Subscribe Bookmark RSS Feed

Peak Prediction

akhst7

Community Trekker

Joined:

Jun 6, 2012

Hi,

I need a ton of help solve this problem.  I have a distribution (blue) that is made of two components.  I know a distribution of the first component (red) but not the second.  Is there any ways to predict or estimate the distribution of the second component ?  

Thanks in advance.

2134_Untitled.jpg

4 REPLIES
pgstats

Community Trekker

Joined:

Aug 30, 2011

The first component looks left censored. Check procedure FMM for finite mixture models estimation.

PG

akhst7

Community Trekker

Joined:

Jun 6, 2012

PG,

Thanks for a suggestion.  I tried FMM but it was not sensitive enough to predict two distinctive distributions even though I manually  played with  some parameters.

Aki

pgstats

Community Trekker

Joined:

Aug 30, 2011

Seems like you have only one parameter to estimate in this problem : the proportion represented by the first component in your distribution. Given a proportion, the shape ofthe second component can be found by subtraction. But then, what would like to optimize? Do you have a hypothesis about the shape of the second component? Is it supposed to be the same as the first component? Is it supposed to be normally distributed?

PG

akhst7

Community Trekker

Joined:

Jun 6, 2012

I can use subtraction to guesstimate the distribution of the second but I don't know if subtraction gives me the accurate estimate of the second component.  I am assuming that a second distribution is near normal although I have no concrete evidence for that, and  a proportion of the first distribution shown in a previous graph is misleading.  It is actually an overly graph of the suspected first component (red) and the distribution in question with proportions being normalized (y-axis). 

Anyway, what I was really hoping was that JMP had some  functions that help me to build a model for two  component normal mixture, by which I could pick the best overall distribution that would resemble the mixed distribution in question.  Maybe what I need is FMM or EM scripts ?