Standard deviation shows how far a value sits from the mean. For roughly normal data, values beyond two standard deviations are often flagged as potential outliers, since about 95% of observations fall within that range. Skewed data may call for another approach, such as the interquartile range (IQR).

Multiple Choice

Any value that is smaller (or larger) than ______ standard deviations below (above) the mean is considered a statistical outlier.

In statistics, you identify unusual values by measuring how far they are from the center of the data in units of standard deviation. The threshold here uses two standard deviations: any value below mean minus two standard deviations or above mean plus two standard deviations is flagged as an outlier. The reason this works well is that, for a roughly normal distribution, about 95% of observations fall within ±2 standard deviations of the mean. So values beyond that range are unusually far from the center and deserve attention as potential outliers. If you used only one standard deviation, you'd label around a third of the data as outliers in a normal distribution; if you used three standard deviations, you'd be much more conservative and fewer observations would be flagged. Four standard deviations would be even rarer. Keep in mind this rule assumes the data are approximately normal; for skewed or non-normal data, other methods (like the IQR rule) may be more appropriate.

Why Two Standard Deviations Can Flag an Outlier

A data point can look unusual without being wrong. A customer who spends far more than the typical shopper may be a valuable regular, not a bad record. A sensor that suddenly reports an extreme temperature might signal a fault—or a real event worth investigating. In data mining, spotting these unusual values is only the first step. The next is working out what they mean.

One familiar way to flag a potential outlier is to measure its distance from the mean in standard deviations. Under a common rule, a value below the mean minus two standard deviations, or above the mean plus two standard deviations, is treated as unusually far from the center. It’s a useful starting point, but it works best when analysts understand what the rule assumes and what it can miss.

A ruler built from the data

The mean is the arithmetic average: add the values, then divide by how many there are. Standard deviation describes how spread out those values are around the mean. A small standard deviation means most observations sit close to the average; a large one means they tend to be more dispersed.

This gives analysts a handy measuring stick. Instead of saying that a value is “a lot higher” than average, they can describe it in standard deviation units. A value two standard deviations above the mean has a standardized score, or z-score, of +2. A value two standard deviations below it has a z-score of –2.

The calculation is simple:

z-score = (value − mean) ÷ standard deviation

A z-score of zero sits right at the mean. A score of +1 is one standard deviation above it; –3 is three standard deviations below. The familiar two-standard-deviation rule therefore flags observations with z-scores below –2 or above +2.

That gives a consistent way to compare values, even when the measurements use different units. For instance, two standard deviations may represent a modest difference in one dataset and a large one in another. The z-score expresses each difference relative to its own data.

Why the two-standard-deviation mark matters

The rule has a useful link to the normal, bell-shaped distribution. In an ideal normal distribution, about 95% of observations fall within two standard deviations of the mean. Roughly 5% fall outside that range—about 2.5% in each tail.

That does not mean every value beyond the boundary is an error. It means that, if the data follow a normal distribution, those values are relatively uncommon. “Uncommon” is a signal to look closer, not a verdict.

The threshold also reflects a trade-off. Set it closer to the mean, and more observations get flagged. In a normal distribution, about 68% of values fall within one standard deviation, so around 32% lie outside. That would mark a large share of ordinary observations as unusual. Set the boundary at three standard deviations, and only about 0.3% of normal observations fall beyond it. That catches fewer values, but it can miss less extreme anomalies.

Two standard deviations sit between those extremes. It’s a convenient screening threshold: sensitive enough to notice many unusual points, yet not so sensitive that a typical bell-shaped dataset sends up alarms everywhere.

A flag is not a diagnosis

Suppose a retailer analyzes daily sales and finds one day far above the mean. The cause might be a data-entry slip, a holiday rush, a flash sale, or a shipment that was recorded on the wrong date. Each possibility calls for a different response.

This is why data teams often treat an outlier flag as the start of an investigation. They may check the original source, compare nearby time periods, or look at related fields. If sales jumped on a holiday and website traffic rose at the same time, the high value may be perfectly plausible. If the reported sales exceed the store’s inventory by a wide margin, a recording issue becomes more likely.

The same point applies in fields such as healthcare, cybersecurity, and manufacturing. An unusually high heart rate, a sudden surge in network requests, or a sensor reading far outside its usual range may be important precisely because it is unusual. Removing it automatically could erase the event that analysts need to understand.

The catch: real data aren’t always bell-shaped

The two-standard-deviation rule depends on a key assumption: the data should be reasonably close to a normal distribution. Real datasets often have a different shape. They may be skewed, contain several clusters, or have a long tail.

Income is a familiar example. Most people earn within a broad middle range, while a smaller number earn much more. The distribution is often right-skewed, with a long tail toward high incomes. In that setting, the mean can be pulled upward by a few very large values, and a symmetric rule around the mean may not describe what counts as unusual very well.

Time-based data can also behave differently. Electricity use changes by hour and season. A value that looks extreme when compared with the annual average may be normal for a cold evening in winter. In those cases, comparing a point with the right group—such as the same hour of the day or the same season—can be more informative than comparing it with every observation.

Before using a standard-deviation threshold, it helps to inspect a histogram, a box plot, or a normal quantile plot. These displays can reveal skew, clusters, and gaps that a single summary number hides. A quick chart can save a team from applying a neat rule to messy data.

When the IQR gives a better view

For skewed data, analysts often use the interquartile range, or IQR. The IQR is the distance between the first quartile (the 25th percentile) and the third quartile (the 75th percentile). It describes the spread of the middle half of the observations and is less affected by extreme values than the mean and standard deviation.

A common screening rule marks values below Q1 − 1.5 × IQR or above Q3 + 1.5 × IQR as potential outliers. This is often shown in box plots as points beyond the whiskers. Like the two-standard-deviation rule, it flags candidates for review; it does not prove that they are incorrect.

Other approaches can help when the data have complex patterns. A median absolute deviation (MAD) score uses the median rather than the mean, making it more resistant to extreme values. In larger data-mining projects, clustering and anomaly-detection methods can identify points that differ from nearby records or from an expected pattern. These methods can be useful, though they may need more explanation than a simple z-score.

Context makes the number useful

Outlier detection is not just a math problem. It is also a question about the data’s source and purpose. A reading that is suspicious in one dataset may be ordinary in another. A high payment could be normal for a large business customer but unexpected for a small one. A network spike could be routine during a scheduled software update.

For that reason, analysts should check how the data were collected, whether different groups have different baselines, and whether values have been transformed. They should also keep a record of any points they remove or change. Silent cleanup can make later results hard to trust.

The two-standard-deviation rule offers a clear, quick way to identify values far from the mean: below mean minus two standard deviations or above mean plus two standard deviations. For data that roughly follow a normal distribution, it marks the outer portion of the range—about 5% in total. But an outlier is a clue, not a conclusion. Look at the distribution, ask what produced the value, and choose a method that fits the shape and meaning of the data.