What quantitative data is, with examples
A definition you can use, and a worked example that puts a 95 per cent interval on a 14-point difference before anyone signs off on it.
The short answer
Quantitative data is information recorded as numbers where the number is an amount or a count: four support tickets, 84 seconds at the checkout, 62 per cent satisfied. Totals, averages and differences on it mean something, because the distance between any two values is known. Data that names a category is qualitative even when it is stored as a digit: a plan code, a postcode, a department filed as 1 to 6.
Ordered categories sit between the two. A satisfaction rating of 4 on a 1-to-5 scale has a real order, but nothing establishes that the gap from 3 to 4 equals the gap from 4 to 5. Counting those ratings and taking their median needs no assumption; averaging them is an analysis assumption you make and state. The Australian Bureau of Statistics puts the split in one line: quantitative data are "measures of values or counts and are expressed as numbers", qualitative data are "measures of types".1
What counts as quantitative data
The test is arithmetic. If totalling the values or averaging them tells you something true about the thing you measured, the data is quantitative. Ticket counts pass. Seconds pass. A postcode does not: 3053 and 2000 are written with digits, but their average is not a place. Order on its own is not enough either. It makes a category rankable, not measurable.
The postcode case catches people out, and the ABS definition is careful about it: qualitative data "may be represented by a name, symbol, or a number code".1 A column of numbers in a spreadsheet proves nothing on its own. What matters is what each number was recorded to mean.
| A survey answer | Quantitative? | Why |
|---|---|---|
| How many times did you contact support last month? 3 | Yes | A count. The total and the average both mean something. |
| How long did checkout take? 84 seconds | Yes | A measurement on a scale with a real zero. |
| How satisfied were you? 4 on a 1-to-5 scale | Ordered category | The order is real. The gap from 3 to 4 is not known to equal the gap from 4 to 5, so counts and the median are safe and the mean is an assumption. |
| Which plan are you on? Team | No | A category. The number of people on each plan is quantitative; the answer is not. |
| What should we fix first? free text | No | Words. They turn quantitative only once you code them into categories and count the categories. |
| What is your postcode? 3053 | No | Digits used as a label. The mean of two postcodes is a number that refers to nothing. |
Examples written for this page. The classification follows the Australian Bureau of Statistics definition of quantitative and qualitative data (2023).
The fifth row is the one worth sitting with. Ninety of 200 people writing about delivery speed is a fact you can track week on week, plot and test. The sentences stayed qualitative. The count of them did not.
Quantitative data vs qualitative data
They are told apart by the variable, not by the method or the sample. A numeric variable always returns quantitative data and a categorical variable always returns qualitative data, so you know which you are getting before a single response arrives.1 That is the useful part: the decision is made when you write the question, not when you open the export.
| Quantitative | Qualitative | |
|---|---|---|
| Records | Amounts and counts | Types and categories |
| Asks | How many, how much, how often | What kind, which one, in what words |
| Example | How many days did you use it? | Which feature did you use most? |
| Summaries | Mean, median, spread, shape | Mode and counts. Median if the categories are ordered |
| Inference | An estimate for the population, with an interval | Only after counting the categories |
Rows follow the Australian Bureau of Statistics treatment of what each data type supports (2023). Survey examples written for this page.
Neither is the better kind. A number tells you how far something moved and how sure you are; it holds no explanation at all. Words carry the reason and nothing you can put a confidence interval on. Most real surveys collect both, which is exactly what our own library does.
Types of quantitative data
Two splits matter. The first is whether the values are counted or measured. The second is what the numbers are entitled to carry.
Discrete data arrives in whole units that cannot be subdivided: 3 support tickets, 12 questions answered, 0 complaints. Between 3 and 4 there is nothing.
Continuous data falls anywhere along a scale, limited only by how finely you measure: 84 seconds, 84.2 seconds, 6.5 hours. Time and revenue stay continuous even when the reporting rounds them to whole units.
The second split is older and more argued over. Stevens sorted measurement into four levels in 1946, and that sort still decides what most textbooks will let you compute.2
| Level | What the numbers carry, and a survey example | Safe to compute |
|---|---|---|
| Nominal | Labels only, no order. Department, stored as 1 to 6 | Counts and shares |
| Ordinal | Order, with no established step. A 1-to-5 agreement item | Counts, shares, median, top-two box. The mean only under a stated assumption |
| Interval | Equal steps, no true zero. A standardised score centred on 100 | The above, plus the mean and differences |
| Ratio | Equal steps and a true zero. Seconds to finish, tickets raised | The above, plus ratios: 60 seconds is twice 30 |
Levels after Stevens (1946). Survey examples written for this page.
Nominal and ordinal are the two categorical levels: a nominal value is a label wearing a number, an ordinal value is a label with a rank. Interval and ratio are the two quantitative levels, because only there does the step between neighbouring values mean the same thing everywhere on the scale. Every real argument sits on the line between ordinal and interval, and a Likert item sits exactly on it. Averaging its 1-to-5 codes treats an ordered category as if it were interval data. That is a legitimate analysis assumption, and it is an assumption, so say it whenever you report the mean.
Velleman and Wilkinson pushed back on the whole typology, arguing it misleads when it is used to rule methods in or out, because a variable's level is a claim about what the numbers mean rather than a property you can read off the file.3 A defensible position for survey work: state the assumption you are making, then also report a figure that does not depend on it, the share in each category or the median.
Characteristics of quantitative data
Six properties do the work, and the last one is the limit that matters.
- It is numeric by construction. The value is a count or a measure, not a code standing in for a word and not a rank standing in for a label.
- Its steps are equal. The distance from 3 to 4 is the distance from 7 to 8. Ordered categories lack this, and it is the property that makes an average mean something.
- It is summarisable. Because the values are numeric they can be ordered, added and counted, so the whole set of descriptive statistics is available.1
- It is reproducible. Two analysts handed the same file and the same coding rules return the same mean. Any disagreement is about the coding, not about the reading.
- It supports inference. A sample carries an estimate for the population it was drawn from, with an interval attached to say how firm the estimate is.
- It carries no reason. A drop from 62 to 48 per cent is a fact with no explanation inside it. The explanation is in the words, which is why cutting the open question to shorten a survey is usually the wrong economy.
Quantitative data in survey research
What a real survey returns, counted from 109 published templates and 1,422 questions.
On 10 September 2026 we read every question in the live SuperSurvey template library and classified each one by its stored answer control: 109 templates, 1,422 questions, a median of 13 questions per template. The mix decides what kind of data comes back.
Rating items produce more survey numbers than anything else
Every question in 109 published templates, classified by its stored answer control.
Add the three closed formats and 1,201 of the 1,422 questions, 84.5 per cent, return something countable the moment somebody answers: a rating position or a category code. The other 221 return words. That ratio is the practical answer to what a survey will produce: roughly five parts arithmetic to one part reading.
Rating items dominate, and they are strikingly uniform. Of the 644 rating questions, 637 use five points and the remaining 7 are 0-to-10 recommendation items, eleven options each. Not one uses three, four, six or seven. Whatever the count, the labels on the two ends are what a 4 or a 9 actually means, so choose meaningful rating labels before deciding how the answers will be scored.
Where the words sit is just as consistent.
The words are asked last, almost every time
Median position of each question type, as a share of the way through its own template. 0 is the first question, 100 the last.
Open text is the last thing asked. All 109 templates carry at least one open question, a median of two each, sitting at 87 per cent of the way through. Demographic questions sit almost as late, at 83 per cent, and only 26 of the 109 templates ask any at all. So the quantitative part of a survey is not one section among several. It is nearly the whole instrument, with the words held back to the end.
These are our own published templates. They show what one survey tool recommends as good practice, not what the world's surveys do. Read 98.9 per cent as "of the rating items in this library", never as "of surveys". Anyone quoting it the second way is quoting it wrongly.
Steps in quantitative data analysis
Five steps, in order. The first two are where most of the damage gets done, and neither of them involves a statistic.
-
Write the decision down first
"Should we roll the new checkout out to everyone?" is a decision. "Look at satisfaction" is not. Name the comparison you will make and the size of difference that would change your mind, and write both down while the survey is still in the field. Deciding afterwards is how a three-point wobble becomes a finding.
-
Build a codebook, then freeze it
One row per respondent, one column per question, one meaning per code. Record the missing-value code explicitly and never let it be 0 unless 0 is a real answer. Averaging a 1-to-5 column together with a 0-to-10 column is the most common way a survey report goes quietly wrong, and nothing downstream will catch it for you. The codebook for the worked example on this page is below the steps, ready to copy.
-
Screen for careless answers
Straightlining down a grid, impossible completion times and duplicate submissions all move an average. Meade and Craig compared the detection methods (instructed items, response consistency indices, multivariate outlier analysis, response time and self-reported diligence) and identified roughly 10 to 12 per cent of undergraduates completing a long survey for course credit as careless responders, in two patterns, random and non-random, that need different checks to catch.5
Do not carry that percentage across to your own data. Report what you removed and the rule you removed it by. Who never answered at all is a separate problem, covered under response bias.
-
Describe before you average
Two groups can both average 3.4 while one is bunched in the middle and the other is split at both ends. Those are different situations and they call for different decisions. Publish the share choosing each option, or at minimum the top-two-box share, beside any mean you report.
-
Put an interval on the difference
A difference without a range is a claim with no size on it. The interval says how much of the difference would survive a repeat, which is the only version of the number a decision can be made on. The worked example further down does one from end to end.
The codebook for the worked example
Five variables, one line each. It is written for the checkout survey in the next section, and it is the reason the four counts there can be trusted: every code has one meaning and the not-applicable answer has a home that is not 0.
Codebook: checkout satisfaction survey
| Variable | Label | Codes | Missing | Denominator |
|---|---|---|---|---|
respondent_id | Completed response | 1 to 200, as stored by the survey tool | None. A row with no id is dropped before anything is counted | Not applicable |
checkout | Checkout version shown | 1 = old checkout; 2 = new checkout | None. Assigned by the system before the survey opened | All 200 completed responses |
sat_1to5 | How satisfied were you with the checkout? | 1 = Very dissatisfied; 2 = Dissatisfied; 3 = Neither; 4 = Satisfied; 5 = Very satisfied | 9 = Not applicable, did not reach checkout. Excluded. Never 0 | Responses coded 1 to 5 only: 100 per group |
satisfied_top2 | Derived: rated 4 or 5 | 1 = sat_1to5 is 4 or 5; 0 = sat_1to5 is 1, 2 or 3 | Blank when sat_1to5 is 9. Never 0 | Responses coded 1 to 5 only: 100 per group |
fix_first | What should we fix first? (open text) | Verbatim text, coded into categories after fielding closes | Empty string = no comment. Excluded | Responses with any text |
Written for this page. The copy button puts the same five rows on the clipboard as comma-separated text with a header row, ready for a spreadsheet or a script.
The third row is the one that matters. Somebody who never reached the checkout cannot rate it, so their answer is coded 9 and leaves the denominator: each satisfied share below is computed over the 100 people per group who answered 1 to 5. Code that answer 0 instead and the software reads it as a score below "very dissatisfied", drags the mean down and sits in the bottom of the top-two-box calculation. The derived column follows the same rule: blank when the source is 9, never 0.
Worked example: quantitative data interpretation
Two hundred responses, four counts, and one number that decides a rollout.
A team tested a new checkout against the old one and asked buyers a single satisfaction question on a 1-to-5 scale. One hundred people completed the survey on each version. Satisfied means a 4 or a 5.
Top box is the share choosing the single highest option, a 5 out of 5. Top-two box is the share choosing either of the highest two, a 4 or a 5. They are different numbers, and top-two box is the one most satisfaction reporting uses because it moves around less when few people answer. Everything below uses top-two box.
| Group | Rated 4 or 5 | Satisfied share |
|---|---|---|
| Old checkout | 48 of 100 | 48.0% |
| New checkout | 62 of 100 | 62.0% |
| Difference | +14 | +14.0 points |
Illustrative counts written for this page, not a measured result. Every figure below is arithmetic on these four numbers and can be repeated.
Fourteen points looks decisive. It came from one sample, so the honest question is how much of the 14 would survive a repeat.
For a difference between two independent proportions the standard error is SE = sqrt(p1(1-p1)/n1 + p2(1-p2)/n2). With p1 = 0.62 and n1 = 100, p2 = 0.48 and n2 = 100, that is sqrt(0.002356 + 0.002496), which is sqrt(0.004852) = 0.069656.
A 95 per cent interval is the difference plus or minus 1.96 standard errors. 1.96 x 0.069656 = 0.13653, so 0.14 either side of that lands at 0.35 to 27.65 percentage points.
The estimate stays at 14 points. The uncertainty is what moves.
The same satisfied shares, 48 per cent against 62 per cent, at three response counts. Each bar is the 95 per cent interval for the difference, in percentage points.
Read the interval first and the point estimate second. The best guess is a 14-point lift. The same data is also consistent with a lift of almost nothing. If your bar for rolling out is five points, this result clears it on the estimate and fails it at the bottom of the range, and that is a decision about risk rather than about statistics.
The width comes from the count, and only from the count. Hold both shares exactly where they are and double the responses to 200 a side: the estimate stays at 14 points and the interval tightens to 4.3 and 23.7. At 400 a side it is 7.2 to 20.8. Nothing about the finding improved. The measurement of it did.
Turned around, a two-sided test at 5 per cent has about a 51 per cent chance of flagging a real 14-point difference when only 100 people answer on each side, and 197 per group, close to the second row of the chart above, would lift that to 80 per cent. Setting the target before fielding is how many responses you need.
The matching two-sided z test on these counts gives p = 0.044. That uses the same unpooled standard error as the interval above, which keeps this page internally consistent. The textbook hypothesis test pools the two groups first (110 satisfied out of 200, so SE = 0.070356) and returns p = 0.047. Both sit under 0.05 and neither is a different result. If you publish an interval and a p value together, compute them the same way and say which you used.
One more caveat, and it is the reason to lead with the interval rather than the test. The formula above gives the Wald interval. Newcombe compared eleven methods for this exact problem and found the Wald version covers the true difference less often than its 95 per cent label claims, particularly at small counts.6 Agresti and Caffo showed that adding one success and one failure to each group before computing the same interval fixes most of the shortfall.7 Doing that here, 63 of 102 against 49 of 102, moves the interval to 0.2 to 27.3 points. The rollout decision does not change. On smaller counts it can be the whole answer.
Analysis methods for quantitative data
Pick the simplest method that answers the question at an acceptable risk of being wrong. Survey data is usually asked five things.
| What you need to know | Method | What to report |
|---|---|---|
| How common is it? | Share of one categorical answer, with its denominator | The per cent and the number behind it |
| Are two groups different on a yes or no outcome? | Difference in proportions with a 95 per cent interval, the example above | The interval, then the point estimate |
| Are two groups different on a rating or a measure? | Test of statistical significance on the difference in means, with its interval | The interval, the p value, and both distributions |
| Do two answers move together? | Correlation between survey items, two ordered or numeric columns | The coefficient with its interval, and the number of pairs |
| What predicts an outcome? | Regression analysis, step by step, one outcome against several inputs | Coefficients, the assumptions, and what moves if you drop an input |
A practical starting point for survey data, written for this page. The right method for a specific dataset depends on how it was collected.
The last two rows are where over-claiming starts. A coefficient between two survey items is one number from one sample and it says nothing about which answer came first. Regression puts several inputs in at once and is worth only as much as its assumptions, so read those before the coefficients.
Frequently asked questions
Is ordinal data qualitative or quantitative?
It is categorical data with an order, and the order is the only numeric property it has. So it is qualitative by type, and it becomes quantitative only through a decision you make: assigning the categories numbers and averaging them. Software will do that without complaint. If you would rather not have to defend it, report the share in each category or the median, because neither needs an assumption about spacing.
Is age quantitative data?
Age in years is, at ratio level: 40 really is twice 20. An age band is an ordered category. "25 to 34" comes after "18 to 24" and that rank is all it carries; the bands need not be the same width, and a mean of bands is a mean of midpoints you chose, which is an assumption rather than a measurement. Ask for the number when you intend to model it and the band when you intend only to cross-tabulate.
Why is quantitative data important?
Because it makes a disagreement resolvable. Two people can argue all afternoon about whether the checkout got better. They cannot argue about 48 against 62 out of 100 once they agree on the coding, and the argument moves on to what to do about it. The trade is that you gain a settled fact and lose the reason behind it.
How much quantitative data do I need?
It follows from the smallest difference you would act on, not from how large the population is. Detecting a 14-point gap between two halves of a sample with an 80 per cent chance takes about 197 responses per group. A 3-point gap takes about 4,360 per group. Fix the difference that matters first and the number falls out of it.
Do I need statistics software?
Not for anything on this page. A spreadsheet handles the counts, the shares and the interval in the example above in about four cells. Move to R, Python or a stats package when you need survey weights, repeated measures on the same people, or a run that goes from raw export to finished chart without a person in the middle.
How we counted
The corpus figures come from the SuperSurvey template library as it stood on 10 September 2026: every live template row, every question card inside it. That is 109 templates and 1,422 questions.
Question type was taken from the stored answer control, not inferred from the wording, so a rating item phrased as a sentence still counts as a rating item. Four controls exist: rating scale, single-select list, checkbox list and free text. Cards the parser could not read a question out of were dropped rather than guessed at.
Position is the question's index, counted from 0, divided by the question count minus 1 and expressed 0 to 100, so the first question in any template scores exactly 0 and the last exactly 100. The figure reported is the median across every item of that type; with an even number of items the two middle values are averaged. Unrounded, the medians are 43.3036 for rating items, 46.1538 for single-select, 61.5385 for multi-select and 86.6667 for open text; the chart shows them to the nearest whole per cent. A template with a single question would be excluded because the figure is undefined for it; this cut contains none.
Demographic items were flagged by a keyword match on the question text (age, gender, role, tenure, location, income and similar), so that count is approximate at the edges, and a demographic question also appears in whichever format count it belongs to. Nothing is double counted in the format totals themselves: 644 plus 516 plus 221 plus 41 is exactly 1,422.
The limit that matters: this is one vendor's published library, not a sample of the surveys people actually run. Every percentage here should be read with "of this library" attached to it.
The worked example is an illustration. Its four counts, 48 and 62 out of 100 each, were written to be worked through; everything derived from them is arithmetic you can repeat, and the unrounded Wald interval is 0.3474 to 27.6526 points. No measured result is reported from those numbers. The two required-response figures and the achieved-power figure use the normal approximation for two independent proportions, with the pooled standard error under the null and the unpooled one under the alternative, at 80 per cent power and a two-sided 5 per cent test; the 3-point case compares 48 against 51 per cent.
References
- Australian Bureau of Statistics (2023). Quantitative and qualitative data. Statistical terms and concepts. abs.gov.au: quantitative and qualitative data
- Stevens, S. S. (1946). On the theory of scales of measurement. Science, 103(2684), 677-680. doi 10.1126/science.103.2684.677
- Velleman, P. F., and Wilkinson, L. (1993). Nominal, ordinal, interval, and ratio typologies are misleading. The American Statistician, 47(1), 65-72. doi 10.1080/00031305.1993.10475938
- SuperSurvey Question Corpus, 109 live templates and 1,422 questions, read 11 September 2026. Method described above.
- Meade, A. W., and Craig, S. B. (2012). Identifying careless responses in survey data. Psychological Methods, 17(3), 437-455. doi 10.1037/a0028085
- Newcombe, R. G. (1998). Interval estimation for the difference between independent proportions: comparison of eleven methods. Statistics in Medicine, 17(8), 873-890. Newcombe (1998) at doi.org
- Agresti, A., and Caffo, B. (2000). Simple and effective confidence intervals for proportions and differences of proportions result from adding two successes and two failures. The American Statistician, 54(4), 280-288. doi 10.1080/00031305.2000.10474560
Michael Hodge: Survey methodologist and editor, SuperSurvey. Bachelor of Science (Psychology), University of Wollongong, with coursework in psychometrics and research methods. Designing surveys since 2003. About the author and how these guides are reviewed
What to read next
The decisions on either side of this one sit in the survey learning centre.
Collect quantitative data of your own
Rating items and single-select lists arrive already coded, one row per respondent, ready for the counts and the interval on this page.
Collect responses for your analysisFree to send, no card. Or start from a ready-made survey template and change the questions.