Introduction

Graphical causal models (GCM's) are directed acyclic graphs (DAG's) that model causal relations between variables. The variables are represented by the nodes in the graph; the causal relations between them are represented by the graph's directed edges, that is, arrows and their direction connecting the nodes; and the absence of causal relations is represented by the absence of edges connecting the nodes. More precisely, the absence of an arrow from a node X to a node Y means that the causal effect of X on Y is certainly zero. The presence of an arrow indicates that the causal effect may or may not be zero. GCM in causal inference should be viewed as assumptions about the causal relations that were or were not operating in reality when the data were collected.

For instance, consider the following graph:

It states that:

  1. X causally affects Y and Z (and these effects may be zero).
  2. Z causally affects Y (and this effect may be zero) but not X (i.e., this effect is zero for sure).
  3. Y does not affect X or Z (i.e., these effects are zero for sure).
  4. There is no unobserved common cause between these variables.

On the other hand, consider these graphs:

This is similar to the above example with one exception. They are stating that:

  1. X causally affects Y and Z.
  2. Z causally affects Y but not X.
  3. Y does not affect X or Z.
  4. U is an unobserved or latent factor that affects both Z and Y.

In the default notation of causalinf, both the left and right graphs above are equivalent to each other. The dashed bidirected arc and the dashed node represent the unobserved/latent common cause (U) of Z and Y.

The assumptions embedded in the graph as a whole are not testable. For any graph, if one or more of its assumptions are incorrect, one of the following things can happen:

  1. Nothing, and the assumption is inconsequential for the estimation of the causal effect;
  2. The causal effect remains identifiable, but the estimator produced can be biased;
  3. The causal effect may change from identifiable to non-identifiable.

Which case occurs depends on the particular situation. But even if the assumptions as a whole are not testable, given a graph and a variation of it, say by removing, adding, or changing the directions of arrows, it is possible to compare which implications would follow by conducting an identification analysis in each instance.

The causalinf package provides functions to conduct causal inference analyses for any given graph. It is possible to:

  1. Create graphical causal models (GCM).
  2. Visualize them.
  3. Check the plausibility of the assumptions of causal relations embedded in the graph
  4. Conduct identification analysis to check if a causal effect is identifiable, given the graph, and determine which identification strategy can be used to estimate the causal effect if it is identifiable.
  5. Estimate and make inferences about causal effects when it is identifiable.
  6. Report the results.

As of now, causalinf does not provide functionalities to search for the graph that best matches the data, which is often called causal search.