Hi,
I have just a general question.
I would like to know if my assumption holds.
Here, I have three data sets – y, x, and z. They are all independent.
If I want to analyze the relationships among them, I can build a linear model such as:
Model 1)
y = a x + bz
Soon after, I found that y is not normally distributed but log(e^{y}) does.
So we can modify Model 1) to:
Model 2)
log(e^{y}+1) = log(e^{ax+bz}+1)
At he same, I found that x has normality. So, I changed the equation to:
Model 3)
x = {1\over a}y - {b \over a}z
Now, we can use the Normal likelihood function to find a and b.
Such as:
with pm.Model() as model_2
a = pm.Normal('a', mu=0, sigma=1)
b = pm.Normal('b', mu=0, sigma=1)
mu = pm.math.log(pm.math.exp(ax+bz)+1)
yt = np.log(np.exp(y)+1)
y_obs = pm.Normal('obs', mu=mu, sigma=1, observed=yt)
pm.sample()
with pm.Model() as model_3
a = pm.Normal('a', mu=0, sigma=1)
b = pm.Normal('b', mu=0, sigma=1)
mu = 1/a*y - b/a*z
x_obs = pm.Normal('obs', mu=mu, sigma=1, observed=x)
pm.sample()
After the samplings from model_2 and model_3, should we get the same or similar distributions of a and b from the trace plots?