assume the ravens call a particular pattern iff it rains the next day.
iow, P(raven's call|rain)P(raven's call|¬rain)=a bajillion, thus observing the raven’s call is strong evidence u ought to hv an umbrella rdy for tomorrow.
“raven’s call” is therefore a v good predictive var.
but bribing the ravens to hush still might not hv any effect on whether it actually rains tomorrow.
it’s therefore a v bad causal var.
it cud even be the case that, up until now, it never not rained unless the raven’s called, and intervening on the var cud still be fruitless if nobody’s ever done that bfr.
for systems u hv 100% accurate & 100% complete predictive maps of, u may still hv a terrible causal map wrt what happens if u try to intervene in ways that take the state of the system out of the distribution u’v been mapping it in.
how then do u build good causal maps?
ig u can still improve ur causal maps wo trial-and-error (empirically testing interventions) if u just do predictive mapping of the system, and u focus in on the predictive power of the simplest vars, and do trial-and-error in ur simulations. or smth.
this is rly good. summary of what i lurned:
assume the ravens call a particular pattern iff it rains the next day.
iow, P(raven's call|rain)P(raven's call|¬rain)=a bajillion, thus observing the raven’s call is strong evidence u ought to hv an umbrella rdy for tomorrow.
“raven’s call” is therefore a v good predictive var.
but bribing the ravens to hush still might not hv any effect on whether it actually rains tomorrow.
it’s therefore a v bad causal var.
it cud even be the case that, up until now, it never not rained unless the raven’s called, and intervening on the var cud still be fruitless if nobody’s ever done that bfr.
for systems u hv 100% accurate & 100% complete predictive maps of, u may still hv a terrible causal map wrt what happens if u try to intervene in ways that take the state of the system out of the distribution u’v been mapping it in.
how then do u build good causal maps?
ig u can still improve ur causal maps wo trial-and-error (empirically testing interventions) if u just do predictive mapping of the system, and u focus in on the predictive power of the simplest vars, and do trial-and-error in ur simulations. or smth.