1
$\begingroup$

Suppose there is $n$ data points $(x_i,y_i)$ and $i=1,...,n$, sampled from a line in 2D modelled by $y = m_n x + b_n$ where $m_n \sim \mathcal{N}(0,\sigma^2_m)$ and $b_n \sim \mathcal{N}(0,\sigma^2_b)$. How can I estimate $\sigma_m$ and $\sigma_b$ very very roughly but in a deterministic way?

I already know the answer is not unique. Namely $\sigma_m$ can be zero and $\sigma_b$ will be variance of data. But what is the most naive-inaccurate way to have a reasonable value for both (both non-zero)?

$\endgroup$

2 Answers 2

1
$\begingroup$

Generally, this seems like a problem where a variation bootstrap would be quite reasonable. Given that $E[b_n] = 0,$ I would recommend performing the following steps: 1. Simulate the data generating process many times to get $(Y^n,X^n)$ each time. 2. Fit a linear model $\hat{Y} = \hat{m}X$ on each of the simulated datasets, and save each $\hat{m}$ for later 3. calculate the sample variance for your distribution of $\hat{m}$'s that you have generated.

$\endgroup$
1
$\begingroup$

$$p(\sigma_n,\sigma_m|D) \propto p(D|\sigma_n,\sigma_m)p(\sigma_n,\sigma_m)$$

$$\sigma_n^*,\sigma_m^* = argmax(p(\sigma_n,\sigma_m|D))$$

Now use a Gaussian likelihood and priors. Differentiate - this should have an analytic solution. Otherwise use a deterministic optimisation and ensure you keep the starting values. Reusing those starting values you have a deterministic solution as required.

See Barber's book, there will be a chapter/section covering Bayesian linear models. Available legally online: web4.cs.ucl.ac.uk/staff/D.Barber/textbook/090310.pdf

$\endgroup$

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .