# The 20% Statistician

A blog on statistics, methods, philosophy of science, and open science. Understanding 20% of statistics will improve 80% of your inferences.

## Thursday, April 13, 2023

### Preventing common misconceptions about Bayes Factors

As more people have started to use Bayes Factors, we should not be surprised that misconceptions about Bayes Factors have become common. A recent study shows that the percentage of scientific articles that draw incorrect inferences based on observed Bayes Factors is distressingly high (Wong et al., 2022), with 92% of articles demonstrating at least one misconception of Bayes Factors. Here I will review some of the most common misconceptions, and how to prevent them.

Misunderstanding 1: Confusing Bayes Factors with Posterior Odds.

One common criticism by Bayesians of null hypothesis significance testing (NHST) is that NHST quantifies the probability of the data (or more extreme data), given that the null hypothesis is true, but that scientists should be interested in the probability that the hypothesis is true, given the data. Cohen (1994) wrote:

What’s wrong with NHST? Well, among many other things, it does not tell us what we want to know, and we so much want to know what we want to know that, out of desperation, we nevertheless believe that it does! What we want to know is “Given these data, what is the probability that Ho is true?”

One might therefore believe that Bayes factors tell us something about the probability that a hypothesis true, but this is incorrect. A Bayes factor quantifies how much we should update our belief in one hypothesis. If this hypothesis was extremely unlikely (e.g., the probability that people have telepathy) this hypothesis might still be very unlikely, even after computing a large Bayes factor in a single study demonstrating telepathy. If we believed the hypothesis that people have telepathy was unlikely to be true (e.g., we thought it was 99.9% certain telepathy was not true) evidence for telepathy might only increase our belief in telepathy to the extent that we now believe it is 98% unlikely. The Bayes factor only corresponds to our posterior belief if we were perfectly uncertain about the hypothesis being true or not. If both hypotheses were equally likely, and a Bayes factor indicates we should update our belief in such a way that the alternative hypothesis is three times more likely than the null hypothesis, only then would we end up believing the alternative hypothesis is exactly three times more likely than the null hypothesis. One should therefore not conclude that, for example, given a BF of 10, the alternative hypothesis is more likely to be true than the null hypothesis. The correct claim is that people should update their belief in the alternative hypothesis by a factor of 10.

Misunderstanding 2: Failing to interpret Bayes Factors as relative evidence.

One benefit of Bayes factors that is often mentioned by Bayesians is that, unlike NHST, Bayes factors can provide support for the null hypothesis, and thereby falsify predictions. It is true that NHST can only reject the null hypothesis, although it is important to add that in frequentist statistics equivalence tests can be used to reject the alternative hypothesis, and therefore there is no need to switch to Bayes factors to meaningfully interpret the results of non-significant null hypothesis tests.

Bayes factors quantify support for one hypothesis relative to another hypothesis. As with likelihood ratios, it is possible that one hypothesis is supported more than another hypothesis, while both hypotheses are actually false. It is incorrect to interpret Bayes factors in an absolute manner, for example by stating that a Bayes factor of 0.09 provides support for the null hypothesis. The correct interpretation is that the Bayes factor provides relative support for H0 compared to H1. With a different alternative model, the Bayes factor would change. As with a signiifcant equivalence tests, even a Bayes factor strongly supporting H0 does not mean there is no effect at all - there could be a true, but small, effect.

For example, after Daryl Bem (2011) published 9 studies demonstrating support for pre-cognition (conscious cognitive awareness of a future event that could not otherwise be known) a team of Bayesian statisticians re-analyzed the studies, and concluded “Out of the 10 critical tests, only one yields “substantial” evidence for H1, whereas three yield “substantial” evidence in favor of H0. The results of the remaining six tests provide evidence that is only “anecdotal”” (2011). In a reply, Bem and Utts (2011) reply by arguing that the set of studies provide convincing evidence for the alternative hypothesis, if the Bayes factors are computed as relative evidence between the null hypothesis and a more realistically specified alternative hypothesis, where the effects of pre-cognition are expected to be small. This back and forth illustrates how Bayes factors are relative evidence, and a change in the alternative model specification changes whether the null or the alternative hypothesis receives relatively more support given the data.

Misunderstanding 3: Not specifying the null and/or alternative model.

Given that Bayes factors are relative evidence for or against one model compared to another model, it might be surprising that many researchers fail to specify the alternative model to begin with when reporting their analysis. And yet, in a systematic review of how psychologist use Bayes factors, van de Schoot et al. (2017) found that “31.1% of the articles did not even discuss the priors implemented”. Where in a null hypothesis significance test researchers do not need to specify the model that the test is based on, as the test is by definition a test against an effect of 0, and the alternative model consists of any non-zero effect size (in a two-sided test), this is not true when computing Bayes factors. The null model when computing Bayes factors is often (but not necessarily) a point null as in NHST, but the alternative model only one of many possible alternative hypotheses that a researcher could test against. It has become common to use ‘default’ priors, but as with any heuristic, defaults will most often give an answer to a nonsensical question, and quickly become a form of mindless statistics. When introducing Bayes factors as an alternative to frequentist t-tests, Rouder et al. (2009) write:

This commitment to specify judicious and reasoned alternatives places a burden on the analyst. We have provided default settings appropriate to generic situations. Nonetheless, these recommendations are just that and should not be used blindly. Moreover, analysts can and should consider their goals and expectations when specifying priors. Simply put, principled inference is a thoughtful process that cannot be performed by rigid adherence to defaults.

The priors used when computing a Bayes factor should therefore be both specified and justified.

Misunderstanding 4: Claims based on Bayes Factors do not require error control.

In a paper with the provocative title “Optional stopping: No problem for Bayesians” Rouder (2014) argues that “Researchers using Bayesian methods may employ optional stopping in their own research and may provide Bayesian analysis of secondary data regardless of the employed stopping rule.” If one would merely read the title and abstract, a reader might come to the conclusion that Bayes factors a wonderful solution to the error inflation due to optional stopping in the frequentist framework, but this is not correct (de Heide & Grünwald, 2017).

There is a big caveat about the type of statistical inferences that is unaffected by optional stopping. Optional stopping is no problem for Bayesians if they refrain from making a dichotomous claim about the presence or absence of an effect, or when they refrain from drawing conclusions about a prediction being supported or falsified. Rouder notes how “Even with optional stopping, a researcher can interpret the posterior odds as updated beliefs about hypotheses in light of data.” In other words, even after optional stopping, a Bayes factor tells researchers who much they should update their belief in a hypothesis. Importantly, when researchers make dichotomous claims based on Bayes factors (e.g., “The effect did not differ significantly between the condition, BF10 = 0.17”) then this claim can be correct, or an error, and error rates become a relevant consideration, unlike when researchers simply present the Bayes factor for readers to update their personal beliefs.

Bayesians disagree among each other about whether Bayes factors should be the basis of dichotomous claims, or not. Those who promote the use of Bayes factors to make claims often refer to thresholds proposed by Jeffreys (1939), where a BF > 3 is “substantial evidence”, and a BF > 10 is considered “strong evidence”. Some journals, such as Nature Human Behavior, have the following requirement for researchers who submit a Registered Report: “For inference by Bayes factors, authors must be able to guarantee data collection until the Bayes factor is at least 10 times in favour of the experimental hypothesis over the null hypothesis (or vice versa).” When researchers decide to collect data until a specific threshold is crossed to make a claim about a test, their claim can be correct, or wrong, just as when p-values are the statistical quantity a claim is based on. As both the Bayes factor and the p-value can be computed based on the sample size and the t-value (Francis, 2016; Rouder et al., 2009), there is nothing special about using Bayes factors as the basis of an ordinal claim. The exact long run error rates can not be directly controlled when computing Bayes factors, and the Type 1 and Type 2 error rate depends on the choice of the prior and the choice for the cut-off used to decide to make a claim. Simulations studies show that for commonly used priors and a BF > 3 cut-off to make claims the Type 1 error rate is somewhat smaller, but the Type 2 error rate is considerably larger (Kelter, 2021).

To conclude this section, whenever researchers make claims, they can make erroneous claims, and error control should be a worthy goal. Error control is not a consideration when researchers do not make ordinal claims (e.g., X is larger than Y, there is a non-zero correlation between X and Y, etc). If Bayes factors are used to quantify how much researchers should update personal beliefs in a hypothesis, there is no need to consider error control, but researchers should also refrain from making any ordinal claims based on Bayes factors in the results section or the discussion section. Giving up error control also means giving up claims about the presence or absence of effects.

Misunderstanding 5: Interpret Bayes Factors as effect sizes.

Bayes factors are not statements about the size of an effect. It is therefore not appropriate to conclude that the effect size is small or large purely based on the Bayes factor. Depending on the priors used when specifying the alternative and null model, the same Bayes factor can be observed for very different effect size estimates. The reverse is also true. The same effect size can correspond to Bayes factors supporting the null or the alternative hypothesis, depending on how the null model and the alternative model are specified. Researchers should therefore always report and interpret effect size measure. Statements about the size of effects should only be based on these effect size measures, and not on Bayes factors.

Any tool for statistical inferences will be mis-used, and the greater the adoption, the more people will use a tool without proper training. Simplistic sales pitches for Bayes factors (e.g., Bayes factors tell you the probability that your hypothesis is true, Bayes factors do not require error control, you can use ‘default’ Bayes factors and do not have to think about your priors) contribute to this misuse. When reviewing papers that report Bayes factors, check if the authors use Bayes factors to draw correct inferences.

Bem, D. J. (2011). Feeling the future: Experimental evidence for anomalous retroactive influences on cognition and affect. Journal of Personality and Social Psychology, 100(3), 407–425. https://doi.org/10.1037/a0021524

Bem, D. J., Utts, J., & Johnson, W. O. (2011). Must psychologists change the way they analyze their data? Journal of Personality and Social Psychology, 101(4), 716–719. https://doi.org/10.1037/a0024777

Cohen, J. (1994). The earth is round (p .05). American Psychologist, 49(12), 997–1003. https://doi.org/10.1037/0003-066X.49.12.997

de Heide, R., & Grünwald, P. D. (2017). Why optional stopping is a problem for Bayesians. arXiv:1708.08278 [Math, Stat]. https://arxiv.org/abs/1708.08278

Francis, G. (2016). Equivalent statistics and data interpretation. Behavior Research Methods, 1–15. https://doi.org/10.3758/s13428-016-0812-3

Jeffreys, H. (1939). Theory of probability (1st ed). Oxford University Press.

Kelter, R. (2021). Analysis of type I and II error rates of Bayesian and frequentist parametric and nonparametric two-sample hypothesis tests under preliminary assessment of normality. Computational Statistics, 36(2), 1263–1288. https://doi.org/10.1007/s00180-020-01034-7

Rouder, J. N. (2014). Optional stopping: No problem for Bayesians. Psychonomic Bulletin & Review, 21(2), 301–308.

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(2), 225–237. https://doi.org/10.3758/PBR.16.2.225

van de Schoot, R., Winter, S. D., Ryan, O., Zondervan-Zwijnenburg, M., & Depaoli, S. (2017). A systematic review of Bayesian articles in psychology: The last 25 years. Psychological Methods, 22(2), 217–239. https://doi.org/10.1037/met0000100

Wagenmakers, E.-J., Wetzels, R., Borsboom, D., & van der Maas, H. L. J. (2011). Why psychologists must change the way they analyze their data: The case of psi: Comment on Bem (2011). Journal of Personality and Social Psychology, 100(3), 426–432. https://doi.org/10.1037/a0022790

Wong, T. K., Kiers, H., & Tendeiro, J. (2022). On the Potential Mismatch Between the Function of the Bayes Factor and Researchers’ Expectations. Collabra: Psychology, 8(1), 36357. https://doi.org/10.1525/collabra.36357

#### 1 comment:

1. Thanks for writing this - after reading a recent tweet of yours, I thought the arguments behind that brief statement must be somewhere in your posts and materials, ooor maybe even there's a new post in the making, and here it comes I guess :)

Two - hopefully not silly - questions came into mind.

1: "The correct claim is that people should update their belief in the alternative hypothesis by a factor of 10." - Doesn't this updating factor require that the personal prior belief of the reader was the same as the prior specified in the analysis? So when I don't agree with (or don't understand) the priors, then I'm probably gonna have a hard time drawing a conclusion for myself... (...which could also apply to a corresponding frequentist analysis as well)

2: "Giving up error control also means giving up claims about the presence or absence of effects." - This is a strong claim, and I think this was basically what the tweet I was referring to was about. You concentrated on Bayes Factors, but I wondered how this would apply to a Bayesian inferring, plotting and characterizing the full Bayesian posterior distribution and make claims and interpretations ba(ye)sed on that. For instance, seeing that moost of the posterior mass encompasses a range of effects of meaningful strength, then they would base their interpretation and further research or other decision based on that. Would you consider some formal "Bayesian error control" possible and necessary in this situation? I'm curious how you look at this from your point of view.

(I unfortunately don't have deep experience with Bayesian methods yet, but I'd think that a) based on the posterior distribution and the specific hypothesis, readers might be able to consider the weight of the evidence for themselves, b) maybe in critical situations, the downstream consequences of dichotomous decisions could even be explicitly modeled in a Bayesian framework.)