Inference for Quantitative Data: Means

A mean summarizes quantitative observations. When the population standard deviation is unknown, the sample standard deviation estimates it. A paired response is one numerical difference per matched pair.

Define μ as the population mean of the measured quantity. The sample mean estimates μ, and the sample standard deviation estimates individual-level variation. The standard error s√n describes the estimated variability of sample means, not the variability of individual observations.

When the population standard deviation is unknown, t procedures account for estimating it. A t distribution is centered at zero with heavier tails than the standard normal, especially at small degrees of freedom. For a one-sample procedure, degrees of freedom are n − 1.

Random sampling or appropriate random assignment supports the inferential model. When sampling without replacement, the 10% condition is commonly used for approximate independence. Examine quantitative data for strong skew and outliers, especially for small samples.

A nearly normal population supports small-sample t methods. With larger samples, the sampling distribution of the mean can become more nearly normal, but extreme outliers, severe heavy tails, or dependence still require care. A large n does not repair selection bias or turn a time-dependent series into independent observations.

A t interval estimates plausible values for μ using x̄ ± t*s√n. A test compares the observed mean with a specified null mean using t = (x̄ − μ₀)(s√n). The alternative determines which tail or tails define the P-value.

Interpret an interval for the population mean, not individual outcomes. A narrow interval can estimate the mean precisely even while individuals vary widely. If the scientific question concerns individual prediction rather than a mean, a confidence interval for μ is not the correct answer.

Before-and-after measurements on the same individuals or deliberately matched pairs are dependent within pairs. Define a difference direction, calculate one difference per pair, and perform a one-sample t procedure on those differences. The relevant standard deviation is the standard deviation of differences.

Check distribution conditions for the differences, not separately for each original measurement set. Pairing can reduce variation by removing stable individual differences, but it does not guarantee a causal conclusion if time trends or other changes are uncontrolled. Randomized order or a suitable control may be needed.

For independent groups, estimate μ₁ − μ₂ with x̄₁ − x̄₂. The standard error combines s₁²n₁ and s₂²n₂ under a square root. The usual unpooled two-sample t procedure does not assume equal population variances. Technology may use approximate Welch degrees of freedom.

Keep the group order consistent in hypotheses, calculations, and conclusions. An interval containing zero does not prove the means are identical; it indicates that zero remains compatible with the interval procedure at that confidence level. A meaningful effect can be missed by a low-power small study.

A small P-value can occur for a very small effect with a large sample, while a large practical effect may be uncertain in a small sample. Report the estimated difference and units as well as evidence against the null. Distinguish no evidence of an effect from evidence that any effect is practically negligible.

A complete response names the procedure, checks conditions using the study, shows the relevant calculation or technology output, and concludes about the population in context. Avoid a generic ending such as “reject H₀” with no explanation of what that means for the original question.

Estimating population spread introduces extra uncertainty. A t procedure accounts for this with heavier tails than a standard Normal distribution, especially at small degrees of freedom. For a one-sample mean, degrees of freedom equal n−1. The standard error is s divided by the square root of n, not s divided by n.

Check randomization and independence, then examine sample size, shape, and outliers. With small samples, strong skewness or outliers can make a t procedure unreliable. A large sample improves robustness to non-Normal shape, but it does not fix dependence or biased data collection. Include the data’s context and units in the conclusion.

Measurements before and after on the same person are paired, as are deliberately matched subjects. Define one difference direction, such as after minus before, and compute a difference for each pair. Analyze those differences with a one-sample t procedure. The sample size is the number of pairs, not the total number of measurements.

Conditions concern the distribution of differences. Even if the two raw measurement distributions are skewed, their differences may behave differently. Reversing the subtraction changes the sign of the estimate, test statistic, and directional hypothesis; it does not change a two-sided p-value.

Two independent groups require a two-sample method. The standard error of their mean difference combines the two estimated variances divided by their sample sizes. Do not treat two independent groups as pairs merely because the sample sizes match. Conversely, discarding genuine pairing can lose useful information.

State whether evidence concerns an association or a causal effect based on the design. Report an estimated difference and uncertainty, not just “significant.” An interval from 0.1 to 0.3 minutes may exclude zero but represent a small practical gain. A wide interval can include both valuable and negligible effects, revealing a need for greater precision.

A quantitative response calls for mean inference when the target is a population average. If each subject contributes before and after measurements, analyze the within-subject differences with a one-sample t procedure. State the subtraction order so that the sign has a clear meaning.

Two independent groups require a two-sample procedure. Equal sample sizes do not create pairing, and unequal sample sizes do not by themselves invalidate independent-group inference. Pairing comes from the design and relationship between observations. Treating paired measurements as independent discards their dependence and changes the standard error.

The t model accounts for estimating population standard deviation from sample data. Degrees of freedom depend on the procedure; for one-sample or paired differences, df=n−1 where n is the number of independent observations or pairs. Use the appropriate software output for a two-sample method rather than inventing a single common n.

Inspect quantitative data for strong skewness and outliers, especially with small samples. A large sample can make mean inference more robust, but cannot repair dependent observations or biased sampling. For paired data, examine the distribution of differences, not merely the two separate measurement distributions.

Standard error s√n measures uncertainty in a sample mean, while s measures variation among individual observations. A confidence interval for a mean does not predict where most individuals fall. Keeping these roles separate prevents a common error in interpreting a narrow interval as evidence that all observations are similar.

A one-sample test compares the sample mean with a null mean using the estimated standard error. The alternative determines the tail direction. Choose that direction from the question before inspecting the result; changing it after seeing the data changes the evidential meaning of the procedure.

A small P-value indicates incompatibility with the null model under the assumptions, not necessarily a large or useful effect. With a very large sample, a tiny difference can be statistically detectable. Report the estimated difference and a confidence interval to discuss practical importance.

A Type I error rejects a true null; a Type II error fails to reject a false null. Power is the probability of rejecting for a specified true alternative. Increasing sample size generally improves power for the same effect and significance level, but the practical meaning depends on which effect matters.

For a study conclusion, separate the numerical inference from the scope of inference. Random assignment can support causality; random sampling can support generalization. A statistically convincing difference from an observational sample still may be confounded.