Descriptive statistics summarise data you already have. They don't predict or prove anything, but nearly every analysis starts with them, and misreading them is the root of many bad decisions. Think of any numeric column in terms of three questions.
1. Where is the centre?
Mean
The mean (average) adds all values and divides by the count. It uses every data point, which is its strength and its weakness. Ten employees earning $50,000 and one earning $1,000,000 have a mean salary of about $136,000, a figure that describes nobody in the room.
Median
The median is the middle value when data is sorted. In the salary example it's $50,000. Medians resist outliers, which is why house prices and household incomes are usually reported as medians.
Mode
The most frequent value. Most useful for categories ("the most common plan is Basic") and for spotting data-entry defaults (a suspicious spike at 1970-01-01 or 0).
2. How spread out is it?
Range
Maximum minus minimum. Simple, but determined entirely by the two most extreme values.
Standard deviation
Roughly, the typical distance of values from the mean, in the same units as the data. For bell-shaped data, about 68% of values lie within one standard deviation of the mean and about 95% within two. For skewed data, those percentages don't hold, so be careful quoting them.
Percentiles and the IQR
The 25th percentile (Q1) and 75th percentile (Q3) bracket the middle half of the data. Their difference, the interquartile range (IQR), is a robust measure of spread. Like the median, it ignores extremes. Percentiles also make great business metrics: "95% of pages load in under 1.8 seconds" is far more useful than "average load time is 0.9 seconds".
3. What shape is it?
- Symmetric: mean ≈ median. Heights and measurement errors often look like this.
- Right-skewed: a long tail of large values, so mean > median. Incomes, order values, time on site and city sizes are typical.
- Left-skewed: a tail of small values, so mean < median. Exam scores on an easy test, for example.
- Bimodal: two peaks, often meaning two different populations are mixed together, such as weekday and weekend traffic, or consumer and business customers. Split them before summarising.
The fastest way to see shape is a histogram. Our Data Profiler draws one for every numeric column automatically.
Outliers
A common convention flags values below Q1 − 1.5×IQR or above Q3 + 1.5×IQR as potential outliers. "Potential" matters. An outlier might be a typo, a test transaction, or your most important customer. Investigate before removing anything, and always say what you excluded.
Summaries that mislead
- Averages of averages. Averaging the conversion rates of a tiny campaign and a huge one gives them equal weight. Compute rates from the underlying totals instead.
- Simpson's paradox. A trend can reverse when data is grouped. Treatment A can beat B in every hospital yet lose overall if it's used more in harder cases. Check whether a hidden grouping variable drives the result.
- Survivorship bias. Summarising only what's left (active customers, surviving companies) hides what dropped out.
- Small samples. A 100% conversion rate from two visitors is not a finding. Always show counts alongside rates.
A quick reporting template
For any important metric, report: count (n), median, mean, IQR or standard deviation, min and max, the number of missing values, and a histogram. That handful of numbers prevents most misunderstandings.