Descriptive vs inferential statistics: the difference with examples
Descriptive statistics summarise observed data; inference makes claims beyond them. Compare examples, confidence intervals and the assumptions behind each.
Descriptive statistics summarise the observations you have. Inferential statistics use observations to make claims about a population or data-generating process, with an account of uncertainty. The difference lies in the claim you want to make, not in whether you use a spreadsheet, Python or a sophisticated model.
A mean can serve either purpose. Reporting the mean spending of survey respondents describes those respondents. Using that mean to estimate spending among all households requires an inferential argument about how the respondents relate to the target population.
The difference at a glance
| Feature | Descriptive statistics | Inferential statistics |
|---|---|---|
| Target | The observed dataset | A population, process or parameter beyond the observed records |
| Typical question | What did these respondents report? | What does this sample suggest about all eligible households? |
| Typical outputs | Counts, proportions, means, medians and charts | Estimates, standard errors, confidence intervals and tests |
| Main checks | Definitions, measurement, coverage and calculation | All descriptive checks plus sampling, model and dependence assumptions |
| Common mistake | Treating an average as a complete description | Treating a precise estimate as protection against selection bias |
Descriptive analysis still requires care. A perfectly calculated average can describe the wrong quantity if currencies, periods or units have been mixed together.
An example using a household survey
Imagine a hypothetical survey of 400 households about monthly electricity spending. Suppose their reported mean is 80 currency units. That statement describes the 400 responses. A table could also show the median, the share with missing bills and how spending varies by household size.
Now consider the statement: “Mean monthly spending among every household in the city is about 80.” It extends beyond the observations. To defend it, ask how the sample was selected, who did not respond and whether survey weights are needed. A survey posted in an energy-saving forum may attract households unlike the rest of the city.
Collecting 40,000 responses from the same selective source can make the reported sample average stable while leaving the population estimate biased. Sample size and sample quality address different problems.
Standard deviation and standard error answer different questions
The standard deviation describes how much individual observations vary. The standard error describes uncertainty in an estimator under a specified sampling or model framework. They can have the same units while answering different questions.
For an arithmetic teaching example, suppose the 400 invented spending observations have sample standard deviation 30. Under independent random sampling from a large population, the estimated standard error of the mean is approximately 30 divided by the square root of 400: 1.5. The household variation is 30; the uncertainty in the estimated population mean is much smaller.
This calculation is an illustration, not a result from a real survey. If households were sampled within neighbourhoods, shared local conditions could make their responses dependent. A design-based or clustered calculation may then be needed. The robust and clustered standard errors guide develops that issue for regression.
What a confidence interval adds
Using the illustrative standard error above, a large-sample 95% interval is approximately 80 plus or minus 1.96 times 1.5, or 77.1 to 82.9. It communicates sampling uncertainty under the assumptions used to construct it.
A frequentist 95% interval comes from a procedure intended to cover the fixed population parameter in about 95% of repeated samples under its assumptions. It does not mean that 95% of household bills lie inside the interval, nor does it repair a biased sampling frame. NIST explains this repeated-sampling interpretation in its confidence-limits reference.
An interval for the population mean is also different from an interval predicting an individual household's spending. Individual outcomes generally vary much more than an estimated average.
Where hypothesis tests fit
A hypothesis test asks how compatible the observations are with a specified claim and model. For example, a city might compare the population mean with a benchmark specified before looking at the survey.
The resulting p-value depends on the benchmark, the test and the uncertainty calculation. It does not measure whether a policy is useful or whether the estimate is economically important. Read p-values in economics before reducing a result to a significance star.
Which approach should you use?
Begin with description in either case. Then write the exact sentence you want to support. If it refers only to the records observed, a transparent descriptive analysis may be enough. If it refers to other people, future outcomes or an underlying relationship, explain what permits that extension.
Practise summaries in the Statistics centre lab and uncertainty in the confidence intervals lab. The central habit is to keep the observed number, the target quantity and the assumptions connecting them separate.
References
- Illowsky, B. and Dean, S. OpenStax, Introductory Statistics 2e. Definitions of statistics, probability and key terms
- NIST/SEMATECH. Confidence limits for the mean