Statistics

Sampling Techniques

Read at your own pace. Use chapter practice to check your understanding.

Chapters and chapter contents
100%
Practise this chapter ↓

Last updated · By the Clerk2Officer editorial team · Report a correction

1. Why a bank samples

A regional bank has 40,000 completed applications and wants an estimate of average processing time. Reviewing every file manually would delay the staffing decision. Sampling makes the investigation manageable by observing selected applications. Its value comes from the selection design and the connection between the sample and population. Taking the first available records is easy, but may systematically select early, uncomplicated cases. A probability sample uses a known random selection mechanism and permits uncertainty to be assessed under that mechanism.

Begin by writing the target population, unit, reference period and parameter. Here the population is all qualifying completed applications during the quarter, the unit is an application and the parameter is population mean processing time. A sample mean is a statistic calculated from the selected records. It estimates the population mean when the design and measurement process support that inference. The sample size is the number of observed units, not the number of columns in the data file. A million transaction rows from ten customers do not constitute a million independently sampled customers. This distinction becomes especially important when observations share branches, officers or common outages.

2. Build a usable frame

The sampling frame is the actual list from which selection occurs. Reconcile it to the intended population before drawing numbers. Remove verified duplicate records, mark ineligible cases and identify missing branches. Each eligible unit should have a stable identifier. If closed accounts have disappeared from the available list, a sample from that list cannot directly describe all accounts active during the year. Random selection within an incomplete frame does not restore missing groups.

Record eligibility rules before inspecting outcomes. Excluding difficult cases after selection can change the population and bias the result. Keep selected units even if they are initially inconvenient to retrieve, and record nonresponse or unavailability. A replacement policy, if needed, should be specified in advance and preserve the intended design as far as possible. Distinguish an ineligible record from an eligible unit whose response is missing. The first indicates a frame problem, while the second can introduce nonresponse bias. A sampling log should retain the frame version, selection method, random seed or draw record, selected identifiers and completion status. This allows another analyst to reproduce the design without exposing unnecessary customer information.

3. Simple random sampling

In simple random sampling without replacement, every possible subset of the chosen size has the same probability of selection. Each population unit appears at most once. Number the eligible records and select distinct random identifiers, or randomly shuffle the full list and take the required number. A random start alone does not make every selection scheme simple random sampling. What matters is the probability structure of the complete sample, not whether one random number appeared somewhere in the procedure.

For a population of four files labelled A, B, C and D, there are six unordered samples of size two. They are AB, AC, AD, BC, BD and CD. Under simple random sampling each has probability one sixth. Each file occurs in three samples, giving inclusion probability one half. With replacement, the same file can be selected again and draws can be independent. Without replacement, selecting one file changes the probabilities for subsequent draws. The difference is negligible for some calculations when the sample is a tiny fraction of a large population, but becomes important as the sampling fraction increases. The finite population multiplier accounts for this reduction in variation for the sample mean.

4. Systematic sampling

Systematic sampling selects records at a fixed interval after a random start. If 1,000 eligible records are in a list and fifty are required, an interval of twenty is convenient. Select a random starting position from one to twenty, then take that position and every twentieth record thereafter. Starting at seven gives positions seven, twenty-seven, forty-seven and so on. This is operationally simple and spreads the sample across the list. It is not generally identical to simple random sampling because only certain combinations of records can occur.

Inspect the ordering for periodicity. If every twentieth record is produced by the same processing batch, systematic selection with interval twenty may repeatedly select that batch. Sorting can also help when it spreads relevant characteristics, but the implications should be understood. A list sorted by increasing balance may provide broad coverage, while a recurring pattern aligned with the interval can be problematic. When population size is not a multiple of sample size, use an explicitly defined systematic procedure rather than improvising the last few selections. For an examination case, identify the random start and fixed interval. For a real audit, document the ordering and any periodic structure that could affect representativeness.

5. Stratified sampling

Stratification divides the population into nonoverlapping groups and draws a probability sample within each group. Risk grades, regions or customer segments may serve as strata. Every unit belongs to one stratum. The design can ensure representation of small but important groups and improve precision when units within strata are relatively similar for the measured outcome. Stratification is especially useful when the overall population contains groups with different means and the group labels are available before selection.

Suppose routine applications constitute eighty percent of the population and specialist applications twenty percent. An audit takes equal sample sizes from each group to study specialists carefully. The observed group means are forty and seventy minutes. The population-weighted mean estimate is 0.8 × 40 + 0.2 × 70 = 46 minutes. The unweighted mean of the two equal-sized samples is fifty-five minutes and overrepresents specialists. For independent stratum samples, the variance of the weighted estimator is the sum of squared population weights times the variances of the stratum mean estimators. Proportional allocation follows population shares. Disproportionate allocation can serve precision or subgroup objectives, but appropriate weights remain essential for the overall estimate.

Population weights matterRoutine and specialist population shares are 80% and 20%. Equal sample counts do not change these population weights. Means 40 and 70 combine to 46.Population weights matterPopulation share (%)80Routine20Specialist
Routine and specialist population shares are 80% and 20%. Equal sample counts do not change these population weights. Means 40 and 70 combine to 46.

6. Cluster and multistage sampling

Cluster sampling selects groups of units, such as branches, and then observes all or some units within selected groups. In a one-stage cluster design, selected branches contribute every eligible application. In a multistage design, branches are selected first and applications are sampled within them. This can reduce travel and administrative cost. It differs from stratification, where each stratum is represented through its own sample. A cluster design may leave many clusters entirely unobserved.

Applications processed in the same branch may share staff, local demand and system conditions. Positive within-cluster correlation means that adding more applications from one branch supplies less independent information than adding applications from many branches. For equal cluster sizes m and intracluster correlation ρ, the approximation 1 + (m−1)ρ illustrates the design effect relative to independent sampling under suitable conditions. With m = 10 and ρ = 0.10, it is 1.9. This is an illustration, not a universal correction for every design. Survey variance should follow the actual selection scheme. A large total record count should not distract from a small number of independently selected branches. Planning must balance cost, coverage and precision.

7. Nonprobability samples and bias

Convenience sampling uses readily available units. Judgment sampling selects units considered informative. Quota sampling seeks specified category totals without necessarily using random selection within categories. These approaches can be useful for exploration, pilots or targeted investigations, but a conventional probability-sampling standard error is not automatically justified. Matching visible population proportions does not ensure that selected and unselected units are comparable on unobserved characteristics.

An online satisfaction survey may attract customers with unusually positive or negative experiences. Increasing responses reduces numerical instability within respondents but may leave self-selection bias unchanged. Similarly, surveying only customers who remain at the branch excludes those who abandoned a long queue. Bias is systematic departure associated with the method, while sampling variability describes differences between repeated samples under the design. A method can have low variability and substantial bias. Compare the target population with the actual observed group before interpreting precision. When the design is nonprobability based, describe results for the observed respondents and explain any additional assumptions needed to generalise. Do not manufacture a familiar margin of error merely because the sample size is large.

8. What a sampling distribution represents

A sampling distribution describes how a statistic varies over repeated samples selected using the same design. It is different from the distribution of individual observations. Suppose the population consists of processing times two, four and six minutes. Under sampling with replacement, samples of size two have nine equally likely ordered pairs. Their means are two, three, four, three, four, five, four, five and six. The average of these sample means is four, equal to the population mean.

The population variance, using denominator three, is eight thirds. The variance of the sample mean for two independent draws is four thirds, half as large. Individual processing times still take values two, four and six. Averaging changes the distribution of the statistic, not the underlying population. In practice, an analyst usually observes one sample rather than every possible sample. Mathematical results or simulation describe the repeated-sample behaviour that supports inference. Keep three concepts separate: population distribution, distribution within the observed sample, and sampling distribution of an estimator. Confusing them causes errors such as treating the standard error as the spread of individual customer waiting times.

9. Mean and standard error

For independent observations sharing population mean μ and finite variance σ², the sample mean has expected value μ and variance σ²/n. Its standard deviation is σ/√n, called the standard error of the mean. The formula follows because the variance of a sum of independent observations is the sum of their variances, and dividing the sum by n divides its variance by n². Independence is doing important work. Correlated observations require covariance terms as well.

If individual service times have standard deviation twelve minutes and n = 36, the standard error is two minutes. At n = 144 it becomes one minute. Quadrupling the sample size halves the standard error. Doubling the sample size multiplies it by approximately 0.707 rather than one half. The units of standard error remain minutes. When σ is unknown, s/√n estimates the standard error under the corresponding assumptions. A smaller estimated standard error indicates a more stable sample mean under the model, not that individual customers receive more consistent service. To describe consistency across customers, examine the standard deviation and other features of the individual outcome distribution.

Averages become more preciseWith population SD 12, independent mean SE falls from 6 at n=4 to 1 at n=144. Quadrupling n halves SE.Averages become more preciseSE (minutes)Sample size n003726144
With population SD 12, independent mean SE falls from 6 at n=4 to 1 at n=144. Quadrupling n halves SE.

10. Sampling from normal populations

If observations are independent and normally distributed with common mean μ and variance σ², their sample mean is normally distributed for every positive sample size. Its mean is μ and variance is σ²/n. No large-sample approximation is needed for that result. Standardising with known σ gives Z = (x̄−μ)/(σ/√n), which follows the standard normal distribution. The denominator is the standard error of the mean, not the individual standard deviation.

Suppose normal processing times have mean fifty minutes and standard deviation ten. For twenty-five independent applications, the sample mean has standard error two minutes. A sample mean of fifty-four is two standard errors above the population mean. The probability of a sample mean exceeding fifty-four is therefore the standard normal upper tail above two, approximately 0.0228. For one individual application, fifty-four is only 0.4 standard deviations above fifty, giving a much larger exceedance probability. If σ is estimated from a normal sample, the corresponding standardised mean uses a Student t distribution with n−1 degrees of freedom. Chapter A07 uses this distinction when constructing confidence intervals.

11. Non-normal populations and the central limit theorem

Many banking measurements are not normally distributed. Transaction values may have a long right tail, and counts are discrete. Under the usual independent, identically distributed finite-variance conditions, the central limit theorem says that the standardised sample mean approaches a standard normal distribution as sample size grows. The individual observations do not become normal. The approximation concerns repeated-sample averages and supports many large-sample calculations even when the original population is skewed.

There is no universal sample size that guarantees an adequate approximation for every distribution and purpose. Strong skewness, rare extreme values or dependence can slow convergence or invalidate the usual argument. Thirty is a common classroom guideline, not a mathematical guarantee. Consider a portfolio dominated by small loans with a few enormous exposures. A sample of thirty may frequently miss the large exposures and produce unstable average balances. Inspect the data-generating process and the sensitivity of the intended calculation. For proportions, the mean of zero-one indicators is the sample proportion. Its approximate normal behaviour requires enough expected successes and failures for the intended tail accuracy. Rare events often need methods beyond a simple normal approximation.

12. Finite population multiplier

When sampling without replacement from a finite population, observing a larger share of the population reduces remaining uncertainty. With population variance defined using denominator N, the standard error of a simple random sample mean is σ/√n multiplied by √((N−n)/(N−1)). This factor is the finite population multiplier. It approaches one when the sample is a very small fraction of the population and reaches zero at a census. Zero sampling error at a census does not imply zero recording error.

For N = 101 and n = 20, the multiplier is √(81/100) = 0.9. With population standard deviation ten, the adjusted standard error is 0.9 × 10/√20, approximately 2.012. State the variance convention because a related formula uses √(1−n/N) with the finite population variance defined using denominator N−1. When estimating that latter variance using the usual sample variance s², the estimated standard error is √((1−n/N)s²/n). These expressions are compatible when conventions are matched. Mixing a denominator-N population variance with the wrong correction produces a small but avoidable inconsistency. Apply the correction only when the finite population and without-replacement design are relevant.

13. A complete sampling plan

A bank wants the quarterly mean handling time and separate results for rural and urban branches. Define all completed eligible applications as the population and reconcile the frame to branch totals. Use strata that support the required comparisons. Within each stratum, randomly select applications and retain the selection record. Decide in advance how missing timestamps will be investigated. Record actual achieved sample sizes, not merely planned sizes. If the stratum sampling fractions differ, use population weights for the overall mean.

Suppose rural and urban applications account for thirty and seventy percent of the population. Their estimated means are sixty and forty minutes. The overall mean estimate is forty-six minutes. If the two independent stratum means have standard errors three and two minutes, the overall estimated variance is 0.3² × 3² + 0.7² × 2² = 2.77. Its standard error is approximately 1.664 minutes. Averaging the two standard errors gives the wrong result. The variance calculation uses squared weights because uncertainty is measured through squared deviations. A final report should distinguish design-based uncertainty from potential measurement and nonresponse problems, so precision is not mistaken for a guarantee of complete accuracy.

14. Diagnose an answer before accepting it

For any sampling question, first identify whether selection is with or without replacement and whether observations can reasonably be treated as independent. Next identify the statistic and the population variance convention. Calculate the uncorrected standard error, then apply a finite population correction if appropriate. Check units and limiting cases. A larger sample should reduce standard error under the same model. A complete census should remove finite-population sampling variation. A claimed correction greater than one is inconsistent with the usual simple-random-sampling finite multiplier.

Try a comparison between a sample of one hundred applications from one branch and one application from each of one hundred randomly selected branches. Neither design is universally superior without the population structure and cost objective. If branch effects are strong, broader branch coverage can provide more information about the bank-wide mean. If the target is that one branch, the first design addresses a different and legitimate question. Good sampling begins with the target, then chooses a design, then computes uncertainty. Memorising the square-root formula without this sequence makes it easy to produce a precise answer to the wrong question. Use the interactive standard-error exercise to see the effect of size and sampling fraction separately.

Guided exercise: compare two selection designs

A bank has two hundred eligible files with population standard deviation fifteen minutes under the denominator-N convention. It selects fifty files by simple random sampling without replacement. The uncorrected standard error is fifteen divided by the square root of fifty, approximately 2.121 minutes. The finite population multiplier is the square root of 150 divided by 199, approximately 0.8682. Multiplying gives an adjusted standard error of about 1.842 minutes. The correction is material because one quarter of the finite population is observed.

Now imagine fifty independent draws with replacement from the same population. The uncorrected standard error applies because a file can recur and observing it once does not remove it from later draws. Do not apply the without-replacement multiplier merely because the population has a known size. The selection mechanism determines the formula. Finally, suppose the analyst takes the first fifty files sorted by completion time. Neither standard error establishes valid random-sampling inference for that selection. The sample deliberately favours short times. The exercise shows three separate decisions: identifying the variance convention, identifying replacement, and checking whether selection is probabilistic. Correct arithmetic begins only after these design questions are settled. If the target changes from this finite batch to future applications, the uncertainty question changes again because observing the current batch does not reveal future outcomes exactly.

Open the interactive chapter laboratory and offline practice

LEARN THROUGH EXPLANATIONS

Chapter practice

This practice is separate from timed mock tests. Answers and explanations appear after you choose an option.

Your reading progress

Sign in to save your progress across devices.

Browse published timed tests for this paper