Bayesian inference for psychology. Part II: Example applications with JASP
 11k Downloads
 20 Citations
Abstract
Bayesian hypothesis testing presents an attractive alternative to p value hypothesis testing. Part I of this series outlined several advantages of Bayesian hypothesis testing, including the ability to quantify evidence and the ability to monitor and update this evidence as data come in, without the need to know the intention with which the data were collected. Despite these and other practical advantages, Bayesian hypothesis tests are still reported relatively rarely. An important impediment to the widespread adoption of Bayesian tests is arguably the lack of userfriendly software for the runofthemill statistical problems that confront psychologists for the analysis of almost every experiment: the ttest, ANOVA, correlation, regression, and contingency tables. In Part II of this series we introduce JASP (http://www.jaspstats.org), an opensource, crossplatform, userfriendly graphical software package that allows users to carry out Bayesian hypothesis tests for standard statistical problems. JASP is based in part on the Bayesian analyses implemented in Morey and Rouder’s BayesFactor package for R. Armed with JASP, the practical advantages of Bayesian hypothesis testing are only a mouse click away.
Keywords
Hypothesis test Statistical evidence Bayes factor Posterior distributionAs demonstrated in part I of this series, Bayesian inference unlocks a series of advantages that remain unavailable to researchers who continue to rely solely on classical inference (Wagenmakers et al. 2017). For example, Bayesian inference allows researchers to update knowledge, to draw conclusions about the specific case under consideration, to quantify evidence for the null hypothesis, and to monitor evidence until the result is sufficiently compelling or the available resources have been depleted. Generally, Bayesian inference yields intuitive and rational conclusions within a flexible framework of information updating. As a method for drawing scientific conclusions from data, we believe that Bayesian inference is more appropriate than classical inference.
Pragmatic researchers may have a preference that is less pronounced. These researchers may feel it is safest to adopt an inclusive statistical approach, one in which classical and Bayesian results are reported together; if both results point in the same direction this increases one’s confidence that the overall conclusion is robust. Nevertheless, both pragmatic researchers and hardcore Bayesian advocates have to overcome the same hurdle, namely, the difficulty in transitioning from Bayesian theory to Bayesian practice. Unfortunately, for many researchers it is difficult to obtain Bayesian answers to statistical questions for standard scenarios involving correlations, the ttest, analysis of variance (ANOVA), and others. Until recently, these tests had not been implemented in any software, let alone userfriendly software. And in the absence of software, few researchers feel enticed to learn about Bayesian inference and few teachers feel enticed to teach it to their students.
To narrow the gap between Bayesian theory and Bayesian practice we developed JASP (JASP Team 2017), an opensource statistical software program with an attractive graphical user interface (GUI). The JASP software package is crossplatform and can be downloaded free of charge from http://www.jaspstats.org. Originally conceptualized to offer only Bayesian analyses, the current program allows its users to conduct both classical and Bayesian analyses.^{1} Using JASP, researchers can conduct Bayesian inference by dragging and dropping the variables of interest into analysis panels, whereupon the associated output becomes available for inspection. JASP comes with default priors on the parameters that can be changed whenever this is deemed desirable.
This article summarizes the general philosophy behind the JASP program and then presents five concrete examples that illustrate the most popular Bayesian tests implemented in JASP. For each example we discuss the correct interpretation of the Bayesian output. Throughout, we stress the insights and additional possibilities that a Bayesian analysis affords, referring the reader to background literature for statistical details. The article concludes with a brief discussion of future developments for Bayesian analyses with JASP.
The JASP philosophy
The JASP GUI is familiar to users of SPSS and has been programmed in C++, html, and javascript. The inferential engine is based on R (R Development Core Team 2004) and –for the Bayesian analyses– much use is made of the BayesFactor package developed by Morey and Rouder (2015) and the conting package developed by Overstall and King (2014b). The latest version of JASP uses the functionality of more than 110 different R packages; a list is available on the JASP website at https://jaspstats.org/rpackagelist/. The JASP installer does not require that R is installed separately.
Our longterm goals for JASP are twofold: the primary goal is to make Bayesian benefits more widely available than they are now, and the secondary goal is to reduce the field’s dependence on expensive statistical software programs such as SPSS.
Example 1: a Bayesian correlation test for the height advantage of US Presidents
For our first example we return to the running example from Part I. This example concerned the height advantage of candidates for the US presidency (Stulp, Buunk, Verhulst, & Pollet, 2013). Specifically, we were concerned with the Pearson correlation ρ between the proportion of the popular vote and the height ratio (i.e., height of the president divided by the height of his closest competitor). In other words, we wished to assess the evidence that the data provide for the hypothesis that taller presidential candidates attract more votes. The scatter plot was shown in Figure 1 of Part I. Recall that the sample correlation r equaled .39 and was significantly different from zero (p = .007, twosided test, 95% CI [.116,.613]); under a default uniform prior, the Bayes factor equaled 6.33 for a twosided test and 12.61 for a onesided test (Wagenmakers et al. 2017).

Descriptives (with the option to display a matrix plot for selected variables).

Reliability analysis (e.g., Cronbach’s α, Gutmann’s λ6, and McDonald’s ω).

Independent samples ttest, paired samples ttest, and one sample ttest. Key references for the Bayesian implementation include Jeffreys (1961), Ly, Verhagen, and Wagenmakers (2016a, 2016b), Rouder, Speckman, Sun, Morey, and Iverson (2009) and Wetzels, Raaijmakers, Jakab, and Wagenmakers (2009).

ANOVA, repeated measures ANOVA, and ANCOVA. Key references for the Bayesian implementation include Rouder, Morey, Speckman, and Province (2012), Rouder, Morey, Verhagen, Swagman, and Wagenmakers (in press), and Rouder, Engelhardt, McCabe, and Morey (in press).

Correlation. Key references for the Bayesian implementation include Jeffreys (1961), Ly et al. (2016b), and Ly, Marsman, and Wagenmakers (in press) for Pearson’s ρ, and van Doorn, Ly, Marsman, and Wagenmakers (in press) for Kendall’s tau.

Linear regression. Key references for the Bayesian implementation include Liang, Paulo, Molina, Clyde, and Berger (2008), Rouder and Morey (2012), and Zellner and Siow (1980).

Binomial test. Key references for the Bayesian implementation include Jeffreys (1961) and O’Hagan and Forster (2004).

Contingency tables. Key references for the Bayesian implementation include Gunel and Dickey (1974) and Jamil et al. (in press).

Loglinear regression. Key references for the Bayesian implementation include Overstall and King (2014a) and (2014b).

Principal component analysis and exploratory factor analysis.
 1.
The Bayes factor expressed as BF_{10} (and its inverse BF_{01} = 1/BF_{10}), grading the intensity of the evidence that the data provide for \(\mathcal {H}_1\) versus \(\mathcal {H}_0\) (for details see Part I).
 2.
A proportion wheel that provides a visual representation of the Bayes factor.
 3.
The posterior median and a 95% credible interval, summarizing what has been learned about the size of the correlation coefficient ρ assuming that \(\mathcal {H}_1\) holds true.
 4.
A figure showing (a) the prior distribution for ρ under \(\mathcal {H}_1\) (i.e., the uniform distribution, which is the default prior proposed by Jeffreys (1961) for this analysis; the user can adjust this default specification if desired), (b) the posterior distribution for ρ under \(\mathcal {H}_1\), (c) the 95% posterior credible interval for ρ under \(\mathcal {H}_1\), and (d) a visual representation of the SavageDickey density ratio, that is, grey dots that indicate the height of the prior and the posterior distribution at ρ = 0 under \(\mathcal {H}_1\); the ratio of these heights equals the Bayes factor for \(\mathcal {H}_1\) versus \(\mathcal {H}_0\) (Dickey & Lientz, 1970; Wagenmakers, Lodewyckx, Kuriyal, & Grasman, 2010).
Before proceeding we wish to clarify the meaning of the proportion wheel or “pizza plot”. The wheel was added to assist researchers who are unfamiliar with the odds formulation of evidence – the wheel provides a visual impression of the continuous strength of evidence that a given Bayes factor provides. In the presidents example BF_{10} = 6.33, such that the observed data are 6.33 times more likely under \(\mathcal {H}_1\) than under \(\mathcal {H}_0\). To visualize this ratio, we transform it to the 01 interval and plot the resulting magnitude as the proportion of a circle (e.g., Tversky, 1969, Figure 1; Lipkus & Hollands, 1999). For instance, the presidents example has a ratio of BF_{10} = 6.33 and a corresponding proportion of 6.33/7.33 ≈ 0.86;^{3} consequently, the red area (representing the support in favor of \(\mathcal {H}_1\)) covers 86% of the circle and the white area (representing the support in favor of \(\mathcal {H}_0\)) covers the remaining 14%.
The proportion wheel underscores the fact that the Bayes factor provides a graded, continuous measure of evidence. Nevertheless, for historical reasons it may happen that a discrete judgment is desired (i.e., an allornone preference for \(\mathcal {H}_0\) or \(\mathcal {H}_1\)). When the competing models are equally likely a priori, then the probability of making an error equals the size of the smaller area. Note that this kind of “error control” differs from that which is sought by classical statistics. In the Bayesian formulation the probability of making an error refers to the individual case, whereas in classical procedures it is obtained as an average across all possible data sets that could have been observed. Note that the longrun average need not reflect the probability of making an error for a particular case (Wagenmakers et al. 2017).
JASP offers several ways in which the present analysis may be refined. In Part I we already showed the results of a onesided analysis in which the alternative hypothesis \(\mathcal {H}_+\) stipulated the correlation to be positive; this onesided analysis can be obtained by ticking the check box “correlated positively” in the input panel. In addition, the twosided alternative hypothesis has a default prior distribution which is uniform from − 1 to 1; a userdefined prior distribution can be set through the input field “Stretched beta prior width”. For instance, by setting this input field to 0.5 the user creates a prior distribution with smaller width, that is, a distribution which assigns more mass to values of ρ near zero.^{4} Additional check boxes create sequential analyses and robustness checks, topics that will be discussed in the next example.
Example 2: a Bayesian ttest for a kitchen roll rotation replication experiment
Across a series of four experiments, the data reported in Topolinski and Sparenberg (2012) provided support for the hypothesis that clockwise movements induce psychological states of temporal progression and an orientation toward the future and novelty. Concretely, in their Experiment 2, one group of participants rotated kitchen rolls clockwise, whereas the other group rotated them counterclockwise. While rotating the rolls, participants completed a questionnaire assessing openness to experience. The data from Topolinski and Sparenberg (2012) showed that, in line with their main hypothesis, participants who rotated the kitchen rolls clockwise reported more openness to experience than participants who rotated them counterclockwise (but see Francis, 2013).
“We will collect aminimum of 20 participants in each betweensubject condition (i.e., the clockwise and counterclockwise condition, for aminimum of 40 participants in total). We will then monitor the Bayes factor and stop the experiment whenever the critical hypothesis test (detailed below) reach aBayes factor that can be considered “strong” evidence (Jeffreys 1961); this means that the Bayes factor is either 10 in favor of the null hypothesis, or 10 in favor of the alternative hypothesis. The experiment will also stop whenever we reach the maximum number of participants, which we set to 50 participants per condition (i.e., amaximum of 100 participants in total). Finally, the experiment will also stop on October 1st, 2013. From aBayesian perspective the specification of this sampling plan is needlessly precise; we nevertheless felt the urge to be as complete as possible.”
In addition, the preregistration form indicated that the Bayes factor of interest is the default onesided ttest as specified in Rouder et al. (2009) and Wetzels et al. (2009). The twosided version of this test was originally proposed by Jeffreys (1961), and it involves a comparison of two hypothesis for effect size δ: the null hypothesis \(\mathcal {H}_0\) postulates that effect size is absent (i.e., δ = 0), whereas the alternative hypothesis \(\mathcal {H}_1\) assigns δ a Cauchy prior centered on 0 with interquartile range r = 1 (i.e., δ ∼Cauchy(0,1)). The Cauchy distribution is similar to the normal distribution but has fatter tails; it is a tdistribution with a single degree of freedom. Jeffreys chose the Cauchy because it makes the test “information consistent”: with two observations measured without noise (i.e., y _{1} = y _{2}) the Bayes factor in favor of \(\mathcal {H}_1\) is infinitely large. The onesided version of Jeffreys’s test uses a folded Cauchy with positive effect size only, that is, \(\mathcal {H}_+: \delta \sim \text {Cauchy}^+(0,1)\).
The specification \(\mathcal {H}_+: \delta \sim \text {Cauchy}^+(0,1)\) is open to critique. Some people feel that this distribution is unrealistic because it assigns too much mass to large effect sizes (i.e., 50% of the posterior mass is on values for effect size larger than 1); in contrast, others feel that this distribution is unrealistic because it assigns most mass to values near zero (i.e., δ = 0 is the most likely value). It is possible to reduce the value of r, and, indeed, the BayesFactor package uses a default value of \(r = \frac {1}{2}\sqrt {2} \approx 0.707\), a value that JASP has adopted as well. Nevertheless, the use of a very small value of r implies that \(\mathcal {H}_1\) and \(\mathcal {H}_0\) closely resemble one another in the sense that both models make similar predictions about tobeobserved data; this setting therefore makes it difficult to obtain compelling evidence, especially in favor of a true \(\mathcal {H}_0\) (Schönbrodt, Wagenmakers, Zehetleitner, & Perugini, in press). In general, we feel that reducing the value of r is recommended if the location of the prior distribution is also shifted away from δ = 0. Currently JASP fixes the prior distribution under \(\mathcal {H}_1\) to the location δ = 0, and consequently we recommend that users deviate from the default setting only when they realize the consequences of their choice.^{5} Note that Gronau, Ly, and Wagenmakers (2017) recently extended the Bayesian ttest to include prior distributions on effect size that are centered away from zero. We plan to add these “informed ttests” to JASP in May 2017.
In order to conduct the analysis, selecting the “Ttest” tab reveals the option “Bayesian Independent Samples Ttest”, the dialog of which is displayed in the middle panel of Fig. 5. We have selected “mean NEO” as the dependent variable, and “Rotation” as the grouping variable. After ticking the box “Descriptives”, the output displayed in the right panel of Fig. 5 indicates that the mean opennesstoexperience is slightly larger in the counterclockwise group (i.e., N = 54;M = .71) than in the clockwise group (i.e., N = 48; M = .64) – note that the effect goes in the direction opposite to that hypothesized by Topolinski and Sparenberg (2012).
For demonstration purposes, at first we refrain from specifying the direction of the test. To contrast our results with those reported by Wagenmakers et al. (2015), we have set the Cauchy prior width to its JASP default r = 0.707 instead of Jeffreys’s value r = 1. We have also ticked the plotting options “Prior and posterior” and “Additional info”. This produces the plot shown in the right panel of Fig. 5. It is evident that most of the posterior mass is negative. The posterior median is − 0.13, and a 95% credible interval ranges from − 0.50 to 0.23. The Bayes factor is 3.71 in favor of \(\mathcal {H}_0\) over the twosided \(\mathcal {H}_1\). This indicates that the observed data are 3.71 times more likely under \(\mathcal {H}_0\) than under \(\mathcal {H}_1\). Because the Bayes factor favors \(\mathcal {H}_0\), in the input panel we have selected “ BF_{01}” under “Bayes Factor” – it is easier to interpret BF_{01} = 3.71 than it is to interpret the mathematically equivalent statement BF_{10} = 0.27.
After this initial investigation we now turn to an analysis of the preregistered orderrestricted test (with the exception of using r = 0.707 instead of the preregistered r = 1). The output of the “Descriptives” option has revealed that “clock” is group 1 (because it is on top), and “counter” is group 2. Hence, we can incorporate the order restriction in our inference by ticking the “Group one > Group two” box under “Hypothesis” in the input panel, as is shown in the middle panel of Fig. 6.
As an aside, note that under \(\mathcal {H}_+\) the posterior distribution is concentrated near zero but does not have mass on negative values, in accordance with the orderrestriction imposed by \(\mathcal {H}_+\). In contrast, the classical onesided confidence interval ranges from − .23 to ∞. This classical interval contrasts sharply with its Bayesian counterpart, and, even though the classical interval is mathematically welldefined (i.e., it contains all values that would not be rejected by a onesided α = .05 significance test, see also Wagenmakers et al., 2017), we submit that most researchers will find the classical result neither intuitive nor informative.
Finally, the middle panel of Fig. 7 also shows that the options “Sequential analysis” and “robustness check” are ticked, and these together produce the lower plot in the right panel of Fig. 7. The sequential analysis is of interest here because it was part of the experiment’s sampling plan, and because it underscores how researchers can monitor and visualize the evidential flow as the data accumulate. Closer examination of the plot reveals that for the preregistered value of r = 1, Wagenmakers et al. (2015) did not adhere to their preregistered sampling plan to stop data collection as soon as BF_{0+} > 10 or BF_{+0} > 10: after about 55 participants, the dotted line crosses the threshold of BF_{0+} > 10 but data collection nonetheless continued. Wagenmakers et al. (2015, p. 3) explain: “This occurred because data had to be entered into the analysis by hand and this made it more difficult to monitor the Bayes factor continually. In practice, the Bayes factor was checked every few days. Thus, we continued data collection until we reached our predetermined stopping criterion at the point of checking.”
One of the advantages of the sequential robustness plot is that it provides a visual impression of when the Bayes factors for the different priors have converged, in the sense that their difference on the log scale is constant (e.g., Gronau & Wagenmakers, in press). For the current situation, the convergence has occurred after testing approximately 35 participants. To understand why the difference between the log Bayes factors becomes constant after an initial number of observations, consider data y that consists of two batches, y _{1} and y _{2}. As mentioned above, from the law of conditional probability we have BF_{0+}(y) =BF_{0+}(y _{1}) ×BF_{0+}(y _{2}∣y _{1}). Note that this expression highlights that Bayes factors for different batches of data (e.g., participants, experiments) may not be multiplied blindly; the second factor, BF_{0+}(y _{2}∣y _{1}), equals the relative evidence from the second batch y _{2}, after the prior distributions have been properly updated using the information extracted from the first batch y _{1} (Jeffreys 1961, p. 333). Rewriting the above expression on the log scale we obtain logBF_{0+}(y) = logBF_{0+}(y _{1}) + logBF_{0+}(y _{2}∣y _{1}). Now assume y _{1} contains sufficient data such that, regardless of the value of prior width r under consideration, approximately the same posterior distribution is obtained. In most situations, this posterior convergence happens relatively quickly. This posterior distribution is then responsible for generating the Bayes factor for the second component, logBF_{0+}(y _{2}∣y _{1}), and it is therefore robust against differences in r.^{6} Thus, models with different values of r will make different predictions for data from the first batch y _{1}. However, after observing a batch y _{1} that is sufficiently large, the models have updated their prior distribution to a posterior distribution that is approximately similar; consequently, these models then start to make approximately similar predictions, resulting in a change in the log Bayes factor that is approximately similar as well.
Bayes factor  Evidence category 

> 100  Extreme evidence for \(\mathcal {H}_1\) 
30  100  Very strong evidence for \(\mathcal {H}_1\) 
10  30  Strong evidence for \(\mathcal {H}_1\) 
3  10  Moderate evidence for \(\mathcal {H}_1\) 
1  3  Anecdotal evidence for \(\mathcal {H}_1\) 
1  No evidence 
1/3  1  Anecdotal evidence for \(\mathcal {H}_0\) 
1/10  1/3  Moderate evidence for \(\mathcal {H}_0\) 
1/30  1/10  Strong evidence for \(\mathcal {H}_0\) 
1/100  1/30  Very strong evidence for \(\mathcal {H}_0\) 
< 1/100  Extreme evidence for \(\mathcal {H}_0\) 
Example 3: a Bayesian oneway ANOVA to test whether pain threshold depends on hair color
The first column of the output table, “Models”, lists the models under consideration. The oneway ANOVA features only two models: the “Null model” that contains the grand mean, and the “Hair Color” model that adds an effect of hair color. The next point of interest is the “ BF_{10}” column; this column shows the Bayes factor for each rowmodel against the null model. The first entry is always 1 because the null model is compared against itself. The second entry is 11.97, which means that the model with hair color predicts the observed data almost 12 times as well as the null model. As was the case for the output of the ttest, the rightmost column, “% error”, indicates the size of the error in the integration routine relative to the Bayes factor; similar to a coefficient of variation, this means that small variability is more important when the Bayes factor is ambiguous than when it is extreme.
Column “P(M)” indicates prior model probabilities (which the current version of JASP sets to be equal across all models at hand); column “P(M data)” indicates the updated probabilities after having observed the data. Column “ BF_{M}” indicates the degree to which the data have changed the prior model odds. Here the prior model odds equals 1 (i.e., 0.5/0.5) and the posterior model odds equals almost 12 (i.e., 0.923/0.077). Hence, the Bayes factor equals the posterior odds. JASP offers the user “Advanced Options” that can be used to change the prior width of the Cauchy prior for the model parameters. As the name suggest, we recommend that the user exercises this freedom only in the presence of substantial knowledge of the underlying statistical framework.
Currently JASP does not offer posthoc tests to examine pairwise differences in oneway ANOVA. Such posthoc tests have not yet been developed in the Bayesian ANOVA framework. In future work we will examine whether posthoc tests can be constructed by applying a Bayesian correction for multiple comparisons (i.e., Scott & Berger, 2006, 2010; Stephens & Balding, 2009). Discussion of this topic would take us too far afield.
Example 4: a Bayesian twoway ANOVA for singers’ height as a function of gender and pitch
Our analysis concerns the extent to which the dependent variable “height” is associated with gender (i.e., male, female) and/or pitch. This question can be examined statistically using a 2 × 4 ANOVA. Consistent with the visual impression from Fig. 10, a classical analysis yields significant results for both main factors (i.e., p < .001 for both gender and pitch) but fails to yield a significant result for the interaction (i.e., p = .52). In order to assess the extent to which the data support the presence and absence of these effects we now turn to a Bayesian analysis.
The first column of the output table, “Models”, lists the five models under consideration: the “Null model” that contains only the grand mean, the “Gender” model that contains the effect of gender, the “Pitch” model that contains the effect of Pitch, the “Gender + Pitch” model that contains both main effects, and finally the “Gender + Pitch + Gender × Pitch” model that includes both main effects and the interaction. Consistent with the principle of marginality, JASP does not include interactions in the absence of the component main effects; for instance, the interactiononly model “Gender × Pitch” may not be entertained without also adding the two main effects (for details, examples, and rationale see Bernhardt & Jung, 1979, Griepentrog, Ryan, & Smith 1982, McCullagh & Nelder, 1989; Nelder, 1998, 2000; Peixoto, 1987, 1990; Rouder, Engelhardt, et al., in press; Rouder, Morey, et al., in press; Venables, 2000).
Now consider the BF _{10} column. All models (except perhaps for Pitch) receive overwhelming evidence in comparison to the Null model. The model that outperforms the Null model the most is the two main effects model, Gender + Pitch. Adding the interaction makes the model less competitive. The evidence against including the interaction is roughly a factor of ten. This can be obtained as 8.192e+39 / 8.864e+38 ≈ 9.24. Thus, the data are 9.24 times more likely under the two main effects model than under the model that adds the interaction.
Column “P(M)” indicates the equal assignment of prior model probability across the five models; column “P(M data)” indicates the posterior model probabilities. Almost all posterior mass is centered on the two main effects model and the model that also includes the interaction. Column “BF _{M}” indicates the change from prior to posterior model odds. Only the two main effects model has received support from the data in the sense that the data have increased its model probability.
In sum, the Bayesian ANOVA reveals that the data provide strong support for the two main effects model over any of the simpler models. The data also provide good support against including the interaction term.
Finally, as described in Cramer et al. (2016), the multiway ANOVA harbors a multiple comparison problem. As for the oneway ANOVA, this problem can be addressed by applying the proper Bayesian correction method (i.e., Scott & Berger 2006, 2010; Stephens & Balding,2009). This correction has not yet been implemented in JASP.
Example 5: a Bayesian twoway repeated measures ANOVA for people’s hostility towards arthropods
To conduct the Bayesian analysis the user first needs to open the data set in JASP.^{11} Next the user selects the “Bayesian Repeated Measures ANOVA” input panel that is nested under the ribbon option “ANOVA”. Next the user needs to name the factors (here “Disgust” and “Fright”) and their levels (here “LD”, “HD”, and “LF”, “HF”). Finally the input variables need to be dragged to the matching “Repeated Measures Cells”.
The “P(M)” column shows the uniform distribution of prior model probabilities across the five candidate models, and the “P(M data)” column shows the posterior model probabilities. Finally, the “BF _{M}” column shows the change from prior model odds to posterior model odds. This Bayes factor also favors the two main effects model, but at the same time indicates mild support in favor of the interaction model. The reason for this discrepancy (i.e., a Bayes factor of 2.6 against the interaction model versus a Bayes factor of 1.5 in favor of the interaction model) is that these Bayes factors address different questions: The Bayes factor of 2.6 compares the interaction model against the two main effects model (which happens to be the model that is most supported by the data), whereas the Bayes factor of 1.5 compares the interaction model against all candidate models, some of which receive almost no support from the data. Both analyses are potentially of interest. Specifically, when the two main effects model decisively outperforms the simpler candidate models then it may be appropriate to assess the importance of the interaction term by comparing the two main effects model against the model that adds the interaction. However, it may happen that the simpler candidate models outperform the two main effects model – in other words, the two main effects model has predicted the data relatively poorly compared to the Null model or one of the single main effects models. In such situations it is misleading to test the importance of the interaction term by solely focusing on a comparion to the poorly performing two main effects model. In general we recommend radical transparency in statistical analysis; an informative report may present the entire table shown in Fig. 15. In this particular case, both Bayes factors (i.e., 2.6 against the interaction model, and 1.5 in favor of the interaction model) are “not worth more than a bare mention” (Jeffreys 1961, Appendix B); moreover, God loves these Bayes factors almost an equal amount, so it may well be argued that the discrepancy here is more apparent than real.
As the number of factors grows, so does the number of models. With many candidate models in play, it may be risky to base conclusions on a comparison involving a small subset. In Bayesian model averaging (BMA; e.g., Etz & Wagenmakers, in press; Haldane 1932; Hoeting, Madigan, Raftery, & Volinsky, 1999) the goal is to retain model selection uncertainty by averaging the conclusions from each candidate model, weighted by that model’s posterior plausibility. In JASP this is accomplished by ticking the “Effects” input box, which results in an output table shown in the bottom panel of Fig. 15.
In our example, the averaging in BMA occurs over the models shown in the Model Comparison table (top panel of Fig. 15). For instance, the factor “Disgust” features in three models (i.e., Disgust only, Disgust + Fright, and Disgust + Fright + Disgust * Fright). Each model has a prior model probability of 0.2, so the summed prior probability of the three models that include disgust equals 0.6; this is known as the prior inclusion probability for Disgust (i.e., the column P(incl)). After the data are observed we can similarly consider the sum of the posterior model probabilities for the models that include disgust, yielding 4.497e9 + 0.712 + 0.274 = 0.986. This is the posterior inclusion probability (i.e., column P(incl data)). The change from prior to posterior inclusion odds is given in the column “BF _{Inclusion}”. Averaged across all candidate models, the data strongly support inclusion of both main factors Disgust and Fright. The interaction only receives weak support. In fact, the interaction term occurs only in a single model, and therefore its posterior inclusion probability equals the posterior model probability of that model (i.e., the one that contains the two main effects and the interaction).
It should be acknowledged that the analysis of repeated measures ANOVA comes with a number of challenges and caveats. The development of Bayes factors for crossedrandom effect structures is still a topic of ongoing research. And in general, JASP currently does not feature an extensive suite of estimation routines to assess the extent to which generic model assumptions (e.g., sphericity) are violated.
Future directions for Bayesian analyses in JASP
The present examples provides a selective overview of default Bayesian inference in the case of the correlation test, ttest, oneway ANOVA, twoway ANOVA, and twoway repeated measures ANOVA. In JASP, other analyses can be executed in similar fashion (e.g., for contingency tables, Jamil, Ly, et al., in press, Jamil, Marsman, et al., in press; Scheibehenne, Jamil, & Wagenmakers, in press; or for linear regression Rouder & Morey, 2012). A detailed discussion of the entire functionality of JASP is beyond the scope of this article.
In the near future, we aim to expand the Bayesian repertoire of JASP, both in terms of depth and breadth. In terms of depth, our goal is to provide more and better graphing options, more assumption tests, more nonparametric tests, posthoc tests, and corrections for multiplicity. In terms of breadth, our goal is to include modules that offer the functionality of the BAS package (i.e., Bayesian model averaging in regression, Clyde, 2016), the informative model comparison approach (e.g., Gu, Mulder, Decović, & Hoijtink, 2014; Gu, 2016; Mulder, 2014, 2016), and a more flexible and subjective prior specification approach (e.g., Dienes, 2011, 2014, 2016; Gronau et al., 2017). By making the additional functionality available as addon modules, beginning users are shielded from the added complexity that such options add to the interface. In the shortterm we also aim to develop educational materials that make JASP output easier to interpret and to teach to undergraduate students. This entails writing a JASP manual, developing course materials, writing course books, and designing a Massive Open Online Course.
Our longterm goal is for JASP to facilitate several aspects of statistical practice. Free and userfriendly, JASP has the potential to benefit both education and research. By featuring both classical and Bayesian analyses, JASP implicitly advocates a more inclusive statistical approach. JASP also aims to assist with data preparation and aggregation; currently, this requires that JASP launches and interacts with an external editor (see our dataediting video at https://www.youtube.com/watch?v=1dTiAU9Zuc&t=70s); in the future, JASP will have its own editing functionality including filtering and outlier exclusion. Finally, by offering the ability to save, annotate, and share statistical output, JASP promotes a transparent way of communicating one’s statistical results. An increase in statistical transparency and inclusiveness will result in science that is more reliable and more replicable.
As far as the continued development of JASP is concerned, our two main software developers and several core team members of the JASP team have tenured positions. The Psychological Methods Group at the University of Amsterdam is dedicated to longterm support for JASP, and in 2017 we have received four million euro to set up projects that include the development of JASP as a key component. The JASP code is opensource and will always remain freely available online. In sum, JASP is here to stay.
Concluding comments
In order to promote the adoption of Bayesian procedures in psychology, we have developed JASP, a free and opensource statistical software program with an interface familiar to users of SPSS. Using JASP, researchers can obtain results from Bayesian techniques easily and without tears. Dennis Lindley once said that “Inside every NonBayesian, there is a Bayesian struggling to get out” (Jaynes 2003). We hope that software programs such as JASP will act to strengthen the resolve of one’s inner Bayesian and pave the road for a psychological science in which innovative hypotheses are tested using coherent statistics.
Footnotes
 1.
Bayesian advocates may consider the classical analyses a Bayesian Trojan horse.
 2.
JASP currently reads the following file formats: .jasp, .txt, .csv (i.e., a plain text file with fields separated by commas), .ods (i.e., OpenDocument Spreadsheet, a file format used by OpenOffice), and .sav (i.e., the SPSS file format).
 3.
With equal prior odds, a ratio of x corresponds to a proportion of x/(x + 1).
 4.
Statistical detail: the stretched beta prior is a beta(a,a) distribution transformed to cover the interval from − 1 to 1. The prior width is defined as 1/a. For instance, setting the stretched beta prior width equal to 0.5 is conceptually the same as using a beta(2, 2) distribution on the 01 interval and then transforming it to cover the interval from − 1 to 1, such that it is then symmetric around ρ = 0.
 5.
For an indication of how Bayes factors can be computed under any proper prior distribution see http://jeffrouder.blogspot.nl/2016/01/whatpriorsshouldiuseparti.html, also available as a pdf file at the OSF project page https://osf.io/m6bi8/.
 6.
This also suggests that one can develop a Bayes factor that is robust against plausible changes in r: first, sacrifice data y _{1} until the posterior distributions are similar; second, monitor and report the Bayes factor for the remaining data y _{2}. This is reminiscent of the idea that underlies the socalled intrinsic Bayes factor (Berger and Pericchi 1996), a method that also employs a “training sample” to update the prior distributions before the test is conducted using the remaining data points. The difference is that the intrinsic Bayes factor selects a training sample of minimum size, being just large enough to identify the model parameters.
 7.
The present authors are not all agreed on the usefulness of such descriptive classifications of Bayes factors. All authors agree, however, that the advantage of Bayes factors is that –unlike for instance p values which are dichotomized into “significant” and “nonsignificant”– the numerical value of the Bayes factor can be interpreted directly. The strength of the evidence is not dependent on any conventional verbal description, such as “strong”.
 8.
The data are available at http://www.statsci.org/data/oz/blonds.html.
 9.
The Cauchy prior width r _{ t } for the independent samples ttests yields the same result as a twogroup oneway ANOVA with a fixed effect scale factor r _{ A } equal to \(r_t/\sqrt {2}\). With the default setting \(r_t=1/2 \cdot \sqrt {2}\), this produces r _{ A } = 0.5. In sum, for the default prior settings in JASP the independent samples ttest and the twogroup oneway ANOVA yield the same result. For examples see https://cran.rproject.org/web/packages/BayesFactor/vignettes/priors.html.
 10.
 11.
The data set is available on the project OSF page and from within JASP (i.e., File → Open → Examples → Bugs).
Notes
Acknowledgements
The development of JASP was supported by the European Research Council grant “Bayes or bust: Sensible hypothesis tests for social scientists”. Supplementary materials are available at https://osf.io/m6bi8/. The JASP team can be reached through GitHub, twitter, Facebook, and the JASP Forum. EricJan Wagenmakers, University of Amsterdam, Department of Psychology, PO Box 15906, 1001 NK Amsterdam, the Netherlands. Email address: EJ.Wagenmakers@gmail.com.
References
 Berger, J. O., & Delampady, M. (1987). Testing precise hypotheses. Statistical Science, 2, 317–352.CrossRefGoogle Scholar
 Berger, J. O., & Pericchi, L. R. (1996). The intrinsic Bayes factor for model selection and prediction. Journal of the American Statistical Association, 91, 109–122.CrossRefGoogle Scholar
 Bernhardt, I., & Jung, B. S. (1979). The interpretation of least squares regression with interaction or polynomial terms. The Review of Economics and Statistics, 61, 481–483.CrossRefGoogle Scholar
 Chambers, J. M., Cleveland, W. S., Kleiner, & Tukey, P. A. (1983). Graphical methods for data analysis. New York: Chapman and Hall.Google Scholar
 Clyde, M. (2016). BAS: Bayesian adaptive sampling for Bayesian model averaging. (R package version 1.4.1).Google Scholar
 Costa, P. T., & McCrae, R. R. (1992). NEO Personality Inventory professional manual. Odessa, FL: Psychological Assessment Resources.Google Scholar
 Cramer, A. O. J., van Ravenzwaaij, D., Matzke, D., Steingroever, H., Wetzels, R., Grasman, R. P. P. P., Waldorp, L. J., & Wagenmakers, E.J. (2016). Hidden multiplicity in multiway ANOVA: Prevalence, consequences, and remedies. Psychonomic Bulletin & Review, 23, 640–647.Google Scholar
 Dickey, J. M., & Lientz, B. P. (1970). The weighted likelihood ratio, sharp hypotheses about chances, the order of a Markov chain. The Annals of Mathematical Statistics, 41, 214–226.CrossRefGoogle Scholar
 Dienes, Z. (2011). Bayesian versus orthodox statistics: Which side are you on? Perspectives on Psychological Science, 6, 274–290.CrossRefPubMedGoogle Scholar
 Dienes, Z. (2014). Using Bayes to get the most out of nonsignificant results. Frontiers in Psycholology, 5, 781.Google Scholar
 Dienes, Z. (2016). How Bayes factors change scientific practice. Journal of Mathematical Psycholology, 72, 78–89.CrossRefGoogle Scholar
 Edwards, W., Lindman, H., & Savage, L. J. (1963). Bayesian statistical inference for psychological research. Psychological Review, 70, 193–242.CrossRefGoogle Scholar
 Etz, A., & Wagenmakers, E. J. (in press). J. B. S. Haldane’s contribution to the Bayes factor hypothesis test. Statistical Science.Google Scholar
 Francis, G. (2013). Replication, statistical consistency, and publication bias. Journal of Mathematical Psychology, 57, 153–169.CrossRefGoogle Scholar
 Griepentrog, G. L., & Ryan, J. M. (1982). Linear transformations of polynomial regression models. The American Statistician, 36, 171–174.Google Scholar
 Gronau, Q. F., Ly, A., & Wagenmakers, E.J. (2017). Informed Bayesian ttests. Manuscript submitted for publication.Google Scholar
 Gronau, Q. F., & Wagenmakers, E.J. (in press). Bayesian evidence accumulation in experimental mathematics: A case study of four irrational numbers. Experimental Mathematics.Google Scholar
 Gu, X. (2016). Bayesian evaluation of informative hypotheses. Utrecht University.Google Scholar
 Gu, X., Mulder, J., Decović, M., & Hoijtink, H. (2014). Bayesian evaluation of inequality constrained hypotheses. Psychological Methods, 19, 511–527.CrossRefPubMedGoogle Scholar
 Gunel, E., & Dickey, J. (1974). Bayes factors for independence in contingency tables. Biometrika, 61, 545–557.CrossRefGoogle Scholar
 Haldane, J. B. S. (1932). A note on inverse probability. Mathematical Proceedings of the Cambridge Philosophical Society, 28, 55–61.CrossRefGoogle Scholar
 Hoekstra, H. A., Ormel, & de Fruyt, F. (1996). Handleiding bij de NEO persoonlijkheids vragenlijsten NEOPIR NEOFFI [manual for the NEO personality inventories NEOPIr and NEOFFI]. Lisse, the Netherlands: Swets & Zeitlinger.Google Scholar
 Hoeting, J. A., Madigan, D., Raftery, & Volinsky, C. T. (1999). Bayesian model averaging: A tutorial. Statistical Science, 14, 382–417.CrossRefGoogle Scholar
 Jamil, T., Ly, A., Morey, R. D., Love, J., Marsman, M., & Wagenmakers, E.J. (in press). Default “Gunel and Dickey” Bayes factors for contingency tables. Behavior Research Methods.Google Scholar
 Jamil, T., Marsman, M., Ly, A., Morey, R. D., & Wagenmakers, E.J. (in press). What are the odds? Modern relevance and Bayes factor solutions for MacAlister’s problem from the 1881 Educational Times. Educational and Psychological Measurement.Google Scholar
 JASP Team (2017). JASP (Version 0.8.1)[Computer software].Google Scholar
 Jaynes, E. T. (2003). Probability theory: The logic of science. Cambridge: Cambridge University Press.CrossRefGoogle Scholar
 Jeffreys, H. (1961). Theory of probability, (3rd ed.) Oxford, UK: Oxford University Press.Google Scholar
 Johnson, V. E. (2013). Revised standards for statistical evidence. Proceedings of the National Academy of Sciences of the United States of America, 110, 19313–19317.CrossRefPubMedPubMedCentralGoogle Scholar
 Lee, M. D., & Wagenmakers, E. J (2013). Bayesian cognitive modeling: A practical course. Cambridge University Press.Google Scholar
 Liang, F., Paulo, R., Molina, G., & Clyde, M. A. (2008). Mixtures of g priors for Bayesian variable selection. Journal of the American Statistical Association, 103, 410–423.CrossRefGoogle Scholar
 Lipkus, I. M., & Hollands, J. G. (1999). The visual communication of risk. Journal of the National Cancer Institute Monographs, 25, 149–163.CrossRefGoogle Scholar
 Ly, A., Marsman, M., & Wagenmakers, E.J. (in press). Analytic posteriors for Pearson’s correlation coefficient. Statistica Neerlandica.Google Scholar
 Ly, A., Verhagen, A. J., & Wagenmakers, E.J. (2016a). An evaluation of alternative methods for testing hypotheses, from the perspective of Harold Jeffreys. Journal of Mathematical Psychology, 72, 43–55.Google Scholar
 Ly, A., Verhagen, A. J., & Wagenmakers, E.J. (2016b). Harold Jeffreys’s default Bayes factor hypothesis tests: Explanation, extension, and application in psychology. Journal of Mathematical Psychology, 72, 19–32.Google Scholar
 Matzke, D., Nieuwenhuis, S., van Rijn, H., Slagter, H. A., van der Molen, M. W., & Wagenmakers, E.J. (2015). The effect of horizontal eye movements on free recall: A preregistered adversarial collaboration. Journal of Experimental Psychology: General, 144, e1–e15.Google Scholar
 McClave, J. T., & Dietrich, F. H. II. (1991). Statistics. San Francisco: Dellen Publishing.Google Scholar
 McCullagh, P., & Nelder, J. A. (1989). Generalized linear models, 2nd ed. London: Chapman & Hall.CrossRefGoogle Scholar
 Morey, R. D., & Rouder, J. N. (2015). BayesFactor 0.9.111. Comprehensive R Archive Network.Google Scholar
 Mulder, J. (2014). Prior adjusted default Bayes factors for testing (in)equality constrained hypotheses. Computational Statistics and Data Analysis, 71, 448–463.CrossRefGoogle Scholar
 Mulder, J. (2016). Bayes factors for testing order–constrained hypotheses on correlations. Journal of Mathematical Psychology, 72, 104–115.CrossRefGoogle Scholar
 Nelder, J. A. (1998). The selection of terms in responsesurface models—how strong is the weakheredity principle? The American Statistician, 52, 315–318.Google Scholar
 Nelder, J. A. (2000). Functional marginality and responsesurface fitting. Journal of Applied Statistics, 27, 109–112.CrossRefGoogle Scholar
 O’Hagan, A., & Forster, J. (2004). Kendall’s advanced theory of statistics vol 2B: Bayesian inference, 2nd ed. London: Arnold.Google Scholar
 Overstall, A. M., & King, R. (2014a). A default prior distribution for contingency tables with dependent factor levels. Statistical Methodology, 16, 90–99.Google Scholar
 Overstall, A. M., & King, R. (2014b). conting: An R package for Bayesian analysis of complete and incomplete contingency tables. Journal of Statistical Software, 58, 1–27.Google Scholar
 Peixoto, J. L. (1987). Hierarchical variable selection in polynomial regression models. The American Statistician, 41, 311–313.Google Scholar
 Peixoto, J. L. (1990). A property of wellformulated polynomial regression models. The American Statistician, 44, 26–30.Google Scholar
 R Development Core Team (2004). R: A language and environment for statistical computing. Vienna, Austria. (ISBN 3–900051–00–3).Google Scholar
 Rosnow, R. L., & Rosenthal, R. (1989). Statistical procedures and the justification of knowledge in psychological science. American Psychologist, 44, 1276–1284.CrossRefGoogle Scholar
 Rouder, J. N., Engelhardt, C. R., McCabe, S., & Morey, R. D. (in press). Model comparison in ANOVA. Psychonomic Bulletin & Review.Google Scholar
 Rouder, J. N., & Morey, R. D. (2012). Default Bayes factors for model selection in regression. Multivariate Behavioral Research, 47, 877–903.CrossRefPubMedGoogle Scholar
 Rouder, J. N., Morey, R. D., Speckman, & Province, J. M. (2012). Default Bayes factors for ANOVA designs. Journal of Mathematical Psychology, 56, 356–374.CrossRefGoogle Scholar
 Rouder, J. N., Morey, R. D., Verhagen, A. J., Swagman, A. R., & Wagenmakers, E. J. (in press). Bayesian analysis of factorial designs. Psychological Methods.Google Scholar
 Rouder, J. N., Speckman, P. L., Sun, D., Morey, R. D., & Iverson, G. (2009). Bayesian t tests for accepting and rejecting the null hypothesis. Psychonomic Bulletin & Review, 16, 225–237.CrossRefGoogle Scholar
 Ryan, R. S., & Wilde, M. (2013). Compared to a small, supervised lab experiment, a large, unsupervised web–based experiment on a previously unknown effect has benefits that outweigh its potential costs. Computers in Human Behavior, 29, 1295– 1301.CrossRefGoogle Scholar
 Scheibehenne, B., Jamil, T., & Wagenmakers, E.J. (in press). Bayesian evidence synthesis can reconcile seemingly inconsistent results: The case of hotel towel reuse. Psychological Science.Google Scholar
 Schönbrodt, F. D., Wagenmakers, E.J., Zehetleitner, M., & Perugini, M. (in press). Sequential hypothesis testing with Bayes factors: Efficiently testing mean differences. Psychological Methods.Google Scholar
 Scott, J. G., & Berger, J. O. (2006). An exploration of aspects of Bayesian multiple testing. Journal of Statistical Planning and Inference, 136, 2144–2162.CrossRefGoogle Scholar
 Scott, J. G., & Berger, J. O. (2010). Bayes and empirical–Bayes multiplicity adjustment in the variable–selection problem. The Annals of Statistics, 38, 2587–2619.CrossRefGoogle Scholar
 Stephens, M., & Balding, D. J. (2009). Bayesian statistical methods for genetic association studies. Nature Reviews Genetics, 10, 681–690.CrossRefPubMedGoogle Scholar
 Stevens, S. S. (1946). On the theory of scales of measurement. Science, 103, 677–680.CrossRefGoogle Scholar
 Stulp, G., Buunk, A. P., Verhulst, & Pollet, T. V. (2013). Tall claims? Sense and nonsense about the importance of height of US presidents. The Leadership Quarterly, 24, 159–171.CrossRefGoogle Scholar
 Topolinski, S., & Sparenberg, P. (2012). Turning the hands of time: Clockwise movements increase preference for novelty. Social Psychological and Personality Science, 3, 308–314.CrossRefGoogle Scholar
 Tversky, A. (1969). Intransitivity of preferences. Psychological Review, 76, 31–48.CrossRefGoogle Scholar
 van Doorn, J., Ly, A., Marsman, M., & Wagenmakers, E.J. (in press). Bayesian inference for Kendall’s rank correlation coefficient. The American Statistician.Google Scholar
 Venables, W. N. (2000). Exegeses on linear models. Paper presented to the SPLUS User’s Conference.Google Scholar
 Wagenmakers, E.J., Lodewyckx, T., Kuriyal, H., & Grasman, R. (2010). Bayesian hypothesis testing for psychologists: A tutorial on the Savage–Dickey method. Cognitive Psychology, 60, 158–189.CrossRefPubMedGoogle Scholar
 Wagenmakers, E. J., Beek, T., Rotteveel, M., Gierholz, A., Matzke, D., Steingroever, H., Ly, A., Verhagen, A. J., Selker, R., Sasiadek, A., & Pinto, Y. (2015). Turning the hands of time again: A purely confirmatory replication study and a Bayesian analysis. Frontiers in Psychology: Cognition, 6, 494.CrossRefGoogle Scholar
 Wagenmakers, E.J., Marsman, M., Jamil, T., Ly, A., Verhagen, A. J., Love, J., Selker, R., Gronau, Q. F., Šmíra, M., Epskamp, S., Matzke, D., Rouder, J. N., & Morey, R. D. (2017). Bayesian statistical inference for psychological science. Part I: Theoretical advantages and practical ramifications. Psychonomic Bulletin & Review, (in this issue).Google Scholar
 Wetzels, R., Raaijmakers, J. G. W., Jakab, E., & Wagenmakers, E.J. (2009). How to quantify support for and against the null hypothesis: A flexible WinBUGS, implementation of a default Bayesian t test. Psychonomic Bulletin & Review, 16, 752–760.CrossRefGoogle Scholar
 Wetzels, R., Matzke, D., Lee, M. D., Rouder, J. N., Iverson, G. J., & Wagenmakers, E. J. (2011). Statistical evidence in experimental psychology: An empirical comparison using 855 t tests. Perspectives on Psychological Science, 6, 291–298.CrossRefPubMedGoogle Scholar
 Zellner, A., & Siow, A. (1980). Posterior odds ratios for selected regression hypotheses. In Bernardo, J. M., DeGroot, M. H., Lindley, D. V., & Smith, A. F. M. (Eds.), Bayesian statistics (pp. 585–603). Valencia: University Press.Google Scholar
Copyright information
Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.