Creating Robust Results
We did not know the best parameter values, nor their distribution, so we
assigned a uniform distribution for each parameter and used a Monte
Carlo approach so that we could produce more robust results (Paxton et
al 2001). We randomly sampled 10,000 values of the six parameters that
were used in the algorithm (Table 1). These varying parameters allowed
us to simulate many combinations of parameter values. Therefore, each
iteration of the algorithm produced a different estimate of our focus
variables (e.g. FSC, PT). Thus, we
created a rule set to decide how to summarize these Monte Carlo
iterations. We narrowed the available pool of iterations for each single
precipitation event based on first, how frequently a distinct
precipitation event was detected, and second, how often that unique
event exhibited a chemograph response.
Subsequently, we considered precipitation events highly likely to
represent the reality of conditions and catchment characteristics only
if the algorithm detected that unique event within at least 70% of its
iterations (7,000 or more instances out of the 10,000). Of the storms
that passed the 70% threshold, only those that resulted in a SC
response 60% of the time were considered highly likely to have a SC
response present (a minimum of between 4,200 responses out of 7,000
existing storm events and 6,000 responses of 10,000 existing storm
events). The precipitation events that survived this culling were
summarized by the median values of each dependent variable:
FSC, DSC, RSC. These
were then compared to the independent, environmental factors surrounding
the event.