I have performed Bayesian Mann-Whitney test in JASP.

When I test H1: 1>2, Bayes factor is inconclusive

When I test H1: 1~=2, BF is moderately on a side of H0

When I test H1: 1<2 BF i strongly on a side of H0.

I think that the proper interpretation would be that either the groups are equal or the 1>2.

There is no conclusion which of these two possibilities should we believe more , but we can quite firmly discard the possibility 1<2. Am I right? Can I draw such a joint conclusions from three tests?

Moreover, my bayes factor and credible interval values of main effect also change when I add a covariate. This has huge consequences for the main effects found, as somehow there occurs an effect but merely when no covariate is added. But again, I do not know why that is the case.

Another question I have also concerns the model averaged posterior summary. How is the mean and standard deviation built here, as these are not the actual means or SD`s of the scores. Is there a certain distribution when bayesian statistics is applied? I could not find that information in the student guide, unfortunately.

I would really appreciate, if you could help me with this problem.

Kind regards,

Niklas

]]>I downloaded open source demographic data from governmental source in .csv.

Once I uploaded it on JASP, the variable "death" which is the number of death should be continuous since it is nominal by default. The problem is that I can't change the nature of nominal variable.

However I can easily change variables from a continuous, to categorical and nominal, and vice versa.

Is it about the file itself, how the data has been transcripted through .csv and than JASP ?

Is it because it is a nominal variable with the white "a" on the icon ? What does "a" mean ?

Also, I can't change the value on the variable. Is it normal ?

Thanks for your help

]]>Beforehand, For my masterthesis, I conducted a field study. The following study design was chosen: 2 experimental groups. Each participant of each group experienced two different systems. Participants of group 1 first experienced system 1 and then system 2, particpants of group 2 first experienced system 2 and afterwards system 1. Trust and Mistrust ratings were collected after each system was experienced. So for group 1, each participant had 2 trust and mistrust ratings, respectively. The same for group 2, logically. Hence, trust and mistrust ratings are the independent variable. In order to eliminate effects due to the measurement time points or order of experienced systems, before starting the actual analysis, I calculated a bayesian repeated measures anova with two factors (Think of a 2x2 table with trial I and II as the measurement time point in columns and the group I and II as the order of system in rows( the cell of trial I and group I would be system 1, the cell of trial II and group I would be System 2. The same for group 2 but the other way around):

Factor 1 = Measurement time point with two levels, namely trial I and trial II

Factor 2 = Trust/Mistrust with two levels, namely trust and mistrust, of course.

Therefore, the four repeated measures cells are:

).Trial I & Trust, 2)Trial I & Mistrust, 3)Trial II & Trust, 4)Trial II & Mistrust

I added the between subject factor Group.

When the interaction effect of measurement time point * group on the mistrust and trust ratings is calculated, I get the outcome of the potential effect the order of systems had on the trust and mistrust ratings. So, it is compared, whether mistrust and trust ratings were different between the four the cells due to order.

Now the actual problem I had: the output values of the credible intervals for mistrust and trust ratings are the same for system 1 in both groups and for system 2 in both groups:

For system 1 in group 1 (Trial 1 and group 1): (*95% CI* [-0.20; 0.11])

For system 1 in group 2 (Trial 2 and group 2) : (*95% CI* [-0.20; 0.11])

For system 2 in group 1 (Trial 2 and group 1): (*95% CI* [-0.11; 0.20])

For system 2 in group 2 (Trial 1 and group 2): (*95% CI* [-0.11; 0.20])

Is this usual? Is this due to the distribution of bayesian parameter estimation? By the way, how are the means and sd`s in the model averaged posterior summery built? I really do not understand this. I also checked my excel table and I could exclude a mistake there.

Considering that the means and sd`s of trust and mistrust ratings for System 1 of group 1 and 2 are nearly the same as well as the ratings for system 2 of group 1 and 2, would this be an explanation for the same credible intervals?

I really have the feeling, that there is a mistake, as the means are differing with a value of .5 approximately.

I hope you can understand the issue I tried to explain to you.

For safety, I added two screenshot of the jasp output, one of the descriptive

Kind regards,

Niklas

Super big fan of your work. Was just curious what bloats JASP files so much.

Example 1: Data file is 27.9 kB, associated JASP file is 47.6 MB (!). There are two 2x3 repeated-measures ANOVAs in there, each with 15 post-hoc tests and 1 graph at 300 dpi.

Example 2: Data file is 173.0 kB, associated JASP file is 917 MB (!!!). This has an embarrassingly big repeated-measures ANOVA in there (2*2*24), with two graphs at 300 dpi.

It can't just be the images, given the huge difference between the files. Are the values saved in a memory-intensive way, or are a lot of in-between steps saved too?

Thanks!

Edwin

]]>Today I wrote a script to compare output from ttestBF and ttest.tstat, as I was confused by the differences, and it looks like the output from ttestBF is a raw odds ratio, whereas the output from ttest.tstat is a log odds ratio.

If I try to work out the odds ratios directly using the t distribution or Cauchy distribution (pt or pcauchy), I get values that are correlated with the Bayes factors from the package, but are very different in magnitude. I think this is almost certainly due to the fact that I am uncertain about how to incorporate the prior. If I could see how this is done in the package, this would be a great help to my understanding. I will append a short script (rmd) I have used to try and test this.

```{r generatedata}

#Generate 2 groups of 20, with mean difference m

makedat<-function(N,m){

mydat1<-rnorm((N/2),m,1)

mydat2<-rnorm((N/2),0,1)

myt<-t.test(mydat1,mydat2,paired=FALSE)$statistic

bf = ttestBF(mydat1,mydat2,mu=0,paired=FALSE,rscale="medium")

bf<-data.frame(bf)$bf # convert to data frame to access bf element

myresults<-c(myt,bf)

return(myresults)

}

```

## Create data frame to hold results for repeated runs

```{r dosums}

nrun<-10

sumdat<-data.frame(matrix(nrow=nrun,ncol=11))

colnames(sumdat)<-c('run','t','ttestBF','log.ttestBF','BFttest.tstat','p.t0','p.t1','logORt','p.cauchy0','p.cauchy1','logORcauchy')

N=40 #total N for both groups

m=.8 # true mean difference in population - can vary this

esteff=1 #effect size for H1 when computing BF

estcauch<-1

for (n in 1:nrun){

sumdat$run[n]<-n

myresults<-makedat(N,m)

sumdat$t[n]<-myresults[1]

sumdat$ttestBF[n]<-myresults[2]

sumdat$log.ttestBF[n]<-log(sumdat$ttestBF[n])

sumdat$BFttest.tstat[n]<-ttest.tstat(sumdat$t[n], N/2, N/2, rscale = 1,

complement = FALSE, simple = FALSE)$bf

sumdat$p.t0[n]<-1-pt(sumdat$t[n],df=(N-2),lower.tail=TRUE)

sumdat$p.t1[n]<-1-pt((esteff-sumdat$t[n]),df=(N-2),lower.tail=TRUE)

sumdat$logORt[n]<-log(sumdat$p.t1[n]/sumdat$p.t0[n])

sumdat$p.cauchy0[n]<-1-pcauchy(sumdat$t[n],lower.tail=TRUE)

sumdat$p.cauchy1[n]<-1-pcauchy((estcauch-sumdat$t[n]),lower.tail=TRUE)

sumdat$logORcauchy[n]<-log(sumdat$p.cauchy1[n]/sumdat$p.cauchy0[n])

}

plot(sumdat$log.ttestBF,sumdat$BFttest.tstat)

plot(sumdat$log.ttestBF,sumdat$logORt,pch=15)

lines(sumdat$log.ttestBF,sumdat$logORcauchy,type='p')

text(0,6,'Filled logOR.t \nand open logORcauchy')

```

]]>I include the factors "Type" (probe vs. irrelevant) and "Hand-position" (index vs. thumb), for the "duration RT mean" variable: "duration_RT_mean_probe_0", "duration_RT_mean_irrelevant_0", "duration_RT_mean_probe_1", "duration_RT_mean_irrelevant_1" (where _0 is index and _1 is thumb). See:

And what's weird is that for the interaction I find a p = .004:

But a BF = 0.182 (inclusion BF based on matched model - when based on all models, it's even 0.077):

So p value supports difference quite strongly (.004), and BF supports equivalence substantially (1 / 0.182 = 5.495). (There is a similar contradiction for the Type main effect too, but there at least the p value is not so clear.)

Any ideas what to make of this? Perhaps any references about interpreting such a case?

]]>One of my collaborators tried this in JASP and gets a BF around ~0.8-1.3. I have tried re-running my model in R using the 'withmain' option (to make it similar to JASP), but I'm still getting a similar BF to the one I got in R originally.

I was wondering if this has something to do with the priors? I know BayesFactor uses JZS priors, but does JASP use the same? If not, does anyone have any ideas why this discrepancy may be emerging?

]]>I would like to consult this forum, and check that I understand how to interpret the error percentage presented with the Bayes factor values.

If I understand the JASP manual correctly, the error percentage is the percentage to which the BF value can increase or decrease due to Monte-Carlo simulation noise randomisation. So, for example, if I have a BF10 value of 14.150 and an 28.539% error percentage, it means that upon recomputation, the BF10 can fluctuate approximately between 10 to 18 (14.150*0.285= approximately 4.05).

Is that a correct interpretation?

Thank you very much,

Lior

]]>I am interested to examine the effect of a three-way interaction with two within-participants variables (2 levels each) and one between-participants variable (2 levels).

I used JASP to examine it with a Bayesian analysis. If I understand correctly, to test the effect of the three-way interaction I need to divide the BF10 of the model with the interaction by the BF10 of the model without the interaction (i.e., the model that includes all the possible effects except for the three-way interaction). Alternatively, I can divide the equivalent B(M|data) indexes of these two models.

The problem is that the two ways give different results (BF10=2 ; P(M|data)=3), and now the question is- which one should I count on?

To complex things further, as a sanity check, I examined the same dataset with the R package "BayesFactor", and there I got a third result (BF10=2.78). Does anyone know what are the (meaningful) differences between the JASP and the BayesFactor algorithms?

Thanks a lot!

Lior

]]>So I have been wondering something lately; is it possible to have JASP give you some sort of indicator of the effect for an AN(C)OVA, or t-test? I'm looking for some sort of Bayesian equivalent of partial eta squared or Cohen's d. Would be amazing if there was a paper somewhere with guidelines on how to qualitatively interpret that Bayesian effect size as small, medium or large.

]]>Sometimes JASP (NHST analyses) reports that p = 1.000. I believe there is no way of finding out what the exact number is, but I assume it is incredibly close to 1.000. What would be the best way to report this?

]]>I'm struggling to color my nodes in the network analysis. I want to create a variable/column for this which includes 6 groups of colors, each group will consist out of several variables.

Can anyone help me to R-code this, I can't find any examples.

Much appreciated!

Regard,

Alex

]]>Any idea?

Thanks!

]]>On my desktop, doing the same exact with 0.9 works fine.

I'm wondering what might be happening. Does version 10 depend on JAVA or something that version 9 does not? Any hint would be appreciated. I tried reinstalling about 5 times over two weeks time...

]]>In a scenario that I run a Bayes linear regression analysis adding say 8 predictors. The analysis may come up with the best regression model incorporating 4 predictors. However, regression coefficients are given for all 8 variables. Should I use these for the regression equation or would it be prudent to re-run the regression just including the best models 4 predictors and then use those regression coefficients?

Cheers

Mark

]]>I understand that if you want to quantify the evidence in favour of the alternative interaction hypothesis, you would compare the BF10 of [Time + Response + Time*Response] to the BF10 of [Time + Response]. However, does the same logic apply if you want to quantify the evidence in favour of the null model, as compared to an interaction model?

In the attached example, the null model is favoured to both a main effects model and a main effects + interaction model. Considering that a model is said to be complete if it contains all lower-order terms (Bernhardt & Jung 1979), I am unsure if the evidence in support of the null model compared to an interaction model should be:

A) Null model vs. Time + Response + Time*Response (BF01 = 13.17, therefore the null model is 13 times more likely than an interaction model)

Or

B) Time + Response + Time*Response vs. Time + Response (BF01 = 0.996, therefore the evidence is inconclusive)?

Thanks in advance for your help.

]]>Under a set of hypotheses, I received as output the attached file. Is it not weird that the BF32 is only 1.36 when the BF of H3 is 24757.71 and that of H2 8.11? I understand that BF32 is not a division of the two numbers, but still I find it weird that the ratio is so small.

Can someone explain to me whether it is possible to calculate CS- coefficient? I mean centrality stability coefficient.

The second... How should I interpret this graph? What's the meaning of the black and red line, and gray area?

im practising using JASP and am interested in the network module

however- i have hit a problem with adding colour to the nodes….i cannot work out how to add the instruction/correct label for the analysis to add in the coloured nodes...and wonder if someone can advise (as i cannot find the solution in the gift/video or manual)

I am trying to add some coloured nodes into my network analyses-

I’ve tried various ways to make this work….but it fails every time…and i cannot work out what is wrong with the variable or the coding to make it work...

The data I am using is in an spss file….so e.g. I cannot include ‘+’ in the column name

But everything I try comes up with the following warning

*Warning: * Data supplied in SEX could not be used to determine variables types. Data should: Start with the column name of the variable. Contain an '=' to distinghuish betweem column name and group.

I wonder if you can advise me what the instruction means (i apologise but I don’t quite understand this sentence)…

And/or can you show me an example of the required labels

Thanking you in advance

]]>finally I analysed my own data for a study, that I'm supposed to replicate. Remeber the orignal effect was really huge (p< .001, ηp^2 = .136; BF10 =800000...)?

Now, in the replication data, the ANOVA shows no evidence for anything: p = .075, np^2 =.057; BF01=1.66 with a power of 60% (N=122).

Now, what would be a sensible answer to the question, whether the replication was successful? On the first glance, the previously set criteria (significant p value; CI of Eta to include original effect size) were not met. But what if I took into account the huge effect that showed in the original study and altered the prior probability of H1 accordingly? Actually, everybody expected the effect to replicate with a 90% probability. Is there any way to do that with JASP?

Thanks so much,

Jan

I am analyzing the effect of two different drugs on cognition. Clinically, we expect both drugs to have an effect, one possibly more than the other. Therefore, I used a 2x2 design with the within-subject factor "time" and the between-subject factor "treatment arm". The treatment arms differed in two other factors, which were included as covariates.

I'd like to use Bayes Statistics, but put the frequentist equivalent into additional material. Interestingly, the two analyses show different results: The RM ANOVA shows a significant if small interaction effect time * treatment arm (F=4.32, p=.04, partial eta² 0.06) but no simple effects of either factor. The Bayes RM ANOVA shows a very strong simple effect of time (BFincl = 1124.9), but no interaction effect.

I don't think the results are necessarily incompatible with our clinical observations but I'm confused by the difference.

I checked the forum for other discussions and found violated assumptions as possible explanation. Sphericity is met as the RM factor has only two levels, but Levene's test for homogeneity is significant. Trying to use the nonparametric addition (and thank you for that!) gave me the error message "Specified ANOVA design is not balanced". I am not sure what this means.

Are there any other possible reasons for the differences in the results? Is there some other procedure I should try?

Thank you very much!

I wanted to ask which underlying R function JASP uses to run the parallel analysis in order to quantify the number of factors for an exploratory factor analysis?

I ran the same EFA in JASP and R using the fa.parallel function (psych package). With the fa.parallel function, I get 6 factors no matter which factor method I use (minres, ml, wls, gls, pa). However, JASP outputs 10 factors when I use the parallel analysis option to extract the number of factors.

Looking at the scree plot, I can see that one could argue for two different points of inflexion, yielding either 6 or 10 factors. However, based on the scree plot, I would think that 6 factors are the more straightforward solution and I am puzzled that the parallel analysis in JASP yields a different result. Therefore, I would really appreciate your help!

]]>I've read in another post that the requirements for performing Bayesian statistics are similar to the requirements in frequentist statistics (such as normal distributions, sphericity etc. for different analyses).

I've seen in JASP (Version 0.10.2) that there are non-parametric options for the frequentist statistics (e.g. Friedman test for the RM ANOVA). But I haven't found non-parametric tests in the Bayesian statistics section. So, I was wondering whether there is e.g. a Bayesian Friedman test in JASP?

Thanks,

SClausen

]]>I just realized, there are two ways in the Bayesian ANOVA to produce R^2: There is model averaged R^2 and Single model inference R^2. What I'm looking for, is Eta^2 produced by a Bayesian ANOVA analysis. Would it be correct to report the model averaged R^2 = Eta^2?

Thanks for your help,

Jan

]]>Thanks to JASP, I decided to work for my next article with Bayesian statistics, and it seems I need some help:

My dataset consists of one continuous independent variable (IV) and a categorical variable with four levels (so M1, M2, M3 and M4, the averages of the groups in the IV). If I understand correctly, in a Bayesian ANOVA, JASP will compare HO: M1 = M2 = M3 = M4 with H1: M1, M2, M3, M4. Is that so?

However, based on the literature, my alternative hypothesis would be H1' = M1 > {M2, M3, M4}. Can the Bayesian ANOVA compare H0 and H1'? How do I specify the H1' so that the analysis compares H0 with H1' and not H1?

In the literature, many studies prefer to group M2, M3 and M4 and compare the resulting group with M1. In older literature, researchers were grouping M1, M3 and M4 and comparing the resulting group with M2. Using a Bayesian analysis, I would like to see which of the two groupings is more informative. This would mean the comparison of two hypotheses:

H2 = M1 > {M2, M3, M4} and H3 = M2 < {M1, M3, M4}

Is a Bayesian t-test of a Bayesian ANOVA better suited to compare H2 and H3? If the answer is an ANOVA, how do I define the direction of difference to be tested? (in the t-test, I should be using a one-sided t-test)

Thank you in advance!

Georgios.

]]>(sorry for my very bad english)

I discover the version 10 and i have a question : why did you remove the Dunn's post hoc test in non parametric ANOVA?

Hervé

]]>