The Normal Distribution and Confidence Intervals

Introduction

Replicate results vary, so a sample mean is not known with infinite precision. A confidence interval uses the observed spread, sample size and a statistical model to express a range of plausible values for the population mean. The normal distribution often supports the calculation, but real analytical data can have skew, outliers or drift. The interval is only as useful as its assumptions and sampling design.

Core explanation

The normal distribution is symmetric around a mean and described by a standard deviation. Many small independent effects can yield approximately normal measurement errors, but that is a model to examine, not a law that every chemistry result obeys. Repeated observations from a stable measurement process may cluster in a bell-shaped pattern. Gross mistakes, detection limits or heterogeneous populations can produce non-normal behaviour.

When population standard deviation is unknown, a two-sided confidence interval for a mean under suitable independent approximately normal observations is x̄ ± t s/√n, where t is a critical value from Student's t distribution with n − 1 degrees of freedom. The t factor is larger for small n because s is estimated from limited data. As n grows, it approaches the corresponding normal critical value. At 95% confidence, the procedure is constructed so that over many independent repetitions about 95% of intervals would cover the fixed population mean under its assumptions.

Once a particular interval is calculated, the underlying fixed mean either lies in it or does not; frequentist 95% confidence refers to the procedure's long-run coverage, not a 95% probability that a fixed mean moves into this one interval. This distinction can feel subtle, but it prevents overstating what data alone prove.

Interval width grows with s and shrinks roughly as 1/√n. Doubling the number of replicates does not halve the uncertainty of the mean; roughly four times as many are needed for that simple scaling, if measurements are independent and variability is stable. More measurements do not remove a shared systematic bias. The interval estimates a mean of the process sampled, not necessarily the composition of an unrepresentative field population.

A confidence interval and a prediction interval answer different questions. The former estimates uncertainty in the mean; a future single result is more variable and generally needs a wider prediction interval. Reporting x̄ ± s as a “95% confidence interval” without the t factor and √n is incorrect unless some separately justified convention is explicitly being used.

Step-by-step reasoning

1. Define the population or measurement process whose mean is sought. 2. Check independence, stability and plausibility of the distribution model. 3. Compute x̄, sample s and n, giving n − 1 degrees of freedom. 4. Choose confidence level and matching two-sided t critical value. 5. Calculate x̄ ± t s/√n and interpret it with its assumptions and units.

Visual explanation

Draw a bell curve of possible measurement results centred on a fixed mean. Then draw several intervals from repeated imagined samples; most cross a vertical mean line and a few miss it. Shade one interval to show that “95%” describes the repeated method, not that the mean itself is randomly moving from sample to sample.

Real-world analogy

Imagine repeatedly casting a net of a fixed design around a stationary marker. A well-designed 95% net catches the marker in about 95 of 100 repeated casts; a particular cast either catches it or misses. A confidence interval works similarly under its model, while a biased sampling process is like casting all nets at the wrong lake.

Real-world example

A laboratory measures a stable reference solution in five independent preparations. It reports a mean and 95% interval for the method's mean response under those conditions. If the interval excludes the certified reference range, bias is worth investigating, but the comparison must consider the reference value's uncertainty and whether the sample matrix is comparable.

Why?

Why use t rather than a fixed normal multiplier for a small sample? The population spread is unknown and estimated by s. That extra uncertainty makes extreme standardised deviations more common, so the t distribution has heavier tails. With more degrees of freedom, s is better determined and the distinction shrinks.

Common misconception

“A 95% confidence interval contains 95% of individual observations” confuses a mean interval with data spread. Another mistake is believing additional replicates eliminate bias; an interval around a biased mean can become very narrow while remaining centred at the wrong value.

Worked example

Five independent results have x̄ = 10.0 mg L⁻¹ and sample s = 0.50 mg L⁻¹. For a two-sided 95% interval with four degrees of freedom, t is about 2.776. The margin is 2.776 × 0.50/√5 ≈ 0.62 mg L⁻¹, giving approximately 9.38 to 10.62 mg L⁻¹. The arithmetic assumes stable independent approximately normal results and addresses the measurement-process mean, not any untested sampling bias.

Quick check

1. Does a narrower confidence interval after more repeated injections show that extraction recovery is unbiased? Answer: No. It shows greater precision for the mean injection response under that setup. A shared extraction loss remains outside what repeated injections can diagnose.

Exam focus

Use n − 1 degrees of freedom and the correct t critical value for a small-sample mean interval. State the long-run coverage interpretation and assumptions. Distinguish standard deviation of individual results, standard error of the mean and interval half-width.

Advanced insight

If data arise in batches or along a time trend, observations may be correlated. Then s/√n can understate uncertainty because n readings do not provide n independent pieces of information. A hierarchical or time-series analysis, or deliberately randomised independent preparations, better matches the actual structure.

Summary

Confidence intervals use replicate mean, spread, count and a statistical model to estimate a process mean. For unknown variance and suitable data, x̄ ± t s/√n gives a stated long-run coverage. The interval does not include unrecognised bias, poor sampling or every future observation.

Practice questions

1. What degrees of freedom are used for s and a one-sample t interval from n = 8 results? Answer: n − 1 = 7 degrees of freedom.

2. If s stays constant, approximately how many times more independent results are needed to halve standard error? Answer: Four times as many, because standard error scales approximately as 1/√n.

3. Does a 95% confidence interval mean there is a 95% frequentist probability that its fixed population mean changes into the interval? Answer: No. The procedure has about 95% coverage over repeated independent samples under its assumptions; a calculated interval either contains the fixed mean or it does not.