I've written a small MATLAB function that (I think) could be useful for others. The idea is to find the distribution that best fits a set of data. The function tries all distributions available (continuous or discrete depending on the data), chooses the one with the highest likelihood, returns its parameters with 95% CI and plots the data and the fitting. Everybody is welcome to try it, and If you find bugs or have suggestions, please comment.
The code is in 'File Exchange' at mathworks.com: http://www.mathworks.com/matlabcentral/fileexchange/40167-fitmethis
Nice piece of code. Could you do ourselves a favor? Place your work on a blog with some documentation. Then, place this link and others there so that we can bookmark it and send it to other people. Um abraço,
Vania
Will it work on multivariate data? I mean if the data has more independent variables
I performed the following test using the allfitdist, fitmethis, and fitdist function using the same statistical toolbox. The results are as follows:
>> data1 = normrnd(5,3,1000,1);
>> data2 = gamrnd(5,0.5,1000,1);
>> x = data1+data2;
>> [D PD] = allfitdist(x); D(1)
>> F= fitmethis(x);
>> pd = fitdist(x,'Normal')
Results of allfitdist: DistName: 'normal'; Params: [7.3265 3.1743]
Results of fitmethis: name: 'normal'; par: [7.4077 3.1510]
Results of the fitdist: mu = 7.51921; sigma = 3.27417
As shown, both allfitdist, and fitmethis found the same best fit; however, the parameters are not the same. In addition, the fitdist results in another set of parameters. Can anyone explain this issue?
Hello,
thank you so much.
I did not see you are included the Poisson distribution
To the question posed by the EL-Shimy.
First, apologies for the delay in answering. I can't really help much, since in my computer I get exactly the same parameters from the three functions. Using your example, I get:
allfitdist: Params: [7.5026 3.0778]
fitmethis: normal 7.503e+00 3.076e+00
fitdist: mu = 7.50264, sigma = 3.07779
I repeated the thing several times, and I always get the same values from all the functions. Sorry I can't help more.
Francisco
Can you please add the F distribution too (with degrees of freedom in some suitable range)? You don't cover F, right?
Answer to Somabha Mukherjee:
The F distribution is not included in MATLAB's 'mle' function, which is what my function uses to do the fitting. So, I can't include it easily.
If you'd like to try, maybe using functions 'fpdf' and 'fminsearch' you can get a ML approach to the problem.
Best,
Comment to Francisco de Castro:
Thanks for clarifying! I found out an useful way of doing this for F-distributions, though. The idea is, if X ~ F_{a,b} , then aX/(b+aX) has a Beta(a/2 , b/2) distribution, which is included in your function. So, what basically needed to do, is to transform my sample (I already knew what the a and b should be, beforehand), and apply your function on it!
Thanks for sharing. Would you please update your code, covering other popular distribution functions like (log) Pearson type 3, Gumbel and Wakeby?
Regards
Sattat. I'm sorry but I can only add distributions that are already included in MATLAB. Those you mention are not, I believe.
- Badges
- Science topic