Not true: it means you shouldn’t use a normal distribution, and when you do you should say so up front. I see no reason not to apply normal distributions if your limit is high (say, greater than 4 sigmas—social science is much fuzzier than physical science). Better yet, make your limit a function of the number of observations you have. As the probability of getting into the long tail gets higher, make the tail longer.
Sentence 1: True, fair point.
Sentence 2: This isn’t obvious to me. Selecting random values from a truncated normal distribution is (slightly) more complex than, say, a uniform distribution over the same range, but it is demonstrably (slightly) less complex than selecting random values from an unbounded normal distribution. Without finite boundaries, you’d need infinite precision arithmetic just to draw a value.
The problem is not with value selection, the problem is with model manipulation. The normal distribution is very well-studied, it has a number of appealing properties which make working with it rather convenient, there is a lot of code written to work with it, etc. Replace it with a truncated normal and suddenly a lot of things break.
As a trivial point, imposing finite bounds means that you can’t use the normal distribution, for example :-)
Not true: it means you shouldn’t use a normal distribution, and when you do you should say so up front. I see no reason not to apply normal distributions if your limit is high (say, greater than 4 sigmas—social science is much fuzzier than physical science). Better yet, make your limit a function of the number of observations you have. As the probability of getting into the long tail gets higher, make the tail longer.
Truncated normal is not the same thing as a plain-vanilla normal. And using it does mean increasing the complexity of the simulation.
Sentence 1: True, fair point. Sentence 2: This isn’t obvious to me. Selecting random values from a truncated normal distribution is (slightly) more complex than, say, a uniform distribution over the same range, but it is demonstrably (slightly) less complex than selecting random values from an unbounded normal distribution. Without finite boundaries, you’d need infinite precision arithmetic just to draw a value.
The problem is not with value selection, the problem is with model manipulation. The normal distribution is very well-studied, it has a number of appealing properties which make working with it rather convenient, there is a lot of code written to work with it, etc. Replace it with a truncated normal and suddenly a lot of things break.
Oh! I see what you’re saying. Definitely can’t argue with that.