The idea behind the diagram is that researchers build their explanations around theoretical constructs, ideas like “trust” or “political polarization,” rather than something you can point to and measure directly. These constructs are the building blocks of an explanation. But because a construct like “trust” is an abstraction, researchers need a way to connect it to something real and measurable. That step is called operationalization. Trust, for example, can’t be observed directly, so a researcher might operationalize it as answers to a short survey, or as how often someone shares posts from a particular source. Once a construct is operationalized this way, you can measure it and use the resulting data to test your hypothesis. In short: explanatory modeling means applying statistical models to data in order to test causal hypotheses about theoretical constructs.

Predictive modeling is a different job entirely which means applying a statistical model or data-mining algorithm to data in order to predict new or future observations. Here the goal isn’t to explain why something happens, it’s to correctly guess an output based on a set of inputs. And unlike explanatory modeling, prediction doesn’t require a theory of why. A weather forecaster is a good illustration: they don’t need a complete theory of atmospheric physics to tell you it will probably rain tomorrow. They just need a model that’s proven reliable under similar conditions before. In social science, this idea is most easily illustrated with time series analysis, where researchers forecast a future value based only on values observed earlier in time.

For a long time, prediction got less respect in social science than explanation did, mostly because researchers put their energy into finding causes for social phenomena. That’s shifted, and scholars like Hofman, Sharma, and Watts (2017) argue that the computational tools now available to social science should let prediction play a bigger role. But they also flag three problems that need fixing first. First, the field needs agreed-upon ways to measure whether a prediction is actually good, what the authors call a “common task framework” which is essentially a shared benchmark problem and dataset that lets different research teams compare results on equal footing, the way computer-vision researchers all test their models against the same image dataset. Without something like that, researchers should at least be upfront about whether their analysis is exploratory (looking for patterns) or confirmatory (testing a specific claim). Second, we need a better sense of prediction’s actual limits, some things are genuinely unpredictable no matter how much data you throw at them, and it matters to know where that ceiling is, rather than assuming more data always means a better forecast. Third, there’s a common belief that simple, easy-to-interpret models and complex, highly accurate models are locked in a trade-off where you always sacrifice one for the other. That trade-off turns out to be smaller than most people assume: simple models don’t automatically generalize better than complicated ones, and “understanding” can mean either making sense of something (explanation) or successfully accounting for what actually happened (prediction).

Descriptive modeling is the third category, and its job is simply to summarize and represent data well, nothing more. It differs from explanatory modeling because it makes no causal claim, and it differs from predictive modeling because it isn’t trying to forecast anything. A typical example of description is a study reporting that posts using angry language get shared twice as often as calmer ones. That’s a descriptive claim, it doesn’t mean anger causes sharing, and it isn’t trying to predict which individual post will go viral. It’s just describing a pattern in the data.

This matters because a lot of published social science actually lives in a gray area. Even though there’s a well-developed toolkit for causal inference, a large share of papers present associations grounded in theory without ever claiming causation. That’s the whole point of the cliché “correlation is not causation”: path models, regression, and structural equation models (SEM) don’t necessarily let you compute a causal effect in the strict statistical sense, but they can still offer an explanatory, theory-grounded association. There’s no hard rule for telling whether a paper is explanatory or merely descriptive, but seeing a path diagram or SEM model is usually a signal that the authors are making a relational — sometimes causal — claim grounded in theory.

With that groundwork in place, the rest of this series introduces a set of methods meant as a starting point for researchers interested in tools for building explanations and predictions. These methods come from a few different fields: causal inference, time series analysis, and information theory. Table 1 summarizes the key features of each; I’ll walk through each one, one post at a time, later in the series.