Numb3rs Season 3 Episode 14: Take Out

In this episode Charlie uses the mathematical theory of outliers to find anomalous money transfers.


Say a researcher is attempting to formulate theories about the way that people move on a particular pedestrian walkway. Our imaginary researcher measures the speed of people at set time intervals and gets as data points the following speeds (all in miles per hour): 0, 2, 2, 0, 2, 3, 1, 2, 0, 2, 3, 1, 2, 2, 2, 17. One of these numbers is not like the others. The 17 mi/hour data point is much higher than one would naively expect to see, simply from the collection of the other data points. Such a data point is termed an outlier.

When compiling the data, the outlying data point puzzles our researcher, and she goes back to her notes and realizes that the 17 mi/hr "pedestrian" was, in fact, the only bicycler she saw that day. Situations like this are really the essential part of labeling a data point as an outlier - an outlier indicates that certain assumptions inherent in a model do not hold or that a given theory is not true in a particular situation (here, our imaginary researcher assumed that the only mode of transportation on the pedestrian walkway would be walking or perhaps jogging). Outliers can oftentimes suggest directions for improvements to working models.

Note that in very large samples, a small portion of data points are expected to be far away from the mean just by random chance. For instance, in a normal distribution (a distribution arising from a bell curve), about 4% of a randomly selected sample will not be within two standard deviations of the mean, and about 0.2% of the sample will not be within three standard deviations of the mean. So in a large sample of several thousand data points, a dozen or so of these data points might be more than three standard deviations from the mean, and when plotted, they may seem isolated. However, in this context, these data points are not considered outliers because they arise from an expected amount of variation in the population and they do not indicate that any models or theories have broken down.

Normal Distribution with mean and standard deviation :

Origins of Outliers

There are many different possible causes for an outlier:

Activity 1: Identifying Outliers

In our example above, 17 was obviously an outlier, but what if instead of 17, the maximum value had been 9, or 6, or 5? Would you still identify the extremal value as an outlier? Where does one draw the line?

Another way that the line can be blurred is if there are fewer "normal" data points. In our example above example, there were 15 data points between 0 and 3. What if we had only 3 data points in that range? Would it make the data point 17 stand out more or less? Would you still consider it to be an outlier?

Thus far we have used only subjective judgements to determine if a given data point is labeled an outlier. Such subjective judgements are necessarily a bit fuzzy, and we will have some ambiguity at the dividing line. Before you go on, try to devise a quantitative or algorithmic means of identifying outliers. If you have some trouble, the next section on visualizing data may be suggestive.

Visualizing Data

There are many ways to visualize a list of numerical values to potentially see outliers. We will discuss three possible visualizations: box and whiskers plots, histograms, and cumulative distribution plots.

Box and Whiskers Plot

One way to visualize data is a box and whiskers plot . A box and whiskers plot is a linear drawing which consists of a rectangle with a vertical line inside it and two "whiskers" coming off of the sides which each have a short vertical segment at the end, as in the following:

In these plots, the vertical ticks, from left to right, indicate, in order, the minimum data point, the 25th percentile, the median, the 75th percentile, and the maximum data point. These are indicated below:

In the above plot, the 0th, 25th, 50th, 75th, and 100th percentiles are, respectively, 9, 18, 35, 48, and 65.

Activity 2: Box and Whiskers Plot

  1. Poll your friends for their heights and record the data in a list
  2. Sort the list of heights from smallest to largest
  3. Identify the 0th, 25th, 50th, 75th, and 100th percentile values
  4. Make a box and whiskers plot from your data
  5. Are there any outliers in your data? What is your criterion for an outlier? Does an outlier indicate a data value in a different paradigm? Does it significantly alter your box and whiskers plot to remove the outliers?
  6. Do you think there is a way to use box and whiskers plots to identify outliers?


Another way to visualize a set of numeric data points is a histogram. Here, the range of possible data points is partitioned into a finite number of contiguous intervals, and the number of data points in each interval is represented by the height of a rectangle whose base is that interval. By convention, intervals include their left-hand endpoint, but not their right-hand endpoint, so a data point exactly on a number dividing two intervals will contribute to the right interval, not the left. The following is an example of a histogram:

Choosing Intervals

Sometimes choosing how to divide the possible data values into intervals is difficult. The smaller the intervals, the more detail the graph shows about precisely where data points are accumulated. However, smaller intervals also increase the amount of noise in the graph. That is to say that random chance is more likely to make some rectangles larger and some smaller when the underlying population has no such differences. Lengthening the intervals tends to smooth out these random variations.

Most histograms are constructed with evenly spaced intervals. This is to preserve the correspondence between the area of the rectangles and the portion of data points which fall in that interval (when the intervals are evenly spaced, the proportion of the area of the graph that is in an interval is precisely the proportion of the data points that fall in that interval). However, sometimes if there are many data points in a particular region where there is much statistically significant detail and fewer data points in other regions, it can be advantageous to plot a histogram with unevenly spaced intervals. The heights of the rectangles in such graphs are meaningless, and instead the areas of the rectangles retain their correspondence to the number of data points in the interval. The height of the rectangle would be the number of data points in the interval divided by the length of the interval and possibly multiplied by some scaling factor.

Even when the intervals are evenly spaced, the choice of their length can accentuate certain trends or hide others. Consider the following histogram of the number of lost library books at a small university library over the last nine years:

In the above plot there are three intervals with three years in each interval, and it appears that the number of lost books is steadily decreasing. The students are learning to be more responsible! To see more detail on this trend, we plot the same data, but instead dividing into nine intervals, one for every one of the last nine years:

By changing the scale of our plot, we've uncovered a disturbing trend in recent years of an increasing number of lost library books. There are many ways to accentuate the data you want to show or to lie with statistics.

Activity 3: Histograms

  1. Use your data from Activity 2 to create a histogram.
  2. How did you choose your sub-intervals? Do you think that the plot will alter significantly if you change your sub-intervals?

Cumulative Distribution Plots

Cumulative distribution plots are superficially similar to histograms. The difference is that the height of a given point in a cumulative distribution plot corresponds not to the number of data points in some interval around a point, but rather to the number of data points less than or equal to that value. Cumulative distribution overcome a major shortcoming of histograms, that of having to make a choice of the sub-interval size.

Roughly speaking, the slope of the cumulative distribution near a point is roughly the proportion of data points near that value. The following is the cumulative distribution plot of the standard normal curve:

A cumulative distribution plot makes some properties of a data set easier to see and makes other harder to see. For instance, it is obvious from the above plot that exactly 50% of the data in a standard normal curve is less than 0. Also, by subtracting, we can see that about or 82% of the values in a standard normal distribution are between and .

Activity 4: Cumulative Distributions

  1. Use your data from Activity 2 to create a cumulative distribution plot.
  2. Outliers can be harder to see with this type of plot. What are the tell-tale signs of the presence of an outlier?

Defining Outliers

Earlier, we loosely spoke of outliers as being unlike other values in the data set. This definition is adequate for many uses, but it is imprecise in the extreme. We will describe here two precise ways of defining outliers. Keep in mind that any judgement as to whether a data point is an outlier is subjective and situationally dependent. That is to say that some data point in a data set can be an outlier or not depending on the nature of the experiment. While mathematical techniques can be utilized to identify potential outliers, outlier is not a mathematically defined term.

Interquartile range

Given a data set, define the following variables, called the quartiles:

The four intervals , , , and , called the quartiles, each contain roughly one quarter of the data points. The interval contains about half of the data points, and it contains the most central half. The length of this interval, the distance , called the inter-quartile range, is a good measure of how spread out the data points are while remaining immune to change as a result of extreme values.

Choose some constant factor . Then we can define points as being outliers if they are more than greater than or more than less than . Equivalently, outliers are points that are outside the interval . The choice of is entirely subjective, but for many applications is used. However, depending on the context, values from to might be reasonable. Why wouldn't or smaller be a reasonable constant factor to use?

Activity 5: Interquartile Ranges

  1. Use the interquartile range method to identify outliers in the pedestrian speed data and also in your data from Activity 2.
  2. Larger values of will identify fewer data points as being outliers. For any given data set, there is a value of so that the inter-quartile method will not identify any outliers. What is the largest value of so that the data point of 17 in the pedestrian speed data is identified as an outlier? This number can give some quantitative measure of how much a given data point diverges from most of the other data points.
  3. We mentioned earlier that choosing , though common, is entirely subjective. However, there is another arbitrary choice that has been made: the choice to base our range on quartiles. How could we formulate a criterion for identification of outliers based on quintiles, deciles, or arbitrary percentiles? Apply your new model to both of the data sets.

Problems With Interquartile Range Method

Unfortunately, the interquartile range method of identifying outliers is a rather blunt instrument and has some very serious liabilities. In many types of populations, the variation that we see in a measured quantity comes from myriad small factors. Based on random chance, most of these factors cancel each other out, and what does not cancel out and is left over is the observed variation of the sample. Rarely, there are samples where more of these factors do not cancel each other, and these samples correspond to extreme data points. These data points, while possibly very extreme, are not outliers in the sense that they do not represent an error in measurement or a distinct paradigm or theory governing that sample. These extreme values are a legitimate part of the population and very likely should not be thrown out in a statistical analysis. We obviously need more refined methods to identify outliers.

Rolling 1000 dice can be a lot of work. You can speed this up by using an online die roller such as the Konkret Dice Roller. This web utility will roll however many dice you choose, sum their outcomes, and repeat the action however many times you choose.

Activity 6: Spurious Outliers

Consider an experiment that consists of rolling ten 6-sided dice and recording the sum of their values, a number ranging from 10 to 60. The sum of six dice is not normally distributed, but it is very close.
  1. Run one hundred trials of this experiment.
  2. Find the first and third quartile values, and construct the interval of non-outlier values using a value of .
  3. Do you have any outliers? If so, do you think that they are they the result of measurement error or of certain dice rolls being different on a fundamental level? What is an alternative explanation for these supposed outliers?

Chauvenet's Criterion

Chauvenet's Criterion is a partial answer to the fundamental problem of the interquartile criterion for outlier identification. This method of identification attempts to take into account precisely how unlikely a given data point is, making some reasonable (though not universal) assumptions about the underlying distribution of the population.

The assumption underlying this criterion is that the population is normally distributed. To execute the test, one first calculates the mean and standard deviation of the sample. To test if a particular extremal value is an outlier, one calculates the probability that, given a normal distribution with the same mean and standard deviation, a randomly selected data point would be that far away from the mean or farther. This probability is multiplied by the total number of data points in the sample. If the resulting product is less than , then the point is deemed an outlier. This extreme data point is eliminated from the sample, and the test is repeated with one fewer data points.

Activity 7: Chauvenet's Criterion

Perform Chauvenet's Criterion to test for outliers on the data collected from the dice rolls in Activity 6. Is this a better test for outliers than the one used earlier?

Problems with Chauvenet's Criterion

If the underlying population distribution differs from normal, then Chauvenet's Criterion can give false positives or false negatives. If the actual distribution has greater kurtosis (the fourth standardized moment about the mean) than a normal distribution, then Chauvenet's criterion will identify these extreme data points as outliers, when in fact they are unremarkable points. If, instead, the distribution has a smaller kurtosis than a normal distribution, then Chauvenet's criterion will tend to fail to identify potential outliers.

Activity 8: Problems with Chauvenet's Criterion

Following the same pattern as Chauvenet's Criterion, can you construct a criterion for identifying outliers that works with non-normal distributions?
Questions? Comments? Email me: