Featured post

Textbook: Writing for Statistics and Data Science

If you are looking for my textbook Writing for Statistics and Data Science here it is for free in the Open Educational Resource Commons. Wri...

Thursday, 20 June 2019

Replication Report - Informative Priors and Bayesian Updating

The following is a report on the reproduction of the statistical work in the paper "The use of informative priors and Bayesian updating: implications for behavioral research" by Charlotte O. Brand et al.

The original paper was accepted for publication by Meta-Psychology, https://open.lnu.se/index.php/metapsychology, a journal focused on methodology and reproductions of existing work. This report continues my attempt to establishing a standard for replication reports according to the Psych Data Standards found here https://github.com/psych-ds/psych-DS

"Informative priors" is a simulation study that examines the differences between analysis methods of a collection of simple synthetic datasets. In each of 60 datasets there are 80 subjects that answer 25 questions each. The response is the number out of 25 questions correct, the covariates are the sex of the respondent, and the condition under which they were put (a neutral control or a stereotype threat). Each respondent also had a hidden latent variable which was his or her aptitude, a log-odds of getting any one of the questions correct in neutral conditions. Each dataset had 40 men and 40 women, and 20 respondents of each sex were subjected to each condition. Men under the stereotype threat condition enjoyed an advantage which the remaining respondents were unaffected.

The same 60 datasets were analyzed under five different methods, ANOVA, GLMM, Bayesian GLMM, Posterior Passing, and "mega" Bayesian GLMM. The simulation was repeated under 25 different combinations of aptitude variance and stereotype threat strength. These analyses are described below as necessary.

For the computational work, I relied heavily on the code provided by the authors at https://github.com/thomasmorgan/freqbayes2. Primarily, I used the analyses.R file to extract details about analyses. Also used was main.R, which is a wrapper and collection of arguments which are passed on to the simulation file (simulation.R), the population as well as calls the analysis (analysis.R) and plotting R  files.

My own version of the code for population simulation, experiment simulation, result extraction, and two of the five described analyses are included in the R code below. These are only meant to provide a shorthand for the author's much more extensive and stable code.

Analysis 1 (ANOVA)

Going from the text alone, there was some minor difficulty related to interpretation of the analysis done. It's not clear that the "effect size" mentioned refers to the interaction's F-statistic, the regression slope coefficient, or the regression t-statistic. Digging further, the analysis.R code uses the aov() function and extracts the sum of squares of the interaction for the effect size. While reproducing this analysis, I opted for the slope-coefficient of the interaction (divided by 25 for normalization purposes) from the lm() function, which seemed more in line with the other analyses. I obtained results very similar to those reported.

Analysis 2 (GLMM)

This analysis was clear and straightforward to reproduce. The version I programmed is quite slow and has some computational hiccups. The author's version is much more stable. Both versions use the well-tested glmer() command from the car package in R to conduct the analyses. As in the first analysis, any missing details necessary for reproduction in the text are made apparent in the supplementary code, such as the use of the Wald confidence interval.

Analysis 3 (Bayesian GLMM using MCMC)

This analysis uses the jags.model() function, as found in the well-tested rjags package. For reference, JAGS stands for "Just Another Gibbs Sampler". A Gibbs Sampler is a standard Monte Carlo-based method for estimating parameters from otherwise intractable distributions. The sampler requires a prior distribution to operate, and diffuse or non-informative prior was used in order to reduce the effect of the prior to a negligible amount. Finally, the jags.model() function in this code calls a BUGS (Bayesian inference Using Gibbs Sampling) script, which is found in "bglmm.R" file in the github link.

Analysis 4 (Posterior Passing)

My commentary on this analysis is from a position of relative ignorance; posterior passing is a new concept to me, at least under this name. It appears to be an iterative process for updating a posterior much as one would do with data becomes available sequentially in time, such as sports results, or, in this case, a sequence of identical experiments on a common phenomenon.

Both this and the Bayesian GLMM analysis use the jags.model() function to make their inferences via Gibbs sampling. The difference between the analyses is in the BUGS script. In "bglmm.R" script for Analysis 3, all four of the parameters (the intercept, the main effects for sex and for condition, and the interaction, respectively) are estimated using the same non-informative prior. In the "pp.R" script for Analysis 4, the first three parameters of interest are estimated this way, but the four parameter, the interaction (the parameter of interest) is estimated using a prior based on the median and (inverse) variance of the interaction parameter estimates of the previous experiment.

Analysis 5 ("Mega" Bayesian GLMM)

This analysis is very similar in structure to analysis 3. The "mega" part refers to the fact that the results from all 60 of the experiments are fed into the Gibbs Sampler together as if they came from one very large sample of size 4800, instead of 60 distinct samples of size 80 analyzed separately. It works the same and produces very similar results, but requires much more computational resources.

Analysis 5 also uses the jags.model() function and calls the same "bglmm.R" BUGS script as Analysis 3.

One final comment of the manuscript: Two sets of effect sizes are given. However, this the same set, just that in one location it is described as a difference in probability from 0.5 (0, 0.12, 0.23, 0.32, 0.38), and later as a difference in the log-odds from 0 (0, 0.5, 1, 1.5, 2).

### Population Generation
protoN = 250000
N = 4*protoN

protoLatent = rnorm(protoN)
sex = rep(c(0,1),each=2*protoN)
latent = c(protoLatent, -protoLatent, protoLatent, -protoLatent)

### Sample generation
Nexp = 60
sampSize = 80

sampleMat = matrix(NA,nrow=60,ncol=80)
for(k in 1:Nexp)
sampleMat[k,1:40] = c(sample(1:(N/2),20), sample((N/2)+(1:(N/2)),20))    
sampleMat[k,41:80] = c(sample(1:(N/2),20), sample((N/2)+(1:(N/2)),20))

sampleIdx = rep(1:Nexp,each=5*5*20)

### Experiment generation
vari = rep(rep(c(0,.25,.50,.75,1),times=5,each=20),times=Nexp)
effect = rep(rep(c(0,.5,1,1.5,2),each=100),times=Nexp)
effect_prob = rep(rep(c(0,.12,.23,.32,.38),each=100),times=Nexp)

Niter = length(sampleIdx)
manifestMat = matrix(NA,nrow=Niter,ncol=80)

for(k in 1:Niter)
     thisSample = sampleMat[sampleIdx[k],]
     thisVari = vari[k]
     thisEffect = effect[k]
     thisSex = sex[sampleMat[sampleIdx[k],]]
     thisCond = rep(c(0,1),each=40)
     thisLatent = latent[thisSample]*thisVari + thisSex*thisCond*thisEffect
     thisProb = exp(thisLatent) / (1 + exp(thisLatent))
     thisManifest = rbinom(n=80, size=25, prob=thisProb)
     manifestMat[k,] = thisManifest

#### ANOVA Analysis
sex = rep(c(0,1),each=20,times=2)
cond = rep(c(0,1),each=40)

ESE1 = rep(NA,Niter) # 1) Effect size estimate;
PRR1 = rep(NA,Niter) # 2) Positive result rate;
UNC1 = rep(NA,Niter) # 3) Uncertainty;
EER1 = rep(NA,Niter) # 4) Estimate error;

Niter = 30000
tcrit = qt(.975,df=76)
for(k in 1:Niter)
     y = manifestMat[k,]
     res = summary(lm(y ~ sex*cond))

     ESE1[k] = res$coef[4,1]/25
     PRR1[k] = res$coef[4,4]  # P-value, we can dimension reduce later
     UNC1[k] = res$coef[4,2]/25 * 2 * tcrit
     EER1[k] = res$coef[4,1]/25 - effect_prob[k]
     if(k %% 1000 == 0){print(k)}

### Results presentation (Analysis 1)
levels_effect = unique(effect)
ESE1grid = matrix(NA,length(levels_vari),length(levels_effect))
PRR1grid = matrix(NA,length(levels_vari),length(levels_effect))
UNC1grid = matrix(NA,length(levels_vari),length(levels_effect))
EER1grid = matrix(NA,length(levels_vari),length(levels_effect))

for(i in 1:length(levels_vari))
     for(j in 1:length(levels_effect))
           ESE1grid[i,j] = mean( ESE1[which(vari == levels_vari[i] & effect == levels_effect[j])])
           PRR1grid[i,j] = mean( PRR1[which(vari == levels_vari[i] & effect == levels_effect[j])])
           UNC1grid[i,j] = mean( UNC1[which(vari == levels_vari[i] & effect == levels_effect[j])])
           EER1grid[i,j] = mean( EER1[which(vari == levels_vari[i] & effect == levels_effect[j])])


sex = rep(c(0,1),each=20,times=2)
cond = rep(c(0,1),each=40)
sexlong = rep(sex,each=25)
condlong = rep(cond,each=25)
IDlong = rep(1:80,each=25)
Niter = 30000

ESE2 = rep(NA,Niter) # 1) Effect size estimate
PRR2 = rep(NA,Niter) # 2) Positive result rate
UNC2 = rep(NA,Niter) # 3) Uncertainty
EER2 = rep(NA,Niter) # 4) Estimate error;

zcrit = qnorm(.975)

for(k in 1:Niter)
     y = manifestMat[k,]
     ylong = rep(0,80*25)

     for(i in 1:80)
           ylong[(i-1)*25+(1:25)] = c(rep(1,y[i]),rep(0,25-y[i]))
     res = summary(glmer(ylong ~ sexlong*condlong + (1|IDlong), family="binomial"))

     ESE2[k] = res$coef[4,1]
     PRR2[k] = res$coef[4,4]  # P-value, we can dimension reduce later
     UNC2[k] = res$coef[4,2] * 2 * zcrit
     EER2[k] = res$coef[4,1] - effect[k]
     if(k %% 1000 == 0){print(k)}

### Results presentation (Analysis 2)
levels_vari = unique(vari)
levels_effect = unique(effect)
ESE2grid = matrix(NA,length(levels_vari),length(levels_effect))
PRR2grid = matrix(NA,length(levels_vari),length(levels_effect))
UNC2grid = matrix(NA,length(levels_vari),length(levels_effect))
EER2grid = matrix(NA,length(levels_vari),length(levels_effect))

for(i in 1:length(levels_vari))
     for(j in 1:length(levels_effect))
           ESE2grid[i,j] = mean( ESE2[which(vari == levels_vari[i] & effect == levels_effect[j])])
           PRR2grid[i,j] = mean( PRR2[which(vari == levels_vari[i] & effect == levels_effect[j])])
           UNC2grid[i,j] = mean( UNC2[which(vari == levels_vari[i] & effect == levels_effect[j])])
           EER2grid[i,j] = mean( EER2[which(vari == levels_vari[i] & effect == levels_effect[j])])


The previous replication report on Jessica Witt's work is found here: https://www.stats-et-al.com/2019/02/replication-report-signal-detection.html

No comments:

Post a comment