ggloor
commented
6 years ago You are correct, I should have plotted perb.rare@matrix instead. Thank you for pointing out the error. I have updated the figure and the code. Note that, in reference to you second point, that the E(rho) for 11 vs 37246 is now even lower.
The point estimate includes much more random noise than does the expected value. Thus, the E(rho) value is more robust. One way to think about it is that the E(rho) is the value that you would observe if you could have averaged the values across multiple technical replicates. When collecting associations, the investigator is interested in those that are not affected by random sampling, but that would have been observed upon replication
erfanshek
In R_block_11, why the variable used in y axis is perb.rare@matrix, I would think if the expected rho value should be plotted rho.rare@matrix should be used instead.
Also, I was a bit concern about the great difference between the point estimator of rho and the expected value of rho, when inquiring correlation of the same OTUs used in R_block_10.. Please see attachment, while point estimators of rho reached proportionalities over 0.95 in many instances, the estimated values reached at maximum ~0.8. I was expecting differences but not so notable ones.
I also tied to investigate the range of correlations obtaining by every different approximation (see attached). As it is also observed on Figure 8, correlation coefficients from E(rho) and pearson tend to be smaller than those obtained by the point estimate (see attached the hist of the sd for the whole dataset), I also saw similar behavior when using in-house datasets. I cannot really formulate a reason for this observation, or even understand which should be consider the "gold standard". dist_rho.pdf correlations_prop.txt