Why Your Descriptive Stats Keep Looking Wrong

I've seen the same issue come up repeatedly in data work. People calculate mean, median, and standard deviation and then just report numbers without thinking about what those numbers actually mean in context. The gap between knowing the formulas and being able to use them correctly is where most mistakes happen. Measures Of Central Tendency And Dispersion Practice isn't just about crunching numbers from a dataset. It's about understanding whether your summary statistics are lying to you. A mean of 50 and a standard deviation of 2 look precise until you realize the data is bimodal, which neither number tells you about. That's the core problem.

What You Actually Need To Know Before Starting

Start with whichever measure matches your distribution shape. If the data is roughly symmetric with no extreme outliers, the mean works fine. If it's skewed or has outliers, the median is your starting point, not a fallback option. I used to default to the mean for everything until I analyzed a response time dataset where a handful of values were 40 times higher than the rest. The mean was completely unrepresentative. The median made sense immediately. For dispersion, standard deviation assumes a roughly normal distribution. When that assumption breaks, interquartile range gives you something more stable. Range is almost useless as a standalone measure because it only looks at two points and is wildly sensitive to outliers.

Calculating Central Tendency Correctly

The arithmetic mean is the sum of all values divided by the count. That part is straightforward. The mistake people make is treating it as equally valid across all data types. For interval and ratio data, it works. For ordinal data like Likert scales, it can produce results that don't correspond to any actual response category. I've worked on surveys where the calculated mean of a 5-point scale came out to 3.47, which no respondent actually chose, and presenting that as a meaningful finding without context was misleading. For the median, order the data and find the middle value. If there's an even number of observations, average the two middle values. With a dataset of 1,000 entries, I sort them and look at positions 500 and 501. That's it. No complicated procedure. The mode is the most frequent value. In practice, it's more useful than people think, especially for categorical data. One project I was on involved customer complaint categories, and the mode identified the single most common issue type faster than any other measure. Mean and median don't exist for nominal data, so the mode was the only option that made sense.

Measuring Dispersion With Actual Purpose

Standard deviation is the square root of the average squared deviation from the mean. Variance is that same calculation without the square root. Both describe spread around the mean. The key distinction is units. Standard deviation is in the same units as your data, which makes it interpretable. Variance is in squared units, which is why it's rarely reported directly in final summaries. When I analyze income data, standard deviation alone doesn't tell me much without the mean for context. That's where the coefficient of variation comes in. It's the standard deviation divided by the mean, expressed as a percentage. It lets you compare variability across datasets with different units or vastly different means. An income distribution with a standard deviation of $30,000 might seem large until you divide by a mean of $75,000 and get a CV of 40%. Now you can compare it to another dataset where a CV of 15% tells a different story about relative spread. Interquartile range is the difference between the 75th and 25th percentiles. It captures the middle 50% of the data. I prefer it for skewed distributions because it ignores the tails entirely. There was a dataset of transaction amounts where the IQR was $45 and the standard deviation was $320. The IQR told me that most transactions clustered tightly while the standard deviation was inflated by a small number of very large purchases. Reporting both would have given a complete picture, but most people only report one.

What Goes Wrong In Real Work

The most common error I see is reporting only the mean and standard deviation for non-normal data. A recent analysis of server response times showed a mean of 200 milliseconds with a standard deviation of 150. On paper that looks reasonable. The actual distribution was heavily right-skewed, and most responses were under 100 milliseconds with a long tail of slow requests pulling the mean up. The median was 85 milliseconds, which was closer to what users actually experienced. I had to explain this distinction to stakeholders who only cared about the mean because it was the number they were familiar with. Another frequent issue is using range in small samples. With only 10 data points, the range depends entirely on whatever two extreme values happen to appear. It has no stability. I found this with a quality control process where the range varied wildly between batches of 12 items, making it impossible to set consistent control limits. Switching to IQR stabilized the analysis immediately.

When These Measures Break Down Completely

Mean and standard deviation become unreliable with heavy-tailed distributions. Cauchy distributions are the textbook example, but in practice, financial return data and network traffic patterns often exhibit similar behavior. In these cases, the mean doesn't converge reliably as sample size increases, and the standard deviation becomes increasingly unstable. I encountered this with latency measurements from a distributed system where occasional network events produced extreme values that didn't disappear with more data. The median and IQR remained stable across samples, while the mean and standard deviation continued to shift unpredictably. Using robust measures like the median absolute deviation instead of standard deviation resolved the issue. For multimodal data, no single measure of central tendency captures the structure. The data might have two distinct clusters, and reporting one mean hides that entirely. In a customer segmentation project, the age distribution had peaks at 25 and 55, and the mean of 40 described nobody. I split the analysis by segment and reported separate statistics for each group, which was the only way the numbers reflected reality.

A Practical Workflow

Plot the data first. Histograms, box plots, or density plots reveal structure that summary statistics alone will miss. I typically run a visual check before calculating anything. This takes about 30 seconds in most tools and prevents most downstream errors. Check for outliers next. Extreme values affect the mean and standard deviation disproportionately. In a dataset of 500 records, I flag anything beyond 1.5 times the IQR from the quartiles using the standard box plot rule. Whether to keep or remove those points depends on whether they represent data entry errors or genuine extreme values. I rarely remove outliers without documenting the decision and running the analysis both ways. Calculate central tendency using the appropriate measure for your distribution. Calculate dispersion using at least two measures. Report the mean with standard deviation for normal data, the median with IQR for skewed data, and the mode for categorical data. If the distribution is unusual, report multiple measures and explain why.

The actual calculation process in most spreadsheet software takes under two minutes per dataset. The time savings come from getting it right the first time instead of catching misinterpreted results later. I've had stakeholders push back on median-based reporting, but the numbers never argue with you. They just sit there and show what's actually happening.