The best measure of center depends on the shape of your data and whether extreme values are present. For symmetric data with no outliers, the mean works well. For skewed data or data with outliers, the median is almost always the better choice. And for categorical data or distributions with two peaks, the mode tells you the most useful story.
Choosing the right measure matters because picking the wrong one can misrepresent what’s “typical” in your dataset, sometimes dramatically. Here’s how to make that call for different situations.
How Data Shape Drives the Decision
When data is symmetric and bell-shaped, the mean, median, and mode all land in roughly the same spot. In that case, the mean is the standard choice because it uses every value in the dataset and works well with further statistical calculations.
The picture changes when data is skewed. In a skewed distribution, the mean gets pulled toward the tail, away from where most of the data actually sits. If your data is skewed to the right (a long tail stretching toward higher values), the mean will be higher than the median. If it’s skewed to the left, the mean drops below the median. Either way, the mean stops representing what’s typical. The median, which simply finds the middle value when everything is lined up in order, stays put regardless of how stretched out one tail is.
A quick way to check: if the mean and median are close together, your data is fairly symmetric and the mean is fine. If they’re noticeably different, your data is skewed and the median is the safer pick.
Why Outliers Distort the Mean
The mean is calculated by adding every value and dividing by the count, which means a single extreme number can drag it far from the center. Consider a simple example: in a dataset where most values cluster around 10 to 30 but one value is 500, the mean jumps to 69.3 while the median stays at 22. The median reflects where the data actually concentrates. The mean reflects a number that doesn’t describe any real pattern in the data at all.
This happens constantly with real-world data. Income is the classic case: a handful of extremely high earners pull the mean income well above what most people actually make. Home prices in a neighborhood work the same way, as do response times, medical costs, and social media engagement metrics. Any dataset where a few values can be dramatically larger (or smaller) than the rest will produce a misleading mean.
The median is called “resistant” or “robust” because changing an extreme value, or even removing it entirely, barely moves the median. It only cares about the position of values, not their magnitude.
When the Mode Is the Right Choice
The mode, the most frequently occurring value, is the only measure of center that works for categorical data. If you’re looking at favorite colors, most common shoe sizes sold, or the most popular response on a survey, neither the mean nor the median makes sense. The mode tells you which category dominates.
The mode also becomes important when data has two distinct peaks, called a bimodal distribution. Imagine plotting the heights of a mixed group of adults and children. You’d see two clusters rather than one smooth curve. In that situation, the mean and median both fall between the two peaks, in a valley where very few data points actually exist. Neither number describes a “typical” value. The mode (or modes, since there are two peaks) at least points you toward where the data actually concentrates. When you encounter bimodal data, it’s often a signal that you’re dealing with two separate groups that should be analyzed independently.
Matching the Measure to Common Data Types
Some practical guidelines for situations you’ll encounter frequently:
- Survey ratings on a 1-to-5 or 1-to-10 scale: The mean works if responses are roughly symmetric. If most people rate something a 4 or 5 and a few rate it a 1, the median better reflects the group’s opinion.
- Incomes, home prices, or costs: Almost always right-skewed. Use the median.
- Test scores in a large class: Often approximately symmetric. The mean is appropriate unless a cluster of zeros or perfect scores creates skew.
- Time-based data (wait times, response times): Usually right-skewed because times can’t go below zero but can stretch very high. The median is typically more informative.
- Categorical data (favorite brand, most common diagnosis): Use the mode. Mean and median don’t apply.
The Trimmed Mean as a Middle Ground
Sometimes you want the mathematical properties of the mean but need protection from outliers. A trimmed mean offers a compromise. It works by sorting your data, removing a fixed percentage of the highest and lowest values, and then calculating the mean of what remains.
A 20% trimmed mean is the most commonly recommended version. If you have a dataset of 10 values, you’d remove the two smallest and two largest, then average the remaining six. Research published in the International Journal of Medical Education demonstrated this with a dataset where the regular mean was 10 but the 20% trimmed mean was 5.1, much closer to the values most people actually reported. The trimmed mean achieves nearly the same statistical power as the regular mean when data is normally distributed, while performing far better when it isn’t.
You won’t encounter trimmed means in most basic statistics courses, but they’re worth knowing about for real-world data analysis where distributions are messy and you need a single representative number.
A Simple Decision Process
Start by looking at your data. Plot it if you can, even a rough sketch helps. Then ask three questions in order:
First, is the data categorical? If yes, use the mode. Mean and median don’t apply to categories.
Second, is the data symmetric with no extreme outliers? If yes, the mean is your best option. It uses all the information in your dataset and plays well with other statistical methods.
Third, is the data skewed or does it contain outliers? If yes, use the median. It resists the pull of extreme values and better represents the center of where your data actually falls. If you need something closer to a mean for computational reasons, consider a 20% trimmed mean instead of the standard version.
And if your data has two peaks, report both modes and consider whether you’re actually looking at two different groups mixed together. No single number will represent bimodal data well.