# I am having trouble getting a normal bell curve out.... - Any help is appreciated. Will pay

#### hlsmith

##### Less is more. Stay pure. Stay poor.
Sounds like you have count data, which can be represented by the Poisson distribution. If the mean is large enough, 8 or higher, the Poisson data can be approximated by the normal distribution. What are your mean and variance values?

#### obh

##### Active Member
Poisson makes sense

> fitdistr(data1, "Poisson")

lambda

1.31425234

(0.03918343)

Code:
library(MASS)
data1<-c(1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,4,4,4,4,4,4,4,4,4,5,5,5,6,6,6,6,8,8,9,11,11,11,12,14
)
fitdistr(data1, "Poisson")

Last edited:

#### katxt

##### Active Member
It looks like Poisson may work. It's definitely not Normal.
The mean and variance of the Poisson are equal which yours are roughly 1.31 and 1.24, but you probably also need the number of days with zero sales included as well.

#### noetsi

##### Fortran must die
The data itself may or may not have to be normal. For instance in regression the raw data does not have to be normal just the residuals. I am not sure what test you are running. There are btw software, including I think some excel ad ons which will superimpose a normal curve over your data.

#### katxt

##### Active Member
Just thinking. Poisson isn't the best because with a mean of around 1 or 2, you should not see anything above about 6.
In the meantime, you can use the data itself as its own probability distribution and use a Monte Carlo technique.
You could also consider zero-truncated negative binomial but that's getting rather obscure.

#### noetsi

##### Fortran must die
of course if it is obscure no one will know if you are right or not

#### hlsmith

##### Less is more. Stay pure. Stay poor.
Just thinking. Poisson isn't the best because with a mean of around 1 or 2, you should not see anything above about 6.
In the meantime, you can use the data itself as its own probability distribution and use a Monte Carlo technique.
You could also consider zero-truncated negative binomial but that's getting rather obscure.
Can you elaborate on the MC

#### katxt

##### Active Member
It depends on what the fitted normal was going to be used for. If you wanted a confidence interval for the mean, for instance, you could do a bootstrap CI using the actual data. You could compare one area or period with another using a permutation test. If you wanted to do a risk analysis which involved sampling from the real fitted distribution, then sample from the data.
However, I think one thing is clear. The Normal is not a suitable distribution to try and fit.