Systematic sampling is a probability sampling method: you number the population, divide its size by the sample size to get an interval k, pick a random start between 1 and k, and take every k-th unit from there. It needs one random number and spreads the sample evenly along the list. This guide covers the steps, what to do when N/n is not a whole number, and the trap that can ruin it.
Key Takeaways
- The sampling interval is k = N/n. Draw one random start between 1 and k, then take units r, r + k, r + 2k and so on.
- When N/n is not whole, round k down and go circular: a random start anywhere from 1 to N, wrapping round to the top of the list.
- If the list cycles in step with k, every possible sample is biased. In our simulation of 100 weeks of daily data with k = 7, no possible sample’s 95% interval contained the true mean.
- On a list sorted by something related to the outcome, a systematic sample beats a simple random one: a tenth of the variance in our trended example.
- One random start gives no unbiased variance estimate. Use several random starts, or argue that the list is in random order.
What is systematic sampling?
A systematic sample takes units at a fixed interval along a list after a random starting point. With 1,200 students and room for 100, the interval is 12; if the random start is 7, the sample is students 7, 19, 31 and so on up to 1,195. Each student has a 1 in 12 chance of selection, so it is a probability method, but only 12 different samples are possible.
Statistics Canada’s teaching module describes a design where “there is a gap, or interval, between each selected unit in the sample.” The selection chance matches a simple random sample’s, but with simple random sampling “any combination of 100 units would have a chance of making up the sample, while with systematic sampling, there are only four possible samples” (its example takes 100 of 400).
Penn State’s STAT 506 course treats each possible sample as one “primary unit” whose members sit far apart on the list, noting that “it is common to have a systematic sample of size 1.” That explains the variance problem further down. The glossary entry on random sampling places the method among the other probability designs.
How do you do systematic sampling, step by step?
Number the units in your sampling frame from 1 to N. Divide N by the sample size n to get the interval k. Pick a random whole number between 1 and k with a random number generator; that unit is the first in your sample. Then keep adding k until you have n units. Record the start and the interval, and check how the list is ordered before you draw.
It also works on a stream you can count instead of a list, such as “every 10th person who enters a particular store” in Statistics Canada’s example. A study of Ethiopian truck drivers needed about 28 interviews a day, planned on at least 300 trucks a day, and so interviewed “every eleventh driver from a random start” (Yosef, 2020). The stream has to be the population you want to describe: every 10th shopper at one store is a probability sample of that store’s shoppers during the hours sampled, but a convenience sample of shoppers in general.
What if N/n is not a whole number?
Round the interval down and use circular systematic sampling: treat the list as a circle, pick a random start anywhere from 1 to N, and step by k, wrapping past the end, until you have n units. Every unit keeps exactly the same chance of selection, n/N. Rounding without the circle leaves you with the wrong sample size or unequal chances.
Training notes from the UN Statistical Institute for Asia and the Pacific (UNSIAP) say a rounded interval makes the ordinary method “no longer epsem” (equal probability of selection), and round k down for the circular fix: “If N = 15 and n = 4, then k is taken as 3 and not 4.” OpenStax’s Introductory Statistics uses the same wrap-around with a phone book of 20,000 listings, warning that “you might have to go back to the beginning of your phone list.”
The shortcuts each cost something:
- Round down, no circle. With 1,450 units and k = 14, starts 1 to 8 give 104 units and starts 9 to 14 give 103. Penn State’s 9,000-student example rounds 7.5 down to 7 and draws the start from 1 to 600 so that exactly 1,200 picks fit. By our count the chances then differ: 1 in 600 for unit 1, 14.2% or 14.3% in the middle of the list, zero for units 8,994 to 9,000, against a flat 13.3%.
- Round up. The sample runs short: an interval of 3 through 1,144 names yields at most 382, the case in the study below.
Worked example: 100 students from a list of 1,450
A student wants to survey 100 of the 1,450 first-year students on an alphabetical enrolment list. The interval is 1,450 / 100 = 14.5, so the design is circular with k = 14 and a random start between 1 and 1,450. Our seeded draw gave a start of 1,126; the picks run to the end of the list and wrap round to the top.
- N = 1,450, n = 100, N/n = 14.5. Round down: k = 14.
- Random start between 1 and 1,450: 1,126 (Python, seed 1450).
- Picks: 1,126, 1,140, 1,154 … 1,448, which is 24 names.
- Wrap: 1,448 + 14 = 1,462, and 1,462 − 1,450 = 12. Continue with 12, 26, 40 … and stop at the 100th name, 1,062.
- Check: 100 different students. Enumerating all 1,450 starts confirmed that each student’s chance is exactly 100/1,450, or 6.9%.
Surnames are unlikely to cycle every 14 names, so an alphabetical list suits this survey. How large n should be is a sample size decision made before any of these steps.
Systematic vs simple random vs stratified vs cluster sampling
All four are probability methods that randomise different things: single units, one start, units within every group, or whole groups. The systematic design sits closest to simple random sampling, and its precision depends on how the list is ordered: about the same on a random order, better on a list sorted by a related variable, worse on a cycling one.
| Design | What is chosen at random | Frame needed | Precision vs a simple random sample of the same size | Main risk |
|---|---|---|---|---|
| Simple random | Every unit, independently | A full list | Baseline | Small groups missed by chance |
| Systematic | One start (or a few) | A full list, or a stream you can count | About the same on a randomly ordered list; better on a list sorted by a related variable; worse if the list cycles in step with k | Periodicity; no variance estimate from one start |
| Stratified | Units within every stratum | A list with the stratum variable | Better when strata differ on the outcome | No gain on an unrelated variable |
| Cluster | Whole groups (schools, blocks) | A list of clusters | Usually worse | Analysis that ignores the clusters |
Sorting the list before the draw, which UNSIAP calls implicit stratification, “guarantees that the sample of units will be spread across the categories,” though without the fixed numbers per group that stratified sampling sets. And where a cluster sample’s units sit together, Penn State notes that a systematic sample’s units “are separate,” so on most lists it avoids the precision loss of cluster sampling.
The periodicity trap: what can go wrong?
If the list repeats a pattern and the interval matches it, every sample hits the same point in the cycle. A list of days with k = 7 gives you only Mondays or only Saturdays; a staff list in sections of 10 employees, each listing its manager first, can give you only managers with k = 10. The estimate is biased, and the sample looks deceptively consistent.
The staff list is Statistics Canada’s example: with an interval of 10, “you could end up selecting only managers or only the newest employees in each section.”
To measure the damage, we simulated 700 consecutive days (100 weeks) of a hospital’s emergency admissions: weekdays average 50, weekends 70, plus noise. The true mean was 55.86. With n = 100 the interval is 7, so only seven samples are possible; for each we computed the usual 95% confidence interval, as if it were a simple random sample.
| Random start | Days in sample | Mean | Naive 95% interval | Contains 55.86? |
|---|---|---|---|---|
| 1 to 5 | One weekday only | 49.74 to 50.52 | about ±1.1 around each | No |
| 6 | Saturdays only | 70.70 | 69.58 to 71.82 | No |
| 7 | Sundays only | 70.11 | 68.87 to 71.35 | No |
| Simple random, 10,000 draws | Any 100 days | 55.88 on average | about ±2.0 around each | 95.0% of intervals |
No systematic sample came within 5 admissions of the truth, and each interval was about half as wide as a simple random sample’s, because 100 Mondays resemble each other more than 100 random days do. Penn State describes this case: when the population is periodic, the textbook formula “will underestimate the variance since if the period k is chosen poorly, then the elements sampled may be too similar to each other.”
Two fixes worked. Shuffling the 700 days before taking every 7th matched simple random sampling (95.1% of intervals held the truth). Ten random starts of 10 days each, Penn State’s repeated design, also gave honest intervals (94.5%), but wider ones: a standard error of 2.74 against 1.03 for a simple random sample, because each start still picked a single weekday.
When is systematic sampling better than simple random sampling?
A systematic sample does better when the list is sorted by something related to what you measure, such as patients by age or firms by size. The fixed interval then spreads the sample evenly from one end of the list to the other, a built-in stratification that a simple random draw lacks. On a randomly ordered list the two do about equally well; on a periodic list the systematic sample does worse.
In Penn State’s example the first 400 chips in a batch are fine and the last 300 defective after a machine fault; a systematic sample “will select uniformly over the defective and non-defective items.”
Our second simulation used 1,000 units sorted on a variable that raised the outcome steadily (a linear trend plus noise), with n = 50 and k = 20. Across all 20 possible starts, the estimates ranged from 43.19 to 45.95 around a true mean of 44.77. The exact standard error was 0.68, against 2.12 for a simple random sample of 50: about a tenth of the variance.
How do you estimate the variance of a systematic sample?
One random start selects one primary unit, a sample of size 1 at the level that matters, so no formula gives an unbiased variance from the data alone. Either assume the list is in random order and use the simple random sample formula, or design in several random starts and measure the spread between them.
Penn State says the shortcut is safe “if the population is randomly ordered.” On a sorted list the simple random formula “will overestimate the variance”: in our trended simulation its standard error averaged 2.14 against a true 0.68, so intervals came out about three times too wide. On a periodic list it underestimates, as the weekday table showed. UNSIAP’s notes put it in one line: “Variance estimates cannot be obtained from a single systematic sample.”
The design fix is repeated systematic sampling. Penn State’s ferry example wants 80 of about 400 cars, so it draws 10 random starts between 1 and 50 and takes every 50th car from each; the variance comes from the spread of the 10 sample means, with 9 degrees of freedom. Report whichever route you took in your methodology section.
A real study: every 3rd patient on a diabetes register
Belsti, Akalu and Animut (2020) studied attitudes and practices around diabetes complications among type 2 diabetes patients at Addis Zemen District Hospital in northwest Ethiopia, from April to June 2019. They took every 3rd patient after a random start from the hospital’s register of 1,144 patients on follow-up, and report that 402 of the 404 patients selected took part.
What they did. A formula gave 384, plus 5% for non-response: 404. The patients were “listed on the diabetic registration chart with their respective medical registration number,” and the authors divided 1,144 by 404, “which became 3(K).” The random start came from the first interval “with simple random sampling.”
The numbers. A 99.5% response rate. 65.2% of patients had a good attitude (95% CI 60.2% to 69.4%) and 48.8% good practice (95% CI 44.0% to 53.5%).
What it can conclude. The register lists the whole target group and the start was random, so the estimates describe patients on follow-up at that hospital during the study period, not only the easiest to reach.
What it cannot. The results do not extend to other hospitals or to patients not on follow-up. The interval also shows the rounding issue: 1,144 / 404 is 2.83, and an interval of 3 through 1,144 names yields at most 382, 22 short of 404. The methods do not say how the rest were found; a circular design with k = 2 would have reached 404 with equal chances. The authors’ own limitations include self-reported answers, which no sampling design fixes.
Where Papero fits
Papero does not draw samples or compute intervals; those stay your decisions. It helps with the reading and writing around them. Find papers searches 250M+ papers through OpenAlex, so you can find published studies in your field that sampled every k-th unit and see how they described the frame, the interval and the start.
Upload those studies as PDFs and ask Chat how a study chose its interval; answers come from your library and the open document (on Plus you can also turn on web search, whose answers list their web sources) and cite the passages they used. Papero’s free course has a lesson on writing a methods section someone could repeat, which is where your frame, start and interval belong.
Frequently Asked Questions
Is systematic sampling a probability sampling method?
Yes. The start is random, so every unit on the list has a known chance of selection, 1 in k when N/n is a whole number; hence the name systematic random sampling. Unlike simple random sampling, only k different samples are possible, so the order of the list decides which units appear together.
How do you calculate the sampling interval?
Divide the population size by the sample size: k = N/n. To sample 100 from a list of 1,200, k = 12, and you take every 12th unit after a random start between 1 and 12. If N/n is not a whole number, round down and use the circular method, with a random start between 1 and N.
What is circular systematic sampling?
It is the version for a population size that is not a multiple of the sample size. Round the interval down, pick a random start from 1 to N, and step through the list by k, going back to the beginning after the end, until you have n units. Every unit keeps the same chance, n/N.
What is the main disadvantage of systematic sampling?
Periodicity. If the list repeats a pattern whose length matches the interval, such as days of the week with k = 7, every possible sample lands on the same point in the cycle and is biased. Also, one random start gives no unbiased variance estimate unless the list is in random order.
What is the difference between systematic and stratified sampling?
Stratified sampling divides the population into groups and draws a separate random sample from each, with a sample size set for every group. A systematic sample takes every k-th unit from one list after one random start. Sorting that list by a grouping variable first spreads the sample across the groups, but without fixed numbers per group.
Sources
- Belsti, Y., Akalu, Y., & Animut, Y. (2020). Attitude, practice and its associated factors towards diabetes complications among type 2 diabetic patients at Addis Zemen District hospital, Northwest Ethiopia. BMC Public Health, 20, Article 785. https://doi.org/10.1186/s12889-020-08953-6
- Illowsky, B., & Dean, S. (2023). Introductory statistics 2e, Section 1.2, Data, sampling, and variation in data and sampling. OpenStax. https://openstax.org/books/introductory-statistics-2e/pages/1-2-data-sampling-and-variation-in-data-and-sampling
- Penn State Eberly College of Science. “7 Part 1 of Cluster and Systematic Sampling.” STAT 506: Sampling Theory and Methods (online.stat.psu.edu, November 2026): https://online.stat.psu.edu/stat506/Lesson07.html
- Penn State Eberly College of Science. “8 Part 2 of Cluster and Systematic Sampling.” STAT 506: Sampling Theory and Methods (online.stat.psu.edu, November 2026): https://online.stat.psu.edu/stat506/Lesson08.html
- Statistics Canada. “3.2.2 Probability sampling.” Statistics: Power from Data! (statcan.gc.ca, November 2026): https://www150.statcan.gc.ca/n1/edu/power-pouvoir/ch13/prob/5214899-eng.htm
- UN Statistical Institute for Asia and the Pacific. (2016). Systematic sampling [Slides, Regional Training Course on Agricultural Cost of Production Statistics, Daejeon] (unsiap.or.jp, November 2026): https://www.unsiap.or.jp/sites/default/files/pdf/e-learning_el_material_5_agri_1611_cost_korsystematic_sampling.pdf
- Yosef, T. (2020). Prevalence and associated factors of chronic non-communicable diseases among cross-country truck drivers in Ethiopia. BMC Public Health, 20, Article 1564. https://doi.org/10.1186/s12889-020-09646-w