Extracting BFs from anovaBF or any Formal class BFBayesFactor object

Comments

• MSB

  December 2019 edited December 2019

  You can use as.vector():

  library(BayesFactor)
  
  data(puzzles)
  
  result <- anovaBF(RT ~ shape*color + ID, data = puzzles,
           whichRandom = "ID", progress = FALSE)
  
  result
  #> Bayes factor analysis
  #> --------------
  #> [1] shape + ID                       : 2.840104 ±0.84%
  #> [2] color + ID                       : 2.841925 ±1.48%
  #> [3] shape + color + ID               : 11.43469 ±1.23%
  #> [4] shape + color + shape:color + ID : 4.873394 ±10.38%
  #> 
  #> Against denominator:
  #>  RT ~ ID 
  #> ---
  #> Bayes factor type: BFlinearModel, JZS
  
  as.vector(result)
  #>            shape + ID            color + ID 
  #>             2.840104             2.841925 
  #>        shape + color + ID shape + color + shape:color + ID 
  #>            11.434689             4.873394
  

  
  
• NaomiHavron

  December 2019

  This works, thank you for such a swift and helpful response!

  ___________________________

  Now maybe you could also answer a more conceptual question?

  My own dataset gives inconclusive results for my effect of interest (BF10 = ~2).

  A reviewer asked how many participants I would have to add to cross my desired strength of evidence (BF10>3), and suggested I run simulations. I thought it makes sense to create a larger sample by creating a fake dataset with the same betas as my actual regression (which I obtained from a non-bayesian analysis):

  # Set coefficients
  alpha = -1.3091
  beta1 = .0556
  beta2 = -1.8874
  beta3 = 6.0730
  
  # Generate 100 trials
  A = c(rep(c(0), 50), rep(c(1), 50)) # '0' 100 times, '1' 100 times
  B = rep(c(rep(c(0), 25), rep(c(1), 25)), 2) # '0'x50, '1'x50, '0'x50, '1'x50
  e = rnorm(100, 0, sd=1) # Random noise, with standard deviation of 1
  
  # Generate data using the regression equation
  y = alpha + beta1*A + beta2*B + beta3*A*B + e
  
  # Join the variables in a data frame
  data = data.frame(cbind(A, B, y))
  
  # Fit a regression
  model = generalTestBF(y ~ A*B, data=data)
  
  

  Now, if I increase the number of rows in my fake data set (let's say from 100 to 300), the strength of evidence should logically increase, am I right? Or maybe it doesn't have to? Because when I do that I get more or less the same result (~700 of my fake datasets give a result around BF10~3, and the rest give a BF~2). Maybe it's my NHT logic at work and it doesn't work this way for a Bayesian analysis?
• MSB

  December 2019

  Generally, BFs are suppose to converge to the "truth" as N increases. So I depends what "truth" your betas are closer to, I guess.

  For determining N, you might also be interested in https://github.com/nicebread/BFDA
• NaomiHavron

  December 2019

  Well, according to the references on the BDFA page (thank you!!) the evidence should become more conclusive as I increase sample size (either towards H1 or H0, but it should still change). So I must be doing something wrong somewhere.
• EJ

  December 2019

  Hi Naomi,

  In general, asymptotically, and on average (!), increasing sample size leads to more conclusive BFs. But for a specific data set, a BF might become less compelling as N grows (up to a point). Also, the BFs may increase (again, on average) only slowly, depending on the details of the model specification -- if the data show a whopper of an effect, evidence will initially increase steeply, but if the effect is modest the change is evidence may only be noticeable for a large increase in sample size.

  Cheers

  E.J.
• NaomiHavron

  December 2019

  Thanks!