What is the minimum number of runs for the simulation experiment to be considered statistically significant?
There is no general rule to this as it depends on the PTV Vissim application:
- In a 10-km stretch of highway with low traffic volume simulated for one hour, travel times will be very stable and depend only on the distribution of desired speeds. Three to five runs should be sufficient. - In an urban network consisting of some complex junctions with traffic actuated signal control, possibly including public transport priority, a lot of 'random' events influence the signal control (e.g. two trams approaching at the same time). In this case, more runs are needed to get a significant result, especially if one or more junctions operate close to capacity.
Formally, you have to estimate the variance coefficient of the measured value, e.g. travel time. You can do this by running the simulation several times with different seeds and computing the variance. Then you define a confidence level, say e.g. 5 %, and a tolerance interval for the result, say 10%. Then you need n simulation runs to be able to say, that 'with probability 95% the real mean value of travel times lies within the interval measured value +/- 10%.' n is given by the formula: n = t² * v² / e², where t is the value from the t-distribution for the given confidence level, v is the variance coefficient (standard deviation / mean) of the measured value, and e is the tolerance (in this example: 10%).