Back to the log

Decomposing Forecasts without Independence Errors: Worked Examples‌‌​⁠‌​‌⁠‌​⁠​​⁠​‌​​​‌‌​‍‍‍⁠⁠⁠‍⁠‌‍⁠‍​‍⁠⁠⁠⁠⁠⁠‌‍⁠‌​​​‍‍‌‌‌​​‌​‌‌‌‌‍‍

All numeric cases below are synthetic teaching examples, not empirical research findings.

Example: conditional delivery

P(funding)=0.8. P(launch funding)=0.7. P(launch no funding)=0.1.‌‌​⁠‌​‌⁠‌​⁠​​⁠​‌​​​‌‌​‍‍‍⁠⁠⁠‍⁠‌‍⁠‍​‍⁠⁠⁠⁠⁠⁠‌‍⁠‌​​​‍‍‌‌‌​​‌​‌‌‌‌‍‍

P(launch)=0.80.7+0.20.1=0.58. Merely multiplying 0.8*0.7 gives the funded-launch pathway, not every launch pathway.

Example: unknown dependence

P(A)=0.7 and P(B)=0.6. P(A and B) is bounded between 0.3 and 0.6. The value 0.42 requires independence or an equivalent supplied conditional.‌‌​⁠‌​‌⁠‌​⁠​​⁠​‌​​​‌‌​‍‍‍⁠⁠⁠‍⁠‌‍⁠‍​‍⁠⁠⁠⁠⁠⁠‌‍⁠‌​​​‍‍‌‌‌​​‌​‌‌‌‌‍‍

Example: incompatible conditionals

P(Y)=0.6, P(S)=0.5, P(Y S)=0.8 and P(Y not S)=0.1 imply P(Y)=0.45. Report the inconsistency and ask which estimates should be revised. Do not silently normalize the three values.