@c_drosten Hi, ich habe kurz mal aus Interesse den Welch-Test, bezüglich Figure 2 durchgezogen (ich habe eine solche Rechnung bisher nicht gemacht) und komme für den p-value auf einen Wert von 0.006. Die Unterschrift sagt dies aber für den Mann-Whitney rank test aus, oder? 
@c_drosten Ich kann mich da durchaus irren! Kann das jemand validieren?
@c_drosten I implemented the Welch test by hand writing my own function.#Mathematica complains that the data sets are not normally distributed. Hence, I visualized in Mathematica QQ-Plots against the normal distribution. Especially,the healthy data set seems to be off.#CoronaDaten #Drosten
@c_drosten Looking only on Figure 2, I am a little bit confused. Why is it possible to mix the results from two machines with different sample numbers? As far as I get it, the mean of the cobas is alway below the mean of the LC480. If I mix more samples of LC480 in the healthy group,...
@c_drosten I artificially increase the mean of the healthy group, or not? Maybe it is explained in the text? Does anyone have an explanation, why the mixing procedure of Figure 2 is valid? #CoronaDaten #Drosten #Corona
@c_drosten To check the statistics, I used the data sets from the link below, which are reversed-engineered from the swarm plots of Figure 2. The mean values coincide well with those given by the figure caption (the difference is below 1%). #CoronaVirusDE
@c_drosten My impression from the Q-Q Plots was confirmed by @hofmann_tech in a objective manner using statistical tests. Hence, the Welch Test cannot be used, because the needed requirement of normality is not fulfilled. #Drosten #Corona
@c_drosten @hofmann_tech I can confirm the result of the Mann-Whitney U test, where I calculate a p-value of 0,032, which might be close enough, however, there is another problem: mixing LC480 and cobas with different fractions for the healthy and preconditioned group, you change the distributions...
@c_drosten @hofmann_tech because both systems exhibit different statistical distributions and need to considered separately as clearly stated in the text. However, the equality of the distributions is a prerequisite for the M-W U test. statistikguru.de/spss/mann-whit… #Drosten #CoronaVirusDE #CoronaDaten
@c_drosten @hofmann_tech It might be speculative: but a more solid statement could be made, if the same fractions (No.LC480/No.cobas) for the healthy and pre-cond. group could be *somehow* provided?? ...
@c_drosten @hofmann_tech ... However, if one chooses a approviate subset, one might has to proof *somehow* - at least in a statistical manner - that the outcome of the hypothesis test is independent from the specific choice of such a subset.
@c_drosten @hofmann_tech Another chance might be a measurement of a kind-of statistical transfer function between the statistics of the LC480 and cobas system by measuring pairs of samples with the same viral load in both instruments. Such a transfer function might bear the potential to *somehow* ...
@c_drosten @hofmann_tech ... correct the statistics, when using different fractions.
Haven tweeted that, I like to empathize that the posterior change of the statistics analysis always harbours the danger of p-hacking... en.m.wikipedia.org/wiki/Data_dred…
Haven tweeted that, I like to empathize that the posterior change of the statistics analysis always harbours the danger of p-hacking... en.m.wikipedia.org/wiki/Data_dred…
@c_drosten @hofmann_tech The Autors should be aware of this problem, which might be IMHO, of course, not completely avoidable, because such a huge statistic might be not provided a 2nd time.
#Drosten #CoronaVirus #Corona
#Drosten #CoronaVirus #Corona
