The proper distribution to use, depends on the real
distribution.
For example a process feature (say, sheet resistance)
which has +/- 3 sigma limits based on large lot sample
history is probably a good bet for a gauss() (normal)
distribution model.
That same process, if acceptance limits are closer-in
(say, process scatter runs to +/-30% but you insist
to screen to +/-10% for production) may well look
more like "uniform" because you're throwing away the
"tails".
And many attributes tend to follow other distributions
(log-normal is especially common).
For a real, relevant Monte Carlo you also need the
correlation coefficient matrix (for example, TOXP and
TOXN would be highly correlated and not allowed to
wander apart; beta and Early voltage, highly anti-
correlated and no point simulating an op amp with
both gm and Rout impossibly high at once). Param
interactions can make screwy results if you let all
of the model params that matter, be random and
entirely independent.