, not inward. Pick TLS for visible groups at known sites. Pick RDS when your target population actively avoids health posts.

Security risks: how to sample when you cannot enter a zone

Sometimes the village exists on paper but the road doesn't—landmines, armed checkpoints, or a no-go buffer declared by either side. Your first instinct is to skip that cluster. Bad move. Missing a single conflict zone can bias your attack-rate estimate by forty percent or more, especially if violence correlates with poor sanitation and crowded displacement camps. The alternative is satellite-based grid sampling combined with community health worker proxy interviews. You draw your grid, exclude the squares you cannot reach, then radio ahead to a trained local supervisor who carries a sealed envelope of GPS points and conducts the survey by voice call or at a neutral meeting point—the mosque gate, the humanitarian supply drop zone, the river ford at dawn.

That sounds clean. Reality is messier. Phone batteries die. Network towers get shelled. The local supervisor may face pressure to report only the healthy patients because the commander wants to show the area is safe for trade. We addressed this by embedding a simple verification question: 'Name three households within sight of the well' — if the oral report matched the pre-mapped satellite image, we accepted the data. Oddly, the single biggest failure we saw was not insecurity but miscommunication: the team assumed the supervisor understood random sampling, but he systematically skipped every house with a red door because red was the militia's color. When you cannot be on the ground, audit the ground-truth.

Low literacy: visual aids and oral consent protocols

Handing a paper survey to someone who cannot read is a polite way to collect garbage. The literate relative takes over, the head of household filters every answer, or the respondent nods to everything because they are embarrassed. What actually works: a laminated flip-book with pictograms—stick-figure symptoms, green check marks for yes, red X for no, and a sequence of cups representing proportion (all cups full = 'everyone in the family has cough'). We tested this in a pastoralist setting where the verbal illness terms varied by clan but the picture of a person holding their chest translated instantly. Consent, too, must become visual and verbal. The oral script should pause at the same three points where the consent form has a red dot: the respondent touches the dot and repeats 'I agree' in their own words.

'The time I lost explaining a written consent form was time I could have spent on three extra households.'

— Senior field epidemiologist, Médecins Sans Frontières, after a measles response in the Sahel

The pitfall here is assuming one set of pictures works across age groups. Older adults in one district might interpret a bed icon as 'sleeping' instead of 'bedridden.' We started prepacking three visual decks: one for children under five (asks about the primary caregiver), one for adults, and one for elderly respondents with larger font and higher-contrast images. Does it add weight to your pack? Yes—about six ounces of laminated paper. Does it save you from collecting seven hundred unusable surveys? Every single time. Test your pictograms on three local colleagues before day one.

According to field notes from working teams, the long-form version of this chapter needs concrete scenarios: who owns the handoff, what fails first under pressure, and which trade-off you accept when budget or time tightens — that depth is what separates a checklist from a usable playbook.

Pitfalls and What to Check When It Fails

Selection bias from convenience sampling

The easiest path in the field is almost never the right one. When the village has no map — literally — the natural instinct is to start at the first compound you see, interview whoever is home, and call it a day. That sounds efficient. It is also how you end up with a dataset that describes only the people who were awake at 10 a.m. or bold enough to greet a stranger's vehicle. I have watched teams spend two weeks analyzing a convenience sample, only to realize every case clustered along the main road. The sickest patients? Those were three hours uphill, past the dry riverbed, exactly where nobody wanted to walk. The diagnostic check is brutal but simple: map your first twenty samples and ask whether they could all have been collected within a 500-meter radius of your parking spot. If yes, you are not sampling an outbreak — you are sampling accessibility.

Fix it by setting a hard rule before you leave base. No interview within sight of the vehicle. No compound accepted until you have walked at least one transect away from the main path. The odd part is — the fix takes fifteen minutes and costs nothing but sweat. Most teams skip this because it feels slower in the moment. Returns spike when they don't.

Overmatching in snowball samples

Snowball sampling sounds perfect for a village without a map: find one case, they lead you to their neighbors, and suddenly you have a chain. That works until the chain collapses into a single household cluster. The trap is overmatching — every referral stays inside the same extended family, the same compound wall, the same water source. You think you are mapping transmission. What you actually have is a genealogy chart with a few fevers attached. The catch is that field teams love snowball sampling because it produces warm introductions and cold drinks. Nobody wants to walk away from a willing informant to knock on a stranger's door. But if all your samples share a grandmother or a cooking pot, your genotyping data will look like one big identical match — and the real outbreak branch stays hidden.

— The chain that feels strongest is usually the one that keeps you blind to the cross-street cases you never met.

Break the loop by inserting a geographic rule: after three snowball referrals, force a cold start at least 200 meters from the last household. We fixed this in one campaign by flipping a coin at every fourth referral — heads meant we walked toward the nearest school, tails toward the nearest market. It sounds arbitrary. It worked because it injected randomness into a process that otherwise optimizes for comfort, not coverage.

Logistical fatigue and how to reset

What usually breaks first is not the sampling plan — it is the samplers. Day three of walking through heat, refusal after refusal, and the team starts shortening interviews, skipping the second backup compound, or recording 'no answer' for houses they never actually approached. That is not malice; it is collapse. The diagnostic check for logistical fatigue is a gut check before the evening data sync: did we hit our target number of samples today, or did we hit our target number of excuses? A single day of drift can cascade into a week of unusable data because the gaps are not random — they are biased toward easy houses, known contacts, and short walks.

Reset protocol: stop sampling at noon. Eat. Sit under a tree. Redraw the day's remaining targets on a paper map — not a phone screen, because phones die and distract. Assign each remaining compound a difficulty rating (close, far, hostile dog). Then send the most rested team member to the hardest one first. The psychological effect is immediate: once the worst sample is done, everything else feels manageable. I have used this reset in four outbreaks and never once regretted the lost hour. The alternative — pushing through fatigue — guarantees a dataset that looks complete but tells a false story. That hurts more than a late lunch.

FAQ: Common Questions from the Field

What if there is literally no sampling frame?

You land in a settlement where the last census was a rumor, roads are goat trails, and every structure looks identical from the air. No roster. No map. No household list pinned to the health post wall. Most teams freeze here. The correct move is not to panic — it's to build a frame by walking. Send two teams in opposite directions along the main footpaths, counting every structure they pass on a phone tally counter or a notepad tally. One pass. Then a perpendicular transect. That gives you a rough perimeter and a density estimate. The catch is that you will miss hidden clusters — compounds tucked behind ridges, seasonal shelters deep in the bush. But you will have a number. A bad number you can adjust is better than no number at all. Use that raw count as your denominator for the day and flag it clearly in your field log: 'Frame built by rapid census walk, ±20% error likely.' I have seen teams waste two full days searching for a perfect frame that never existed — the outbreak moved on without them.

How do I estimate sample size without a denominator?

You have a suspected syndrome, a clinical case definition, and absolutely no idea how many people live in the catchment. Textbook formulas bleed out here. The pragmatic fix: flip the logic. Instead of calculating a sample from a known population, work backward from a feasible daily maximum. Decide how many households your team can actually reach in one day — given terrain, daylight, and transport. That number becomes your sample size. Not elegant. But it stops you from designing a survey that requires 400 interviews when your three-person team can only do 18 before the afternoon rains wash out the trail. The trade-off is precision: your confidence intervals widen, and your findings become directional rather than definitive. That is acceptable for an outbreak response. One trick I have used: if you can count the number of visible structures from a high point, multiply that by an average household size from a nearby district (or a regional demographic estimate) and use that as a soft denominator. Label it 'provisional' in every report. The odd part is that local teachers or shopkeepers often know the settlement size better than any printed document — ask them first.

You can always correct a bad estimate later. You cannot correct a delay in starting.

— field protocol lead, Mozambique, 2022

How do I validate representativeness after the fact?

You have collected your samples. The data sits in a tablet. But you suspect you oversampled the main road and undersampled the deep compounds — a common bias when the village lacks addresses. What now? Run a quick spatial check: plot each sampled household on a paper map or a phone screenshot. Mark the ones you missed. If your samples cluster along one axis — the road, the river, the market square — your coverage is skewed. Immediate fix: return the next morning for a targeted mop-up of the empty zones. Not the whole village, just the gaps. Most teams skip this step, assuming randomization solved the problem. It did not. The real-world chaos of field logistics — impassable tracks, hostile dogs, absent households — breaks randomness every time. Validate by comparing your sample's age-sex distribution against any available community data (clinic registers, vaccination tallies, even school attendance records). If your sample shows 80% adults and the clinic says 60% of the population is under 15, your adult-heavy sample will mislead attack rate calculations. Correct it by weighting during analysis, not by discarding data. That hurts less than restarting.