diff --git a/.gitignore b/.gitignore index 6a614e4700196b4417bff5d3d479c1b13b43ba1f..646b22721ee4d0a15416a72f866f761a84747a3b 100644 --- a/.gitignore +++ b/.gitignore @@ -5,6 +5,5 @@ old -.backup -.log - +*.backup +*.log diff --git a/bf_materials.zip b/bf_materials.zip index 1fb8605bbd9f5551a9ae2c3bb25e6a90fd126211..55cace2f8b96129197385f6fc3418725dc288f74 100644 Binary files a/bf_materials.zip and b/bf_materials.zip differ diff --git a/docs/guide.pdf b/docs/guide.pdf index e3821c4d451ae738150553524e6b23be3bda6967..857925d4124d02a785fabdaae603fbdc011e7a38 100644 Binary files a/docs/guide.pdf and b/docs/guide.pdf differ diff --git a/docs/presentation.Rmd b/docs/presentation.Rmd index ac9fc67d7ecf91a21e07e969ae641c689f464720..2f4429bbfb5caf4a58b8396442b212fa98331a52 100644 --- a/docs/presentation.Rmd +++ b/docs/presentation.Rmd @@ -2,6 +2,7 @@ title: "Introduction to Bayes factors" author: "Jeffrey R. Stevens" date: "9 May 2019" +institute: "https://osf.io/h38sx/" output: beamer_presentation: theme: "CambridgeUS" @@ -174,7 +175,9 @@ $\frac{P(D|H_0)}{P(D|H_1)}$ \pause * $BF_{10}=1/BF_{01}$ \pause -* So $BF_{10}$=10 means that there is 10 times more evidence for $H_1$ than $H_0$ (and is equivalent to a $BF_{01}=0.1$) +* $BF_{10}$=8 means that there is 8 times more evidence for $H_1$ than $H_0$ (and is equivalent to a $BF_{01}=1/8=0.125$) +\pause +* Bayes factors range from 0 to $\infty$ ## Bayes factor cutoffs @@ -245,6 +248,7 @@ $\frac{P(D|H_0)}{P(D|H_1)}$ - Set pre-determined BF thresholds (e.g., 10 or 1/10) - Start with at least 20 samples per group - Adjust based on prior knowledge +\pause * It is possible to conduct something like a Bayesian power analysis\footnote{\tiny{(Schönbrodt \& Wagenmakers, 2018, \emph{Psychonomic Bulletin \& Review})}} @@ -279,15 +283,17 @@ $\frac{P(D|H_0)}{P(D|H_1)}$ ## Model selection -\Large -If $BF_{10} = \frac{P(D|H_1)}{P(D|H_0)}$ and $BF_{20} = \frac{P(D|H_2)}{P(D|H_0)}$, +Let's say you are interested in comparing two models in a regression: $H_1$ and $H_2$. First, you will compare them to an intercept-only model $H_0$ to generate their Bayes factors: -\pause -then $BF_{12} = \frac{P(D|H_1)}{P(D|H_2)} = \frac{BF_{10}}{BF_{20}}$. +\Large +$BF_{10} = \frac{P(D|H_1)}{P(D|H_0)}$ and $BF_{20} = \frac{P(D|H_2)}{P(D|H_0)}$ \pause \normalsize -You can divide Bayes factors of two models (compared to null) to find the Bayes factor comparing those two models. +Then, you can divide the Bayes factors of two models (compared to null) to find the Bayes factor comparing those two models: + +\Large +$\frac{BF_{10}}{BF_{20}} = \frac{P(D|H_1)}{P(D|H_2)} = BF_{12}$ ## Model selection @@ -333,11 +339,15 @@ $BF_{interaction} = \frac{16}{14} = 1.14$ * Methods - Define Bayes factor (I cite Wagenmakers, 2007) - - Describe levels of evidence (I cite Wagenmakers et al., 2018) + - Describe cutoffs for evidence (I cite Wagenmakers et al., 2018) - Describe priors/assumptions (I cite journal articles cited by packages) \pause * Results - Clarify direction (alternative/null) + - Describe effect in "levels of evidence" terms + + "There is moderate evidence for a difference between ..." + + "There is very strong evidence for no difference between ..." + + "There is no evidence for a difference between ..." - Give Bayes factors as you would p-values (use 1-2 decimal places for > 1; 2-3 decimal for < 1; cutoffs for > 100, < 0.01) \pause * Supplementary materials @@ -354,7 +364,7 @@ $BF_{interaction} = \frac{16}{14} = 1.14$ ## Take home \Large -**With both JASP and R, BF is as easy as frequentist---and you can convert between them!** +**With both JASP and R, Bayes factors are as easy as frequentist---and you can convert between them!** # Resources diff --git a/docs/presentation.pdf b/docs/presentation.pdf index f05744edaaebdcd04803a51c5a9b587d55c6e1fa..9a60757e39528b11780205c9bdc4dbe9d635f4c2 100644 Binary files a/docs/presentation.pdf and b/docs/presentation.pdf differ