Written and maintained by CASRAI Editorial Board
Last updated
Probability proportional to size (PPS) sampling is a method for selecting sampling units — typically clusters such as clinics, schools, villages, or facilities — where each unit’s chance of being selected is proportional to its size, rather than every unit having an equal chance. A clinic with 5,000 registered patients is roughly five times more likely to be selected than a clinic with 1,000 patients. This is the standard first-stage selection method in large multistage surveys (national health surveys, household demographic surveys, facility-based studies) precisely because, combined with a fixed sample size per selected cluster, it produces a self-weighting sample: every individual in the population ends up with the same overall probability of selection, regardless of which cluster they belong to.
This guide covers the logic behind PPS sampling, how a PPS sample is actually drawn (the cumulative-size and systematic-selection mechanics), a fully worked selection example with an independently verified Monte Carlo check, why PPS beats simple random sampling of clusters when cluster sizes vary, and the practical pitfalls that erode the self-weighting property in real fieldwork.
The Core Logic: Why Size Should Drive Selection Probability
PPS sampling exists to solve a specific problem in multistage cluster designs. Picture a two-stage design: first select a sample of clusters (clinics, schools, villages), then select a fixed number of individuals from within each selected cluster. If clusters are selected with equal probability — simple random sampling (SRS) of clusters — but cluster sizes vary substantially, something goes wrong at the individual level:
- A clinic with 500 patients and a clinic with 5,000 patients are equally likely to be selected under SRS of clusters.
- But if the same fixed number of patients (say, 25) is then sampled from whichever clinics happen to be selected, a patient at the 500-patient clinic has a much higher chance of ending up in that within-clinic sub-sample (25/500 = 5%) than a patient at the 5,000-patient clinic (25/5,000 = 0.5%).
- The net effect: individuals in small clusters are systematically over-represented in the final sample, and individuals in large clusters are systematically under-represented — even though clusters themselves were selected “fairly.”
PPS sampling corrects this at the selection stage rather than the analysis stage. By giving larger clusters a proportionally higher chance of being selected in the first place, PPS offsets the fact that a fixed-size sub-sample drawn from a large cluster represents a smaller fraction of that cluster. The two effects cancel — that cancellation is exactly what “self-weighting” means, and the worked example below shows it numerically.
What “Self-Weighting” Means, Precisely
A sample is self-weighting when every unit in the final sample has the same overall probability of selection, so a simple unweighted mean or total is already an unbiased estimate of the population quantity — no analysis-stage sampling weights are required to correct for unequal selection probabilities. (Design-based variance still needs to account for the clustering itself — see the cluster sampling guide‘s design effect section — self-weighting removes the need for bias-correcting weights, not the need to model clustering.)
For a two-stage design where clusters are selected by PPS and a fixed number of individuals k is sampled within each selected cluster, the overall selection probability for any individual in cluster i (size Mi, sampling n clusters from a frame totaling M individuals) is:
| Stage | Selection probability |
|---|---|
| Stage 1 — cluster i selected by PPS | π1 = n × Mi / M |
| Stage 2 — individual selected within cluster i (fixed k) | π2 = k / Mi |
| Overall | π1 × π2 = (n × Mi / M) × (k / Mi) = n × k / M |
The Mi terms cancel algebraically — the overall probability no longer depends on cluster size at all. That is the entire mechanism behind self-weighting, and it only holds under two conditions: clusters are selected proportional to size, and a fixed (not proportional) number of individuals is sampled within each selected cluster. Vary either condition and the sample stops being self-weighting — see the pitfalls section below.
How to Actually Draw a PPS Sample
Drawing a PPS sample requires a measure of size (MOS) for every unit on the sampling frame — a count reasonably correlated with what you actually care about (registered patients, enrolled students, households, prior-year visit volume). Two selection mechanics are used in practice:
- Cumulative-total (Lahiri’s) method. List units with their cumulative size totals, draw a uniform random number between 1 and the maximum unit size (or between 1 and the frame total, in some variants), and accept or reject candidate draws against each unit’s size. Typically used to draw one unit per random number, repeated independently — which allows the same large unit to be drawn more than once (sampling with replacement).
- Systematic PPS selection. The more common method for drawing n distinct clusters without replacement, and the method used in the worked example below: order the frame, compute each unit’s cumulative size range, divide the frame total M by the number of clusters wanted (n) to get a sampling interval I = M / n, pick one random start point r drawn uniformly between 0 and I, then select the units whose cumulative range contains r, r + I, r + 2I, …, r + (n−1)I. Because the points are spaced exactly I apart and no single unit’s size exceeds I (see the pitfall below when it does), this selects exactly n distinct units, each with probability n × Mi / M.
Worked Example: Selecting Four Clinics From Twelve by PPS
The following is an illustrative composite example, computed independently with a seeded pseudo-random number generator (mulberry32, seed 20260829) so every number below is reproducible, not asserted. A sampling frame of 12 primary-care clinics, with each clinic’s registered-patient count as its measure of size:
| Clinic | Registered patients (Mi) | Cumulative range |
|---|---|---|
| A | 3,200 | 0 – 3,200 |
| B | 850 | 3,200 – 4,050 |
| C | 5,100 | 4,050 – 9,150 |
| D | 1,400 | 9,150 – 10,550 |
| E | 2,200 | 10,550 – 12,750 |
| F | 640 | 12,750 – 13,390 |
| G | 4,300 | 13,390 – 17,690 |
| H | 990 | 17,690 – 18,680 |
| I | 1,750 | 18,680 – 20,430 |
| J | 3,600 | 20,430 – 24,030 |
| K | 720 | 24,030 – 24,750 |
| L | 2,460 | 24,750 – 27,210 |
Frame total M = 27,210 patients across 12 clinics. Selecting n = 4 clinics gives a sampling interval I = M / n = 27,210 / 4 = 6,802.5. (Every clinic’s size is below this interval — 5,100 at most — so systematic selection is valid without a certainty-selection adjustment; see the pitfalls section.) A single uniform random draw between 0 and 6,802.5 gave a random start of r = 6,368.76. The four selection points are r, r + I, r + 2I, r + 3I:
| Selection point | Value | Falls in clinic |
|---|---|---|
| r | 6,368.76 | C (4,050–9,150) |
| r + I | 13,171.26 | F (12,750–13,390) |
| r + 2I | 19,973.76 | I (18,680–20,430) |
| r + 3I | 26,776.26 | L (24,750–27,210) |
The selected clinics are C, F, I, and L — note the method does not simply pick the four largest clinics; a small clinic (F, 640 patients) can be selected in the same draw as the largest clinic on the frame (C, 5,100 patients), because each clinic occupies exactly one interval-width “slice” of the cumulative line regardless of where that slice happens to fall relative to unit boundaries.
Verifying the selection probabilities aren’t just a plausible-looking story
The theoretical selection probability for each clinic is πi = n × Mi / M. To confirm the systematic-selection mechanism actually produces these probabilities (rather than asserting the formula and hoping), the same draw was repeated 200,000 times with independent random starts, and each clinic’s empirical selection frequency was tabulated against the theoretical value:
| Clinic | Size | Theoretical πi (n×Mi/M) | Empirical (200,000 draws) |
|---|---|---|---|
| A | 3,200 | 0.4704 | 0.4708 |
| B | 850 | 0.1250 | 0.1246 |
| C | 5,100 | 0.7497 | 0.7509 |
| D | 1,400 | 0.2058 | 0.2041 |
| E | 2,200 | 0.3234 | 0.3241 |
| F | 640 | 0.0941 | 0.0941 |
| G | 4,300 | 0.6321 | 0.6321 |
| H | 990 | 0.1455 | 0.1454 |
| I | 1,750 | 0.2573 | 0.2575 |
| J | 3,600 | 0.5292 | 0.5288 |
| K | 720 | 0.1058 | 0.1064 |
| L | 2,460 | 0.3616 | 0.3613 |
Every empirical value lands within 0.002 of its theoretical counterpart — the mechanism behaves exactly as the formula predicts, including for clinic C, whose πi of 0.75 confirms that above a certain size a unit is selected in the large majority of possible random starts, not merely “more often than a small unit.”
Why PPS Beats Simple Random Sampling of Clusters
The self-weighting formula above already shows this algebraically — the same worked frame makes it concrete. Suppose 25 patients are sampled from each selected clinic (fixed k = 25), under two different cluster-selection designs:
| Clinic | Size | PPS design: overall probability per patient | SRS-of-clusters design: overall probability per patient |
|---|---|---|---|
| A | 3,200 | 0.003675 | 0.002604 |
| B | 850 | 0.003675 | 0.009804 |
| C | 5,100 | 0.003675 | 0.001634 |
| D | 1,400 | 0.003675 | 0.005952 |
| E | 2,200 | 0.003675 | 0.003788 |
| F | 640 | 0.003675 | 0.013021 |
| G | 4,300 | 0.003675 | 0.001938 |
| H | 990 | 0.003675 | 0.008418 |
| I | 1,750 | 0.003675 | 0.004762 |
| J | 3,600 | 0.003675 | 0.002315 |
| K | 720 | 0.003675 | 0.011574 |
| L | 2,460 | 0.003675 | 0.003388 |
Under PPS with a fixed within-cluster sample size, every patient’s overall probability of selection is identical — 0.003675, regardless of which clinic they happen to attend. Under SRS of clusters with the same fixed within-cluster sample size, overall probability ranges from 0.001634 (a patient at the largest clinic, C) to 0.013021 (a patient at the smallest clinic, F) — nearly an 8-fold difference in who is more likely to end up in the sample, driven entirely by which clinic they happen to attend rather than anything about them.
That 8-fold imbalance is not a hypothetical downside: left uncorrected, it biases every unweighted estimate toward whatever is disproportionately true of small-clinic patients. Correcting it after the fact means constructing and applying analysis weights (the inverse of each unit’s overall selection probability — see the survey weighting guide), and unequal weights carry their own statistical cost: they inflate the variance of weighted estimates, quantified by the weighting component of the design effect, DEFFweighting ≈ 1 + CV(w)², where CV(w) is the coefficient of variation of the final weights. PPS sidesteps that cost by preventing the imbalance from arising in the first place, rather than fixing it after the fact.
PPS in Multistage Designs
PPS sampling is rarely the whole design — it is typically the first stage of a larger multistage structure. A national household health survey, for example, commonly selects primary sampling units (districts or census enumeration areas) by PPS at stage one specifically so that a fixed number of households can be selected within each at later stages while keeping the overall design close to self-weighting. See the cluster sampling guide for the full one-stage/two-stage/multistage framework this fits into, and its sampling-weight construction section for how weights are actually calculated when a real design departs from perfect self-weighting (which nearly every real design does, for the reasons below).
Practical Pitfalls That Erode the Self-Weighting Property
- An outdated measure of size. PPS selection probabilities are only correct relative to whatever size figures are on the frame. A clinic’s patient roster measured two years before fieldwork can materially misstate its current size — the selection math is unbiased with respect to the frame’s numbers, not with respect to reality, if the frame is stale.
- A unit larger than the sampling interval. If a single unit’s size exceeds I = M/n, systematic selection breaks — that unit would occupy more than one interval and could be hit by more than one selection point, or the interval logic simply doesn’t apply cleanly. The standard fix is to designate oversized units as self-representing (selected with certainty, π = 1), remove their size from the frame total, and apply PPS selection to the remaining units for the remaining number of clusters to select.
- Non-fixed within-cluster sample sizes. If the number of individuals sampled per cluster varies (proportional to cluster size, or determined by fieldwork constraints rather than a fixed target), the Mi cancellation in the self-weighting formula no longer holds, and the design is no longer self-weighting even though clusters were still selected by PPS — analysis weights become necessary again.
- Non-response. Even a perfectly executed PPS design loses its exact self-weighting property once non-response is unevenly distributed across clusters — a non-response adjustment factor is layered onto the design weight in practice, which is why real survey weights are rarely just “1” even in a nominally self-weighting design.
- PPS is not a substitute for a good measure of size. If no variable reasonably correlated with true unit size is available, PPS selection using a poor proxy can introduce more bias than it removes; simple random sampling of clusters, corrected with weights at analysis, is sometimes the more defensible choice when size data is unreliable.
Frequently Asked Questions
What counts as “size” in PPS sampling?
Any measure reasonably correlated with the quantity the study cares about: registered patients, enrolled students, households, employees, or (when a direct headcount is unavailable) a defensible proxy like prior-year visit volume or building square footage. The measure of size does not need to be the outcome variable itself — it needs to correlate with cluster size well enough that proportional selection actually approximates proportional representation.
Is every PPS sample automatically self-weighting?
No. Self-weighting requires PPS selection at the cluster stage and a fixed (not size-proportional) number of individuals sampled within each selected cluster. PPS selection alone, paired with a variable within-cluster sample size, does not cancel the Mi terms and still requires analysis weights.
How is PPS sampling different from stratified sampling?
They solve different problems. Stratified sampling divides the population into homogeneous groups and samples from every group to guarantee representation and reduce variance. PPS sampling is a selection rule applied within cluster sampling — it changes each cluster’s chance of being picked based on size, and most clusters are still excluded entirely, unlike stratification where every stratum contributes.
Can PPS be combined with stratification?
Yes, and this is common in large national surveys: the frame is first stratified (by region, urban/rural status, or facility type), and PPS selection is then applied independently within each stratum. Stratification controls which strata are guaranteed representation; PPS controls how selection probability varies by size within each stratum.
What if I can’t get an accurate measure of size for every unit on the frame?
A partially missing or unreliable measure of size is a real practical constraint, not a reason to fabricate one. Options include using the best available proxy with its limitations documented, falling back to simple random sampling of clusters with analysis-stage weighting instead, or conducting a rapid enumeration pass to establish current sizes before selection — whichever is defensible depends on how much the study’s conclusions would change if the size figures are wrong.
