The formula
EDA is a discipline, not one equation, but it leans on a handful of summary numbers you compute and plot.
five-number summary = { min, Q1, median, Q3, max }
Reading: the smallest value, the 25th percentile, the middle, the 75th percentile, and the largest — the skeleton a box plot draws.
IQR = Q3 - Q1
Reading: the interquartile range, the spread of the middle half of the data. Robust — it ignores the tails, so one wild point can't move it.
lower fence = Q1 - 1.5·IQR
upper fence = Q3 + 1.5·IQR
Reading: Tukey's outlier rule. Anything past a fence gets plotted as its own dot and asks for an explanation. Points inside the fences form the whiskers.
bin_count ≈ ceil( √n ) (histogram, a common starting rule)
Reading: one way to pick how many bars a histogram gets from n samples. It is a starting point you adjust, not a law.
Everything else in EDA is a plot: the histogram (shape of the distribution), the box plot (five-number summary plus outliers), the scatter plot (relationship between two variables), the run-sequence plot (value versus sample order), the lag plot (value versus the previous value), and the normal probability plot (data versus a straight line it should follow if it is Gaussian).
Where you meet it
Vibration test stand, before the report. You pull 50,000 accelerometer samples off a shaker run. Before you quote an RMS number, you plot the time history (run-sequence) and a histogram. The run-sequence shows a level shift halfway through — a fixture bolt backed off. That RMS would have been garbage, and no summary statistic alone would have told you.
RF bench, characterizing a batch of amplifiers. Twenty units, gain measured on each. A box plot puts them side by side. Nineteen sit in a tight box; one hangs below the lower fence. That is your rework candidate, spotted in one glance instead of scanning a spreadsheet column.
Design review board, someone's calibration data. A presenter shows a fitted line through voltage-versus-temperature points and an R² of 0.98. Ask for the scatter plot. Half the time the "line" is really two clusters with a gap, and the fit is an artifact of the gap, not a real trend.
Reliability pull, life-test times. A histogram of failure times is right-skewed with a long tail — the tell that you are looking at a Weibull or lognormal process, not a normal one, before you ever open the reliability math.
How it works
The whole idea is old-fashioned and stubborn: plot the data before you model it. A summary statistic compresses a dataset into one number, and compression throws information away. The mean, the standard deviation, and the correlation coefficient all have the same values for datasets that look completely different when you draw them — the classic demonstration is four small datasets that share nearly identical summary statistics yet show a clean line, a curve, a line with one outlier, and a vertical stack with one leverage point. If you only computed, you would call all four the same. If you plotted, you would never confuse them.
The box plot earns its keep because it is robust. The median and the IQR don't flinch when one point is off by a factor of ten, unlike the mean and standard deviation, which a single fat-fingered decimal will drag across the room. That is why Tukey's 1.5·IQR fence is a good first outlier screen — it uses only quartiles. Run the math on a perfect normal distribution and the fences land at ±2.698 standard deviations, flagging about 0.70% of the data as "outliers" by chance alone. That number matters: on 10,000 clean samples, expect roughly 70 points past the fences with nothing actually wrong. The fence is a prompt to investigate, not a verdict to delete.
The NIST/SEMATECH Engineering Statistics Handbook frames EDA around four assumptions every measurement process quietly makes: the data are (1) random draws, (2) from a fixed distribution, (3) with fixed location, and (4) with fixed variation. Their "4-plot" checks all four at once — run-sequence plot (is location and variation drifting?), lag plot (is each point independent of the last, or is there autocorrelation?), histogram (what's the distribution shape?), and normal probability plot (is it Gaussian?). If the run-sequence plot trends, your location isn't fixed. If the lag plot shows a pattern instead of a shapeless blob, your samples aren't independent and every confidence interval you compute downstream is too narrow.
The gotchas are all about not looking, or looking wrong:
- Trusting the mean of skewed data. Failure times, repair costs, and particle sizes are right-skewed. The mean sits out in the tail; the median is the honest center. Report the wrong one and you'll under-quote a warranty reserve.
- Binning a histogram badly. Too few bins hides a bimodal split (two populations mixed together); too many bins turns real structure into noise. Always try a couple of bin widths before you believe the shape.
- Autocorrelation masquerading as clean data. Sensor data sampled fast is correlated point-to-point. A histogram looks fine, but you have far fewer independent samples than rows, and your error bars are lies. The lag plot is the only cheap way to catch it.
- Deleting outliers because a fence flagged them. The fence found a point; it did not diagnose it. Sometimes the outlier is the measurement error. Sometimes it is the one unit that's about to fail in the field — the most important data point you have.
EDA has a hard limit worth stating: it generates hypotheses, it does not confirm them. Seeing a trend in a scatter plot is a reason to test for one, not proof one exists. Tukey's own split was between exploratory analysis (what might be going on?) and confirmatory analysis (is it real, with what confidence?). Skip the confirmatory step and you are pattern-matching noise. Skip the exploratory step and you are modeling data you never actually looked at — which is how a level shift, a second population, or a stuck sensor rides straight into the final report.
History
The field has one clear father: John Wilder Tukey, born in 1915, died in 2000 — a Bell Labs and Princeton statistician who also co-invented the Fast Fourier Transform algorithm in 1965 and is credited with coining the word "bit" and popularizing the word "software" [1][2]. For decades statistics had been dominated by the confirmatory tradition: state a hypothesis, collect data, run a significance test. Tukey's complaint was that this skipped the most useful step — actually looking at the numbers first to see what questions were worth asking [1][3].
In 1977 he published Exploratory Data Analysis, the book that named the field and shipped its toolkit: the box-and-whisker plot, the stem-and-leaf display, and the 1.5·IQR fence rule for flagging outliers, all designed to be drawn by hand with a pencil and to keep the raw data visible rather than compressed into a single statistic [1][2]. The 1.5 multiplier is a pragmatic compromise — tight enough to catch genuine strays, loose enough that it doesn't cry wolf on every clean batch.
The engineering world adopted it wholesale. When the U.S. National Institute of Standards and Technology and SEMATECH co-authored their Engineering Statistics Handbook, they opened it with EDA and built its first chapter around Tukey's premise that you validate the four assumptions of a measurement process graphically before you trust any calculation on it [4]. That is why an engineer meets EDA in chapter one and not chapter twelve: it is the step everything else depends on.
Related tools
- /tools/rms-peak
- /tools/crest-factor
- /tools/strain-gauge-bridge
- /tools/snr-enob
- /tools/thermal-noise-floor