Answering Causal Queries About Singular Cases - An Evaluation of a New Computational Model
von Simon Stephan
Datum der mündl. Prüfung:2019-02-28
Erschienen:2019-03-04
Betreuer:Prof. Dr. Michael Waldmann
Gutachter:Prof. Dr. Michael Waldmann
Gutachter:Prof. Dr. Hannes Rakoczy
Gutachter:Prof. Dr. Ralph Hertwig
Zum Verlinken/Zitieren:
http://dx.doi.org/10.53846/goediss-7323
Dateien
Name:Diss_SimonStephan_final.pdf
Size:19.0Mb
Format:PDF
Lizenzbestimmungen:
Zusammenfassung
Englisch
This thesis addresses the question of how causal queries about singular cases can be answered. Singular causation queries refer to causal connections between actually occurred events that can be localized in space and time. “Did the storm last night cause the flower pot to break apart?” or “Was it the gunshot instead of the poisoning that caused the victim’s death?” are vivid examples. Singular causation queries can be contrasted with general causation queries, which refer to causal connections between re-instantiable types (e.g., “Does smoking cause lung cancer?”). Singular causation queries are prevalent in our daily lives and important in many professional disciplines, such as the law, medicine, or engineering. But how can we assess whether an event c caused another event e? Given that causal connections are not directly perceivable, how can a causal co-occurrence of events be discriminated from a mere coincidental one? In this thesis, I propose a new computational model that is intended to help answering this question. The model is based on Cheng and Novick’s (2005) Power Model of Causal Attribution, according to which general-level causal knowledge about the potential causes’ powers (Cheng, 1997) is crucial. The causal power of a cause is the probability with which it brings about the effect. I will argue that more than that is needed: The assessment of singular causation also needs to take temporal information into account. By focusing solely on causal power, Cheng and Novick’s model assigns singular causal responsibility to a target cause whenever the cause was sufficiently powerful for the effect. It thus fails to take into consideration that causes can be preempted in their efficacy by alternative causes. To account for the problem of causal preemption, the new model combines information about causal power and temporal information. Two different types of temporal information are identified as relevant: information about causal latency, which is the time it takes a cause to unfold its power, and information about the onset difference between the potential causes. A second problem of the original model is that it relies solely on point estimates of the power parameters, and thus does not take into account that reasoners incorporate uncertainty about the underlying general causal structure and the parameters. To account for this problem of inferential uncertainty, the new model is embedded into the structure induction framework (Meder, Mayrhofer, & Waldmann, 2014). The new model was experimentally tested across two research articles (Stephan, Mayrhofer, & Waldmann, submitted; Stephan & Waldmann, 2018). The experiments revealed that the new model better accounts for people’s singular causation judgments than the original model.
Keywords: Singular Causation; Computational Modeling; Causal Learning; Cognitive Science