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

[SanDiegoREStats]

I am a mid level Real Estate Executive and just trying to pencil out a normal distribution. Attached is my excel file. I am trying to take the number of deals column and compute what a bell curve looks like for the number of deals agents are doing in San Diego.

Dropbox link to the file here:

https://www.dropbox.com/s/24k73q14mnkremw/Normal Distribution Not Working .xls?dl=0

 

[obh]

Hi SDR,

The number of deals per day? 1 day 14 one day 12 two days 8?

Ps, did you think about the option that the distribution is no Normal? as you can have a negative number of deals, and the mode is 1 deal per day, not around 7?

For example https://en.wikipedia.org/wiki/Half-normal_distribution

 

[hlsmith]

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]

Poisson makes sense

> fitdistr(data1, "Poisson")

lambda

1.31425234

(0.03918343)

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")

 

[katxt]

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]

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]

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]

of course if it is obscure no one will know if you are right or not

 

[hlsmith]

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]

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.