Why summary statistics come before anything else
Before building a model, running a statistical test, or drawing any conclusion from a dataset, a quick pass of descriptive statistics catches a surprising share of real problems — data entry errors, unexpected missing values, unit mismatches, or an assumption about the data that turns out to be wrong. This step, often called exploratory data analysis (EDA), isn't optional groundwork — it's frequently where the most important discoveries about a dataset's quality actually happen.
The five numbers that catch the most problems
- Count: how many valid values actually exist in this column? A count much lower than the total row count reveals missing or non-numeric data you might not have expected.
- Mean and median together: when these differ substantially, the data is skewed — a few extreme values are pulling the mean away from the "typical" value the median represents. This mismatch alone is often the first sign of outliers worth investigating.
- Standard deviation: how spread out is the data? A surprisingly large standard deviation relative to the mean can indicate genuine high variability, or a data entry error creating an extreme outlier.
- Min and max: the fastest sanity check available — an age column with a max of 200 or a min of -5 immediately flags a data quality problem, often before any other analysis.
Why mean vs. median divergence is worth specifically checking
A perfectly symmetric distribution has mean equal to median. Any meaningful gap between them signals skew — and the direction tells you which way: mean noticeably higher than median suggests a right skew (a few unusually large values pulling the average up, common in income or price data), while mean noticeably lower than median suggests left skew. Spotting this early changes which further statistics and visualizations are actually appropriate — many analyses assume roughly symmetric data, and applying them blindly to skewed data can produce misleading results.
Why min/max checks catch bugs before deeper analysis does
Extreme value checks are the cheapest, fastest data validation step available, and they catch an outsized share of real bugs: a percentage column showing 150%, a date column with a year of 1900 where none was intended, or a price column with a negative value all surface immediately from a min/max scan — no statistical sophistication required, just a plausibility check against what the values are supposed to represent.
What summary statistics don't tell you
These five numbers describe each column in isolation — they say nothing about relationships between columns (that's what a correlation matrix is for), nor about a distribution's actual shape beyond skew direction (a histogram reveals whether data is unimodal, bimodal, or has other structure a single number can't capture). Summary statistics are the fast first pass, not the complete picture.
Common mistakes
- Skipping straight to modeling without checking basic summary statistics first. A surprising number of downstream problems trace back to a data quality issue that a two-minute summary-statistics pass would have caught immediately.
- Ignoring a mean/median gap. This is one of the cheapest signals available that a distribution isn't symmetric, and it changes which subsequent analyses are actually appropriate.
- Trusting a dataset's row count without checking per-column counts. A column with far fewer valid values than the total row count reveals missing data that a simple row count alone won't surface.
FAQ
Why check mean and median together instead of just one?
Because their relationship reveals skew — a symmetric distribution has them roughly equal; a meaningful gap signals the data is pulled in one direction by extreme values, changing what further analysis is appropriate.
Is checking min/max really useful, or too simple to matter?
It's simple and still one of the most effective data-quality checks available — implausible extreme values (negative prices, ages over 150) surface immediately, often catching real bugs before any deeper analysis.
Do summary statistics replace the need for visualizations?
No — they're a fast first pass; visualizing the actual distribution (histogram) or relationships between columns (scatter plot, correlation matrix) reveals structure that summary numbers alone can't show.
Upload a CSV and get instant summary statistics for every numeric column with the CSV Summary Statistics tool — entirely in your browser.