empirio.ai

Descriptive Statistics & Inferential Statistics: Guide

Descriptive statistics tells you what is in your data. When that is enough, when you need to go further, and how to keep it apart from inference.

Author at empirio.ai - Maria Malzewby Maria MalzewUpdated 5 November 2026Reading time 18 min

Your questionnaire is back and 87 rows are sitting in a spreadsheet. Now what? Most people reach straight for the mean without first checking whether these answers allow a mean at all.

Descriptive statistics summarises the cases you actually collected, using measures such as the mean, the median and the standard deviation. Inferential statistics goes one step further and uses that sample to estimate what is probably true of the whole population, together with a statement of how uncertain the estimate is. By the end you will know which measure your data allow and when you can skip the second step entirely.


📌 Key points at a glance

  • Descriptive statistics describes the sample, inferential statistics estimates the population.
  • The mode works at every level of measurement, the mean only from interval upwards.
  • Describing only your own group needs no significance test.
  • An outlier drags the mean noticeably, the median barely moves.
  • The p-value does not say how likely your hypothesis is.

Create a survey for free

With empirio.ai you can create a modern online survey in minutes — with hosting in the EU.

  • AI-built survey
  • Adjust by drag & drop
  • Real-time analysis
Start for free

What is descriptive statistics?

Descriptive statistics is the branch of statistics that summarises and presents an existing data set without going beyond it. Every statement applies to the cases you collected, and to not a single case beyond them.

That restriction sounds modest, yet it carries half the analysis. Nobody learns anything from 200 rows of raw data. Only when those rows become a handful of figures and one chart does a picture emerge of where the middle lies, how widely the answers spread and whether two variables move together. Those three questions are exactly what the three groups of measures answer.

Presentation belongs to descriptive statistics as well: frequency tables, bar and pie charts, histograms, box plots. The form follows the point you want to make, not the first option in your software menu. Nicola Döring places data analysis in the same order in her textbook Forschungsmethoden und Evaluation in den Sozial- und Humanwissenschaften (6th edition, Springer 2023, chapter 12): describe first, test second. For the wider process, see the overview of empirical research.

Descriptive and inferential statistics: the difference

The difference between descriptive and inferential statistics is a question of reach. Descriptive statistics only claims something about the people who answered. Inferential statistics claims something about everyone those people are meant to represent, and quantifies its own uncertainty while doing so.

Inferential statistics is therefore also called inductive statistics. It does two things: it estimates population values, usually as a confidence interval, and it tests hypotheses about differences or relationships. Any data set can be described. Only a sample drawn by a proper selection procedure supports inference.

Tip

You don't have to work out the confidence interval around a percentage from your survey by hand. Enter your sample size and result in our calculator to calculate the margin of error. With 42 per cent agreement from 500 responses, that gives you an interval of roughly 38 to 46 per cent.

Both halves depend on each other more than they first appear to. Without a clean description you will not notice that a distribution is heavily skewed, that a single value distorts everything, or that a comparison group contains three people. All three findings decide which test is admissible at all. That is why description opens the results chapter.

The measures used in descriptive statistics

The measures used in descriptive statistics fall into three groups: central tendency for the middle, dispersion for the spread, association for the relationship between two variables. Reporting only one group leaves half the analysis undone.

One example runs through this chapter. You asked 15 people on your course how many hours a week they spend writing their dissertation. Sorted, the answers are 2, 3, 3, 4, 4, 4, 5, 5, 6, 6, 7, 8, 9, 10, 20. One person stands out, and that single answer demonstrates every measure.

Central tendency: mode, median and mean

Measures of central tendency say where the middle of a distribution lies. The mode is the most frequent value, here 4 hours, which appears three times. If two values occur equally often there are two modes, and if every value differs the mode says nothing at all. The median sits in the middle of the ordered series, so with 15 values it is the eighth: 5 hours. The mean is the total divided by the count, here 96 divided by 15, so 6.4 hours.

Three measures, three different answers to one question. The value of 20 is responsible. Remove it and the mean falls from 6.4 to 5.4 hours while the median stays at 5. The median is robust against outliers, the mean is not. With skewed variables such as income, waiting times or workloads you should report both.

Dispersion: range, variance and standard deviation

Measures of dispersion say how far apart the values lie. The range is the distance between the largest and the smallest value, here 20 minus 2, so 18 hours. The variance gathers up the deviations from the mean: you square each deviation, add them all together and divide by n minus 1, which gives 19.4 here. Subtracting one is the usual route as long as your data are a sample and not the whole population. The standard deviation is the square root of the variance, roughly 4.4 hours.

Variance and standard deviation are often confused, although the distinction is simple: the variance is expressed in squared units and cannot be read intuitively, the standard deviation is expressed in hours and belongs next to the mean. The outlier pulls hard here too. Without the value of 20 the standard deviation drops from 4.4 to 2.4 hours.

Association: which coefficient suits which data

Measures of association describe how strongly two variables move together. The level of measurement decides which coefficient applies: Pearson's r for two metric variables, Spearman's rank correlation as soon as one of them is merely ordinal, Cramér's V for two nominal variables. Pearson and Spearman run between minus 1 and plus 1, Cramér's V between 0 and 1. The contingency coefficient, also widely used, reaches 1 only in its corrected form, which makes Cramér's V the safer choice. With two nominal variables it pays to look at the cross tabulation before the coefficient, because that is where you can see which combination carries the relationship.

Careful

A high correlation coefficient proves no causal link. It only shows that two variables vary together. Write “is associated with” rather than “leads to” unless your design can establish a direction of effect. Pearson's r also detects straight-line relationships only, so a value near zero does not rule out a strong but curved relationship. Always plot a scatter diagram before you read a correlation.

Which measure may you use at which level of measurement?

What is permitted depends on the level of measurement of your variable, not on whether your software returns a number. Excel will happily calculate a mean for a variable in which 1 stands for Cardiff, 2 for Leeds and 3 for Glasgow. The result is still meaningless.

The classification goes back to the psychologist Stanley Smith Stevens, who distinguished four types of scale in Science in 1946 and set out which measures remain meaningful for each (Stevens 1946). The table below gives the reading in common use today, which extends Stevens in places. Statisticians have criticised the scheme since the 1990s: read as a rigid rule, the four types mislead in borderline cases.

Level of measurementPermissible central tendencyPermissible dispersion
Nominal scaleModenone in the strict sense, report frequencies and proportions
Ordinal scaleMode, medianQuartiles, percentiles
Interval scaleplus the arithmetic meanRange, variance, standard deviation
Ratio scaleplus the geometric meanplus the coefficient of variation

Read the table cumulatively: whatever a lower level permits stays permissible at every higher one. The mode therefore works everywhere, the mean only from the interval scale upwards. To work out the level of your own variables, see the article on the level of measurement.

The recurring argument concerns agreement scales running from “strongly disagree” to “strongly agree”. Strictly they are ordinal, because nobody can show that all respondents perceive the gaps between the options as equal. Research practice nevertheless often treats such items as interval data, especially where several items are combined into one scale. That is defensible only with symmetrical, fully labelled options, and the decision belongs in your methods section with a justification. The article on the Likert scale covers it in full.

Create a survey for free

With empirio.ai you can create a modern online survey in minutes — with hosting in the EU.

  • AI-built survey
  • Adjust by drag & drop
  • Real-time analysis
Start for free

Inferential statistics: how the step from sample to population works

Inferential statistics works in two moves. It estimates a population value from your sample and states how uncertain that estimate is, and it tests hypotheses about differences or relationships. Both moves rest on the same foundation, namely the variation that arises from nothing more than which people happened to be selected.

That variation is the whole point. Had you asked 15 different people on the same course, a different mean would have come out, with nothing about the population having changed. Inferential statistics prices that randomness in instead of ignoring it. Standard error, confidence interval and p-value are three answers to one question: how far from the true value can your result sit?

The standard error: where the uncertainty comes from

The standard error states how much a measure such as the mean would vary from sample to sample. For a mean it is the standard deviation divided by the square root of the number of cases. For the 15 answers in the course example that is 4.4 divided by the square root of 15, so roughly 1.1 hours.

Standard deviation and standard error are confused constantly, although they measure different things. The standard deviation describes how far the individual respondents lie apart, and it stays about the same if you ask more people. The standard error describes how precisely you have located the middle, and it shrinks as the number of cases grows. The square root in the denominator is the unwelcome part: halving the standard error takes four times as many responses.

The confidence interval: an estimate with room to move

A confidence interval is the range around your estimate produced by a procedure that, at a 95 per cent level, captures the true value in 95 out of 100 samples. It replaces a single figure with an honest span. For the 6.4 hours in the course example the t distribution on 14 degrees of freedom gives an interval of roughly 4.0 to 8.8 hours.

The commonest misreading is that the true value lies within this particular interval with 95 per cent probability. The claim belongs to the procedure across many samples, not to your one result. For practical purposes little changes: a wide interval means little information, a narrow one means plenty. This is also why the Office for National Statistics publishes intervals alongside its survey estimates rather than bare figures. If you compare two groups and their intervals overlap heavily, you already know without a test that the data carry no clear difference.

Which test suits which data

Three questions decide which test applies, in this order: are you showing a difference or a relationship? What is the level of measurement of the variable in question? And how many groups are you comparing? Only then does the distribution matter. The table covers the cases that come up in almost every dissertation and almost every survey analysis.

QuestionLevel of measurementTest assuming approximate normalityTest without a distribution assumption
Difference between two independent groupsmetricindependent-samples t-testMann-Whitney U test
Difference between two measurements of the same peoplemetricpaired-samples t-testWilcoxon signed-rank test
Difference between three or more groupsmetricone-way analysis of variance (ANOVA)Kruskal-Wallis test
Relationship between two variablesmetricPearson correlationSpearman rank correlation
Relationship between two variablesnominalchi-squared testFisher's exact test for sparse cells

Every test in the third column has a counterpart in the fourth that assumes nothing about the distribution, and in case of doubt the second is the safer choice. It loses only a little power on normally distributed data but stays valid on skewed distributions and small groups. The same logic as with the descriptive measures applies to the level of measurement: an ordinal variable is treated as ordinal, even where the software adds the numbers up without complaint.

What your data have to satisfy first

Every test above assumes two things that have nothing to do with the arithmetic: a selection procedure that involves chance, and observations that do not influence each other. Passing a questionnaire round a group whose members talk to each other while filling it in breaks the second condition, and no procedure repairs that afterwards.

The distribution assumptions come next. The t-test and analysis of variance assume that values within the groups are approximately normally distributed and that the spread is similar across groups. Unequal spreads are handled by the Welch correction, small or skewed groups by the procedures in the right-hand column. The larger the groups, the less the normality assumption matters, because the distribution of sample means approaches normality as the number of cases grows. The chi-squared test has a condition of its own, namely adequately filled cells. The usual convention asks for an expected frequency of at least five in nearly every cell; below that, Fisher's exact test is the right call.

When you need no inferential statistics at all

Inferential statistics becomes unnecessary the moment nobody wants to generalise to a larger group. A significance test exists purely to move from a sample to a population. Remove that purpose and the reason for the test disappears with it.

The clearest case is a census of your own population. If all 42 members of your student society answer and you want to say something about exactly those 42 people, the population is in front of you and there is nothing left to infer. That 19 of 42 prefer the later meeting slot is not an estimate with uncertainty attached, it is simply the result. If only some of them answer, it is no longer a census but a group that selected itself.

One limit applies nonetheless, and it belongs in your methods section. The rule holds as long as your claim is confined to exactly this group at exactly this moment. As soon as you read the result as evidence of something underlying it, future members or a cause, you are working with a model again and confidence intervals are back on the table. The methods literature files this case under superpopulation.

Tip

Put one sentence in your methods section explaining why no test follows: “As this is a census of the population, no inferential procedures were applied.” In a viva that reads far more confidently than a test with no target.

The convenience sample: calculate yes, generalise no

A convenience sample is what you get by asking whoever was reachable instead of drawing people at random. Sharing the questionnaire across three WhatsApp groups and one Instagram profile is exactly that. A test can still be calculated on such data, but its result does not travel beyond the respondents, because the assumption of random selection is missing.

That limitation belongs openly in your limitations section, and that is precisely where marks are won. What makes a sample methodologically sound and when a survey counts as representative is covered in the two linked articles.

What the p-value says and what it does not

The p-value states how likely a result like yours, or a more extreme one, would be if the null hypothesis were true and every other assumption in your model held, random selection among them. It is a statement about the data under those assumptions, not about the assumptions themselves. A small p-value can therefore also mean that one of those assumptions is wrong.

The American Statistical Association published six principles on p-values in 2016 because misinterpretations had become widespread in the research literature. Principle two leaves no room: “P-values do not measure the probability that the studied hypothesis is true …” (Wasserstein and Lazar 2016). The sentence continues by ruling out the second common misreading: a p-value says just as little about how likely it is that your result came about by chance alone. A p-value of 0.03 does not mean your hypothesis is 97 per cent likely. What your own two groups produce is shown by the p-value calculator, and the reading of it is still your job.

Nor does a p-value measure the size or the importance of an effect, which is principle five of the same statement. With a very large sample even a tiny, practically irrelevant difference can become significant. Report an effect size alongside every p-value, and a confidence interval wherever possible.

How official statistics handles uncertainty

The Office for National Statistics shows what taking uncertainty seriously looks like in practice. Its employment estimate for July to September 2019 was 32.75 million, published with a confidence interval of plus or minus 177,000 people, and it separates sampling error from non-sampling error, which would remain even in a full census (ONS, Uncertainty and how we measure it). The lesson for your own work: a figure without a statement about its precision is not a finished figure. For formulating your assumptions beforehand, see the article on developing hypotheses.

Descriptive statistics in your results chapter

The results chapter opens with descriptive statistics, and specifically with a description of the sample: number of cases, response rate, distribution by age, gender, year of study or whatever your research question requires. Only then come the measures for the substantive variables, and finally the hypothesis tests.

A simple rule of thumb governs presentation. Anything expressible in two sentences stays in the text. Anything covering more than three values goes into a table. Anything meant to show a distribution or a trend becomes a chart. Every table reports n and the measures that suit the level of measurement, so mean and standard deviation for metric variables and frequencies for nominal ones. Every chart has labelled axes with units.

What to calculate with

A spreadsheet is entirely sufficient for central tendency and dispersion. SPSS is standard at many universities but requires a licence; JASP and jamovi are free, work in a similar way and cover the procedures an undergraduate dissertation usually calls for. R and Python are worth it if you already use them. Your marks depend on a transparent calculation, not on the software.

Things get considerably easier when the data arrive already structured. Online surveys let you download responses as an Excel or CSV file, which saves you retyping and the typing errors that come with it. empirio.ai, an online survey tool from Germany, shows frequencies and means directly in its results view and also exports the raw data as a file for your own analysis; EU hosting and a data processing agreement are included in the free plan. If you have not collected anything yet, you can create your survey online and let the analysis run alongside it. The steps that take you from raw data to a finished results chapter are set out in the article on analysing a questionnaire.

Common mistakes in the analysis

Most weaknesses in a results chapter are not arithmetic errors but matching errors: a measure is applied to data it was never meant for, or a figure is read as saying more than it can. Four patterns come up again and again.

All four go unnoticed while you write and stand out immediately when someone reads. Going through the results chapter for them before submission takes a quarter of an hour, and they are exactly what an examiner asks about.

A mean for a nominal variable

A mean of 1.6 for a variable whose codes stand for degree subjects is not information, it is a by-product of the coding. For nominal variables you report frequencies and proportions, nothing else. The test is quick: if the statement changes when you reassign the codes, the measure is inadmissible.

Reporting only the mean

A mean without a measure of dispersion withholds half the information. Some 6.4 hours a week can mean that everyone works about six hours, or that some respondents work two hours and others eleven. Standard deviation and sample size therefore belong on the same line as the mean.

Percentages without a base

“60 per cent of respondents agreed” sounds solid until you learn that five people answered. Below roughly twenty cases a single person is worth more than five percentage points, which makes the absolute figure the more honest one. Always give it, so “three out of five”.

Confusing significance with relevance

A significant difference is not automatically an important difference. Whether an effect matters in practice is shown by the effect size, set against what is usual in your field and judged by what would actually make a difference in your case. Phrases such as “highly significant” imply a magnitude that the p-value simply does not measure.

Conclusion

Descriptive statistics is not a warm-up for the real analysis. In many student projects it is the whole analysis. Anyone who surveyed everybody, or who could not draw a random sample, should describe carefully and state the limits rather than run a test their data cannot carry. And anyone who does test should place an effect size next to the p-value.

An honestly described sample is worth more than an asterisk next to a number.

Where to go next


Would you rather not retype your data?

With empirio.ai you build your survey free of charge and get frequencies, means and a raw data export without an extra step. Start for free

Frequently asked questions

Descriptive statistics summarises and presents an existing data set without going beyond it. It covers measures of central tendency such as the mode, median and mean, measures of dispersion such as the range, variance and standard deviation, plus measures of association and charts. Every figure applies to the cases you collected.

Descriptive statistics describes the sample, inferential statistics draws conclusions about the population from that sample. It estimates population values, usually as a confidence interval, and tests hypotheses, always with a statement of uncertainty. Any data set can be described, only a properly drawn sample supports inference.

Descriptive statistics uses three groups of measures. Central tendency covers the mode, median and mean. Dispersion covers the range, interquartile range, variance and standard deviation. Association covers Pearson's correlation coefficient, Spearman's rank correlation and Cramér's V. Which of them you may use depends first on the level of measurement and then on the distribution.

The procedures of inferential statistics divide into estimation and testing. Estimation uses point estimates, the standard error and confidence intervals. Testing uses the t-test, analysis of variance, the chi-squared test or a correlation test, and for skewed distributions and small groups the distribution-free alternatives: the Mann-Whitney U test, the Wilcoxon signed-rank test and the Kruskal-Wallis test.

The p-value states how likely a result as extreme as yours would be if the null hypothesis were true. It does not measure the probability that your hypothesis is true, and it does not measure the size of an effect. The American Statistical Association states both points explicitly in its 2016 statement.

Inferential statistics is needed only if you want to generalise from your sample to a wider population. If every member of that population answered, a census, the step falls away. With a convenience sample a test can be calculated, but its result does not travel beyond the people who answered.

The Office for National Statistics reports uncertainty because its headline figures come from samples, not from everyone. Its employment estimate for July to September 2019 was 32.75 million with a confidence interval of plus or minus 177,000 people. Sampling error shrinks with larger samples, non-sampling error does not.

The standard deviation describes how far the individual values lie apart, and it stays about the same as the number of cases grows. The standard error describes how precisely a measure such as the mean has been estimated, and it shrinks as the sample grows. For a mean it is the standard deviation divided by the square root of the number of cases.

You might also be interested in

Empirical Research

Fundamentals of Empirical Research | empirio

Empirical does not mean numbers, and an existing dataset is not a shortcut. Find out which route fits your research question and how to justify it in your methods chapter.