Interpreting data in statistics means moving beyond raw numbers to understand what they actually tell you about patterns, differences, and relationships. The process follows a predictable sequence: start by summarizing what your data looks like, check whether the patterns hold up under scrutiny, then determine how meaningful those patterns really are. Each step builds on the one before it, and skipping ahead is one of the most common mistakes people make.
Start With Descriptive Statistics
Before running any complex analysis, your first job is to describe the data you have. Descriptive statistics give you a general sense of trends and can reveal errors, like values that fall outside an accepted range. This step includes looking at averages, frequencies, minimums, and maximums. It also helps you verify whether your data meets the assumptions required for more advanced tests.
The two most important summary numbers are a measure of center (where most values cluster) and a measure of spread (how far apart the values are). The mean works well when your data is roughly symmetrical, but it can be misleading when the distribution is lopsided. A positive skew means a long tail stretches to the right, pulling the mean higher than where most values sit. A negative skew means the opposite. In either case, the median gives you a more honest picture of the “typical” value because it splits the data in half regardless of extreme values.
Standard deviation tells you how tightly values cluster around the mean. A small standard deviation means the data points are bunched together; a large one means they’re spread out. If you’re comparing two groups and both have wide spread, it becomes harder to tell whether a difference in their averages is real or just noise.
How Data Shape Affects Your Analysis
Many statistical tests assume your data follows a roughly bell-shaped (normal) distribution. When it doesn’t, the results of those tests can be unreliable. You can check this by looking at skewness and kurtosis values. Skewness measures asymmetry. Kurtosis measures whether the data has unusually heavy tails or a sharp peak compared to a normal curve.
When your data shows a substantial departure from normality, you have two options: transform the data (taking logarithms is the most common approach) or switch to a nonparametric method that doesn’t require the normality assumption. Choosing the wrong test for your data’s shape is a quiet way to produce results that look precise but aren’t trustworthy.
Reading Box Plots and Visual Summaries
Box plots pack a surprising amount of information into a simple graphic. The box itself represents the middle 50% of your data, called the interquartile range (IQR). The left (or bottom) edge of the box marks the 25th percentile, meaning 25% of the data falls below that point. The line inside the box is the median, the 50th percentile. The right (or top) edge is the 75th percentile.
The “whiskers” extending from the box show the range of non-outlier data. Outliers are plotted as individual dots beyond the whiskers, defined as any value more than 1.5 times the IQR below the 25th percentile or above the 75th percentile. If you see many outlier dots on one side, your data is skewed in that direction. Comparing box plots side by side across groups is one of the fastest ways to spot differences in both center and spread before you run a single test.
Understanding P-Values
A p-value answers one specific question: if there were truly no effect or no difference, how often would you see results at least this extreme just by chance? A p-value of 0.04 means that if you repeated the study many times under identical conditions, and there truly was no difference, you’d see a result this large or larger about 4% of the time.
The conventional threshold is 0.05 (5%), which roughly corresponds to values more than two standard deviations from the mean in a normal distribution. Results below this threshold are typically called “statistically significant.” But this cutoff is not sacred. In genetics research, for instance, the bar is set astronomically higher, sometimes requiring p-values below 0.00000001 to account for the massive number of comparisons being made. Some statisticians have argued that even general research should use a stricter threshold like 0.005 to reduce false positives.
The American Statistical Association released a landmark statement clarifying what p-values can and cannot do. Three points are especially important for anyone interpreting data. First, a p-value does not tell you the probability that your hypothesis is true. Second, it does not measure the size or importance of an effect. Third, scientific conclusions should never rest on whether a p-value crosses a single arbitrary line. A result with p = 0.049 is not fundamentally different from one with p = 0.051, even though one is labeled “significant” and the other isn’t.
Why Effect Size Matters More Than You Think
Statistical significance tells you whether an effect is likely real. Effect size tells you whether it’s meaningful. A study with thousands of participants can produce a statistically significant result for a difference so tiny it has no practical importance.
The most widely used benchmark is Cohen’s d, which measures the difference between two group averages in standard deviation units. The traditional guidelines classify 0.20 as a small effect, 0.50 as medium, and 0.80 as large. For correlation strength, the equivalent benchmarks using Pearson’s r are 0.10 (small), 0.30 (medium), and 0.50 (large).
These benchmarks are useful starting points, but they aren’t universal. Research in specific fields often finds that real-world effects tend to be smaller than Cohen’s original guidelines suggest. In aging research, for example, empirical analysis of over 1,100 studies found that a more accurate set of benchmarks would be 0.16 for small effects, 0.38 for medium, and 0.76 for large. The lesson: always interpret effect sizes within the context of your field rather than applying a one-size-fits-all label.
How Sample Size Shapes Your Results
Sample size directly controls how precise your results are. A useful rule of thumb: the margin of error for a sample is roughly 100 divided by the square root of the sample size. For 10 people, that gives you a margin of error around 31.6%, which is enormous. For 1,000 people, it drops to about 3.2%. This is why polls with a few hundred respondents report margins of plus or minus 4 to 5 percentage points, while a study with 50 participants can’t pin down much of anything.
Small sample sizes also reduce statistical power, which is your ability to detect a real effect when one exists. When sample sizes are small and effect sizes are moderate or low, you face a high risk of Type II error: concluding there’s no effect when one actually exists. A nonsignificant result from an underpowered study doesn’t mean nothing is happening. It means the study wasn’t equipped to find it.
At the other extreme, very large sample sizes can make trivially small effects appear statistically significant. This is why it’s critically important not to rely only on the p-value and regression or correlation coefficients. Evaluate results alongside effect size and confidence intervals to get the full picture.
Interpreting Common Statistical Tests
Different tests answer different questions, but the interpretation logic follows a consistent pattern: check whether the overall result is statistically significant, then look at the details.
For tests comparing two groups (t-tests), check the t statistic and its p-value. If significant, report which group’s average is higher and by how much. If not significant (p greater than 0.05, or a confidence interval that crosses zero), you conclude there’s no detectable difference between the groups.
When comparing three or more groups (ANOVA), start with the overall F statistic. If it’s not significant, stop there. There’s no point in digging into which specific groups differ when the overall test found nothing. If the F statistic is significant, then use follow-up comparisons to pinpoint which groups differ from each other.
For correlations, check whether the r value is significant before interpreting its direction or strength. A positive r means both variables increase together; a negative r means one increases as the other decreases. Only after confirming significance should you assess the magnitude using the benchmarks above.
For regression models, the F statistic tells you whether the model as a whole predicts anything. If it does, look at the R-squared value to see how much of the variation in your outcome the model explains. Then examine each individual predictor’s p-value to see which variables contribute meaningfully.
Correlation Does Not Mean Causation
This is the most repeated warning in statistics, yet it still trips people up constantly. Two variables can move together for three distinct reasons: one causes the other, both are caused by a third variable you haven’t measured, or the relationship is pure coincidence. Correlation alone cannot tell you which of these is true.
Spurious correlations are relationships that hold up mathematically but have no logical connection. They often arise because a hidden third variable drives both trends, or simply because with enough data, some patterns will emerge by chance. Even when a genuine causal relationship does exist, correlation doesn’t tell you which direction it runs. Does stress cause poor sleep, or does poor sleep cause stress? The correlation between them looks identical either way.
Biases That Distort Interpretation
Even well-designed studies can produce misleading results when bias creeps in. Selection bias occurs when the people in your sample don’t represent the population you’re trying to study, often because of how participants were recruited or who agreed to participate. Confounding bias happens when a third variable is connected to both your cause and your effect, making it look like one thing drives another when the real driver is something you didn’t account for.
Analysis bias is subtler and more personal. Researchers naturally gravitate toward findings that confirm what they expected, sometimes overlooking data that contradicts their hypothesis. This connects to a practice the ASA statement explicitly called out: running many statistical tests on the same dataset and reporting only the ones that produced significant results. This goes by names like cherry-picking, p-hacking, or data dredging, and it’s one of the main causes of false positives in published research. Valid interpretation requires knowing how many tests were conducted and how the reported results were selected.
Reporting and publication bias add another layer. Studies with positive or significant results are more likely to get published, which means the available literature can paint an overly optimistic picture of an effect. When interpreting any statistical finding, consider not just what the numbers say, but what conditions produced them, what was tested but not reported, and whether the sample genuinely represents the group you care about.