248180_en-US

liliconcom

Exponential Response(Y) in Generalized Regression

I want to make a model using Generalized Regression.

Response(Y) follow exponential distribution, but include many '0' value,

 so I think that I could not use exponential as link function.

 Is this right ?

If right, Is there any good method ?

If transform this response(Y) to SHASH ?

 

Re: Exponential Response(Y) in Generalized RegressionExponential distribution is a very restrictive distribution. A more flexible one would be Gamma. In fact, exponential is a special case of gamma where the shape parameter=1 (using JMP’s notion: mu=sigma). There is a ZI (Zero Inflated) Gamma option in GenReg that allows you to model Y with zeros.
Re: Exponential Response(Y) in Generalized Regression

Thank you Jiancao,

Your comment is very appropriate, I think.

And also, I am thinking about other  2 practical(?) methods.

Method 1 : slightly adjusting   every values of Response(Y) equally to avoid problem of  '0'

    i.e. adding very small number  to every values of Response(Y) 

Method 2 : Using Poisson since Response(Y) have many 0 values like defect count data.

But, I'm not sure if these could be a good alternative or not.

Re: Exponential Response(Y) in Generalized Regression

Ikju,

Adjusting values so there are no zeros in the data is extremely dangerous. I have seen situations where the results of a statistical test can be changed (from significant to not significant) by choosing a different adjustment value. In both cases the adjustments were smaller than the limit of detection value. You are much better using the appropriate distribution. Both the gamma and Poisson have zero inflated options to account for zeros in the data.

Re: Exponential Response(Y) in Generalized Regression

Regarding Poisson (and Negative Binomial distribution for that matter), it's for Y taking on discrete values, 0, 1, 2, ..... 

Re: Exponential Response(Y) in Generalized Regression

Hard to know if this is actually useful without more context, but it's always good to reflect on why the 0 value was recorded. But it might be worth replacing these values with '.' and trying the 'Informative Missing' option in Generalised Regression (under the 'Model Specification' LRT)..