Hello everyone
I have precipitation data at 5 different meteorological stations. All of these stations have missing data. I have tried MCMC method using XLSTAT.
It gives negative values for some of the observations. (plz see attached file)
Can anyone guide me what to do with these values?
Should these be replaced with zero or considered as missing again?
Or any other way?
Thanks in anticipation
Are you really serious? You tried MCMC WITH EXCEL? IT'S been known to get arithmetic wrong. Please 🙏 get a reputable package. You can find this in R as freeware for sure. Probably other places as well. Google it and your question and you should find something appropriate.. Good luck, D Booth
You should ask yourself the question why data are missing. This will help you to decide if you impute with reasonable values using imputation packages in R such as VIM, Amelia, ... or to set it to zero.
Is a missing meaning that there was no precipitation: then its a zero. If a missing caused by a erroneous measurement or measurement failure: impute.
missing data means measurement error not zero. that's why I tried using imputation with xlstat but the result gives me negative values which is not possible in case of precipitation. Matthias Templ
You may have constraints and assumptions (e.g., Gaussian distribution, spatial/temporal similarity) on your dataset that you are neglecting. I recommend that instead of using Markov chain Monte Carlo, try other scenarios to tackle the imputation problem. For practical examining of these, you may refer to R packages such as "accelmissing", "ForImp", "imputeMissings", "missRanger", and "VIM".
Having a quick look at XLSTAT it appears the MCMC methods used do not have any transforms or latent forms and thus has support from negative to positive infinity. Unless you can change this, you are going to risk getting negative samples in your data.
In fact, I would really want the underlying MCMC to be over the latent parameters of another model that represented the precipitation data. I.E. Take a continuous distribution with [0, +inf) support and use the MCMC to find the parameters of that instead of the data directly (and then evaluate the model with the MAP estimate or otherwise to find the missing data).
The more that I think about it, having a latent model you can MCMC over (or do MLE / Variational Bayes) is probably the right approach here.
Article Measuring the ecological preference for growth of150 of the ...
https://www.cell.com/heliyon/fulltext/S2405-8440(19)35821-9
- Badges
- Science topic