This question is about intuition more than the actual lines of code. Consider the normal linear regression setup as follows: data { int<lower=1> D; int<lower=1> N; matrix[N, D] X; vector[N] Y; } parameters { real a; vector[D] b; real<lower=0> s; } model { vector[N…

Seems up @jonah , @lauren and @bgoodri ’s alleys.

Most of those weighting schemes are intended to make point estimates have known properties across datasets. My view is the only valid weights for a Bayesian analysis is like in a binomial when you actually observe N_j cases with the same success probability in the j-th group. Multiplying contributio…

So say, in my context, I think that at each time step there is a probability that the process that generates my data undergoes some kind of parameter shift. I.e. that the parameters aren’t really constant through time but some appropriately-parameterized gaussian process or random-walk type thing. …

[image] 8one6: Is there a happy middle ground? You could remove all but the last few observations from a regression model and accept the fact that there is not enough information to estimate anything precisely.

If more recent observations are more useful for inference, then it seems necessarily the case that the parameters themselves should be thought of as changing over time. If b is static, what makes more recent observations more useful? If b is static, then it seems it must be either that the noise i…

Just wanted to quickly add: weighting the target by an integer w[n] (what you call “target hacking” and also how weights are implemented in brms) is exactly the same as putting the same datapoint w[n] times in the data. I think this intuition generalizes to any non-negative real w[n], so if you redu…

If you have k observations your are trying to model, each with t temporally ordered predictors p, you could calculate a weighted average of p to obtain only one predictor. If you use e.g. exponentially decaying weights, you can parameterize the weighting scheme and make the decay rate part of the m…

I have a similar problem and I think option 3 might make the most sense. The reason is that changing the standard deviation/variance will actually implicitly weight each observation. To give an intuition, suppose you have one observation y and you put n different sampling statements on y. So then,…

I agree with @martinmodrak Even if parameters change over time, if we are only interested in parameters now, I think there is no problem using weights, assuming the weights are logical or plausible. My question is, to my knowledge, using weights or using ratio of variance is almost the same i…

Would it not be better to use the log weight target += log(w[n]) + normal_lpdf(Y[n] | mu[n], s); otherwise we actually exponentiate them?

Bayesian parallels of weighted regression

Modeling

martinmodrak Split this topic July 29, 2021, 6:54am 13

3 posts were split to a new topic: Using weights to handle varying regret across data points

Topic		Replies	Views
Survey weighted regression Modeling	34	9359	May 27, 2022
The effect of weights on the resulting estimates Modeling	1	615	May 22, 2020
Weighted loglikelihoods in mixture model PyStan	1	1018	February 3, 2020
Assistance with an exponential model with weights Modeling	6	883	January 6, 2020
Weights in brm brms	12	11534	February 21, 2022

Bayesian parallels of weighted regression

Related topics