A method is described for simultaneously simulating maximum and minimu
m temperatures and daily precipitation amounts in a physically consist
ent manner. The method ''chains'' actual days from a historical datase
t by defining a ''discriminant space'' using multiple discriminant ana
lysis. A set of analogous days is selected from discriminant space usi
ng a nearest-neighbor search. The next day in the chain is the day sub
sequent to a randomly selected day from the set of analogous days. The
method was tested on data for Tucson and Safford, Arizona. A high deg
ree of similarity between the simulated and observed data was found. A
slight tendency to underestimate the variance of monthly average temp
eratures was noted. The distribution of monthly temperature extremes w
as quite well reproduced with the exception of a tendency to be conser
vative in predicting the warmest minimum temperatures and the coolest
maximum temperatures. Very little difference between the simulated and
observed distributions of diurnal temperature range was found. The me
dian and 90th percentile of monthly precipitation totals were well rep
roduced. A tendency to underestimate the frequency of dry months was n
oted. The frequency of runs of wet and dry days of different lengths w
as found to be not significantly different for the simulated and obser
ved data. Reproduction of wet-day run frequency for the first-order mu
ltivariate chain model was comparable to that using a two-state, first
-order Markov chain.