# Hierarchical shift function: a powerful alternative to the t-test

In this post I introduce a simple yet powerful method to compare two dependent groups: the hierarchical shift function. The code is on GitHub. More details are in Rousselet & Wilcox (2019), with a reproducibility package on figshare.

Let’s consider different situations in a hierarchical setting: we’ve got trials from 2 conditions in several participants. Imagine we collected data from one participant and the results look like this: These fake reaction time data were created by sampling from ex-Gaussian distributions. Here the two populations are shifted by a constant, so we expect a uniform shift between the two samples. Later we’ll look at examples showing  differences most strongly in early responses, late responses, and in spread.

To better understand how the distributions differ, let’s look at a shift function, in which the difference between the deciles of the two conditions are plotted as a function of the deciles in condition 1 – see details in Rousselet et al. (2017). The decile differences are all negative, showing stochastic dominance of condition 2 over condition 1. The function is not flat because of random sampling and limited sample size. Now, let’s say we collected 100 trials per condition from 30 participants. How do we proceed? There are a variety of approaches available to quantify distribution differences. Ideally, such data would be analysed using a multi-level model, including for instance ex-Gaussian fits, random slopes and intercepts for participants, item analyses… This can be done using the lme4 or brms R packages. However, in my experience, in neuroscience and psychology articles, the most common approach is to collapse the variability across trials into a single number per participant and condition to be able to perform a paired t-test: typically, the mean is computed across trials for each condition and participant, then the means are subtracted, and the distribution of mean differences is entered into a one-sample t-test. Obviously, this strategy throws away a huge amount of information! And the results of such second-tier t-tests are difficult to interpret: a positive test leaves us wondering exactly how the distributions differ; a negative test is ambiguous – beside avoiding the ‘absence of evidence is not evidence of absence’ classic error, we also need to check if the distributions do not differ in other aspects than the mean. So what can we do?

Depending on how conditions differ, looking at other aspects of the data than the mean can be more informative. For instance, in Rousselet & Wilcox (2019), we consider group comparisons of individual medians. Considering that the median is the second quartile, looking at the other quartiles can be of theoretical interest to investigate effects in early or later parts of distributions. This could be done in several ways, for instance by making inferences on the first quartile (Q1) or the third quartile (Q3). If the goal is to detect differences anywhere in the distributions, a more systematic approach consists in quantifying differences at multiple quantiles. Here we consider the case of the deciles, but other quantiles could be used. First, for each participant and each condition, the sample deciles are computed over trials. Second, for each participant, condition 2 deciles are subtracted from condition 1 deciles – we’re dealing with a within-subject (repeated-measure) design. Third, for each decile, the distribution of differences is subjected to a one-sample test. Fourth, a correction for multiple comparisons is applied across the 9 one-sample tests. I call this procedure a hierarchical shift function. There are many options available to implement this procedure and the example used here is not the definitive answer: the goal is simply to demonstrate that a relatively simple procedure can be much more powerful and informative than standard approaches.

In creating a hierarchical shift function we need to make three choices: a quantile estimator, a statistical test to assess quantile differences across participants, and a correction for multiple comparisons technique. The deciles were estimated using type 8 from the base R `quantile()` function (see justification in Rousselet & Wilcox, 2019). The group comparisons were performed using a one-sample t-test on 20% trimmed means, which performs well in many situations, including in the presence of outliers. The correction for multiple comparisons employed Hochberg’s strategy (Hochberg, 1988), which guarantees that the probability of at least one false positive will not exceed the nominal level as long as the nominal level is not exceeded for each quantile.

In Rousselet & Wilcox (2019), we consider power curves for the hierarchical shift function (HSF) and contrast them to other approaches: by design, HSF is sensitive to more types of differences than any standard approach using the mean or a single quantile. Another advantage of HSF is that the location of the distribution difference can be interrogated, which is impossible if inferences are limited to a single value.

Here is what the hierarchical shift function looks like for our uniform shift example: The decile differences between conditions are plotted for each participant (colour coded) and the group 20% trimmed means are superimposed in black. Differences are pretty constant across deciles, suggesting a uniform shift. Most participants have shift functions entirely negative – a case of stochastic dominance of one condition over the other. There is growing uncertainty as we consider higher deciles, which is expected from measurements of right skewed distributions. P values are available in the GitHub code.

Instead of standard parametric confidence intervals, we can also consider percentile bootstrap confidence intervals (or highest density intervals), as done here: Distributions of bootstrap estimates can be considered cheap Bayesian posterior distributions. They also contain useful information not captured by simply reporting confidence intervals.

Here we plot them using `geom_halfeyeh()` from tidybayes. The distributions of bootstrap estimates of the group 20% trimmed means are shown in orange, one for each decile. Along the base of each distribution, the black dot marks the mode and the vertical lines mark the 50% and 90% highest density intervals.

Nice hey?! Reporting a figure like that is dramatically more informative than reporting a P value and a confidence interval from a t-test!

A bootstrap approach can also be used to perform a cluster correction for multiple comparisons – see details on GitHub. Preliminary simulations suggest that the approach can provide substantial increase in power over the Hochberg’s correction – more on that in another post.

Let’s look at 3 more examples, just for fun…

# Example 2: early difference

Example participant: Shift function: Hierarchical shift function with confidence intervals: Percentile bootstrap estimate densities: # Example 3: difference in spread

Example participant: Shift function: Hierarchical shift function with confidence intervals: Percentile bootstrap estimate densities: # Example 4: late difference

Example participant: Shift function: Hierarchical shift function with confidence intervals: Percentile bootstrap estimate densities: # Conclusion

The hierarchical shift function can be used to achieve two goals:

• to screen data for potential distribution differences using p values, without limiting the exploration to a single statistics like the mean;
• to illustrate and quantify how distributions differ.

I think of the hierarchical shift function as the missing link between t-tests and multi-level models. I hope it will help a few people make sense of their data and maybe nudge them towards proper hierarchical modelling.

R functions for the parametric hierarchical shift function are available in the `rogme` package. I also plan bootstrap functions. Then I’ll tackle the case of 2 independent groups, which requires a third level quantifying differences of differences.

# Planning for measurement precision, an alternative to power analyses

When we estimate power curves, we ask this question: given some priors about the data generating process, the nature of the effect and measurement variance, what is the probability to detect an effect for a given statistical test (say using an arbitrary p<0.05 threshold) for various sample sizes and effect sizes. While there are very good reasons to focus on power estimation, this is not the only or the most important aspect of an experimental procedure to consider (Gelman & Carlin, 2014). Indeed, finding the number of observations needed so that we get p<0.05 in say 87% of experiments, is not the most exciting part of designing an experiment.

The relevant question is not “What is the power of a test?” but rather is “What might be expected to happen in studies of this size?” (Gelman & Carlin, 2014)

A related but more important question is that of measurement precision: given some priors and a certain number of participants, how close can we get to the unknown population value (Maxwell et al., 2008; Schönbrodt & Perugini, 2013; Peters & Crutzen, 2018; Trafimow, 2019)? Not surprisingly, measurement precision depends on sample size. As we saw in previous posts, sampling distributions get narrower with increasing sample sizes:

And with narrower sampling distributions, measurement precision increases. To illustrate, let’s consider an example from a lexical decision task – hundreds of reaction times (RT) were measured in hundreds of participants who had to distinguish between words and non-words presented on a computer screen.

Here are examples of RT distributions from 100 participants for each condition: Reaction time distributions from 100 participants. Participants were randomly selected among 959. Distributions are shown for the same participants (colour coded) in the Word (A) and Non-Word (B) conditions.

If we save the median of each distribution, for each participant and condition, we get these positively skewed group level distributions:

The distribution of pairwise differences between medians is also positively skewed:

Notably, most participants have a positive difference: 96.4% of participants are faster in the Word than the Non-Word condition – a potential case of stochastic dominance (Rouder & Haaf, 2018; see also this summary blog post).

Now let say we want to estimate the group difference between conditions. Because of the skewness at each level of analysis (within and across participants), we estimate the central tendency at each level using the median: that is, we compute the median for each participant and each condition, then compute the medians of medians across participants (a more detailed assessment could be obtained by performing hierarchical modelling or multiple quantile estimation for instance).

Then we can assess measurement precision at the group level by performing a multi-level simulation. In this simulation, we can ask, for instance, how often the group estimate is no more than 10 ms from the population value across many experiments. To simplify, in each iteration of the simulation, we draw 200 trials per condition and participant, compute the median and save the Non-Word – Word difference. Group estimation of the difference is then based on a random sample of 10 to 300 participants, with the group median computed across participants’ differences between medians. Because the dataset is very large at the two level of analysis, we can pretend we have access to the population values, and define them by first computing, for each condition, the median across all available trials for each participant, second by computing across all participants the median of the pairwise differences.

Having defined population values (the truth we’re trying to estimate, here a group difference of 78 ms), we can calculate measurement precision as the proportion of experiments in which the group estimate is no more than X ms from the population value, with X varying from 5 to 40 ms. Here are the results: Group measurement precision for the difference between the Non-Word and Word conditions. Measurement precision was estimated by using a simulation with 10,000 iterations, 200 trials per condition and participant, and varying numbers of participants.

Not surprisingly, the proportion of estimates close to the population value increases with the number of participants. More interestingly, the relationship is non-linear, such that a larger gain in precision can be achieved by increasing sample size for instance from 10 to 20 compared to from 90 to 100.

The results also let us answer useful questions for planning experiments (see the black arrows in the above figure):

• So that in 70% of experiments the group estimate of the median is no more than 10 ms from the population value, we need to test at least 56 participants.

• So that in 90% of experiments the group estimate of the median is no more than 20 ms from the population value, we need to test at least 38 participants.

Obviously, this is just an example, about a narrow problem related to lexical decisions. Other aspects could be considered too, for instance the width of the confidence intervals (Maxwell, Kelley & Rausch, 2008; Peters & Crutzen, 2017; Rothman & Greenland, 2018). And for your particular case, most likely, you won’t have access to a large dataset from which to perform a data driven simulation. In this case, you can get estimates about plausible effect sizes and their variability from various sources (Gelman & Carlin 2014):

• related data;

• (systematic) literature review;

• meta-analysis;

• outputs of a hierarchical model;

• modelling.

To model a range of plausible effect sizes and their consequences on repeated measurements, you need priors about a data generating process and how distributions differ between conditions. For instance, you could use exGaussian distributions to simulate RT data. For research on new effects, it is advised to consider a large range of potential effects, with their plausibility informed by the literature and psychological/biological constraints.

Although relying on the literature alone can lead to over-optimistic expectations because of the dominance of small n studies and a bias towards significant results (Yarkoni 2009; Button et al. 2013), methods are being developed to overcome these limitations (Anderson, Kelley & Maxwell, 2017). In the end, the best cure against effect size over-estimation is a combination of pre-registration/registered reports (to diminish literature bias) and data sharing (to let anyone do their own calculations and meta-analyses).

# Code

The code is on figshare: the simulation can be reproduced using the `flp_sim_precision` notebook, the illustrations of the distributions can be reproduced using `flp_illustrate_dataset`.

# Shiny app by Malcolm Barrett (@malco_barrett)

https://malcolmbarrett.shinyapps.io/precisely/

# References

Anderson, S.F., Kelley, K. & Maxwell, S.E. (2017) Sample-Size Planning for More Accurate Statistical Power: A Method Adjusting Sample Effect Sizes for Publication Bias and Uncertainty. Psychol Sci, 28, 1547-1562.

Bland J.M.. The tyranny of power: is there a better way to calculate sample size? https://www.bmj.com/content/339/bmj.b3985)

Button, K.S., Ioannidis, J.P., Mokrysz, C., Nosek, B.A., Flint, J., Robinson, E.S. & Munafo, M.R. (2013) Power failure: why small sample size undermines the reliability of neuroscience. Nature reviews. Neuroscience, 14, 365-376.

Ferrand, L., New, B., Brysbaert, M., Keuleers, E., Bonin, P., Meot, A., Augustinova, M. & Pallier, C. (2010) The French Lexicon Project: lexical decision data for 38,840 French words and 38,840 pseudowords. Behav Res Methods, 42, 488-496.

Gelman, A. & Carlin, J. (2014) Beyond Power Calculations: Assessing Type S (Sign) and Type M (Magnitude) Errors. Perspect Psychol Sci, 9, 641-651.

Maxwell, S.E., Kelley, K. & Rausch, J.R. (2008) Sample size planning for statistical power and accuracy in parameter estimation. Annu Rev Psychol, 59, 537-563.

Peters, G.-J.Y. & Crutzen, R. (2017) Knowing exactly how effective an intervention, treatment, or manipulation is and ensuring that a study replicates: accuracy in parameter estimation as a partial solution to the replication crisis. PsyArXiv. doi:10.31234/osf.io/cjsk2.

Rothman, K.J. & Greenland, S. (2018) Planning Study Size Based on Precision Rather Than Power. Epidemiology, 29, 599-603.

Rouder, J.N. & Haaf, J.M. (2018) Power, Dominance, and Constraint: A Note on the Appeal of Different Design Traditions. Advances in Methods and Practices in Psychological Science, 1, 19-26.

Rousselet, G.A. & Wilcox, R.R. (2018) Reaction times and other skewed distributions: problems with the mean and the median. bioRxiv. doi: https://doi.org/10.1101/383935

Rousselet, G.; Wilcox, R. (2018): Reaction times and other skewed distributions: problems with the mean and the median. figshare. Fileset. https://doi.org/10.6084/m9.figshare.6911924.v1

Schönbrodt, F.D. & Perugini, M. (2013) At what sample size do correlations stabilize? J Res Pers, 47, 609-612.

Trafimow, D. (2019) Five Nonobvious Changes in Editorial Practice for Editors and Reviewers to Consider When Evaluating Submissions in a Post p < 0.05 Universe, The American Statistician, 73:sup1, 340-345,

Yarkoni, T. (2009) Big Correlations in Little Studies: Inflated fMRI Correlations Reflect Low Statistical Power‚ Commentary on Vul et al. (2009). Perspectives on Psychological Science, 4, 294-298.

# Reaction times and other skewed distributions: problems with the mean and the median (part 4/4)

This is part 4 of a 4 part series. Part 1 is here.

In this post, I look at median bias in a large dataset of reaction times from participants engaged in a lexical decision task. The dataset was described in a previous post.

After removing a few participants who didn’t pay attention to the task (low accuracy or too many very late responses), we’re left with 959 participants to play with. Each participant had between 996 and 1001 trials for each of two conditions, Word and Non-Word.

Here is an illustration of reaction time distributions from 100 randomly sampled participants in the Word condition: Same in the Non-Word condition: Skewness tended to be larger in the Word than the Non-Word condition. Based on the standard parametric definition of skewness, that was the case in 80% of participants. If we use a non-parametric estimate instead (mean – median), it was the case in 70% of participants.

If we save the median of every individual distribution, we get the two following group distributions, which display positive skewness: The same applies to distributions of means: So we have to worry about skewness at 2 levels:

• individual distributions

• group distributions

Here I’m only going to explore estimation bias as a result of skewness and sample size in individual distributions. From what we learnt in previous posts, we can already make predictions: because skewness tended to be stronger in the Word than in the Non-Word condition, the bias of the median will be stronger in the former than the later for small sample sizes. That is, the median in the Word condition will tend to be more over-estimated than the median in the Non-Word condition. As a consequence, the difference between the median of the Non-Word condition (larger RT) and the median of the Word condition (smaller RT) will tend to be under-estimated. To check this prediction, I estimated bias in every participant using a simulation with 2,000 iterations. I assumed that the full sample was the population, from which we can compute population means and population medians. Because the Non-Word condition is the least skewed, I used it as the reference condition, which always had 200 trials. The Word condition had 10 to 200 trials, with 10 trial increments. In the simulation, single RT were sampled with replacements among the roughly 1,000 trials available per condition and participant, so that each iteration is equivalent to a fake experiment.

Let’s look at the results for the median. The figure below shows the bias in the long run estimation of the difference between medians (Non-Word – Word), as a function of sample size in the Word condition. The Non-Word condition always had 200 trials. All participants are superimposed and shown as coloured traces. The average across participants is shown as a thicker black line. As expected, bias tended to be negative with small sample sizes. For the smallest sample size, the average bias was -11 ms. That’s probably substantial enough to seriously distort estimation in some experiments. Also, variability is high, with a 80% highest density interval of [-17.1, -2.6] ms. Bias decreases rapidly with increasing sample size. For n=60, it is only 1 ms.

But inter-participant variability remains high, so we should be cautious interpreting results with large numbers of trials but few participants. To quantify the group uncertainty, we could measure the probability of being wrong, given a level of desired precision, as demonstrated here for instance.

After bootstrap bias correction (with 200 bootstrap resamples), the average bias drops to roughly zero for all sample sizes: Bias correction also reduced inter-participant variability.

As we saw in the previous post, the sampling distribution of the median is skewed, so the standard measure of bias (taking the mean across simulation iterations) does not provide a good indication of the bias we can expect in a typical experiment. If instead of the mean, we compute the median bias, we get the following results: Now, at the smallest sample size, the average bias is only -2 ms, and it drops to near zero for n=20. This result is consistent with the simulations reported in the previous post and confirms that in the typical experiment, the average bias associated with the median is negligible.

What happens with the mean? The average bias of the mean is near zero for all sample sizes. Individual bias values are also much less variable than median values. This difference in bias variability does not reflect a difference in variability among participants for the two estimators of central tendency. In fact, the distributions of differences between Non-Word and Word conditions are very similar for the mean and the median. Estimates of spread are also similar between distributions:

IQR: mean RT = 78; median RT = 79

MAD: mean RT = 57; median RT = 54

VAR: mean RT = 4507; median RT = 4785

This suggests that the inter-participant bias differences are due to the shape differences observed in the first two figures of this post.

Finally, let’s consider the median bias of the mean. For the smallest sample size, the average bias across participants is 7 ms. This positive bias can be explained easily from the simulation results of post 3: because of the larger skewness in the Word condition, the sampling distribution of the mean was more positively skewed for small samples in that condition compared to the Non-Word condition, with the bulk of the bias estimates being negative. As a result, the mean tended to be more under-estimated in the Word condition, leading to larger Non-Word – Word differences in the typical experiment.

I have done a lot more simulations and was planning even more, using other datasets, but it’s time to move on! Of particular note, it appears that in difficult visual search tasks, skewness can differ dramatically among set size conditions – see for instance data posted here.

# Concluding remarks

The data-driven simulations presented here confirm results from our previous simulations:

• if we use the standard definition of bias, for small sample sizes, mean estimates are not biased, median estimates are biased;
• however, in the typical experiment (median bias), mean estimates can be more biased than median estimates;

• bootstrap bias correction can be an effective tool to reduce bias.

Given the large differences in inter-participant variability between the mean and the median, an important question is how to spend your money: more trials or more participants (Rouder & Haaf 2018)? An answer can be obtained by running simulations, either data-driven or assuming generative distributions (for instance exGaussian distributions for RT data). Simulations that take skewness into account are important to estimate bias and power. Assuming normality can have disastrous consequences.

Despite the potential larger bias and bias variability of the median compared to the mean, for skewed distributions I would still use the median as a measure of central tendency, because it provides a more informative description of the typical observations. Large sample sizes will reduce both bias and estimation variability, such that high-precision single-participant estimation should be easy to obtain in many situations involving non-clinical samples. For group estimations, much larger samples than commonly used are probably required to improve the precision of our inferences.

Although the bootstrap bias correction seems to work very well in the long run, for a single experiment there is no guarantee it will get you closer to the truth. One possibility is to report results with and without bias correction.

For group inferences on the median, traditional techniques use incorrect estimations of the standard error, so consider modern parametric or non-parametric techniques instead (Wilcox & Rousselet, 2018).

# References

Miller, J. (1988) A warning about median reaction time. J Exp Psychol Hum Percept Perform, 14, 539-543.

Rouder, J.N. & Haaf, J.M. (2018) Power, Dominance, and Constraint: A Note on the Appeal of Different Design Traditions. Advances in Methods and Practices in Psychological Science, 1, 19-26.

Wilcox, R.R. & Rousselet, G.A. (2018) A Guide to Robust Statistical Methods in Neuroscience. Curr Protoc Neurosci, 82, 8 42 41-48 42 30.

# Bayesian shift function

Two distributions can differ in many ways, yet the standard approach in neuroscience & psychology is to assume differences in means. That’s why the first step in exploratory data analysis should always be detailed graphical representations (Rousselet et al. 2016, 2017). To help quantify how two distributions differ, a fantastic tool is the shift function – an example is provided below. It consists in plotting the difference between group quantiles, as a function of the quantiles in one group. The technique was first described in the 1970s by Doksum (1974; Doksum & Sievers, 1976), and later refined by Wilcox using the Harrell-Davis quantile estimator (Harrell & Davis, 1982) in conjunction with two percentile bootstrap methods (Wilcox 1995; Wilcox et al. 2014). The technique is related to delta plots and relative distribution methods (see details in Rousselet et al. 2017).

The goal of this post is to get feedback on my first attempt to make a Bayesian version of the shift function. Below I describe three potential strategies. My main motivation for making a Bayesian version is that the shift function comes with frequentist confidence intervals and p values. Although I still use confidence intervals to describe sampling variability, they are inherently linked to p values and tend to be associated with major flaws in interpretation (Morey et al. 2016). And my experience is that if p values are available, most researchers will embark on lazy and irrational decision making (Wagenmakers 2007). P values would not be a problem if they were just used as one of many pieces of evidence, without any special status (McShane et al. 2017).

Let’s consider a toy example: two independent exGaussian distributions, with n = 100 in each group. The two groups clearly differ, as expected from the generative process (see online code). A t-test on means suggests a large uncertainty about the group difference:

difference = – 65 ms [-166, 37]

t = -1.26

p = 0.21 The x-axis shows the quantiles of group 1 (here only the 9 deciles, which is a good default). The y-axis shows the difference between deciles of group 1 and group 2. Intuitively, the difference shows by how much group 2 would need to be shifted to match group 1, for each decile. The coloured labels show the quantile differences. I let you go back and forth between the density estimates and the shift function to understand what’s going on. Another detailed description is provided here if needed.

# Strategy 1: Bayesian bootstrap

A simple strategy to make a Bayesian shift function is to use the Bayesian bootstrap instead of a percentile bootstrap. The percentile bootstrap can already be considered a very cheap way to create Bayesian posterior distributions, if we make the strong (and wrong) assumption that our observations are the only possible ones. Rasmus Bååth provides a detailed introduction to the Bayesian bootstrap, and it’s R implementation on his blog. There is also a video.

Using the Bayesian bootstrap, and 95% highest density intervals (HDI) of the posterior distributions, the shift function looks very similar to the original version, as expected. Except now we’re dealing with credible intervals, and there are no p values, so users have to focus on quantification!

# Strategy 2: Bayesian quantile regression

Another strategy is to use quantile regression, which comes in a Bayesian flavour using the asymmetric Laplacian likelihood family. To do that, I’m using the amazing `brms` R package by Paul Bürkner.

To get started with Bayesian statistics and the `brms` package, I recommend this excellent blog post by Matti Vuorre. Many other great posts are available here.

Using the default priors from `brms`, we can fit a quantile regression line for each group and for each decile. The medians (or means, or modes) and 95% HDIs of the posterior distributions can then be used to create a shift function. Again, the results are quite similar to the original ones, although this time the quantiles do differ a bit from those in the previous versions because we use the medians of the posterior distributions to estimate them. Also, I haven’t looked in much detail at how well the model fits the data. The posterior predictive samples suggest the fits could be improved, but I have too little experience to make a call.

# Strategy 3: Bayesian model with exGaussians

The third strategy is to fit a descriptive model to the distributions, generate samples from the posterior distributions, and compute quantiles from these predicted values. Here, since our toy model simulates reaction time data using exGaussian distributions, it makes sense to fit an exGaussian family to the data. More generally, exGaussian distributions are very good at capturing the shape of RT data (Matzke & Wagenmakers, 2009). Again, the results look similar to the original ones. This strategy has the advantage to require fitting only one model, instead of a new model for each quantile in strategy 2. In addition, we get an exGaussian fit of the data, which is very useful in itself. So that would probably be my favourite strategy.

Does this make even remotely sense?

Which strategy seems more promising?

The clear benefit of strategies 2 and 3 is that they can easily be extended to create hierarchical shift functions, by fitting simultaneously observations from multiple participants. Whereas for the original shift function using the percentile bootstrap, I don’t see any obvious way to make a hierarchical version.

You might want to ask, why not simply focus on modelling the distributions instead of looking at the quantiles? Modelling the shape of the distributions is great of course, but I don’t think it can achieve the same fine-grained quantification achieved by the shift function. Another approach of course would be to fit a more psychologically motivated model, such as a diffusion model. I think these three approaches are complementary.

# References

Doksum, K. (1974) Empirical Probability Plots and Statistical Inference for Nonlinear Models in the two-Sample Case. Annals of Statistics, 2, 267-277.

Doksum, K.A. & Sievers, G.L. (1976) Plotting with Confidence – Graphical Comparisons of 2 Populations. Biometrika, 63, 421-434.

Harrell, F.E. & Davis, C.E. (1982) A new distribution-free quantile estimator. Biometrika, 69, 635-640.

Blakeley B. McShane, David Gal, Andrew Gelman, Christian Robert, Jennifer L. Tackett (2017) Abandon Statistical Significance. arXiv:1709.07588

Matzke, D. & Wagenmakers, E.J. (2009) Psychological interpretation of the ex-Gaussian and shifted Wald parameters: a diffusion model analysis. Psychon Bull Rev, 16, 798-817.

Morey, R.D., Hoekstra, R., Rouder, J.N., Lee, M.D. & Wagenmakers, E.J. (2016) The fallacy of placing confidence in confidence intervals. Psychon Bull Rev, 23, 103-123.

Rousselet, G.A., Foxe, J.J. & Bolam, J.P. (2016) A few simple steps to improve the description of group results in neuroscience. The European journal of neuroscience, 44, 2647-2651.

Rousselet, G., Pernet, C. & Wilcox, R. (2017) Beyond differences in means: robust graphical methods to compare two groups in neuroscience, figshare.

Wagenmakers, E.J. (2007) A practical solution to the pervasive problems of p values. Psychonomic bulletin & review, 14, 779-804.

Wilcox, R.R. (1995) Comparing Two Independent Groups Via Multiple Quantiles. Journal of the Royal Statistical Society. Series D (The Statistician), 44, 91-99.

Wilcox, R.R., Erceg-Hurn, D.M., Clark, F. & Carlson, M. (2014) Comparing two independent groups via the lower and upper quantiles. J Stat Comput Sim, 84, 1543-1551.

# Reaction times and other skewed distributions: problems with the mean and the median (part 3/4)

Bias is defined as the distance between the mean of the sampling distribution (here estimated using Monte-Carlo simulations) and the population value. In part 1 and part 2, we saw that for small sample sizes, the sample median provides a biased estimation of the population median, which can significantly affect group comparisons. However, this bias disappears with large sample sizes, and it can be corrected using a bootstrap bias correction. In part 3, we look in more detail at the shape of the sampling distributions, which was ignored by Miller (1988).

# Sampling distributions

Let’s consider the sampling distributions of the mean and the median for different sample sizes and ex-Gaussian distributions with skewness ranging from 6 to 92 (Figure 1). When skewness is limited (6, top row), the sampling distributions are symmetric and centred on the population values: there is no bias. As we saw previously, with increasing sample size, variability decreases, which is why studies with larger samples provide more accurate estimations. The flip side is that studies with small samples are much noisier, which is why their results tend not to replicate… Figure 1

When skewness is large (92, middle row), sampling distributions get more positively skewed with decreasing sample sizes. To better understand how the sampling distributions change with sample size, we turn to the last row of Figure 1, which shows 50% highest-density intervals (HDI). Each horizontal line is a HDI for a particular sample size. The labels contain the values of the interval boundaries. The coloured vertical tick inside the interval marks the median of the distribution. The red vertical line spanning the entire plot is the population value.

Means For small sample sizes, the 50% HDI is offset to the left of the population mean, and so is the median of the sampling distribution. This demonstrates that the typical sample mean tends to under-estimate the population mean – that is to say, the mean sampling distribution is median biased. This offset reduces with increasing sample size, but is still present even for n=100.

Medians With small sample sizes, there is a discrepancy between the 50% HDI, which is shifted to the left of the population median, and the median of the sampling distribution, which is shifted to the right of the population median. This contrasts with the results for the mean, and can be explained by differences in the shapes of the sampling distributions, in particular the larger skewness and kurtosis of the median sampling distribution compared to that of the mean (see code on github for extra figures). The offset between 50% HDI and the population reduces quickly with increasing sample size. For n=10, the median bias is already very small. From n=15, the median sample distribution is not median bias, which means that the typical sample median is not biased.

Another representation of the sampling distributions is provided in Figure 2: 50% HDI are shown as a function of sample size. For both the mean and the median, bias increase with increasing skewness and decreasing sample size. Skewness also increases the asymmetry of the sampling distributions, but more for the mean than median. Figure 2

So what’s going on here? Is the mean also biased?  According to the standard definition of bias, which is based on the distance between the population mean and the average of the sampling distribution of the mean, the mean is not biased. But this definition applies to the long run, after we replicate the same experiment many times. In practice, we never do that. So what happens in practice, when we perform only one experiment? In that case, the median of the sampling distribution provides a better description of the typical experiment than the mean of the distribution. And the median of the sample distribution of the mean is inferior to the population mean when sample size is small. So if you conduct one small n experiment and compute the mean of a skewed distribution, you’re likely to under-estimate the true value.

Is the median biased after all? The median is indeed biased according to the standard definition. However, with small n, the typical median (represented by the median of the sampling distribution of the median) is close to the population median, and the difference disappears for even relatively small sample sizes.

# Is it ok to use the median then?

If the goal is to accurately estimate the central tendency of a RT distribution, while protecting against the influence of outliers, the median is far more efficient than the mean (Wilcox & Rousselet, 2018). Providing sample sizes are large enough, bias is actually not a problem, and the typical bias is actually very small, as we’ve just seen. So, if you have to choose between the mean and the median, I would go for the median without hesitation.

It’s more complicated though. In an extensive series of simulations, Ratcliff (1993) demonstrated that when performing standard group ANOVAs, the median can lack power compared to other estimators. Ratcliff’s simulations involved ANOVAs on group means, in which for each participant, very few trials (7 to 12) are available for each condition. Based on the simulations, Ratcliff recommended data transformations or computing the mean after applying specific cut-offs to maximise power. However, these recommendations should be considered with caution because the results could be very different with more realistic sample sizes. Also, standard ANOVA on group means are not robust, and alternative techniques should be considered (Wilcox 2017). Data transformations are not ideal either, because they change the shape of the distributions, which contains important information about the nature of the effects. Also, once data are transformed, inferences are made on the transformed data, not on the original ones, an important caveat that tends to be swept under the carpet in articles’ discussions… Finally, truncating distributions introduce bias too, especially with the mean – see next section (Miller 1991; Ulrich & Miller 1994)!

At this stage, I don’t see much convincing evidence against using the median of RT distributions, if the goal is to use only one measure of location to summarise the entire distribution. Clearly, a better alternative is to not throw away all that information, by studying how entire distributions differ (Rousselet et al. 2017). For instance, explicit modelling of RT distributions can be performed with the excellent `brms` R package.

# Other problems with the mean

In addition to being median biased, and a poor measure of central tendency for asymmetric distributions, the mean is also associated with several other important problems. Standard procedures using the mean lack of power, offer poor control over false positives, and lead to inaccurate confidence intervals. Detailed explanations of these problems are provided in Field & Wilcox (2017) and Wilcox & Rousselet (2018) for instance. For detailed illustrations of the problems associated with means in the one-sample case, when dealing with skewed distributions, see the companion reproducibility package on figshare.

If that was not enough, common outlier exclusion techniques lead to bias estimation of the mean (Miller, 1991). When applied to skewed distributions, removing any values more than 2 or 3 SD from the mean affects slow responses more than fast ones. As a consequence, the sample mean tends to underestimate the population mean. And this bias increases with sample size because the outlier detection technique does not work for small sample sizes, which results from the lack of robustness of the mean and the SD. The bias also increases with skewness. Therefore, when comparing distributions that differ in sample size, or skewness, or both, differences can be masked or created, resulting in inaccurate quantification of effect sizes.

Truncation using absolute thresholds (for instance RT < 300 ms or RT > 1,200 ms) also leads to potentially severe bias of the mean, median, standard deviation and skewness of RT distributions (Ulrich & Miller 1994). The median is much less affected by truncation bias than the mean though.

In the next and final post of this series, we will explore sampling bias in a real dataset, to see how much of a problem we’re really dealing with. Until then, thanks for reading.

[GO TO POST 4/4]

# References

Field, A.P. & Wilcox, R.R. (2017) Robust statistical methods: A primer for clinical psychology and experimental psychopathology researchers. Behav Res Ther, 98, 19-38.

Miller, J. (1988) A warning about median reaction time. J Exp Psychol Hum Percept Perform, 14, 539-543.

Miller, J. (1991) Reaction-Time Analysis with Outlier Exclusion – Bias Varies with Sample-Size. Q J Exp Psychol-A, 43, 907-912.

Ratcliff, R. (1993) Methods for dealing with reaction time outliers. Psychol Bull, 114, 510-532.

Rousselet, G.A., Pernet, C.R. & Wilcox, R.R. (2017) Beyond differences in means: robust graphical methods to compare two groups in neuroscience. The European journal of neuroscience, 46, 1738-1748.

Ulrich, R. & Miller, J. (1994) Effects of Truncation on Reaction-Time Analysis. Journal of Experimental Psychology-General, 123, 34-80.

Wilcox, R.R. (2017) Introduction to Robust Estimation and Hypothesis Testing. Academic Press, 4th edition., San Diego, CA.

Wilcox, R.R. & Rousselet, G.A. (2018) A Guide to Robust Statistical Methods in Neuroscience. Curr Protoc Neurosci, 82, 8 42 41-48 42 30.

# Reaction times and other skewed distributions: problems with the mean and the median (part 1/4)

UPDATE: this series of posts, along with much more material, is now part of this article:

Reaction times and other skewed distributions: problems with the mean and the median

In this series of 4 posts, I replicate, expand and discuss the results from

Miller, J. (1988) A warning about median reaction time. J Exp Psychol Hum Percept Perform, 14, 539-543.

Part 1 = replicate Miller’s simulations + apply bootstrap bias correction

Part 2 = expand Miller’s simulations to group comparison

Part 3 = problems with the mean

Part 4 = application to a large dataset

Data & code are available on github. The content of the 4 posts is also described in this article.

Reaction times (RT) and many other quantities in neuroscience & psychology are skewed. This asymmetry tends to differ among experimental conditions, such that a measure of central tendency and a measure of spread are insufficient to capture how conditions differ. Instead, to understand the potentially rich differences among distributions, it is advised to consider multiple quantiles of the distributions (Doksum 1974; Pratte et al. 2010; Rousselet et al. 2017), or to model the shapes of the distributions (Heathcote et al. 1991; Rouder et al. 2005; Palmer et al. 2011; Matzke et al. 2013). Yet, it is still common practice to summarise reaction time distributions using a single number, most often the mean: that one value for each participant and each condition can then be entered into a group ANOVA to make statistical inferences. Because of the skewness of reaction times, the mean is however a poor measure of central tendency: skewness shifts the mean away from the bulk of the distribution, an effect that can be amplified by the presence of outliers or a thick right tail. For instance, in the figure below, the median better represents the typical observation than the mean because it is closer to the bulky part of the distribution. Mean and median for a very right skewed distribution. The distribution is bounded to the left and has a long right tail. This is an ex-Gaussian distribution, which is popular to model the shape of reaction time distributions, even though the matching between its parameters and mental processes can be debated (Sternberg, 2014).

So the median appears to be a better choice than the mean if the goal is to have a single value to tell us about the location of most observations in a skewed distribution. The choice between the mean and the median is however more complicated. It could be argued that because the mean is sensitive to skewness, outliers and the thickness of the right tail, it is better able to capture changes in the shapes of the distributions among conditions. But the use of a single value to capture shape differences will lead to intractable analyses because the same mean could correspond to various shapes. Instead, a multiple quantile approach or explicit shape modelling should be used.

The mean and the median differ in another important aspect: for small sample sizes, the sample mean is unbiased, whereas the sample median is biased (see illustrations in previous post). Concretely, if we perform many times the same RT experiment, and for each experiment we compute the mean and the median, the average mean will be very close to the population mean. As the number of experiments increases, the average sample mean will converge to the exact population mean. This is not the case for the median when sample size is small.

The reason for this bias is explained by Miller (1988):

“Like all sample statistics, sample medians vary randomly (from sample to sample) around the true population median, with more random variation when the sample size is smaller. Because medians are determined by ranks rather than magnitudes of scores, the population percentiles of sample medians vary symmetrically around the desired value of 50%. For example, a sample median is just as likely to be the score at the 40th percentile in the population as the score at the 60th percentile. If the original distribution is positively skewed, this symmetry implies that the distribution of sample medians will also be positively skewed. Specifically, unusually large sample medians (e.g., 60th percentile) will be farther above the population median than unusually small sample medians (e.g., 40th percentile) will be below it. The average of all possible sample medians, then, will be larger than the true median, because sample medians less than the true value will not be small enough to balance out the sample medians greater than the true value. Naturally, the more the distribution is skewed, the greater will be the bias in the sample median.”

To illustrate the sample median’s bias, Miller (1988) employed 12 ex-Gaussian distributions that differ in skewness. The distributions are illustrated in the next figure, and colour-coded using the difference between the mean and the median as a non-parametric measure of skewness. Here are the parameters of the 12 distributions with the associated population parameters. The first figure in this post used the most skewed distribution of the 12, with parameters (300, 20, 300).

To estimate bias, we run a simulation in which we sample with replacement 10,000 times from each of the 12 distributions. We take random samples of sizes 4, 6, 8, 10, 15, 20, 25, 35, 50 and 100, as did Miller. For each random sample, we compute the mean and the median. For each sample size and ex-Gaussian parameter, the bias is then defined as the difference between the mean of the 10,000 sample estimates and the population value.

First, we check that the mean is not biased: Each line shows the results for one type of ex-Gaussian distribution: the mean of 10,000 simulations for different sample sizes. The grey area marks the 50% highest-density interval (HDI) of the 10,000 simulations for the least skewed distribution (the same interval is shown in the next two figures for comparison). The interval shows the variability across simulations and highlights an important aspect of the results: bias is a long-run property of an estimator; there is no guarantee that one value from a single experiment will be close to the population value. Also, the variability among samples increases with decreasing sample size, which is why results across small n experiments can differ substantially.

Here are the median bias estimates in table format: Columns = sample sizes; rows = skewness

The values are very close to the values reported in Miller 1998: An illustration is easier to grasp: As reported by Miller (1988), bias can be quite large and it gets worse with decreasing sample sizes and increasing skewness.

Based on these results, Miller made this recommendation:

“An important practical consequence of the bias in median reaction time is that sample medians must not be used to compare reaction times across experimental conditions when there are unequal numbers of trials in the conditions.”

According to Google Scholar, Miller (1988) has been cited 172 times. A look at some of the oldest and most recent citations reveals that his advice has been followed.

For instance, Lavie (1995) noted:

“In the following experiments I used the mean RTs for each participant rather than the medians, as the increase in number of go/no-go errors under the high-load conditions resulted in a different number of responses in between conditions (see Miller, 1988).”

Tipper et al. (1992):

Analysis by Miller (1988) shows that for large numbers of trials, differences in the numbers between conditions […] has no impact on the medians obtained.

More recently, Robinson et al. (2018):

“[…] comparing medians among conditions with an unequal number of trials may lead to false positives (Miller, 1988)”

Du et al. (2017):

“We chose the mean rather than median of RT as […] the sample median may provide a biased estimation of RT (Miller, 1988)”

The list goes on. Also, in a review paper, Whelan (2008), cited 324 times, reiterates the advice:

“The median should never be used on RT data to compare conditions with different numbers of trials.”

However, there are several problems with Miller’s advice. In particular, using the mean leads to many issues with estimation and statistical inferences, as we will see in the 3rd post. In this post, we tackle one key omission from Miller’s assessment: the bias of the sample median can be corrected, using a percentile bootstrap bias correction, as described in this previous post.

For each iteration in the simulation, bias correction was performed using 200 bootstrap samples. Here are the bias corrected results: The bias correction works very well on average, except for the smallest sample sizes. The failure of the bias correction for very small n is not surprising, because the shape of the sampling distribution cannot be properly estimated by the bootstrap from so few observations. From n = 10, the bias values are very close to those observed for the mean. So it seems that in the long-run, we can eliminate the bias of the sample median by using a simple bootstrap procedure. As we will see in the next post, the bootstrap bias correction is also effective when comparing two groups.

Update: 06 Feb 2018

Finally, let’s have a look at how well bias correction works as a function of sample size and skewness. The smaller the sample size, the less the bootstrap can estimate the sampling distribution and its bias. Ideally, after bias correction, we would expect the sample medians to be centred around zero, with limited variability. This ideal situation could look something like that: Here we’re considering the most skewed distribution with the smallest sample size. The x-axis shows the median bias before bias correction, whereas the y-axis shows the median bias after bias correction. In this ideal case, the correction works extremely well irrespective of the original bias (average bias is the red vertical line). As a result, the average bias after bias correction is very near zero (horizontal green line).

The reality is very different. The bias correction is only partially effective and is inhomogeneous. Let’s add some markup to help understand what’s going on. If the original bias is negative, after correction, the median tends to be even more negative, so over corrected in the wrong direction (lower left triangle).

If the original bias is positive, after correction, the median is either:
– over corrected in the right direction (lower right triangle)
– under corrected in the right direction (middle right triangle)
– over corrected in the wrong direction (upper right triangle)

This pattern remains, although attenuated, if we consider the largest sample size. Or if we consider the least skewed distribution. We can look at the different patterns as a function of sample size and skewness.

The figure below shows the probability of over correcting in the wrong direction given that the original bias is negative (lower left triangle of the marked up figure). In the ideal situation illustrated previously, the expected proportion of over correction in the wrong direction, given an original negative bias, is 6.7%. So here we clearly have an overrepresentation of these cases. When the original bias is negative, in most cases the bootstrap is unable to correct in the right direction. The situation gets worse with increasing skewness and smaller sample sizes.

The figure below shows the probability of under correcting in the right direction given that the original bias is positive (middle right triangle of the marked up figure) In the ideal situation illustrated previously, the expected proportion of under correction in the right direction, given an original positive bias, is 44.7 %. So here, we have an overrepresentation of these cases. When the original bias is positive, in too many cases the bootstrap corrects in the right direction, but it under-corrects. The situation gets worse with increasing skewness and smaller sample sizes.

Ok, so bias correction is imperfect and it varies a lot, it part depending on whether the sample median fell above or below the unknown population median. Think using the mean is safer? There are several strong arguments to the contrary – more in another post.

[GO TO POST 2/4]

# References

Doksum, K. (1974) Empirical Probability Plots and Statistical Inference for Nonlinear Models in the two-Sample Case. Annals of Statistics, 2, 267-277.

Du, Y., Valentini, N.C., Kim, M.J., Whitall, J. & Clark, J.E. (2017) Children and Adults Both Learn Motor Sequences Quickly, But Do So Differently. Frontiers in Psychology, 8.

Heathcote, A., Popiel, S.J. & Mewhort, D.J.K. (1991) Analysis of Response-Time Distributions – an Example Using the Stroop Task. Psychol Bull, 109, 340-347.

Lavie, N. (1995) Perceptual Load as a Necessary Condition for Selective Attention. J Exp Psychol Human, 21, 451-468.

Matzke, D., Love, J., Wiecki, T.V., Brown, S.D., Logan, G.D. & Wagenmakers, E.J. (2013) Release the BEESTS: Bayesian Estimation of Ex-Gaussian STop Signal reaction time distributions. Front Psychol, 4.

Miller, J. (1988) A warning about median reaction time. J Exp Psychol Hum Percept Perform, 14, 539-543.

Palmer, E.M., Horowitz, T.S., Torralba, A. & Wolfe, J.M. (2011) What Are the Shapes of Response Time Distributions in Visual Search? J Exp Psychol Human, 37, 58-71.

Pratte, M.S., Rouder, J.N., Morey, R.D. & Feng, C.N. (2010) Exploring the differences in distributional properties between Stroop and Simon effects using delta plots. Atten Percept Psycho, 72, 2013-2025.

Robinson, M. M., Clevenger, J., & Irwin, D. E. (2018). The action is in the task set, not in the action. Cognitive psychology, 100, 17-42.

Rouder, J.N., Lu, J., Speckman, P., Sun, D.H. & Jiang, Y. (2005) A hierarchical model for estimating response time distributions. Psychon B Rev, 12, 195-223.

Sternberg, S. (2014) Reaction times and the ex-Gaussian distribution: When is it appropriate? Retrieved from http://www.psych.upenn.edu//~saul/

Tipper, S.P., Lortie, C. & Baylis, G.C. (1992) Selective Reaching – Evidence for Action-Centered Attention. J Exp Psychol Human, 18, 891-905.

Whelan, R. (2008) Effective analysis of reaction time data. Psychol Rec, 58, 475-482.

# What can we learn from 10,000 experiments?

## The code and a notebook for this post are available on github.

Before we conduct an experiment, we decide on a number of trials per condition, and a number of participants, and then we hope that whatever we measure comes close to the population values we’re trying to estimate. In this post I’ll use a large dataset to illustrate how close we can get to the truth – at least in a lexical decision task. The data are from the French lexicon project:

Ferrand, L., New, B., Brysbaert, M., Keuleers, E., Bonin, P., Meot, A., Augustinova, M. & Pallier, C. (2010) The French Lexicon Project: lexical decision data for 38,840 French words and 38,840 pseudowords. Behav Res Methods, 42, 488-496.

After discarding participants who clearly did not pay attention, we have 967 participants who performed a word/non-word discrimination task. Each participant has about 1000 trials per condition. I only consider the reaction time (RT) data, but correct/incorrect data are also available. Here are RT distributions for 3 random participants (PXXX refers to their ID number in the dataset):   The distributions are positively skewed, as expected for RT data, and participants tend to be slower in the non-word condition compared to the word condition. Usually, a single number is used to summarise each individual RT distribution. From 1000 values to 1, that’s some serious data compression. (For alternative strategies to compare RT distributions, see for instance this paper). In psychology, the mean is often used, but here the median gives a better indication of the location of the typical observation. Let’s use both. Here is the distribution across participants of median RT for the word and non-word conditions: And the distribution of median differences: Interestingly, the differences between the medians of the two conditions is also skewed: that’s because the two distributions tend to differ in skewness.

We can do the same for the mean: Mean differences: With this large dataset, we can play a very useful game. Let’s pretend that the full dataset is our population that we’re trying to estimate. Across all trials and all participants, the population medians are:

Word = 679.5 ms
Non-word = 764 ms
Difference = 78.5

the population means are:

Word = 767.6 ms
Non-word = 853.2 ms
Difference = 85.5

Now, we can perform experiments by sampling with replacement from our large population. I think we can all agree that a typical experiment does not have 1,000 trials per condition, and certainly does not have almost 1,000 participants! So what happens if we perform experiments in which we collect smaller sample sizes? How close can we get to the truth?

# 10,000 experiments: random samples of participants only

In the first simulation, we use all the trials (1,000) available for each participant but vary how many participants we sample from 10 to 100, in steps of 10. For each sample size, we draw 10,000 random samples of participants from the population, and we compute the mean and the median. For consistency, we compute group means of individual means, and group medians of individual medians. We have 6 sets of results to consider: for the word condition, the non-word condition, their difference; for the mean and for the median. Here are the results for the group medians in the word condition. The vertical red line marks the population median for the word condition. With increasing sample size, we gain in precision: on average the result of each experiment is closer to the population value. With a small sample size, the result of a single experiment can be quite far from the population. This is not surprising, and also explains why estimates from small n experiments tend to disagree, especially between an original study and subsequent replications.

To get a clearer sense of how close the estimates are to the population value, it is useful to consider highest density intervals (HDI). A HDI is the shorted interval that contains a certain proportion of observations: it shows the location of the bulk of the observations. Here we consider 50% HDI: In keeping with the previous figure, the intervals get smaller with increasing sample size. The intervals are also asymmetric: the right side is always closer to the population value than the left side. That’s because the sampling distributions are asymmetric. This means that the typical experiment will tend to under-estimate the population value.

We observe a similar pattern for the non-word condition, with the addition of two peaks in the distribution from n = 40. I’m not sure what’s causing it, or if it is a common shape. One could check by analysing the lexicon project data from other countries. Any volunteers? One thing for sure: the two peaks are caused by the median, because they are absent from the mean distributions (see notebook on github). Finally, we can consider the distributions of group medians of median differences: Not surprisingly, it has a similar shape to the previous distributions. It shows a landscape of expected effects. It would be interesting to see where the results of other, smaller, experiments fall in that distribution. The distribution could also be used to plan experiments to achieve a certain level of precision. This is rather unusual given the common obsession for statistical power. But experiments should be predominantly about quantifying effects, so a legitimate concern is to determine how far we’re likely to be from the truth in a given experiment. Using our simulated distribution of experiments, we can determine the probability of the absolute difference between an experiment and the population value to be larger than some cut-offs. In the figure below, we look at cut-offs from 5 ms to 50 ms, in steps of 5. The probability of being wrong by at least 5 ms is more than 50% for all sample sizes. But 5 ms is probably not a difference folks studying lexical decision would worry about. It might be an important difference in other fields. 10 ms might be more relevant: I certainly remember a conference in which group differences of 10 ms between attention conditions were seriously discussed, and their implications for theories of attention considered. If we care about a resolution of at least 10 ms, n=30 still gives us 44% chance of being wrong…

(If someone is looking for a cool exercise: you could determine the probability of two random experiments to differ by at least certain amounts.)

We can also look at the 50% HDI of the median differences: As previously noted, the intervals shrink with increasing sample size, but are asymmetric, suggesting that the typical experiment will under-estimate the population difference. This asymmetry disappears for larger samples of group means: Anyway, these last figures really make me think that we shouldn’t make such a big deal out of any single experiment, given the uncertainty in our measurements.

# 10,000 experiments: random samples of trials and participants

In the previous simulation, we sampled participants with replacement, using all their 1,000 trials. Typical experiments use fewer trials. Let say I plan to collect data from 20 participants, with 100 trials per condition. What does the sampling distribution look like? Using our large dataset, we can sample trials and participants to find out. Again, for consistency, group means are computed from individual means, and group medians are computed from individual medians. Here are the sampling distributions of the mean and the median in the word condition: The vertical lines mark the population values. The means are larger than the medians because of the positive skewness of RT distributions.

In the non-word condition: And the difference distributions: Both distributions are slightly positively skewed, and their medians fall to the left of the population values. The spread of expected values is quite large.

Similarly to simulation 1, we can determine the probability of being wrong by at least 5 ms, which is 39%. For 10 ms, the probability is 28%, and for 20 ms 14%. But that’s assuming 20 participants and 100 trials per condition. A larger simulation could investigate many combinations of sample sizes…

Finally, we consider the 50% HDI, which are both asymmetric with respect to the population values and with similar lengths: # Conclusion

To answer the question from the title, I think the main things we can learn from the simulated sampling distributions above are:
– to embrace uncertainty
– modesty in our interpretations
– healthy skepticism about published research
One way to highlight uncertainty in published research would be to run simulations using descriptive statistics from the articles, to illustrate the range of potential outcomes that are compatible with the results. That would make for interesting stories I’m sure.