1. A number about a population
A bank has processed 24,000 loan applications this quarter. The operations head wants to know the mean time between receiving a complete application and issuing a decision. A full timestamp audit would be expensive, so the team draws a sample. Estimation turns the sample into a statement about the population and quantifies one source of uncertainty in that statement. Begin by fixing the population, the unit of observation and the outcome. Here the unit is an application, the outcome is elapsed processing time, and the population is the 24,000 qualifying applications. Measuring time since first enquiry would answer a different question.
Write μ for the population mean and x̄ for the sample mean. The parameter μ is fixed for that defined population, even though the bank does not know its value. The sample statistic x̄ changes when the sample changes. The rule “add the observations and divide by their number” is the estimator. Its numerical result in one particular sample, such as 48 minutes, is the estimate. Similarly, the rule “divide the number of late transactions by the number inspected” estimates the population proportion p. If 80 of 400 payments were late, the resulting estimate p̂ is 0.20.
2. What makes a useful estimator
Imagine drawing many random samples under exactly the same design. An unbiased estimator has an average across those repetitions equal to the target parameter. It can still miss badly on a particular sample. For independent observations with a common mean μ, the expected sample mean equals μ because expectation is linear. If the team chooses only the fastest branches, the expected value under that selection mechanism need not be μ. Arithmetic correctness cannot repair a selection problem.
Precision describes how much the estimator varies from sample to sample. An estimator with lower variance is more precise under the stated design. Consistency means that the estimator converges in probability to the parameter as sample size grows under suitable assumptions. These ideas answer different questions. A small sample mean can be unbiased but imprecise. A rule that shrinks a small sample estimate toward a reference value can have bias yet lower average squared error. Mean squared error combines variance and squared bias. If an estimator has variance 4 and bias 1, its mean squared error is 5. Compare estimators for the same target and under the same assumptions.
3. Spread of observations versus spread of estimates
Suppose individual processing times have a population standard deviation of ten minutes. That number describes variation among applications. Under independent sampling, a sample of 100 has a sample-mean standard deviation of ten divided by the square root of 100, or one minute. The second quantity is called the standard error. Averaging dampens fluctuations because long and short individual times can offset each other.
The general independent-sample formula is SE(x̄) = σ/√n. If σ is unknown, use s/√n as an estimated standard error when the inference procedure supports that substitution. The units remain minutes. Squaring gives variance in squared minutes. Quadrupling n halves the standard error under the same variance and independence assumptions. Doubling n multiplies it by about 0.707. If the audit includes multiple applications processed by each officer during the same outage, positive dependence can inflate the true standard error. Treating all these applications as independent can make an interval deceptively narrow.
4. From a point to an interval
A point estimate gives one plausible value. An interval reports a range constructed by a specified procedure. Many large-sample intervals have the form estimate ± critical value × standard error. The critical value determines how much of the reference sampling distribution is covered. Under a standard normal model, a central 95% region leaves 2.5% in each tail and uses approximately 1.96. A central 90% region uses about 1.645 and a central 99% region uses about 2.576.
Suppose 100 eligible applications yield x̄ = 48 minutes and σ = 10 minutes. A two-sided 95% normal interval is 48 ± 1.96 × 10/√100, which gives 46.04 to 49.96 minutes. The lower endpoint subtracts the margin of error and the upper endpoint adds it. Under the stated design and model, the procedure would cover the fixed population mean in 95% of repeated applications. In frequentist inference, the realized endpoints are fixed once observed and the population mean is fixed as well. The coverage statement applies to the procedure.
5. What the interval is for
The range 46.04 to 49.96 minutes concerns the mean, which is one population parameter. An individual application can take 80 minutes and coexist with that interval. Individual prediction requires an allowance for individual variation. A service target that 95% of applications finish within an hour is a question about the distribution of individual times or a population proportion, rather than the precision of the mean.
Confidence level and precision trade off. Using the same data, the 99% interval is 48 ± 2.576, or approximately 45.424 to 50.576 minutes. It is wider because it offers higher repeated-sample coverage. Report the unit, the estimate, the endpoints, the confidence level and how cases entered the sample. An interval alone cannot tell a manager whether a delay is costly enough to act on. A narrow interval for a difference of 0.2 minutes could be statistically precise but operationally minor. Conversely, a wide interval spanning substantial savings and losses can justify more evidence before an expensive process change.
6. Unknown variability and choice of reference distribution
If individual times are independent and normally distributed but the population standard deviation is unknown, replacing σ by s introduces additional uncertainty. The classical one-sample interval then uses the Student t distribution with n−1 degrees of freedom. For n=16, x̄=48 and s=8, the estimated standard error is 2. A two-sided 95% critical value near 2.131 gives endpoints 43.738 and 52.262. The t multiplier is larger than 1.96 because the sample also estimates the variance. The degrees of freedom follow from using one fitted mean when computing sample deviations.
For larger samples, t multipliers approach the corresponding normal multipliers. An examination question that explicitly requests a large-sample normal approximation with s as a substitute uses the supplied z multiplier. The population itself need not be normal for the large-sample mean approximation, but the sampling distribution must be sufficiently close for the intended inference. Strong right skew, extreme outliers or dependence warrant checking that assumption. Avoid a universal threshold that treats every sample of a given size as adequate. Inspect the data and the process that produced it.
7. Sampling from a finite frame
An audit of a fixed set of 101 branches randomly selects 20 without replacement. Taking one branch removes it from the pool. Under simple random sampling, the exact variance of the mean is reduced compared with independent sampling with replacement from that frame. If σ² is defined with denominator N, the standard error multiplies σ/√n by √((N−n)/(N−1)). Here the multiplier is √(81/100)=0.9. With σ=10 and n=20, the standard error is about 2.012. If population variance is instead defined with denominator N−1, the equivalent formula uses √(1−n/N) with that variance convention.
For a census of the fixed frame, sampling error is zero because every unit is included. Measurement and coverage errors can still occur. The correction matters when the sampling fraction is substantial. It is close to one when n is very small relative to N. Before applying it, determine whether the target is this fixed frame or a broader process generating future observations. A census of last month does not eliminate uncertainty about next month. Also check that the sampling design is simple random sampling without replacement. A cluster or stratified design generally needs a variance estimate suited to its own selection and weighting rules.
8. Estimating a proportion
Encode a late payment as one and an on-time payment as zero. The sample mean of these indicators is the sample proportion. With 80 late payments among 400 independently selected payments, p̂=80/400=0.20. Under a Bernoulli model, the plug-in standard error is √(p̂(1−p̂)/n), which here is √(0.20×0.80/400)=0.02. Using the two-sided 95% normal multiplier gives a margin of 1.96×0.02=0.0392 and endpoints 0.1608 to 0.2392. In percentage form this is 16.08% to 23.92%. The margin is 3.92 percentage points. Multiplying 20% by 1.0392 would instead compute a relative change and gives the wrong endpoint.
This large-sample interval is often called the Wald interval. Its normal approximation depends on having enough events and nonevents and an appropriate sampling design. Suppose a branch sees zero complaints in 20 observations. The plug-in formula would produce standard error zero and a degenerate interval at zero. The sample still provides limited evidence about rare complaints. A Wilson or exact binomial interval can handle such cases more sensibly. In a classroom question, follow the requested method and flag the limitation when the inputs make it questionable. A negative lower endpoint also signals approximation problems. Clipping an endpoint does not restore the nominal coverage.
9. Planning sample size
For a normal mean interval with known σ, choose a target margin E and solve zσ/√n ≤ E. Thus n ≥ (zσ/E)², rounded upward. If the audit wants a margin of two minutes at 95% confidence and expects σ=12 minutes, it needs n ≥ (1.96×12/2)²=138.2976, so plan at least 139 completed observations. Rounding to 138 falls below the requested precision. If the plan anticipates 10% unusable records, drawing 155 records gives approximately 139.5 usable records in expectation. Actual usable count can be lower, so monitor collection and plan reserve capacity.
For an independently sampled proportion, a planning approximation is n ≥ z²p₀(1−p₀)/E², where p₀ is a preliminary proportion from comparable data. If no preliminary estimate exists, 0.5 maximizes p₀(1−p₀) and gives a conservative planning size within this formula. For a margin of five percentage points at 95%, n ≥ 1.96²×0.25/0.05²=384.16, so choose 385. Use 0.05 for five percentage points. If the sample will be clustered or weighted, a simple independent-sample calculation may understate required effort. A pilot can estimate variance, response rate and field costs.
10. Read a banking case before choosing the formula
A service manager asks whether average processing time exceeds 50 minutes. The audit reports a two-sided 95% interval of 46.04 to 49.96 minutes under the independent-sample assumptions. The entire interval lies below 50, giving evidence that the population mean is below 50 at the corresponding two-sided test level. If the upper bound were 50.4, the observed mean could still be below 50 while the data were too imprecise for that conclusion. A value inside the interval remains compatible with the data and model at that level. It does not establish exact equality.
A payments manager asks whether the late share is above 15%. The sample proportion is 20% and the two-sided interval from Section 8 starts at 16.08%. Under the model this is evidence of a share above 15%. Before changing policy, confirm that the sample includes all relevant channels. A data feed omitting failed and retried payments can make the interval precise for the wrong population. When the decision rule specifically uses a one-sided bound, use the corresponding one-sided critical value. A one-sided 95% normal bound uses about 1.645 and differs from the 95% two-sided interval.
11. Reconcile design, measure and reporting
Consider a bank with 80% routine applications and 20% specialist applications. The audit intentionally takes equal sample sizes from each group to study specialists more closely. Suppose estimated group means are 40 and 70 minutes. A simple average of the group means is 55 minutes because the sample composition is balanced. For the overall population mean, weight by the population shares. The estimate is 0.80×40+0.20×70=46 minutes. The difference is material for staffing. For independent stratum samples the variance of the weighted mean is the sum of the stratum estimator variances multiplied by their squared weights. Apply finite population corrections within strata when the design calls for them.
The same logic applies to proportions. If the ordinary channel has late rate 5% and specialist channel has late rate 20%, the overall estimate is 0.8×0.05+0.2×0.20=0.08, or 8%, for that population mix. Distinguish the channel-specific question from the overall question. Keep a note of the population counts used for the weights. If the mix changes, the overall estimate can change even when the within-channel rates are stable. In a dashboard, display the denominator alongside the rate and interval so the reader can see the scope and sampling assumptions.
12. A worked audit from question to decision
A regional team receives a file of 5,000 completed service requests and draws 100 at random without replacement. It defines turnaround as business minutes between two validated timestamps. The sample mean is 32 minutes and the sample SD is 15 minutes. The uncorrected estimated standard error is 1.5 minutes. Under the sample-variance convention, the finite population factor is √(1−100/5000)=√0.98, approximately 0.98995. The adjusted standard error is about 1.485 and the large-sample 95% margin is about 2.91 minutes. The resulting interval is approximately 29.09 to 34.91 minutes.
The team finds 12 of the sampled requests breached a separate service target. The estimated breach rate is 12%. The independent-sample plug-in standard error is √(0.12×0.88/100), or about 0.0325 before any finite population adjustment. That spread is measured in proportion units, so 0.0325 is roughly 3.25 percentage points. The margin for this proportion and the margin for the mean answer separate questions. A manager who reads a mean of 32 as evidence that every request meets a 40-minute target can miss a significant minority of breaches. Report both estimates, with their denominators and the scope of the sampling frame.
13. Use an answer check before moving on
Start with the requested parameter. A mean has the unit of the underlying measurement. A proportion lies between zero and one. Identify the design and the relevant standard error, then choose the confidence level and a compatible reference distribution. Calculate the margin before constructing endpoints and round only after the arithmetic. Finally, state what the result estimates in the original banking language. If the interval is for a proportion, convert both endpoints to percentages consistently. If the question asks for a planning size, round the result upward.
Try three variations on the 48-minute example. If n increases from 100 to 400 at the same SD, the standard error falls from one minute to half a minute and the 95% margin falls from 1.96 to 0.98. If the confidence increases from 95% to 99% at n=100, the margin rises to 2.576. If the observed mean rises to 49 and the standard error stays one, both endpoints move up by one. These comparisons isolate precision, coverage and location. Use them to spot mistakes such as dividing by n instead of √n, using a central 95% multiplier for a 95% one-sided bound, or assigning individual observations the standard error of a mean.
Reading an interval as a decision tool
An interval can be narrow yet unsuitable for the decision if it estimates the wrong population. Suppose an audit samples only approved loans but the manager asks about processing time for all applications. Rejected and withdrawn applications may follow different paths. The confidence calculation describes sampling uncertainty within the observed frame. It does not automatically extend to excluded groups. Define the target before collecting data and carry that definition into the final sentence. A report that says “mean time among completed eligible applications this quarter” is more informative than one that simply says “average time”.
Distinguish statistical precision from practical importance. An enormous sample can make a very small difference statistically detectable, although the difference may have little effect on staffing or customer experience. Conversely, a large estimated difference may have a wide interval because the data is limited. A manager should consider the plausible range, the cost of acting and the cost of waiting for more evidence. The interval helps organise that judgement. It does not supply the business threshold by itself. State any service target as an illustrative assumption when it is invented for a teaching case.
A second worked precision comparison
Two teams estimate the same mean under independent sampling with the same population SD of twenty minutes. Team One observes one hundred applications and Team Two observes four hundred. Their standard errors are two and one minute respectively. At 95% confidence the normal margins are 3.92 and 1.96 minutes. If both sample means happen to be fifty, the intervals are 46.08 to 53.92 and 48.04 to 51.96. The larger sample narrows the interval around the same point estimate. It does not imply that individual processing times have become less variable.
Now keep the first sample size but increase confidence from 95% to 99%. Using 2.576 gives a margin of 5.152 minutes. The wider interval reflects a procedure designed for higher long-run coverage. It is incorrect to claim that greater confidence makes the estimate more precise. Precision and coverage must be balanced. If management wants both higher confidence and a smaller margin, additional observations or a more efficient valid design may be needed. Reducing the margin from four minutes to two at unchanged confidence requires approximately four times the sample size under the same variance assumptions.
Finally, check a proportion boundary case. If none of twenty inspected files has a defect, the plug-in Wald standard error is zero because the observed proportion is zero. A zero-width interval would misleadingly suggest that the population defect rate is known exactly. The data only says that no defects were observed in this small sample. An interval method designed for boundary proportions is needed. This example explains why checking formula conditions is part of solving the problem rather than a separate theoretical exercise after the numerical answer.
Mean and proportion intervals
Independent normal/large-sample illustration. Mean and proportion use the same n. The proportion uses a Wald interval, with a warning for small cell counts.
Practice and review
25 questions. One answer per question, no negative marking. Explanations remain readable without JavaScript.
Local browser storage only; no account synchronisation.
Ready: 25 questions.
References and scope
- IIBF CAIIB Rules & Syllabus 2026, PDF page 5: Official Module A scope; A01 to A09 are Clerk2Officer editorial identifiers.
- NIST confidence limits for a mean: Mean interval and repeated-sample confidence interpretation.
Original exposition, cases and questions. Formula conventions are stated in the lesson. Automated checks support arithmetic and functionality; they are not independent academic certification.