Differential weighting of illness signs and symptoms has surfaced recu
rrently in psychiatric nosology. Six alternately weighted algorithms f
or diagnosing alcohol dependence in accordance with the Diagnostic and
Statistical Manual of Mental Disorders (4th ed.; DSM-IV; American Psy
chiatric Association, 1994), based on statistical, unit, rational and
random criterion weighting systems, were used to predict an array of c
oncurrent validators and 6-month drinking outcomes in a regional clini
cal sample of 365 participants. Comparable predictive efficiency acros
s all algorithms, including the randomly weighted versus statistical b
est-fit model, was observed. Further analyses and geometric modeling s
uggested that this was due to the extremely high internal consistency
of the DSM-IV criteria. An alternative strategy that favors factoriall
y complex, less homogeneous criteria was used to develop an experiment
al DSM-IV algorithm from an array of 39 candidate criteria. This algor
ithm had extremely low internal consistency, high difficulty, and comp
lex factor loadings. Differential weighting of its criteria produced a
good range of efficiencies, predictive power for rational models exce
eding the random weight model, and a best-fit algorithm with substanti
al surplus predictive power. These results illustrate an emerging conf
lict in nosology between 2 opposing trends: a press for the promulgati
on of criterion arrays with high internal consistency and a clear desi
re to assign some criteria extra weight for prognosis or decision maki
ng. Both cannot be had in the same algorithm. An alternative approach
emphasizing diagnostic criteria with complex structures can satisfy th
e multiple demands of brevity, validity and weighting performance.