Closed bob-carpenter closed 9 years ago
See my answer to the previous issue.
Although I understand that it is better to use each replicate separately, we don't have the data in this format. I do have a large amount of files and scripts that takes raw data to replicate the values of eCO2mean
and eCO2sd
, but this may take more than a day of work to figure out. The work was done by a student who already left our lab.
I would suggest to try to find an alternative dataset to implement this hierarchical model. Not only is my dataset difficult to transform to individual measurements, but I also have concerns about the leaks we had in our jars. Maybe Charlotte or someone else from the workshop has a more useful dataset of incubation data.
OK. Thanks. That answers my main question about whether this came from replicated experiments.
Leaks in jars would be a great kind of measurement error to measure --- you could get a mixture model of leaky and non-leaky jars. Anyway, more of a stats geek problem than one that will matter going forward.
On Nov 27, 2014, at 6:18 AM, Carlos A. Sierra notifications@github.com wrote:
See my answer to the previous issue.
Although I understand that it is better to use each replicate separately, we don't have the data in this format. I do have a large amount of files and scripts that takes raw data to replicate the values of eCO2mean and eCO2sd, but this may take more than a day of work to figure out. The work was done by a student who already left our lab.
I would suggest to try to find an alternative dataset to implement this hierarchical model. Not only is my dataset difficult to transform to individual measurements, but I also have concerns about the leaks we had in our jars. Maybe Charlotte or someone else from the workshop has a more useful dataset of incubation data.
— Reply to this email directly or view it on GitHub.
I'd like to build a hierarchical model of the raw data that was aggregated into
eCO2mean
andeCO2sd
in the SoilR data seteCO2
.Background
In a hierarchical model, each replicate
i
gets its own parameter vectortheta[i]
(such as initial carbon and initial mixture between compartments or even decomposition and transfer rates). The item-specific random effects can be given a simple multivariate normal prior:So
mu_theta
gives you the population average for the parameters andSigma_theta
the covariance among the parameters. This unfolds the mean/sd measurement error model based on the aggregated data into something more flexible.We'd put an informative prior on
mu_theta
, the mean population response, in the same way we set priors in the aggregated model. We can put an informative prior on the covarianceSigma_theta
based on expectations for parameter scales, and we can use Stan's LKJ prior to concentrate mass to a flexible degree around a uniform correlation matrix.