Many of the experiment on children's causal inferences have also been done with adult participants, using a devise similar to the Blicket Detector . In addition, one of the strengths of the causal graphical model framework is that it provides algorithms for learning causal structure from observing events co-occur, but also from observing the results of interventions of events. Can adult learners recognize the causal structure among a novel set of events based on the ability to make interventions? I have found that adults' ability to learn causal structures is better when they can intervene on events, rather than simply observing the events co-occur.
My recent investigations have been on the source of intervention data. Many causal learning algorithms do not differ between a learner generating interventions himself and observing the result of another's interventions. However, at least on a qualitative level, the Bayesian approach does, since it emphasizes that each learner constructs an individual hypothesis space. If a learner can generate their own interventions, they can test their own causal hypotheses. If a learner observes another generate interventions, with no control, they may not have the same hypothesis space, and thus the data might not be as effective. I have found that causal learning is superior when learners can generate their own interventions as oppose to observe another intervene, even though the data are identical. Further, learning from one's own interventions is superior to being sat at a computer and forced to make those identical interventions.
I have also used this paradigm to investigate adults' ability to make inferences about hidden causes. I've found that with the right kinds of information, adults are no better or worse at learning causal structures with hidden events than structures with all observable events. Further, under the right circumstances, I've found that learning models with hidden causes shares similar properties as making category judgments.