Causal imprinting in causal structure learning

Abstract

Suppose one observes a correlation between two events, B and C, and infers that B causes C. Later one discovers that event A explains away the correlation between B and C. Normatively, one should now dismiss or weaken the belief that B causes C. Nonetheless, participants in the current study who observed a positive contingency between B and C followed by evidence that B and C were independent given A, persisted in believing that B causes C. The authors term this difficulty in revising initially learned causal structures "causal imprinting." Throughout four experiments, causal imprinting was obtained using multiple dependent measures and control conditions. A Bayesian analysis showed that causal imprinting may be normative under some conditions, but causal imprinting also occurred in the current study when it was clearly non-normative. It is suggested that causal imprinting occurs due to the influence of prior knowledge on how reasoners interpret later evidence. Consistent with this view, when participants first viewed the evidence showing that B and C are independent given A, later evidence with only B and C did not lead to the belief that B causes C.

Figure 1

Potential causal relations among events A, B, and C that participants in Experiments 1–4 had to judge by viewing contingency data.

Figure 2

Contingencies used in Experiments 1–4 for the BC and ABC blocks. The table on the left shows presence (value 1) or absence (value 0) of the three events on each of the 20 trials. The BC and ABC blocks were identical except that in the BC block, values of A were missing (indicated by grey shading in the table). Calculations of ΔP are provided for all causal relations depicted in Figure 1. The subscripts used with ΔP (e.g., ΔP_BC) indicate that it is a measure of the second factor (C) being contingent on the first factor (B). For the ABC block, ΔP for the B-causes-C relation is calculated separately for trials with A present and with A absent, indicating the lack of evidence for this relation when conditionalizing on A.

Figure 3

A summary of the descriptive predictions according to belief revision or for the ABC-ABC condition (above) and causal imprinting (below). The dotted line represents a specifically weak causal relation.

Figure 4

Pictures used to illustrate the presence and absence of events A, B, and C in Experiments 1–4. (Note: Red font used in the experiments is shown in bold.)

Figure 5

Causal strength ratings derived from the normative Bayesian analyses of the ABC-ABC and BC-ABC conditions presented in Table 1, separately for B-causes-C (B→C), A-causes-B (A→B), and A-causes-C (A→C). Sub-headings beneath the graph panels indicate for what conditions each set of predictions should be considered normative. The bounded Bayesian analysis of the BC-ABC condition may also be considered normative for the BC-ABC different tokens condition (though not the same tokens condition), but only when taking into account cognitive limitations. The bounded Bayesian analysis also shows a pattern very similar to what we would expect if causal imprinting occurs.

Figure 6

Average causal strength ratings for the three conditions in Experiment 2, separately for B-causes-C (B→C), A-causes-B (A→B), and A-causes-C (A→C).

Figure 7

Average causal strength ratings for the three conditions in Experiment 3, separately for A-causes-B (A→B), A-causes-B (A→B), and A-causes-C (A→C).

Figure 8

Average causal strength ratings for the two conditions in Experiment 4, separately for A-causes-B (A→B), A-causes-B (A→B), and A-causes-C (A→C).