Articles

Comparison using WAIC

  P ( Model 1 better than Model 2 ) = Φ ( elpd_diff ​/ se_diff) prob_model_better <- function(elpd_diff, se_diff) {   z <- elpd_diff / se_diff   p <- pnorm(z)   return(p) } elpd_diff <- 5 se_diff <- 3 prob_model_better ( elpd_diff , se_diff ) Result: [1] 0.952 So there is approximately  95% probability that model 1 predicts better than model 2 . It is  not  the probability that the model is  better  — only that its  predictive accuracy is higher . You can use also the function Probability_Best_Model_WAIC() in HelpersMG packages How to cite this method: Model predictive performance was evaluated using leave-one-out cross-validation (PSIS-LOO). Uncertainty in model comparison was assessed using resampling of the pointwise log predictive densities following the approach described by Vehtari, Gelman and Gabry (2017) and Yao et al. (2018). Vehtari, A., Gelman, A., & Gabry, J. (2017). Practical...
Enregistrer un commentaire
Lire la suite

When your autocorrelation becomes negative

The negative values appearing at larger lags in the sample autocorrelation function are usually not evidence of true negative dependence but arise from the finite-sample constraint that the estimated autocorrelations must balance around zero once the sample mean has been removed. Positive autocorrelation at small lags forces the estimated ACF at larger lags to become slightly negative  because the correlations must balance around zero in a finite sample .
Enregistrer un commentaire
Lire la suite

Error in brm compilation of a Stan model

 Try this  # Compile packages using all cores Sys.setenv ( MAKEFLAGS = paste0 ( "-j" , parallel :: detectCores ( ) ) ) install.packages ( c ( "StanHeaders" , "rstan" ) , type = "source" )
Enregistrer un commentaire
Lire la suite

Install ggiraph in MacosX

 nano ~/.R/Makevars Ajouter : CPPFLAGS=-I/opt/homebrew/include LDFLAGS=-L/opt/homebrew/lib Sauvegarder puis relancer R et installer : install.packages("ggiraph", type="source")
Enregistrer un commentaire
Lire la suite

Update rJava in Ubuntu 22.04

 First, I have been forced to delete all old installations of open-jdk: sudo apt-get purge openjdk-\* sudo apt autoremove It can be necessary to do: sudo rm -rf /usr/lib/jvm Install most recent version for Ubuntu 22.04 sudo apt install openjdk-25-jdk Then update the variables using: sudo R CMD javareconf And the update of rJava was installing
Enregistrer un commentaire
Lire la suite

Pourquoi la p-value est fausse et produit des faux-positifs

 Parce qu’une  p-value  est calculée  conditionnellement à l’hypothèse nulle étant vraie , elle ne représente pas la probabilité de commettre une erreur de type I dans la situation réelle dans laquelle on se trouve. Lorsqu’elle est interprétée de cette façon, elle  surestime systématiquement  le « risque de premier ordre ». Voici le raisonnement précis. 1. Ce qu’est réellement le « risque de premier ordre » Le taux d’erreur de type I (risque de premier ordre) est : α = P ( rejeter  H 0 ∣ H 0  est vraie ) Il s’agit d’une propriété  à long terme , fixée  a priori , d’une règle de décision (par exemple : « rejeter si  p < 0,05 »). Ce n’est  pas  une probabilité concernant l’expérience en cours. 2. Ce qu’est réellement une  p-value Une  p-value  est : p = P ( T ≥ t obs ∣ H 0 ) Points clés : Elle est conditionnelle au fait que  H 0  soit vraie Ce n’est pas  P ( H 0 ∣ donné e s ) Ce n’est pa...
Enregistrer un commentaire
Lire la suite

Why p-values over-estimate first order risk ?

 The short answer is: 👉  Because a p-value is computed  conditional on the null hypothesis being true , it does  not  represent the probability of making a Type I error in the situation you are actually in. When it is interpreted as such, it systematically overstates (over-estimates) the “first-order risk”. Below is the precise reasoning. 1. What “first-order risk” really is The  Type I error rate  (first-order risk) is: α = P ( reject  H 0 ∣ H 0  is true ) This is a  long-run, pre-specified  property of a decision rule (e.g. “reject if  p < 0.05 ”). It is  not a probability about the current experiment . 2. What a p-value actually is A p-value is: p = P ( T ≥ t obs ∣ H 0 ) Key points: It is  conditional on  H 0 ​  being true It is  not   P ( H 0 ∣ data ) It is  not   P ( Type I error ) 3. Where the over-estimation comes from The common (incorrect) interpretation “If...
Enregistrer un commentaire
Lire la suite