Power analysis for default Bayesian t-tests

January 14, 2016
By

(This article was first published on The 20% Statistician, and kindly contributed to R-bloggers)

One important benefit of Bayesian statistics is that you can provide relative support for the null hypothesis. When the null hypothesis is true, p-values will forever randomly wander between 0 and 1, but a Bayes factor has consistency (Rouder, Speckman, Sun, Morey, & Iverson, 2009), which means that as the sample size increases, the Bayes Factor will tell you which of two hypotheses has received more support.
Bayes Factors can express relative evidence for the null hypothesis (H0) compared to the alternative hypothesis (H1), referred to as a BF01, or relative evidence for H1 compared to H0 (a BF10). A BF01 > 3 is sometimes referred to as substantial evidence for H0 (see Wagenmakers, Wetzels, Borsboom, & Van Der Maas, 2011), but what they really mean is substantial evidence for H0 compared to H1.


Table 1 from Wagenmakers et al., 2011, #fixedthatforyou
Since a Bayes Factor provides relative support for one over another hypothesis, the choice for an alternative hypothesis is important. In Neyman-Pearson Frequentist approaches, researchers have to specify the alternative hypothesis to be able to calculate the power of their test. For example, let’s assume a researcher no idea what to expect, and decided to use the average effect size in psychology of d = 0.43 as the expected effect when the alternative hypothesis is true. It then becomes easy to calculate that you need 115 participants in each group of your experiment, if you plan to perform a two-sided independent t-test with an alpha of 0.05, and you want to have 90% power.
When calculating the Bayes Factor, the alternative is specified through the r-scale. This is a scale for the Cauchy prior, which is chosen in such a way that the researcher expects there is a 50% chance of observing an absolute effect larger than the scale value chosen. As Rouder and colleagues say: “A well-known and attractive alternative to placing priors on the mean µ is to place them on effect size, where effect size is denoted by δ and given as δ = µ/σ. Thus, the researcher above would choose an r-scale of 0.43.
The default Bayesian t-test originally used a r-scale of 1 (a very large effect), but the updated default test uses a r-scale of 0.707. This means that whenever you perform a study where you calculate the default Bayes Factor, and find a BF10 < 1/3, you have observed support for the null-hypothesis, relative to an alternative hypothesis of an effect of d = 0.707.
Let’s assume the default Bayesian t-test answers a question you are interested in, and you want to know whether your data is stronger support for a d = 0 compared to a d = 0.707. You know that to have 90% power to observe a d = 0.707 in an independent t-test, you need 44 participants in each condition, and you have heard people recommend to use at least 50 participants in each condition, so you set out to collect 50 participants in each condition. You are also completely fine with a null-result: You have submitted your study to a journal that accepts pre-registered reports, and the reviewers there agreed the data will be interesting whether they support the null hypothesis or the alternative hypothesis. What is the probability you will observe support for H0 relative to H1 (see this paper by Felix Schonbrodt and colleagues for similar Frequentist calculations of Bayes Factors)?
There are no Bayesian power calculators (yet, as far as I know). Bayes Factors express the relative evidence in your data, and you don’t need a threshold to interpret a Bayes Factor. In principle, any BF10 < 1 provides stronger support for the null hypothesis than for the alternative hypothesis. But no one will be convinced if you argue for the null hypothesis based on a BF10= 0.98, and it’s practical to have some sort of agreed upon threshold where we all agree data can be interpreted as ‘support’ – hence tables with Bayes Factor thresholds as the one reported in Wagenmakers et al, 2011 (see above). So what are the chances of actually observing relative support for the null hypothesis, compare to an alternative hypothesis of d = 0.707, with 50 participants in each condition, when you want a BF10 < 1/3?
Using the R code below, you can easily calculate this probability yourself. It also provides the probability of finding support for the alternative hypothesis. This means you can run the simulations twice, once setting the true effect size to 0, once setting the true effect size to 0.707, and see whether your sample size is large enough to get what you want from the data you plan to collect.Note you need quite some simulations to be accurate, and this takes minutes)
Running these simulations, we see there’s about a 69% probability of finding support for the null-hypothesis with a BF10< 1/3 if the true effect size is 0. In the histogram of Bayes Factors below, this is the distribution to the left of the left vertical red line. Approximately 1.6% of the Bayes Factors give misleading evidence (a BF > 3, to the right of the right vertical red line) and around 30% of the Bayes Factors are inconclusive evidence.

The second set of simulations reveals there’s approximately a 85% chance of finding a BF10 > 3 when the true effect size is 0.707. Based on these data, and your explicit interest in providing support for the null-hypothesis, you might want to collect some more participants.

Let’s look at some other scenarios. Imagine that in the previous situation, the true effect size was d = 0.43. It turns out that 13% of our experiments will find relative support for the null-hypothesis, 38% of the tests will find relative support for the alternative hypothesis, and 48% of the tests will be inconclusive, with a Bayes Factor between 1/3 and 3. Luckily, with Bayesian statistics you can continue collecting data, but since no one has unlimited resources, it’s important to know whether it’s feasible to collect sufficient data to answer your question in the first place. For example, in the default Bayesian t-test, don’t bother trying to find ‘strong evidence’ (BF < 1/10) for the null if you are planning to collect less than 200 participants. When the null is really true, and you plan to collect data from 100 to 200 participants in each condition, you’ll never find the evidence you are looking for. With 300 participants in each condition, you’ll start to get around 33% probability of finding what you want, indicating a threshold of 10 might be unreasonable for single studies, and more suitable for meta-analyses.  
The simulations also show that the probability of an incorrect conclusion (observing support for H0 when H1 is true, or support for H1 when H0 is true) is very low. Indeed, it’s so low, that we might want to reconsider used a Bayes Factor of 3 as a threshold.
For example, if we set a threshold as low as 1.2, instead of 3, the simulations show that, when the null-hypothesis is true, approximately 5% of the Bayes Factors we observe will provide relative support for the alternative hypothesis, and over 90% of the Bayes Factors will provide relative support for the null hypothesis, with 50 participants in each condition (the remainder, 5% fall in the small inconclusive evidence area). This may not be ‘substantial evidence’ but you won’t be wrong very often, in the long run, when you decide to accept the null hypothesis over the alternative hypothesis of d = 0.707 when you collect 50 participants in each condition.
You may not like the use of a cut-off value to indicate ‘relative support’ as proposed by some Bayesians (Dienes, 2016; Wagenmakers et al., 2011). I think these cut-offs are most important when planning an experiment. It makes sense to aim for support most people will find convincing. When the data are in hand, you can interpret it any way you like. If you agree, then important questions are which cut-off values are acceptable under which circumstances? Can we develop formulas for Bayesian power analysis instead of having to rely on simulations?
I learned something by playing around with these simulations, and I expect others moving to Bayes Factors will have similar questions as I tried to examine here. Power analysis is  Frequentist concept, and it might not be close to the heart of Bayesians, but I expect it meets a future need of researchers switching to Bayesian statistics, especially when they want to design studies that provide support for the null.


Dienes, Z. (2016). How Bayes factors change scientific practice. Journal of Mathematical Psychology. http://doi.org/10.1016/j.jmp.2015.10.003

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. http://doi.org/10.3758/PBR.16.2.225

Wagenmakers, E.-J., Wetzels, R., Borsboom, D., & Van Der Maas, H. L. (2011). Why psychologists must change the way they analyze their data: the case of psi: comment on Bem (2011). Retrieved from http://psycnet.apa.org/journals/psp/100/3/426/

 

To leave a comment for the author, please follow the link and comment on their blog: The 20% Statistician.

R-bloggers.com offers daily e-mail updates about R news and tutorials on topics such as: Data science, Big Data, R jobs, visualization (ggplot2, Boxplots, maps, animation), programming (RStudio, Sweave, LaTeX, SQL, Eclipse, git, hadoop, Web Scraping) statistics (regression, PCA, time series, trading) and more...



If you got this far, why not subscribe for updates from the site? Choose your flavor: e-mail, twitter, RSS, or facebook...

Comments are closed.

Sponsors

Mango solutions



RStudio homepage



Zero Inflated Models and Generalized Linear Mixed Models with R

Quantide: statistical consulting and training

datasociety

http://www.eoda.de





ODSC

ODSC

CRC R books series





Six Sigma Online Training









Contact us if you wish to help support R-bloggers, and place your banner here.

Never miss an update!
Subscribe to R-bloggers to receive
e-mails with the latest R posts.
(You will not see this message again.)

Click here to close (This popup will not appear again)