Bayes Factor Calculation In Jasp

Use this premium calculator to estimate Bayes factors either from prior and posterior model probabilities or from a BIC approximation commonly used in model comparison workflows. The tool also visualizes prior odds, posterior odds, and Bayes factor strength to help you interpret evidence the way JASP users typically report it.

Interactive Bayes Factor Calculator

How to do Bayes factor calculation in JASP

Bayes factors are one of the most practical ways to summarize evidence in Bayesian statistics, and JASP has helped make them accessible for everyday research use. If you have ever run a Bayesian t test, ANOVA, regression, contingency table analysis, or model comparison in JASP, you have probably seen output such as BF10, BF01, posterior model probabilities, and inclusion Bayes factors. Understanding how these quantities relate to each other is essential if you want to report Bayesian results confidently and interpret them correctly.

At a high level, a Bayes factor compares how well two competing hypotheses or models predict the observed data. In its common form, BF10 means the evidence for the alternative hypothesis H1 relative to the null hypothesis H0. If BF10 equals 6, the observed data are six times more likely under H1 than under H0, given the specified prior assumptions. If BF01 equals 6 instead, that means the data are six times more likely under H0 than under H1.

The core relationship is simple: posterior odds = prior odds × Bayes factor. JASP often displays Bayes factors directly, but in some situations you may also derive them from prior and posterior model probabilities or approximate them from BIC values.

What JASP usually reports

JASP commonly reports one or more of the following:

• BF10: evidence for the alternative over the null.
• BF01: evidence for the null over the alternative.
• Posterior model probabilities: the probability assigned to each model after seeing the data.
• Prior model probabilities: the probability assigned before observing the data.
• Inclusion Bayes factors: often used in Bayesian ANOVA or regression to assess whether a predictor or effect should be included.

If you have prior and posterior model probabilities, you can compute a Bayes factor exactly through the odds formulation. For two competing hypotheses H1 and H0:

1. Compute prior odds: P(H1) / P(H0)
2. Compute posterior odds: P(H1|D) / P(H0|D)
3. Compute BF10 = posterior odds / prior odds

When prior model probabilities are equal, such as 0.50 for H1 and 0.50 for H0, the prior odds equal 1. In that special case, BF10 is numerically the same as the posterior odds. This is why equal priors make interpretation especially straightforward in many JASP model comparison examples.

Bayes factor formula and intuition

The formal Bayes factor for H1 versus H0 is:

BF10 = P(D|H1) / P(D|H0)

This expression compares the marginal likelihood of the data under each model. The phrase marginal likelihood matters because in Bayesian testing, especially for composite hypotheses, the data likelihood is averaged over the prior distribution of parameter values. In other words, the Bayes factor rewards models that predict the data well while also penalizing models that spread probability too widely over implausible parameter regions. That balance is one reason Bayes factors are often viewed as evidence measures rather than just significance thresholds.

In practical JASP use, you do not usually hand calculate marginal likelihoods from scratch. JASP does that internally for the selected test and prior settings. But you still need to know how to read the output. A Bayes factor is not a probability of truth by itself. BF10 = 8 does not mean there is an 80% chance that H1 is true. It means the data are eight times more compatible with H1 than H0, given the prior setup.

Converting between BF10 and BF01

The two common Bayes factor directions are reciprocals:

• BF01 = 1 / BF10
• BF10 = 1 / BF01

This matters because different software outputs or articles may emphasize different directions. JASP often reports BF10 by default, but many papers also mention BF01 when discussing evidence in favor of the null.

Evidence interpretation benchmarks

Although Bayes factors are continuous evidence measures and should not be treated as rigid cutoffs, many researchers use broad interpretive categories. These are often inspired by Jeffreys style conventions.

BF10	Interpretation	Practical reading
1.00	No preference	The data are equally compatible with H1 and H0.
1.00 to 3.00	Anecdotal evidence for H1	Weak support for H1, often not persuasive on its own.
3.00 to 10.00	Moderate evidence for H1	Meaningful support for H1, often worth reporting prominently.
10.00 to 30.00	Strong evidence for H1	The data substantially favor H1.
30.00 to 100.00	Very strong evidence for H1	H1 predicts the observed data far better than H0.
Greater than 100.00	Extreme evidence for H1	The evidence strongly favors H1 under the chosen prior assumptions.

The same logic applies in reverse for BF01. For example, BF01 = 12 means strong evidence for H0 relative to H1. Always state which direction you are using so readers do not misinterpret the conclusion.

Using posterior and prior probabilities to calculate Bayes factors

Suppose JASP reports two models with equal prior probabilities of 0.50 each. After observing the data, JASP reports a posterior model probability of 0.80 for H1 and 0.20 for H0. Then:

1. Prior odds = 0.50 / 0.50 = 1.00
2. Posterior odds = 0.80 / 0.20 = 4.00
3. BF10 = 4.00 / 1.00 = 4.00

That means the data are four times more likely under H1 than under H0. In evidence language, this is typically described as moderate evidence for H1.

Now consider unequal prior model probabilities, which can arise in model selection settings. Assume:

• Prior P(H1) = 0.25 and P(H0) = 0.75
• Posterior P(H1|D) = 0.60 and P(H0|D) = 0.40

Then:

1. Prior odds = 0.25 / 0.75 = 0.333
2. Posterior odds = 0.60 / 0.40 = 1.50
3. BF10 = 1.50 / 0.333 = 4.50

Even though the prior started out favoring H0, the data shifted the odds enough to produce moderate evidence for H1. This is a useful reminder that posterior probabilities depend on both prior beliefs and the data, while the Bayes factor isolates the evidential contribution of the data given the prior parameter structure of each model.

BIC approximation for Bayes factor calculation

Sometimes you do not have a direct Bayes factor from software output, but you do have BIC values for competing models. A commonly used approximation is:

BF10 ≈ exp((BIC0 – BIC1) / 2)

Here, BIC0 is the Bayesian Information Criterion for the null or reference model, and BIC1 is the BIC for the alternative model. Lower BIC indicates a better tradeoff between fit and complexity. If the alternative has a smaller BIC, then BIC0 – BIC1 is positive and BF10 will be greater than 1.

BIC for H0	BIC for H1	Difference (BIC0 – BIC1)	Approximate BF10	Interpretation
130.0	129.0	1.0	1.65	Anecdotal evidence for H1
130.0	126.0	4.0	7.39	Moderate evidence for H1
130.0	122.0	8.0	54.60	Very strong evidence for H1
130.0	118.0	12.0	403.43	Extreme evidence for H1

These values are real numerical approximations from the exponential formula above. While useful, the BIC approximation is not identical to a full Bayes factor computed under JASP’s exact Bayesian model and prior specification. It is best thought of as a convenient approximation for model comparison rather than a complete replacement for dedicated Bayesian estimation.

How JASP computes Bayes factors in common tests

Bayesian t tests

In Bayesian t tests, JASP often uses a default Cauchy prior on effect size, commonly with scale r = 0.707 for the alternative hypothesis. The resulting BF10 depends on the observed t statistic, sample size, and prior width. A wider prior spreads belief over larger effect sizes and may change the Bayes factor meaningfully. That is why sensitivity analyses are important when reporting Bayesian t test results.

Bayesian ANOVA and regression

For ANOVA and regression, JASP often reports model comparison tables and inclusion Bayes factors. Inclusion Bayes factors summarize the evidence for including a predictor or effect across the model space, rather than comparing just one pair of models. Researchers often misread these as ordinary BF10 values, so it is essential to label them accurately.

Bayesian contingency tables and correlation analyses

For contingency table tests and Bayesian correlation, JASP again relies on likelihood based evidence integrated over prior distributions. The interpretation remains the same: values larger than 1 support the numerator model, values smaller than 1 support the denominator model, and reciprocal conversion allows you to present the result in your preferred direction.

Step by step workflow in JASP

1. Open your data in JASP.
2. Select the relevant Bayesian procedure such as Bayesian Independent Samples T Test, Bayesian ANOVA, or Bayesian Regression.
3. Choose the prior settings or use the default prior if justified for your field.
4. Run the analysis and inspect BF10, BF01, posterior probabilities, and model comparison output.
5. If necessary, use the calculator above to derive Bayes factors from posterior and prior model probabilities or from BIC summaries.
6. Report the direction clearly, for example: “The data provided moderate evidence for H1, BF10 = 4.23.”
7. When possible, add a sensitivity analysis showing whether the Bayes factor is robust to reasonable prior changes.

Common reporting mistakes to avoid

• Confusing evidence with posterior probability. A Bayes factor is not itself the probability that a hypothesis is true.
• Failing to state the direction. Always clarify whether you are reporting BF10 or BF01.
• Ignoring prior assumptions. Bayes factors depend on the specified prior distributions, especially for H1.
• Using thresholds mechanically. Interpretation categories are rough guides, not hard rules.
• Treating BIC approximations as exact. They are useful shortcuts, but they are not always equivalent to a full Bayesian analysis in JASP.

Example APA style reporting

You might write:

A Bayesian independent samples t test indicated moderate evidence for a group difference, BF10 = 6.84, suggesting that the observed data were about 6.8 times more likely under the alternative hypothesis than under the null hypothesis.

Or for evidence favoring the null:

The Bayesian analysis supported the null model, BF01 = 8.21, indicating that the data were approximately eight times more likely under the null than under the alternative model.

Why this calculator is useful

This calculator helps in two common scenarios. First, if JASP reports prior and posterior model probabilities, you can compute Bayes factors from the odds ratio relationship directly. Second, if you are comparing model summaries and only have BIC values, you can produce an evidence approximation that often aligns with broad Bayesian model selection logic. The accompanying chart makes it easier to communicate how the evidence shifted from prior odds to posterior odds.

Authoritative references for deeper study

If you want more technical background on Bayesian reasoning, model comparison, and statistical evidence, these sources are useful starting points:

• Penn State STAT 414: Bayes’ Rule
• NIST Engineering Statistics Handbook
• NIH NCBI resource on Bayesian methods and evidence

Final takeaway

Bayes factor calculation in JASP becomes much easier once you recognize that every output is essentially about comparing predictive performance between models. Whether you read BF10 directly, derive it from prior and posterior model probabilities, or approximate it from BIC values, the logic is the same: quantify how strongly the data shift the balance of evidence. If you also report priors, the model direction, and a plain language interpretation, your Bayesian results will be much clearer, more transparent, and more persuasive.