Statistical analysis falls into a few major categories, each designed to answer a different kind of question. Descriptive analysis summarizes what already happened. Inferential analysis draws conclusions about a larger group from a smaller sample. Exploratory analysis hunts for hidden patterns. Predictive analysis forecasts what might happen next. And causal analysis tries to determine whether one thing actually caused another. Understanding these categories helps you pick the right approach for any data question you face.
Descriptive Analysis
Descriptive analysis is the starting point. Its job is to summarize a set of data so you can see what’s in front of you before doing anything more complex. It answers straightforward questions: What’s the average? How spread out are the values? How often does something occur?
The core tools of descriptive analysis are measures of central tendency and measures of spread. Central tendency tells you where the middle of your data sits, using three common metrics. The mean is the arithmetic average: add up all values and divide by the number of values. The median is the middle value when you line everything up from lowest to highest. The mode is whichever value shows up most often. Each one is useful in different situations. The mean is sensitive to extreme values (a single billionaire can skew the average income of a small town), while the median resists that kind of distortion.
Measures of spread tell you how much variation exists. Variance captures the average squared distance of each data point from the mean. Standard deviation is the square root of variance, which puts the number back into the same units as the original data and makes it easier to interpret. A small standard deviation means values cluster tightly around the average. A large one means they’re scattered.
Frequency counts, percentages, and simple charts like bar graphs or histograms also fall under descriptive analysis. None of these tools make predictions or test theories. They just help you see the shape of your data clearly.
Inferential Analysis
Inferential analysis is where you move beyond what you can directly observe. Instead of describing your entire dataset, you use a sample to draw conclusions about a larger population. Every political poll, clinical trial, and customer survey relies on this logic: measure a small, representative group and generalize the findings outward, while quantifying how uncertain that generalization is.
Hypothesis Testing
The most common framework in inferential statistics is hypothesis testing. You start with a null hypothesis, which is a default assumption that nothing interesting is going on (no difference between groups, no relationship between variables). Then you collect data and calculate a test statistic to see how well the data fits that assumption. The result is expressed as a p-value: the probability of seeing results as extreme as yours if the null hypothesis were actually true. A small p-value suggests your data is unlikely under the null hypothesis, giving you reason to reject it.
The traditional threshold for “statistical significance” has been a p-value below 0.05. That convention, however, has come under serious scrutiny. The American Statistical Association issued a landmark statement in 2016 cautioning against mechanical reliance on p-value cutoffs, and some prominent journals, including the New England Journal of Medicine and Nature, have moved toward reducing or eliminating the use of the phrase “statistically significant” entirely. Some researchers have proposed lowering the threshold to 0.005 or 0.01 to reduce false positives. Others argue for flexible, context-dependent thresholds. The current best practice is to report p-values alongside effect sizes (how large the difference or relationship actually is) and confidence intervals (a range of plausible values for the true result), rather than treating any single cutoff as a pass/fail gate.
Confidence Intervals
A 95% confidence interval gives you a range that, if you repeated the study many times, would contain the true population value 95% of the time. It’s more informative than a p-value alone because it tells you not just whether a result is likely real, but how big or small it might plausibly be. A drug study might report that a treatment lowered blood pressure by 8 points on average, with a 95% confidence interval of 3 to 13 points. That range gives you a much richer picture than a simple “p = 0.02.”
Exploratory Data Analysis
Exploratory data analysis, often called EDA, is less about testing a specific theory and more about asking “What’s going on in this data?” It’s the investigative phase where you look for patterns, relationships, and anomalies that you didn’t necessarily expect to find. EDA is especially useful early in a project when you’re still forming hypotheses rather than testing them.
The techniques are heavily visual. Box plots show the spread and skew of a variable by displaying the minimum, first quartile, median, third quartile, and maximum. Scatter plots reveal relationships between two variables. Heatmaps can highlight correlations across many variables at once. For high-dimensional data with many variables, techniques like clustering (grouping similar observations together) and dimension reduction (condensing many variables into a smaller set that captures most of the information) help you see structure that would otherwise be invisible.
One important caveat: relationships discovered during EDA are not proof of anything. Correlation found during exploration doesn’t imply causation. EDA generates hypotheses. Confirming those hypotheses requires a separate, more rigorous analysis.
Predictive Analysis
Predictive analysis uses historical or current data to forecast future outcomes. Rather than asking “why did this happen?” it asks “what is likely to happen next?” It powers credit scoring, weather forecasting, recommendation engines, and medical risk calculators.
The simplest predictive tool is regression. Linear regression models the relationship between one or more input variables and a numerical outcome (like predicting home prices from square footage, location, and age). Logistic regression does the same for yes/no outcomes (like predicting whether a patient will be readmitted to the hospital). More complex approaches, including decision trees, random forests, and neural networks, can capture nonlinear patterns that regression misses.
Predictive models are judged by how accurately they perform on new data they haven’t seen before, not on how well they fit the data they were trained on. A model that perfectly memorizes its training data but fails on new observations is overfitting. Techniques like cross-validation, where you repeatedly hold out a portion of data for testing, help guard against this.
Causal Analysis
Causal analysis tries to answer the hardest question: does changing one thing actually cause a change in another? Prediction models can identify that two things tend to move together, but they cannot tell you whether intervening on one will change the other. A model might predict that people who carry lighters are more likely to develop lung cancer, but the lighter isn’t the cause. Smoking is.
The gold standard for establishing causation is a randomized controlled experiment, where participants are randomly assigned to different conditions so that the only systematic difference between groups is the variable being tested. When experiments aren’t possible (you can’t randomly assign people to smoke for decades), researchers use observational methods that try to approximate experimental conditions. These include techniques that control for confounding variables, natural experiments where an external event creates quasi-random variation, and statistical frameworks specifically designed for causal inference. The key distinction from predictive analysis is that causal analysis is specifically structured to isolate the effect of one variable while holding everything else constant.
Choosing the Right Type of Analysis
The type of analysis you need depends on your question, your data, and how you plan to use the results. If you just need to summarize what happened last quarter, descriptive analysis is enough. If you want to know whether a new process improved outcomes, you need inferential analysis. If you’re exploring a new dataset looking for leads, EDA is the right starting point. If you need to forecast next month’s demand, you want predictive modeling. And if you need to know whether a specific change caused a specific outcome, you need a causal framework.
Your data type also matters. Whether your measurements are categorical (like yes/no or color categories) or numerical (like weight or temperature) determines which statistical tests are appropriate. Within numerical data, whether values follow a bell-shaped distribution affects the choice between parametric tests (which assume normal distribution and are more powerful at detecting real differences) and nonparametric tests (which make no assumptions about distribution but are less efficient). Time-to-event data, like how long until a machine fails or a patient recovers, requires its own specialized approach called survival analysis.
Common Pitfalls in Statistical Analysis
Even technically correct analyses can produce misleading results when the process around them is flawed. The most widely discussed problem in recent years is p-hacking: the practice of trying multiple analyses, data subsets, or variable combinations until a statistically significant result appears. Common forms include checking results partway through data collection and stopping early if significance appears, testing many outcome variables and reporting only the significant ones, dropping outliers after seeing the results, or adding and removing control variables until the p-value dips below 0.05. Research has shown that p-hacked studies produce a telltale signature: an unnatural cluster of p-values just below 0.05.
The broader problem is publication bias. Studies with positive, statistically significant findings are far more likely to be published, especially in high-profile journals. Studies with null results often sit in the “file drawer” and never see print. The combination of publication bias and p-hacking means that a portion of published significant findings are likely false positives. This is one reason the scientific community has been pushing for preregistration of study designs (declaring your analysis plan before collecting data), reporting all results rather than just significant ones, and emphasizing effect sizes and confidence intervals over bare p-values.
Multivariate Analysis
Many real-world questions involve multiple variables measured simultaneously, and multivariate techniques are designed to handle that complexity. Factor analysis takes a large number of variables and identifies a smaller set of underlying dimensions that explain the patterns in the data. If you measured twenty different personality traits, factor analysis might reveal that they cluster into five core factors. Cluster analysis works in the opposite direction: instead of grouping variables, it groups observations (people, products, cells) into relatively homogeneous categories based on their characteristics, without you needing to specify the groups in advance.
When you have multiple outcome variables to compare across groups, a multivariate extension of standard group-comparison tests lets you analyze them simultaneously rather than running many separate tests (which inflates the risk of false positives). These techniques are common in fields from marketing segmentation to genomics, anywhere the data has many moving parts and you need to find structure within the complexity.