What Is Sampling Without Replacement?
Sampling without replacement is a selection method where each unit drawn from the population is removed from the pool, preventing it from being selected again. Once a person, household, or record enters your sample, it's excluded from all subsequent draws. This is the standard approach in virtually all survey research, you interview each respondent once, not multiple times. The result is that each successive draw has a slightly different selection probability because the remaining pool shrinks by one unit after each selection. This creates a mild statistical dependence between draws that doesn't exist in with-replacement sampling, but it also means every unit in your final sample is unique, and your sample captures as much distinct information as possible for its size. For any given sample size, sampling without replacement is more statistically efficient than sampling with replacement.
Why Sampling Without Replacement Matters
If you're conducting a survey, you're almost certainly sampling without replacement, and that has practical implications for your variance estimates. Standard textbook formulas for sampling variance assume with-replacement sampling, which overestimates uncertainty when you've sampled a meaningful fraction of the population. Applying the finite population correction accounts for the reduced variance that without-replacement sampling provides, tightening your confidence intervals and potentially letting you achieve your target precision with a smaller sample.
How Sampling Without Replacement Works
The process is what you'd expect from the name, but the statistical implications are worth understanding in detail.
The Basic Process
Start with a population of N units. Randomly select one unit and set it aside, it's in your sample. The remaining population now has N-1 units. Select another at random from the remaining N-1, set it aside. Continue until you've selected n unique units.
Every unit appears in the sample at most once. The final sample contains exactly n unique observations, which means n unique pieces of information about the population.
Selection Probabilities
In simple random sampling without replacement, every unit has the same overall inclusion probability: n/N. But the conditional probabilities change across draws. On the first draw, each unit has a 1/N probability. On the second draw (given the first unit was removed), each remaining unit has a 1/(N-1) probability. These conditional probabilities differ from draw to draw, but the unconditional (marginal) inclusion probability for every unit is still n/N.
This equal marginal probability is what makes simple random sampling without replacement a self-weighting design, every unit has the same chance of ending up in the sample.
Variance Reduction
Sampling without replacement is more precise than sampling with replacement because removing selected units from the pool reduces redundancy. With-replacement sampling can select the same unit multiple times, wasting draws on duplicate information. Without-replacement sampling guarantees that every draw adds a new unit and new information.
The efficiency gain is captured by the finite population correction (FPC): (N - n) / (N - 1), which is sometimes approximated as (1 - n/N). Multiply the with-replacement variance by the FPC to get the without-replacement variance.
When n is tiny relative to N (sampling fraction under 5%), the FPC is close to 1 and the efficiency gain is negligible. When n is a substantial fraction of N, say, 20% or more, the FPC meaningfully reduces variance. At the extreme (n = N, a census), variance drops to zero.
Practical Implementation
In practice, sampling without replacement is implemented through random number generators, systematic selection with a random start, or algorithmic methods that draw unique indices from the population list. Most survey software and statistical packages default to without-replacement selection.
For list-based frames, the implementation is straightforward: assign random numbers, sort, and take the top n. For multi-stage designs, without-replacement selection happens at each stage, select PSUs without replacement, then select households within each PSU without replacement.
Estimation
The standard estimator for a population mean from a simple random sample without replacement is the sample mean, with variance estimated as (s^2 / n)(1 - n/N), where s^2 is the sample variance and (1 - n/N) is the FPC. Confidence intervals use this adjusted variance.
For complex designs (stratified, clustered), without-replacement selection at each stage affects the variance calculations. Software packages like Stata's svy commands and R's survey package handle the multi-stage without-replacement variance estimation automatically when you specify the design.
Dependence Between Draws
Without-replacement sampling creates negative covariance between units' inclusion indicators. If Unit A is in the sample, Unit B's conditional inclusion probability drops slightly (from n/N to (n-1)/(N-1) if we condition only on A being selected). This negative dependence is what reduces variance compared to with-replacement sampling, knowing one unit is included slightly reduces the uncertainty about the remaining population composition.
When to Use Sampling Without Replacement
- Virtually all survey research: it's the default and appropriate choice whenever you want each respondent to contribute exactly once
- Studies where the sampling fraction is large enough (above 5-10%) for the finite population correction to meaningfully improve precision
- Designs where duplicate respondents would create analytical problems: repeated-measures confusion, inflated sample sizes, weighting complications
- Any study where maximizing unique information per interview matters: which is every budget-constrained study
- Official statistics and government surveys where without-replacement designs with known inclusion probabilities are required for methodological rigor
Common Mistakes to Avoid
- Ignoring the finite population correction when sampling a large fraction of the population. If you've sampled 15% or more of the population, the FPC tightens your confidence intervals noticeably. Leaving it out makes your estimates look less precise than they actually are.
- Applying the FPC to non-probability samples. The finite population correction is a probability sampling concept. If your sample isn't drawn with known selection probabilities, the FPC doesn't apply and shouldn't be used to artificially shrink your standard errors.
- Forgetting to specify without-replacement selection in complex survey analysis software. Survey analysis packages need to know the sampling design to calculate correct variances. Specifying with-replacement when you sampled without replacement produces conservative (larger) standard errors.
How Quali-Fi Supports Sampling Without Replacement
Quali-Fi's survey platform ensures each respondent completes the survey exactly once through unique invitation links, respondent deduplication, and device fingerprinting, implementing without-replacement sampling at the data collection level. The platform's export tools include sampling fraction metadata so your analysis team can apply the finite population correction when warranted.
Frequently Asked Questions
When does the finite population correction matter enough to use?
The conventional threshold is when the sampling fraction (n/N) exceeds 5%. Below that, the FPC is close to 1 and can be ignored. Above 10%, it starts making a noticeable difference, a sampling fraction of 20% reduces variance by about 20%.
Is systematic sampling equivalent to simple random sampling without replacement?
Systematic sampling (every kth unit from a random start) produces a sample that looks similar to simple random sampling without replacement, but the variance properties can differ depending on how the population is ordered. If the list has a periodic pattern that aligns with your sampling interval, systematic sampling can be more or less efficient than simple random sampling.
Can I convert a without-replacement sample to a with-replacement analysis?
Yes, and some multi-stage variance estimation methods do this. Treating PSUs as if they were selected with replacement produces a conservative (upward-biased) variance estimate, which is sometimes preferred for its simplicity and because it errs on the side of caution.
Related Topics
- Sampling With Replacement
- Finite Population Correction
- Design Effect (DEFF)
- Proportionate Stratified Sampling
- Undersampling
Maximize the information in every interview. Start a free trial with Quali-Fi and use unique respondent links, deduplication, and proper sampling design to get the most from your survey data.