I built a model with parameters that have very different orders of magnitude.

Stan has a lot of problems sampling from the original problem, to remedy this I let Stan sample from the percentage deviation of my initial guess instead. It seems to work well, but it feels sort of awkward. I was just wondering if this approach might be a bad idea for some reason?

My actual model is a bit more complex than this one and has lots of issues with sampling without the transformed parameters. But this just serves as an example of what I’m doing:

```
data {
int<lower=1> N;
vector[N] windspeed;
vector[N] time;
vector[N] ar;
vector[3] inits;
}
parameters{
real p_wind;
real p_time;
real p_sigma;
}
transformed parameters{
real th_wind;
real th_time;
real th_sigma;
th_wind = (p_wind/100 + 1) * inits[2];
th_time = (p_time/100 + 1) * inits[2];
th_sigma = (p_sigma/100 + 1) * inits[3];
}
model {
vector[N] ar_model;
for (i in 1:N)
ar_model[i] = th_wind * windspeed[i]^2 + th_time * time[i];
ar ~ normal(ar_model,th_sigma);
p_wind ~ normal(0,35);
p_time ~ normal(0,35);
p_sigma ~ normal(0,35);
}
```

I guess alternatively I could try to choose the units of my data carefully, and nondimensionalize the parameters in such a way that they are all of order one.