2
$\begingroup$

Background: I've modeled a project effort prediction as a Google Spreadsheet template. Details of the Model: http://sites.google.com/site/effortprediction/methodology .Google Spreadsheet does not implement beta distribution functions. In PERT a beta distribution is used to estimate the effort of a task in Ideal Person Days (i.e.: one worker 8h a day without distraction).

Problem: What consequences has use of normal distributions instead of beta distributions when the distribution are summed up to a single distribution.

Details A project consists of a list of tasks. Each task is beta distributed and independed, the sum of the tasks therefore should obey to the central limit theorem. The project therefore is normal distributed. How does Expected Value and SD change for the project when I use normal instead of beta distributions for each task? I assume that the tasks distribution is skewed more often to the right than to the left.

Questions

  • Are the tails fater?
  • Is the Expected Value higher or lower with normal distributions?
  • Is the dispersion higher or lower?
Improve this question
$\endgroup$
17
  • 1
    $\begingroup$ If you match the mean and variance of each task with the normal distribution, then the expected value and standard deviation of the sum is unchanged. You will loose all information about skewness though - and tails are likely to be thinner with normal $\endgroup$
    – probabilityislogic
    Commented Apr 16, 2011 at 11:00
  • $\begingroup$ mean sounds correct, but not sure with beta, i forgot to say how i calculate sd. I start form 90% confidence interval and mean. now i calculate the sd from a norm dist via z-score for 90%. here an error should already be introduced. but what can I say about the error of sd in comparision with beta-dist? $\endgroup$
    – Roland Kofler
    Commented Apr 16, 2011 at 11:25
  • $\begingroup$ so you don't have a "standard deviation" specified, you have specified the quantiles - 5% and 95% I take it. But in this case - if you use a beta distribution to fit the quantiles, and then use a normal to approximate that beta distribution, my comment above applies. $\endgroup$
    – probabilityislogic
    Commented Apr 16, 2011 at 12:06
  • 1
    $\begingroup$ An exponential distribution may be a better idea for modeling instead of the beta distribution. one because it is highly right skewed, two because it is always positive, three because it has "fat tails", and four because the sum of exponentially distributed quantities has an Erlang or gamma or scaled chi-square distribution - which is no more difficult to handle than a normal, plus it is precise. And exponential arises as a limiting form of a beta distribution. Also exponential has maximum entropy (uncertainty) for a given mean - so you only need to ask for expected time to complete $\endgroup$
    – probabilityislogic
    Commented Apr 16, 2011 at 23:06
  • 1
    $\begingroup$ @roland - the "memoryless" property is a bit confusing because it is conditional, not absolute. More work/effort would result in a higher rate parameter, and thus a smaller amount of time to finish a project. the memorylessness would be assessed by looking at the distribution of the "late tasks". Now why would a task be "late"? One way I can think of is that the amount of work required to finish it was under-specified - adding in these additional tasks - seems like we are back to the start - so memorylessness may be a good property. $\endgroup$
    – probabilityislogic
    Commented Apr 17, 2011 at 14:29

1 Answer 1

Reset to default
2
$\begingroup$

Just go ahead and calculate a random uniform, then use the inverse cdf method (http://en.wikipedia.org/wiki/Inverse_transform_sampling) to get a random Beta. Here is the formula for the inverse beta cdf: http://www.mathworks.com/help/toolbox/stats/betainv.html

If you truly want to just not calculate the beta random number, you can be comfortable that beta converges to normal asymptotically, however the folks at PERT feel very strongly that the normal is not appropriate and list several reasons at http://laserlightnetworks.com/Documents/Modeling%20Schedule%20Uncertainty%20without%20Monte%20Carlo%20Methods.pdf

They suggest the triangle distribution is a better approximation because of steeper decent in cdf and defined support, whereas normal has tails that go to infinity, the triangle and beta distributions have fixed range.

Improve this answer
$\endgroup$
1
  • $\begingroup$ Your idea is great and I will take it into consideration. But I tried also to make my question valuable for others by abstracting from my case and comparing normal with beta distribution. So I believe your answer is more a "comment" than a real answer. A moderator might explain if I am right or not. I just don't want to be selfish and adhere to the spirit of stackexchange... $\endgroup$
    – Roland Kofler
    Commented Apr 18, 2011 at 13:14

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.