How to Draw a Sample from Your Sampling Frame

Knowing your sample size is only half the job. Selecting which units actually end up in your sample is the other half, and it depends entirely on having an accurate, well structured sampling frame to begin with. This guide first explains what a sampling frame is and how to build one, then walks through exactly how each sampling design works, then shows how to apply it in AnalyZ Solutions.

01What a sampling frame is?

A sampling frame is the complete list of every unit in the population you intend to sample from, the actual list a design is applied to. If the population is households in a district, the frame is a list of every household, or every enumeration area that can be used to reach them. If the population is health facilities, the frame is a list of every facility. The frame is not the population itself, it is the operational list that stands in for the population when a sample is drawn, and the quality of the sample can never exceed the quality of the frame it was drawn from.

02How to develop a sampling frame?

A usable sampling frame typically comes from one of a few sources, often combined.

• Administrative or census data. Government census enumeration lists, health facility registries, school censuses, or civil registration data. These are often the most complete starting point where they exist and are reasonably current.
• A dedicated listing exercise. Where no adequate administrative list exists, field teams physically visit an area and record every household, structure, or unit present, building the frame from the ground up. This is common in informal settlements or areas with poor administrative coverage.
• Programme or service records. Where the population of interest is defined by contact with a service, for example clinic attendees or programme beneficiaries, the service's own records can form the frame, with appropriate attention to how complete and current those records are.
• Satellite imagery or geospatial data. For household or structure level frames in areas without a recent census, dwelling structures can sometimes be identified and enumerated from recent satellite imagery, then used as the frame for a subsequent field listing.

Whichever source is used, the frame should be as current as practical, since populations move, grow, and change between when a list was made and when it is used to draw a sample.

03The pitfalls of an inaccurate sampling frame

No adjustment made after the sample is drawn can fully correct for a flawed frame. This makes frame quality one of the highest leverage points in the entire study design, often more consequential than the sample size calculation itself.

• Undercoverage. Units that exist in the population but are missing from the frame have zero chance of selection, no matter how the sample is drawn. If the missing units differ systematically from those included, for example newer informal settlements missing from an outdated administrative list, the resulting sample is biased in a way statistical adjustment cannot fix.
• Overcoverage. Units on the frame that no longer exist, are duplicated, or do not actually belong to the population, for example demolished structures still listed, or the same household appearing twice under different spellings. This wastes sampling effort and can distort estimates if not identified and removed.
• Outdated information. A frame built from data that is months or years old may no longer reflect the current population, particularly in areas with rapid population growth, displacement, or urban development.
• Missing information needed for the design. A frame without a strata or cluster column cannot support a stratified or cluster design without additional work first. Discovering this after fieldwork has begun is far more costly than checking for it in advance.

04What columns your frame should have

At minimum, a frame needs a unique identifier for each unit. Beyond that, which columns are needed depends on the sampling design being used, strata and cluster designs need the relevant grouping column present before the draw, not added afterward.

An example household frame prepared for a cluster survey might look like this.

Here, village supports a cluster or PPS design, region supports a stratified design, and household_id is what the sample selection actually operates on and reports back.

05The five designs, and when to use each

06How each design actually works

Understanding the mechanics behind each design, not just what it is called, is what allows the resulting sample to be explained and defended later, to a supervisor, a reviewer, or a donor asking how a specific unit ended up selected. Each example below uses a small illustrative frame to make the mechanism concrete.

Simple random

Every unit is given an equal chance of selection, with no grouping or ordering involved. Conceptually, this is equivalent to writing every unit's identifier on an identical slip of paper, mixing them thoroughly, and drawing the required number without looking.

Because every unit had the same probability of selection, defending the output only requires confirming the frame was complete and the draw was genuinely random, not why any specific household was or was not chosen.

Systematic

A fixed interval, k, is calculated by dividing the frame size by the desired sample size. A single random starting point between 1 and k is chosen, and every kth unit after that is selected.

Only the starting point is random, everything after that is fully determined by the interval. This is why systematic sampling requires the frame to already be in an order that has no hidden pattern matching the interval, otherwise the fixed step size can accidentally and repeatedly land on, or skip, a particular subgroup.

Stratified

The frame is divided into subgroups, strata, and a sample is drawn independently within each. In proportionate allocation, each stratum's share of the sample matches its share of the total frame.

Every unit within a given stratum still has an equal chance of selection, the allocation across strata is what differs from simple random sampling. Defending the output here means being able to state the allocation rule, proportionate to population share in this example, and confirming it was applied consistently.

Cluster

Whole groups are selected first, and either all units within a selected group, or a fixed number sampled within it, become part of the sample.

Households within the same village are not independently selected from each other, they are included as a block once their village is chosen. This is precisely why a design effect adjustment is needed when calculating the sample size for a cluster design, since households within the same selected village tend to be more alike than households picked completely at random. See Design Effect and ICC Explained for the statistical reasoning behind this.

PPS, probability proportional to size

Clusters are selected with a probability proportional to a size measure, so a village with more households has a proportionally higher chance of being chosen than a smaller village. This is implemented by assigning each cluster a range of cumulative population, then drawing random numbers across the full cumulative range.

Village D, the largest, occupies 200 of the 500 possible numbers, a 40% chance of being hit by any single random draw. Village C, the smallest, occupies only 30 of 500, a 6% chance. This is the entire mechanism behind why PPS gives larger clusters a proportionally higher chance of selection, and it is what makes the resulting sample self weighting when a consistent number of units is then sampled within each selected cluster.

1. Upload the frame. Click to upload a CSV or Excel file. If there are multiple sheets, select the appropriate one. Once loaded, the row and column count confirms it read correctly.
2. Filter a sub-group (optional). If you want to draw the sample for a sub-group of the frame, then define the filter criteria.
3. Remove duplicates (optional). if your data contains duplicate rows and you forgot to remove it before uploading you can remove them here. On selecting a column name, it will display the number of duplicates. On clicking Remove Duplicates, they will be removed.
4. Choose a design. Pick from the five options above, based on the study design and which columns the frame already contains.
5. Set parameters. Enter the target sample size, plus any design specific settings, which column holds strata, a cluster identifier, or a size measure for PPS.
6. Optionally set a random seed. A seed makes the draw exactly reproducible. The same seed, frame, and design will always produce the same sample, useful for audit trails or regenerating a lost enumeration list.
7. Draw and export. Review the preview table, then download the selected sample as CSV or Excel, either the selected rows alone, or the full frame with a selected flag column added.

07About random seeds

A seed is simply a starting number for the random selection process. Setting one means anyone with the same frame, design, and seed will get an identical sample, which is genuinely useful for reproducibility without meaningfully increasing bias risk, since the frame is fixed before selection and there is no outcome data yet to influence.

08Reading the output

Every export includes a sample identifier column giving each selected unit a sequential number, useful directly as an enumeration list for field teams. Design-specific columns are added automatically, a stratum column for stratified draws, and a cluster column for cluster or PPS draws.

09Common mistakes to avoid

• Drawing a sample before the frame has been checked for completeness. A well executed random draw cannot fix a frame that was missing units to begin with. Verify coverage before, not after, drawing the sample.
• Using systematic sampling on a frame with a hidden periodic pattern. If the frame happens to be ordered in a way that repeats at the same interval as the sampling step, the sample can systematically miss or overrepresent a subgroup without this being obvious from the output alone.
• Applying proportionate stratified allocation when subgroup level precision is actually needed. Proportionate allocation can leave a small stratum with very few sampled units. If separate, reliable estimates are needed for each stratum, an equal or disproportionate allocation should be considered instead.
• Forgetting to record the random seed used. Without it, an audit or a request to regenerate a lost enumeration list cannot reproduce the original sample exactly, even with the same frame and design.