Nancy Cartwright has argued that directed acyclic graphs are such poor models of causal structure that they should not be used as guides to policy. A directed acyclic graph (DAG) is a collection of dots and lines where the dots represent variables and the lines represent how these variable influence each other. These graphs should satisfy the Markov Condition. Roughly speaking, the Markov condition states the lines are a usefully complete set of relations between the various variables. Hence, someone looking at a DAG should be able to tell when useful interventions could be made and what there effects might be. If one had a graph for the operation of a power generating dam, for instance, they should be able to see that increasing the rate at which water flows through the dam would have influence on area wild life both because of increased flow down stream and reduced water levels behind the dam.

Those influences that aren't captured in the graph are generally assumed to be meaningfully modeled by error terms, or variables that capture the unrepresented influences on each variable without explicitly listing what these different variables are. The Markov condition however, may be violated if these error terms aren't independent. That is, if the implicit influences on different variables influence each other. There are various methods for controlling this sort of problem. Obviously, these background dependencies can havemore profound implications for social situations where there are likely to be far more implicit or background dependencies.

However, this does not free us from the responsibility of creating well-informed social policy. In order to do this, we need some was of talking and reasoning about what effects various interventions may have. It seems like being able to draw dots for the various points at which interventions at which the various interventions could be made is a simple and effective method of communicating about these topics.

These graphs can also be used for some inferential purposes and, the big pay off, they are formal enough to be used in the design of computer code.   ¶ #

0 comments Post a Comment