Author Archives: garstats

About garstats

Senior Lecturer at the University of Glasgow, in the Institute of Neuroscience and Psychology. I study perception, ageing, and brain dynamics using EEG. Section editor at the European Journal of Neuroscience. I'm not a statistician but when I review and edit papers, I mostly comment on statistics. So here it goes.

Power estimation for correlation analyses

Following the previous posts on small n correlations [post 1][post 2][post 3], in this post we’re going to consider power estimation (if you do not care about power, but you’d rather focus on estimation, this post is for you). 

To get started, let’s look at examples of n=1000 samples from bivariate populations with known correlations (rho), with rho increasing from 0.1 to 0.9 in steps of 0.1. For each rho, we draw a random sample and plot Y as a function of X. The variances of the two correlated variables are independent – there is homoscedasticity. Later we will look at heteroscedasticity, when the variance of Y varies with X.

demo_homo_dist

For the same distributions illustrated in the previous figure, we compute the proportion of positive Pearson’s correlation tests for different sample sizes. This gives us power curves (here based on simulations with 50,000 samples). We also include rho = 0 to determine the proportion of false positives.

figure_power_homo

Power increases with sample size and with rho. When rho = 0, the proportion of positive tests is the proportion of false positives. It should be around 0.05 for a test with alpha = 0.05. This is the case here, as Pearson’s correlation is well behaved for bivariate normal data.

For a given expected population correlation and a desired long run power value, we can use interpolation to find out the matching sample size.

To achieve at least 80% power given an expected population rho of 0.4, the minimum sample size is 46 observations.

To achieve at least 90% power given an expected population rho of 0.3, the minimum sample size is 118 observations.

figure_power_homo_arrows

Alternatively, for a given sample size and a desired power, we can determine the minimum effect size we can hope to detect. For instance, given n = 40 and a desired power of at least 90%, the minimum effect size we can detect is 0.49.

So far, we have only considered situations where we sample from bivariate normal distributions. However, Wilcox (2012 p. 444-445) describes 6 aspects of data that affect Pearson’s r:

  • outliers

  • the magnitude of the slope around which points are clustered

  • curvature

  • the magnitude of the residuals

  • restriction of range

  • heteroscedasticity

The effect of outliers on Pearson’s and Spearman’s correlations is described in detail in Pernet et al. (2012) and Rousselet et al. (2012).

Next we focus on heteroscedasticity. Let’s look at Wilcox’s heteroscedasticity example (2012, p. 445). If we correlate variable X with variable Y, heteroscedasticity means that the variance of Y depends on X. Wilcox considers this example:

X and Y have normal distributions with both means equal to zero. […] X and Y have variance 1 unless |X|>0.5, in which case Y has standard deviation |X|.”

Here is an example of such data:

demo_wilcox_dist

Next, Wilcox (2012) considers the effect of this heteroscedastic situation on false positives. We superimpose results for the homoscedastic case for comparison. In the homoscedastic case, as expected for a test with alpha = 0.05, the proportion of false positives is very close to 0.05 at every sample size. In the heteroscedastic case, instead of 5%, the number of false positives is between 12% and 19%. The number of false positives actually increases with sample size! That’s because the standard T statistics associated with Pearson’s correlation assumes homoscedasticity, so the formula is incorrect when there is heteroscedasticity.

figure_power_hetero_wilcox

As a consequence, when Pearson’s test is positive, it doesn’t always imply the existence of a correlation. There could be dependence due to heteroscedasticity, in the absence of a correlation.

Let’s consider another heteroscedastic situation, in which the variance of Y increases linearly with X. This could correspond for instance to situations in which cognitive performance or income are correlated with age – we might expect the variance amongst participants to increase with age.

We keep rho constant at 0.4 and increase the maximum variance from 1 (homoscedastic case) to 9. That is, the variance of Y linear increases from 1 to the maximum variance as a function of X.

demo_hetero_dist

For rho = 0, we can compute the proportion of false positives as a function of both sample size and heteroscedasticity. In the next figure, variance refers to the maximum variance. 

figure_power_hetero_rho0

From 0.05 for the homoscedastic case (max variance = 1), the proportion of false positives increases to 0.07-0.08 for a max variance of 9. This relatively small increase in the number of false positives could have important consequences if 100’s of labs are engaged in fishing expeditions and they publish everything with p<0.05. However, it seems we shouldn’t worry much about linear heteroscedasticity as long as sample sizes are sufficiently large and we report estimates with appropriate confidence intervals. An easy way to build confidence intervals when there is heteroscedasticity is to use the percentile bootstrap (see Pernet et al. 2012 for illustrations and Matlab code).

Finally, we can run the same simulation for rho = 0.4. Power progressively decreases with increasing heteroscedasticity. Put another way, with larger heteroscedasticity, larger sample sizes are needed to achieve the same power.

figure_power_hetero_rho04

We can zoom in:

figure_power_hetero_rho04_zoom

The vertical bars mark approximately a 13 observation increase to keep power at 0.8 between a max variance of 0 and 9. This decrease in power can be avoided by using the percentile bootstrap or robust correlation techniques, or both (Wilcox, 2012).

Conclusion

The results presented in this post are based on simulations. You could also use a sample size calculator for correlation analyses – for instance this one.

But running simulations has huge advantages. For instance, you can compare multiple estimators of association in various situations. In a simulation, you can also include as much information as you have about your target populations. For instance, if you want to correlate brain measurements with response times, there might be large datasets you could use to perform data-driven simulations (e.g. UK biobank), or you could estimate the shape of the sampling distributions to draw samples from appropriate theoretical distributions (maybe a gamma distribution for brain measurements and an exGaussian distribution for response times).

Simulations also put you in charge, instead of relying on a black box, which most likely will only cover Pearson’s correlation in ideal conditions, and not robust alternatives when there are outliers or heteroscedasticity or other potential issues.

The R code to reproduce the simulations and the figures is on GitHub.

References

Pernet, C.R., Wilcox, R. & Rousselet, G.A. (2012) Robust correlation analyses: false positive and power validation using a new open source matlab toolbox. Front Psychol, 3, 606.

Rousselet, G.A. & Pernet, C.R. (2012) Improving standards in brain-behavior correlation analyses. Frontiers in human neuroscience, 6, 119.

Wilcox, R.R. (2012) Introduction to robust estimation and hypothesis testing. Academic Press, San Diego, CA.

Advertisements

Small n correlations + p values = disaster

Previously, we saw that with small sample sizes, correlation estimation is very uncertain, which implies that small n correlations cannot be trusted: the observed value in any experiment could be very far from the population value, and the sign could be wrong too. In addition to the uncertainty associated with small sample sizes, the selective report of results based on p values < 0.05 (or some other threshold), can lead to massively inflated correlation estimates in the literature (Yarkoni, 2009 ☜ if you haven’t done so, you really should read this excellent paper).

Let’s illustrate the problem (code is on GitHub). First, we consider a population rho = 0. Here is the sampling distribution as a function of sample size, as we saw in an earlier post. 

figure_rpval_ori

Figure 1: Sampling distribution for rho=0.

Now, here is the sampling distribution conditional on p < 0.05. The estimates are massively inflated and the problem gets worse with smaller sample sizes, because the smaller the sample size, the larger the correlations must be by chance for them to be significant.

figure_rpval_cond

Figure 2: Sampling distribution for rho=0, given p<0.05

So no, don’t get too excited when you see a statistically significant correlation in a paper…

Let’s do the same exercise when the population correlation is relatively large. With rho = 0.4, the sampling distribution looks like this:

figure_rpval_ori_04

Figure 3: Sampling distribution for rho=0.4.

If we report only those correlations associated with p < 0.05, the distribution looks like this:

figure_rpval_cond_04

Figure 4: Sampling distribution for rho=0.4, given p<0.05

Again, with small sample sizes, the estimates are inflated, albeit in the correct direction. There is nevertheless a small number of large negative correlations (see small purple bump around -0.6 -0.8). Indeed, in 0.77% of simulations, even though the population value was 0.4, a large and p < 0.05 negative correlation was obtained.

Correlations in neuroscience: are small n, interaction fallacies, lack of illustrations and confidence intervals the norm?

As reviewer, editor and reader of research articles, I’m regularly annoyed by the low standards in correlation analyses. In my experience with such articles, typically:

  • Pearson’s correlation, a non-robust measure of association, is used;
  • R and p values are reported, but not confidence intervals;
  • sample sizes tend to be small, leading to large estimation bias and inflated effect sizes in the literature;
  • R values and confidence intervals are not considered when interpreting the results;
  • instead, most analyses are reported as significant or non-significant (p<0.05), leading to the conclusion that an association exists or not (frequentist fallacy);
  • often figures illustrating the correlations are absent;
  • the explicit or implicit comparison of two correlations is done without a formal test (interaction fallacy).

To find out if my experience was in fact representative of the typical paper, I had a look at all papers published in 2017 in the European Journal of Neuroscience, where I’m a section editor. I care about the quality of the research published in EJN, so this is not an attempt at blaming a journal in particular, rather it’s a starting point to address a general problem. I really hope the results presented below will serve as a wake-up call for all involved and will lead to improvements in correlation analyses. Also, I bet if you look systematically at articles published in other neuroscience journals you’ll find the same problems. If you’re not convinced, go ahead, prove me wrong 😉 

I proceeded like this: for all 2017 articles (volumes 45 and 46), I searched for “correl” and I scanned for figures of scatterplots. If either searches were negative, the article was categorised as not containing a correlation analysis, so I might have missed a few. When at least one correlation was present, I looked for these details: 

  • n
  • estimator
  • confidence interval
  • R
  • p value
  • consideration of effect sizes
  • figure illustrating positive result
  • figure illustrating negative result
  • interaction test.

164 articles reported no correlation.

7 articles used regression analyses, with sample sizes as low as n=6, n=10, n=12 in 3 articles.

48 articles reported correlations.

Sample size

The norm was to not report degrees of freedom or sample size along with the correlation analyses or their illustrations. In 7 articles, the sample sizes were very difficult or impossible to guess. In the others, sample sizes varied a lot, both within and between articles. To confirm sample sizes, I counted the observations in scatterplots when they were available and not too crowded – this was a tedious job and I probably got some estimations and checks wrong. Anyway, I shouldn’t have to do all these checks, so something went wrong during the reviewing process. 

To simplify the presentation of the results, I collapsed the sample size estimates across articles. Here is the distribution: 

figure_ejn_sample_sizes

The figure omits 3 outliers with n= 836, 1397, 1407, all from the same article.

The median sample size is 18, which is far too low to provide sufficiently precise estimation.

Estimator

The issue with low sample sizes is made worse by the predominant use of Pearson’s correlation or the lack of consideration for the type of estimator. Indeed, 21 articles did not mention the estimator used at all, but presumably they used Pearson’s correlation.

Among the 27 articles that did mention which estimator was used:

  • 11 used only Pearson’s correlation;
  • 11 used only Spearman’s correlation;
  • 4 used Pearson’s and Spearman’s correlations;
  • 1 used Spearman’s and Kendall’s correlations.

So the majority of studies used an estimator that is well-known for its lack of robustness and its inaccurate confidence intervals and p values (Pernet, Wilcox & Rousselet, 2012).

R & p values

Most articles reported R and p values. Only 2 articles did not report R values. The same 2 articles also omitted p values, simply mentioning that the correlations were not significant. Another 3 articles did not report p values along with the R values.

Confidence interval

Only 3 articles reported confidence intervals, without mentioning how they were computed. 1 article reported percentile bootstrap confidence intervals for Pearson’s correlations, which is the recommended procedure for this estimator (Pernet, Wilcox & Rousselet, 2012).

Consideration for effect sizes

Given the lack of interest for measurement uncertainty demonstrated by the absence of confidence intervals in most articles, it is not surprising that only 5 articles mentioned the size of the correlation when presenting the results. All other articles simply reported the correlations as significant or not.

Illustrations

In contrast with the absence of confidence intervals and consideration for effect sizes, 23 articles reported illustrations for positive results. 4 articles reported only negative results, which leaves us with 21 articles that failed to illustrate the correlation results. 

Among the 40 articles that reported negative results, only 13 illustrated them, which suggests a strong bias towards positive results.

Interaction test

Finally, I looked for interaction fallacies (Nieuwenhuis, Forstmann & Wagenmakers 2011). In the context of correlation analyses, you commit an interaction fallacy when you present two correlations, one significant, the other not, implying that the 2 differ, but without explicitly testing the interaction. In other versions of the interaction fallacy, two significant correlations with the same sign are presented together, implying either that the 2 are similar, or that one is stronger than the other, without providing a confidence interval for the correlation difference. You can easily guess the other flavours… 

10 articles presented only one correlation, so there was no scope for the interaction fallacy. Among the 38 articles that presented more than one correlation, only one provided an explicit test for the comparison of 2 correlations. However, the authors omitted the explicit test for their next comparison!

Recommendations

In conclusion, at least in 2017 EJN articles, the norm is to estimate associations using small sample sizes and a non-robust estimator, to not provide confidence intervals and to not consider effect sizes and measurement uncertainty when presenting the results. Also, positive results are more likely to be illustrated than negative ones. Finally, interaction fallacies are mainstream.

How can we do a better job?

If you want to do a correlation analysis, consider your sample size carefully to assess statistical power and even better, your long-term estimation precision. If you have a small n, I wouldn’t even look at the correlation. 

Do not use Pearson’s correlation unless you have well-behaved and large samples, and you are only interested in linear relationships; otherwise explore robust measures of associations and techniques that provide valid confidence intervals (Pernet, Wilcox & Rousselet, 2012; Wilcox & Rousselet, 2018).

Reporting

These details are essential in articles reporting correlation analyses:

  • sample size for each correlation;
  • estimator of association;
  • R value;
  • confidence interval;
  • scatterplot illustration of every correlation, irrespective of the p value;
  • explicit comparison test of all correlations explicitly or implicitly compared;
  • consideration of effect sizes (R values) and their uncertainty (confidence intervals) in the interpretation of the results.

 Report p values if you want but they are not essential and should not be given a special status (McShane et al. 2018).

Finally, are you sure you really want to compute a correlation?

“Why then are correlation coefficients so attractive? Only bad reasons seem to come to mind. Worst of all, probably, is the absence of any need to think about units for either variable. Given two perfectly meaningless variables, one is reminded of their meaninglessness when a regression coefficient is given, since one wonders how to interpret its value. A correlation coefficient is less likely to bring up the unpleasant truth—we think we know what r = —.7 means. Do we? How often? Sweeping things under the rug is the enemy of good data analysis. Often, using the correlation coefficient is “sweeping under the rug” with a vengeance. Being so disinterested in our variables that we do not care about their units can hardly be desirable.”
Analyzing data: Sanctification or detective work?

John W. Tukey.
 American Psychologist, Vol 24(2), Feb 1969, 83-91. http://dx.doi.org/10.1037/h0027108

 

References

McShane, B.B., Gal, D., Gelman, A., Robert, C. & Tackett, J.L. (2018) Abandon Statistical Significance. arxiv.

Nieuwenhuis, S., Forstmann, B.U. & Wagenmakers, E.J. (2011) Erroneous analyses of interactions in neuroscience: a problem of significance. Nat Neurosci, 14, 1105-1107.

Pernet, C.R., Wilcox, R. & Rousselet, G.A. (2012) Robust correlation analyses: false positive and power validation using a new open source matlab toolbox. Front Psychol, 3, 606.

Rousselet, G.A. & Pernet, C.R. (2012) Improving standards in brain-behavior correlation analyses. Frontiers in human neuroscience, 6, 119.

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.

[preprint]

Small n correlations cannot be trusted

This post illustrates two important effects of sample size on the estimation of correlation coefficients: lower sample sizes are associated with increased variability and lower probability of replication. This is not specific to correlations, but here we’re going to have a detailed look at what it means when using the popular Pearson’s correlation (similar results are obtained using Spearman’s correlation, and the same problems arise with regression). The R code is available on github.


UPDATE: 2018-06-02

In the original post, I mentioned non-linearities in some of the figures. Jan Vanhove replied on Twitter that he was not getting any, and suggested a different code snippet. I’ve updated the simulations using his code, and now the non-linearities are gone! So thanks Jan!

Johannes Algermissen mentioned on Twitter that his recent paper covered similar issues. Have a look! He also reminded me about this recent paper that makes points very similar to those in this blog.

Gjalt-Jorn Peters mentioned on Twitter that “you can also use the Pearson distribution in package suppdists. Also see pwr.confintR to compute the required sample size for a given desired accuracy in parameter estimation (AIPE), which can also come in handy when planning studies”.

Wolfgang Viechtbauer‏ mentioned on Twitter “that one can just compute the density of r directly (no need to simulate). For example: link. Then everything is nice and smooth”.

UPDATE: 2018-06-30

Frank Harrell wrote on Twitter: “I’ll also push the use of precision of correlation coefficient estimates in justifying sample sizes. Need n > 300 to estimate r. BBR Chapter 8″


Let’s start with an example, shown in the figure below. Nice scatterplot isn’t it! Sample size is 30, and r is 0.703. It seems we have discovered a relatively strong association between variables 1 and 2: let’s submit to Nature or PPNAS! And pollute the literature with another effect that won’t replicate!

figure_random_correlation

Yep, the data in the scatterplot are due to chance. They were sampled from a population with zero correlation. I suspect a lot of published correlations might well fall into that category. Nothing new here, false positives and inflated effect sizes are a natural outcome of small n experiments, and the problem gets worse with questionable research practices and incentives to publish positive new results. 

To understand the problem with estimation from small n experiments, we can perform a simulation in which we draw samples of different sizes from a normal population with a known Pearson’s correlation (rho) of zero. The sampling distributions of the estimates of rho for different sample sizes look like this: 

figure_sampling_distributions

 

Sampling distributions tell us about the behaviour of a statistics in the long run, if we did many experiments. Here, with increasing sample sizes, the sampling distributions are narrower, which means that in the long run, we get more precise estimates. However, a typical article reports only one correlation estimate, which could be completely off. So what sample size should we use to get a precise estimate? The answer depends on:

  • the shape of the univariate and bivariate distributions (if outliers are common, consider robust methods);

  • the expected effect size (the larger the effect, the fewer trials are needed – see below);

  • the precision we want to afford.

For the sampling distributions in the previous figure, we can ask this question for each sample size:

What is the proportion of correlation estimates that are within +/- a certain number of units from the true population correlation? For instance:

  • for 70% of estimates to be within +/- 0.1 of the true correlation value (between -0.1 and 0.1), we need at least 109 observations;

  • for 90% of estimates to be within +/- 0.2 of the true correlation value (between -0.2 and 0.2), we need at least 70 observations. 

These values are illustrated in the next figure using black lines and arrows. The figure shows the proportion of estimates near the true value, for different sample sizes, and for different levels of precision. The bottom-line is that even if we’re willing to make imprecise measurements (up to 0.2 from the true value), we need a lot of observations to be precise enough and often enough in the long run.  

figure_precision

 

The estimation uncertainty associated with small sample sizes leads to another problem: effects are not likely to replicate. A successful replication can be defined in several ways. Here I won’t consider the relatively trivial case of finding a statistically significant (p<0.05) effect going in the same direction in two experiments. Instead, let’s consider how close two estimates are. We can determine, given a certain level of precision, the probability to observe similar effects in two consecutive experiments. In other words, we can find the probability that two measurements differ by at most a certain amount. Not surprisingly, the results follow the same pattern as those observed in the previous figure: the probability to replicate (y-axis) increases with sample size (x-axis) and with the uncertainty we’re willing to accept (see legend with colour coded difference conditions).  

 

figure_replication

In the figure above, the black lines indicates that for 80% of replications to be at most 0.2 apart, we need at least 83 observations.

So far, we have considered samples from a population with zero correlation, such that large correlations were due to chance. What happens when there is an effect? Let see what happens for a fixed sample size of 30, as illustrated in the next figure. 

figure_sampling_distributions_rho

 

As a sanity check, we can see that the modes of the sampling distributions progressively increase with increasing population correlations. More interestingly, the sampling distributions also get narrower with increasing effect sizes. As a consequence, the larger the true effect we’re trying to estimate, the more precise our estimations. Or put another way, for a given level of desired precision, we need fewer trials to estimate a true large effect. The next figure shows the proportion of estimates close to the true estimate, as a function of the population correlation, and for different levels of precision, given a sample size of 30 observations.

figure_precision_rho

 

Overall, in the long run, we can achieve more precise measurements more often if we’re studying true large effects. The exact values will depend on priors about expected effect sizes, shape of distributions and desired precision or achievable sample size. Let’s look in more detail at the sampling distributions for a generous rho = 0.4.

figure_sampling_distributions_rho04

 

The sampling distributions for n<50 appear to be negatively skewed, which means that in the long run, experiments might tend to give biased estimates of the population value; in particular, experiments with n=10 or n=20 are more likely than others to get the sign wrong (long left tail) and to overestimate the true value (distribution mode shifted to the right). From the same data, we can calculate the proportion of correlation estimates close to the true value, as a function of sample size and for different precision levels.

figure_precision_rho04

 

We get this approximate results:

  • for 70% of estimates to be within +/- 0.1 of the true correlation value (between 0.3 and 0.5), we need at least 78 observations;

  • for 90% of estimates to be within +/- 0.2 of the true correlation value (between 0.2 and 0.6), we need at least 50 observations. 

You could repeat this exercise using the R code to get estimates based on your own priors and the precision you want to afford.

Finally, we can look at the probability to observe similar effects in two consecutive experiments, for a given precision. In other words, what is the probability that two measurements differ by at most a certain amount? The next figure shows results for differences ranging from 0.05 (very precise) to 0.4 (very imprecise). The black arrow illustrates that for 80% of replications to be at most 0.2 apart, we need at least 59 observations.

figure_replication_rho04

 

We could do the same analyses presented in this post for power. However, I don’t really see the point of looking at power if the goal is to quantify an effect. The precision of our measurements and of our estimations should be a much stronger concern than the probability to flag any effect as statistically significant (McShane et al. 2018).

There is a lot more to say about correlation estimation and I would recommend in particular these papers from Ed Vul and Tal Yarkoni, from the voodoo correlation era. More recently, Schönbrodt & Perugini (2013) looked at the effect of sample size on correlation estimation, with a focus on precision, similarly to this post. Finally, this more general paper (Forstmeier, Wagemakers & Parker, 2016) about false positives is well worth reading.

Test-retest reliability assessment using graphical methods

UPDATE (2018-05-17): as explained in the now updated previous post, the shift function for pairwise differences, originally described as a great tool to assess test-retest reliability, is completely flawed. The approach using scatterplots remains valid. If you know of other graphical methods, please leave a comment.


Test-retest reliability is often summarised using a correlation coefficient, often without illustrating the raw data. This is a very bad idea given that the same correlation coefficient can result from many different configurations of observations. Graphical representations are thus essential to assess test-retest reliability, as demonstrated for instance in the work of Bland & Altman.

The R code for this post is on github.

Example 1: made up data

Let’s look at a first example using made up data. Imagine that reaction times were measured from 100 participants in two sessions. The medians of the two distributions do not differ much, but the shapes do differ a lot, similarly to the example covered in the previous post.

figure_kde

The kernel density estimates above do not reveal the pairwise associations between observations. This is better done using a scatterplot. In this plot, strong test-retest reliability would show up as a tight cloud of points along the unity line (the black diagonal line).

figure_scatter

Here the observations do not fall on the unity line: instead the relationship leads to a much shallower slope than expected if the test-retest reliability was high. For fast responses in session 1, responses tended to be slower in session 2. Conversely, for slow responses in condition 1, responses tended to be faster in condition 2. This pattern would be expected if there was regression to the mean [wikipedia][ Barnett et al. 2005], that is, particularly fast or particularly slow responses in session 1 were due in part to chance, such that responses from the same individuals in session 2 were closer to the group mean. Here we know this is the case because the data are made up to have that pattern.

We can also use a shift function for dependent group to investigate the relationship between sessions, as we did in the previous post.

figure_sf_dhd

The shift function reveals a characteristic  difference in spread between the two distributions, a pattern that is also expected if there is regression to the mean. Essentially, the shift function shows how  the distribution in session 2 needs to be modified to match the distribution in session 1: the lowest deciles need to be decreased and the highest deciles need to be increased, and these changes should be stronger as we move towards the tails of the distribution. For this example, these changes would be similar to an anti-clockwise rotation of the regression slope in the next figure, to align the cloud of observations with the black diagonal line.  

figure_scatter_regline

To confirm these observations, we also perform a shift function for pairwise differences. 

 

This second type of shift function reveals a pattern very similar to the previous one. In the [previous post], I wrote that this “is re-assuring. But there might be situations where the two versions differ.” Well, here are two such situations…

Example 2: ERP onsets

Here we look at ERP onsets from an object detection task (Bieniek et al. 2016). In that study, 74 of our 120 participants were tested twice, to assess the test-retest reliability of different measurements, including onsets. The distributions of onsets across participants is positively skewed, with a few participants with particularly early or late onsets. The distributions for the two sessions appear quite similar.   

figure_ERP_kde

With these data, we were particularly interested in the reliability of the left and right tails: if early onsets in session 1 were due to chance, we would expect session 2 estimates to be overall larger (shifted to the right); similarly, if late onsets in session 1 were due to chance, we would expect session 2 estimates to be overall smaller (shifted to the left). Plotting session 2 onsets as a function of session 1 onsets does not reveal a strong pattern of regression to the mean as we observed in example 1. 

figure_ERP_scatter1

Adding a loess regression line suggests there might actually be an overall clockwise rotation of the cloud of points relative to the black diagonal.

figure_ERP_scatter1_regline

The pattern is even more apparent if we plot the difference between sessions on the y axis. This is sometimes called a Bland & Altman plot (but here without the SD lines).

figure_ERP_scatter2_regline

However, a shift function on the marginals is relatively flat.

figure_ERP_sf_dhd

Although there seems to be a linear trend, the uncertainty around the differences between deciles is large. In the original paper, we wrote this conclusion (sorry for the awful frequentist statement, I won’t do it again):

“across the 74 participants tested twice, no significant differences were found between any of the onset deciles (Fig. 6C). This last result is important because it demonstrates that test–retest reliability does not depend on onset times. One could have imagined for instance that the earliest onsets might have been obtained by chance, so that a second test would be systematically biased towards longer onsets: our analysis suggests that this was not the case.”

That conclusion was probably wrong, because the shift function for dependent marginals is inappropriate to look at test-retest reliability. Inferences should be made on pairwise differences instead. So, if we use the shift function for pairwise differences, the results are very different! A much better diagnostic tool is to plot difference results as a function of session 1 results. This approach suggests, in our relatively small sample size:

 

  • the earlier the onsets in session 1, the more they increased in session 2, such that the difference between sessions became more negative;
  • the later the onsets in session 1, the more they decreased in session 2, such that the difference between sessions became more positive. 

This result and the discrepancy between the two types of shift functions is very interesting and can be explained by a simple principle: for dependent variables, the difference between 2 means is equal to the mean of the individual pairwise differences; however, this does not have to be the case for other estimators, such as quantiles (Wilcox & Rousselet, 2018).

Also, tThe discrepancy shows that I reached the wrong conclusion in a previous study because I used the wrong analysis. Of course, there is always the possibility that I’ve made a terrible coding mistake somewhere (that won’t be the first time – please let me know if you spot a fatal mistake). So l Let’s look at another example using published clinical data in which regression to the mean was suspected.

Example 3: Nambour skin cancer prevention trial

The data are from a cancer clinical trial described by Barnett et al. (2005). Here is Figure 3 from that paper:

barnett-ije-2005

“Scatter-plot of n = 96 paired and log-transformed betacarotene measurements showing change (log(follow-up) minus log(baseline)) against log(baseline) from the Nambour Skin Cancer Prevention Trial. The solid line represents perfect agreement (no change) and the dotted lines are fitted regression lines for the treatment and placebo groups”

Let’s try to make a similarly looking figure.

figure_nambour_scatter

Unfortunately, the original figure cannot be reproduced because the group membership has been mixed up in the shared dataset… So let’s merge the two groups and plot the data following our shift function convention, in which the difference is session 1 – session 2.

figure_nambour_scatter2

Regression to the mean is suggested by the large number of negative differences and the negative slope of the loess regression: participants with low results in session 1 tended to have higher results in session 2. This pattern can also be revealed by plotting session 2 as a function of session 1.

figure_nambour_scatter3

The shift function for marginals suggests increasing differences between session quantiles for increasing quantiles in session 1.

figure_nambour_sf_dhd

This result seems at odd with the previous plot, but it is easier to understand if we look at the kernel density estimates of the marginals. Thus, plotting difference scores as a function of session 1 scores probably remains the best strategy to have a fine-grained look at test-retest results.

figure_nambour_kde

A shift function for pairwise differences shows a very different pattern, consistent with the regression to the mean suggested by Barnett et al. (2005).

 

Conclusion

To assess test-retest reliability, it is very informative to use graphical representations, which can reveal interesting patterns that would be hidden in a correlation coefficient. Unfortunately, there doesn’t seem to be a magic tool to simultaneously illustrate and make inferences about test-retest reliability.

It seems that the shift function for pairwise differences is an excellent tool to look at test-retest reliability, and to spot patterns of regression to the mean. The next steps for the shift function for pairwise differences will be to perform some statistical validations for the frequentist version, and develop a Bayesian version.

That’s it for this post. If you use the shift function for pairwise differences to look at test-retest reliability, let me know and I’ll add a link here.

References

Barnett, A.G., van der Pols, J.C. & Dobson, A.J. (2005) Regression to the mean: what it is and how to deal with it. Int J Epidemiol, 34, 215-220.

Bland JM, Altman DG. (1986). Statistical methods for assessing agreement between two methods of clinical measurement. Lancet, i, 307-310.

Bieniek, M.M., Bennett, P.J., Sekuler, A.B. & Rousselet, G.A. (2016) A robust and representative lower bound on object processing speed in humans. The European journal of neuroscience, 44, 1804-1814.

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.

A new shift function for dependent groups?

UPDATE (2018-05-17): the method suggested here is completely bogus. I’ve edited the post to explain why. To make inferences about differences scores, use the difference asymmetry function or make inferences about the quantiles of the differences (Rousselet, Pernet & Wilcox, 2017).


The shift function is a graphical and inferential method that allows users to quantify how two distributions differ. It is a frequentist tool that also comes in several Bayesian flavours, and can be applied to independent and dependent groups. The version for dependent groups uses differences between the quantiles of each group. However, for paired observations, it would be also useful to assess the quantiles of the pairwise differences. This is what the this new shift function does was supposed to do.

Let’s consider the fictive reaction time data below, generated using exGaussian distributions (n = 100 participants).

figure_kde

The kernel density estimates suggest interesting differences: condition 1 is overall more spread out than condition 2; as a result, the two distributions differ in both the left (fast participants) and right (slow participants) tails. However, this plot does not reveal the pairwise nature of the observations. This is better illustrated using a scatterplot.

figure_scatter

The scatterplot reveals more clearly the relationship between conditions:
– fast participants, shown in dark blue on the left, tended to be a bit faster in condition 1 than in condition 2;
– slower participants, shown in yellow on the right, tended to be slower in condition 1 than in condition 2;
– this effect seems to be more prominent for participants with responses larger than about 500 ms, with a trend for larger differences with increasing average response latencies.

A shift function can help assess and quantify this pattern. In the shift function below, the x axis shows the deciles in condition 1. The y axis shows the differences between deciles from the two conditions. The difference is reported in the coloured label. The vertical lines show the 95% percentile bootstrap confidence intervals. As we travel from left to right along the x axis, we consider progressively slower participants in condition 1. These slower responses in condition 1 are associated with progressively faster responses in condition 2 (the difference condition 1 – condition 2 increases).

figure_sf_dhd

So here the inferences are made on differences between quantiles of the marginal distributions: for each distribution, we compute quantiles, and then subtract the quantiles.

What if we want to make inferences on the pairwise differences instead? This can be done by computing the quantiles of the differences, and plotting them as a function of the quantiles in one group. A small change in the code gives us a new shift function for dependent groups.

figure_sf_pdhd

The two versions look very similar, which is re-assuring, but does not demonstrate anything (except confirmation bias and wishful thinking on my part). But there might be situations where the two versions differ. Also, the second version makes explicit inferences about the pairwise differences, not about the differences between marginal distributions: so despite the similarities, they afford different conclusions.

Let’s look at the critical example that I should have considered before getting all excited and blogging about the “new method”. A simple negative control demonstrates what is wrong with the approach. Here are two dependent distributions, with a clear shift between the marginals.

figure_kde2

The pairwise relationships are better illustrated using a scatterplot, which shows a seemingly uniform shift between conditions.

figure_scatter2_1

Plotting the pairwise differences as a function of observations in condition 1 confirms the pattern: the differences don’t seem to vary much with the results in condition 1. In other words, differences don’t seem to be particularly larger or smaller for low results in condition 1 relative to high results.

figure_scatter2_2

The shift function on marginals does a great job at capturing the differences, showing a pattern characteristic of stochastic dominance (Speckman, Rouder, Morey & Pratte, 2008): one condition (condition 2) dominates the other at every decile. The differences also appear to be a bit larger for higher than lower deciles in condition 1.

figure_sf_dhd2

The modified shift function, shown next, makes no sense. That’s because the deciles of condition 1 and the deciles of the difference scores necessarily increase from 1 to 9, so plotting one as a function of the other ALWAYS gives a positive slope. The same positive slope I thought was capturing a pattern of regression to the mean! So I fooled myself because I was so eager to find a technique to quantify regression to the mean, and I only used examples that confirmed my expectations (confirmation bias)! This totally blinded me to what in retrospect is a very silly mistake.

figure_sf_pdhd2

Finally, let’s go back to the pattern observed in the previous shift function, where it seemed that the difference scores were increasing from low to high quantiles of condition 1. The presence of this pattern can better be tested using a technique that makes inferences about pairwise differences. One such technique is the difference asymmetry function. The idea from Wilcox (2012, Wilcox & Erceg-Hurn, 2012) goes like this: if two distributions are identical, then the difference scores should be symmetrically distributed around zero. To test for asymmetry, we can estimate sums of lower and higher quantiles; for instance, the sum of quantile 0.05 and quantile 0.95, 0.10 + 0.90, 0.15 + 0.85… For symmetric distributions with a median of zero, these sums should be close to zero, leading to a flat function centred at zero. If for instance the negative differences tend to be larger than the positive differences, the function will start with negative sums and will increase progressively towards zero (see example in Rousselet, Pernet & Wilcox). In our example, the difference asymmetry function is negative and flat, which is characteristic of a uniform shift, without much evidence for an asymmetry. Which is good because that’s how the fake data were generated! So using  graphical representations such as scatterplots, in conjunction with the shift function and the difference asymmetry function, can provide a very detailed and informative account of how two distributions differ.figure_daf2

Conclusion

I got very excited by the new approach because after spending several days thinking about test-retest reliability assessment from a graphical perspective, I thought I had found the perfect tool, as explained in the next post. So the ingredients of my mistake are clear: statistical sloppiness and confirmation bias.

The code for the figures in this post and for the new bogus shift function is available on github. I’ll will not update the rogme package, which implements the otherwise perfectly valid shift functions and difference asymmetry functions.

References

Speckman, P.L., Rouder, J.N., Morey, R.D. & Pratte, M.S. (2008) Delta plots and coherent distribution ordering. Am Stat, 62, 262-266.

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. [preprint] [reproducibility package]

Wilcox, R.R. (2012) Comparing Two Independent Groups Via a Quantile Generalization of the Wilcoxon-Mann-Whitney Test. Journal of Modern Applied Statistical Methods, 11, 296-302.

Wilcox, R.R. & Erceg-Hurn, D.M. (2012) Comparing two dependent groups via quantiles. J Appl Stat, 39, 2655-2664.

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:

figure_flp_w_100_kde

Same in the Non-Word condition:

figure_flp_nw_100_kde

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:

figure_flp_all_p_median

The same applies to distributions of means:

figure_flp_all_p_mean

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. 

figure_flp_bias_diff_md

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:

figure_flp_bias_diff_md_bc

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:

figure_flp_mdbias_diff_md

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?

figure_flp_bias_diff_m

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. 

figure_flp_all_p_diff

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.

figure_flp_mdbias_diff_m

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.