Outlier

Reducing many numbers to one is convenient. Asked about response time, you answer with an average. The mean — every value added up and divided by the count — is easy to compute and looks fair because every value contributes something.

That convenience cost Pagedeck two weeks. The average response time on the document list came out at 1.5 seconds, which read as slow, so two weeks got assigned to performance work.

Laying the values out one by one after starting told a different story. Four of five requests were around 0.3 seconds and one was 6.2. That single request had pulled the average up fivefold. And that request was an administrator downloading the entire dataset at once — something that never belonged in customer-facing response time to begin with.

That one distant value can dominate a summary is why this concept carries its own name. Looking only at the average, you do not know the value is there — the summary is doing its job and misleading you anyway.

Lay the values out and judge by how far one sits from the rest. Three tests are in common use.

1. Distance from the mean. Compute the standard deviation, the spread around the mean expressed in the original unit, and treat anything beyond three times that as an outlier. This only applies when the values are spread in a bell shape with a thick middle.
2. The width of the middle half. Sort the values, take the width the middle 50 percent occupies, and treat anything beyond 1.5 times that width as an outlier. For response times or payment amounts, where one tail runs long, this test is the safer one.
3. Plotting it. Scattering the values on a chart comes before applying any automatic test. A test tells you which value is distant; why it is distant comes from the chart and the original records.

Once found, there are four options.

1. Correct it. If the value came from a typo, fix it back. An amount with three extra zeros belongs here.
2. Drop it. Only when the measurement itself failed — a value that was never part of what you were measuring, like the administrator request above.
3. Keep it. If it really happened and it was inside what you meant to measure, leave it. If somebody genuinely experienced 6.2 seconds, that value is the most important signal present, and dropping it removes the only evidence of the problem.
4. Split the data. When the distant values cluster rather than appearing once or twice, you have two populations rather than an outlier. Free accounts and enterprise accounts in one table produce two clumps, and the response is splitting the table, not removing values.

Whatever you did, write it down. Without a record of which test was applied and how many values came out, the next person produces a different number from the same data. There is also the option of changing how you summarize instead of removing anything: the median, the middle value when sorted, and the 95th percentile both resist distant values without erasing them.

The place this bites most often is performance metrics. The convention of reading the 95th percentile instead of average response time came out of this problem. An average hides the people who had the worst experience, and the worst experience is what people cite when they say something is slow.

The second is experiment results. In an A/B test, where two versions are shown to randomly split groups and compared, one person making a large payment flips that group's average order value. It happens routinely at sample sizes of a few hundred, so money experiments get read as mean and median together, with a version excluding the largest few values alongside.

The third is revenue. When one customer accounts for 30 percent of the total, average revenue per customer explains nothing. That customer gets set aside and the average is computed from the rest.

The fourth is data quality checks. A pipeline gets a step that validates value ranges and raises a warning on a day when signups arrive at twenty times normal. Whether that was real growth or a collection job running twice and loading everything in duplicate is something a person has to confirm; the check itself cannot tell the difference.

The fifth is choosing who to talk to. Interview only average users and problems occurring at both extremes stay invisible. Deliberately mixing in someone who opens a hundred documents a day and someone who opens two a month comes from this reasoning.

"The average is 1.5 seconds and one value is 6. Let's look again without it."

A distant value is dominating the summary.

"I don't think that's an outlier. I think we have two populations."

Split the table rather than remove the values.

"Please record the count dropped and the test used."

Stopping the same data from producing different numbers.

At Pagedeck, Mia keeps one rule when reporting numbers: wherever an average appears, the median appears next to it, and when the two are far apart, one line explains why. If the pair reads 1.5 seconds and 0.3 seconds, that gap is the story, and reporting only one of the two removes it.

Theo uses it from the other side. Range checks sit in the pipeline, and on a day one trips, he opens the original records. Most turned out to be collection problems and some were genuinely a large customer arriving.

Assuming that finding one means removing it is the most common mistake. The word suggests something wrong, so it reads as a bad value. But this value is frequently the most important signal in the set. A payment 300 times the usual size might be a typo or it might be a newly signed enterprise contract, and if it is the second, that one record is half the month's revenue. Skip the step of opening the original record before dropping it and what should be deleted gets the same treatment as what should be studied.

The second is treating a test result as the verdict. Three standard deviations out is computed automatically, so it looks like a conclusion. But that test only means something under the assumption that the values are spread in a bell shape. Apply it as-is to response times or payment amounts, which run with one long tail, and healthy values get flagged in bulk, sometimes a tenth of the dataset; drop all of them and what remains is a shape that does not match reality.