GENERAL COMMENTARY published: 02 October 2013 doi: 10.3389/fpsyg.2013.00700 Good things peak in pairs: a note on the bimodality coefficient Roland Pfister1* , Katharina A. Schwarz2 , Markus Janczyk1 , Rick Dale3 and Jonathan B. Freeman 4 1 Department of Psychology III, Institute of Psychology, Julius Maximilians University of Würzburg, Würzburg, Germany 2 University Medical Center Hamburg-Eppendorf, Hamburg, Germany 3 Cognitive and Information Sciences, University of California, Merced, Merced, CA, USA 4 Department of Psychological & Brain Sciences, Dartmouth College, Hanover, NH, USA *Correspondence: roland.pfi[email protected] Edited by: Holmes Finch, Ball State University, USA Keywords: distribution analysis, bimodality A commentary on Institute Inc, 1990, 2012; Knapp, 2007; “Miscellaneous Formulas” of the SAS Bimodal distribution, 2013; Freeman User’s Guide (SAS Institute Inc, 1990,p. Assessing bimodality to detect the pres- and Dale, 2013)—certainly a potential 561), the original formulation of the BC is ence of a dual cognitive process source of confusion among researchers by Freeman, J. B., and Dale, R. (2013). using the BC.1 Additionally, the Appendix m2 + 1 BC = 3 , Behav. Res. Methods 45, 83–97. doi: of Freeman and Dale (2013) gives a (n − 1)2 m4 + 3· ( − )( − ) 10.3758/s13428-012-0225-x slightly ambiguous formula for the BC n 2 n 3 because their approach used non-standard with m3 referring to the skewness of the MATLAB functions that are not widely distribution and m4 referring to its excess DISTRIBUTION ANALYSES AND accessible. The present article aims at clar- kurtosis (see Knapp, 2007, for critical BIMODALITY ifying and correcting these issues in an remarks about this notation), with both Distribution analyses are becoming attempt to prevent misunderstanding and moments being corrected for sample bias increasingly popular in the psychological confusion. Further, methodological issues (cf. Joanes and Gill, 1998). The BC of a literature as they promise invaluable infor- in using this measure are sketched to pro- given empirical distribution is then com- mation about hidden cognitive processes vide an intuition about its behavior. Note pared to a benchmark value of BC = (e.g., Ratcliff and Rouder, 1998; Ratcliff crit that the current paper does not intend to 5/9 ≈ 0.555 that would be expected for et al., 1999; Wagenmakers et al., 2005; argue in favor of the BC as compared to a uniform distribution; higher numbers Miller, 2006; Freeman and Dale, 2013). One other measures (see Freeman and Dale, point toward bimodality whereas lower particular approach probes distributions 2013, for a thorough comparison). Rather, numbers point toward unimodality. for uni- vs. bi-modality, because bimodal- we want to point out pitfalls and limita- Freeman and Dale (2013) gave infor- ity often results from the contribution of tions of this measure that can easily be mation about computation of the BC with dual processes underlying the observed data overlooked. Matlab, but unfortunately two problems (Larkin, 1979; Freeman and Dale, 2013;see likely arise from using their code (for more Knapp, 2007, for a historical overview). THE BC AND ITS CAVEATS information and examples of calculation Although several statistical tools for this The computation of the BC is easy and with different software packages, see the purpose exist, it remains unclear which one straightforward as it only requires three online material): First, the call can be considered as a gold standard for numbers: the sample size n,theskew- assessing bimodality in practice. ness of the distribution of interest, and m3 = skew(x); Freeman and Dale (2013) have recently its excess kurtosis2 (see DeCarlo, 1997, shed some light on the utility of three and Joanes and Gill, 1998, for a detailed likely results in an error, as skew() is not different measures of bimodality known description of the latter two statistics). a native Matlab function. The correct call as the bimodality coefficient (BC; SAS First appearing as part of the SAS pro- should be Institute Inc, 1990), Hartigan’sdip cedure CLUSTER under the headline m = skewness(x, 0); statistic (HDS; Hartigan and Hartigan, 3 1985), and Akaike’s information criterion where the second input parameter 0 1 (AIC; Akaike, 1974) as applied to one- The corresponding Wikipedia article (Bimodal distri- prompts the necessary correction for component and two-component Gaussian bution, 2013) used a wrong formula throughout, but has been corrected as part of preparing this article. sample bias. Secondly, kurtosis() com- mixture distribution models (McLachlan 2Excess kurtosis and Pearson’s original kurtosis dif- putes Pearson’s original kurtosis (The and Peel, 2000). Overall, their analyses fer only as to whether the distribution’s fourth scaled MathWorks Inc., 2012). To obtain the favored the HDS but also credited the BC moment is normalised to a value of 0 for normal dis- correct and sample-bias corrected value, with considerable utility. Notably, how- tributions or not (with “excess” indicating that a value of three has been subtracted for normalisation). The the call should be ever, rather different formulas for the present article assumes all statistics to represent excess BC can be found in the literature (SAS kurtosis if not explicitly indicated otherwise. m4 = kurtosis(x, 0) − 3; www.frontiersin.org October 2013 | Volume 4 | Article 700 | 1 Pfister et al. Bimodality coefficient in case of convergent results. Should the results not converge, it seems the best strategy to investigate distributions for other measures, such as skewness and kurtosis individually (as well as their appearance when inspected by eye), to determine whether the result of the BC might be biased in one or the other direction. ACKNOWLEDGMENTS WearegratefultoEdHuddlestonofthe SAS Institute Inc. for providing detailed information about the evolution of the BC. SUPPLEMENTARY MATERIAL The Supplementary Material for this article can be found online at: http://www.frontiersin.org/Quantitative_Ps ychology_and_Measurement/10.3389/fpsyg. 2013.00700/full FIGURE 1 | Histograms for four hypothetical distributions, their skewness (m3) and kurtosis (m4), as well as the corresponding BCs (values exceeding 0.555 are taken to indicate REFERENCES bimodality). Panel (A) shows a clearly unimodal distribution whereas the distribution in Panel (B) is Akaike, H. (1974). A new look at the statistical clearly bimodal. Both distributions are classified correctly by the BC. Panel (C) shows a skewed model identification. IEEE Trans. Autom. Control unimodal distribution that is classified erroneously as bimodal by the BC. The distribution in Panel 19, 716–723. doi: 10.1109/TAC.1974.1100705 (D) is correctly classified as bimodal, even though its BC is lower than that of distribution C. See Bimodal distribution. (2013). In Wikipedia.Available . the text for a detailed comparison of the distributions. online at: http://en wikipedia org/wiki/Bimodal distribution. [A correction to the listed formula for the BC has been submitted on February, 12, 2013 as part of writing this article, Retrieved: January 4, Irrespective of these computational issues, CONCLUSIONS 2013]. the above-mentioned formula reveals that As described above, empirical values of DeCarlo, L. T. (1997). On the meaning and use > of kurtosis. Psychol. Methods 2, 292–307. doi: the BC is directly influenced by both, BC 0.555 are taken to indicate bimodal- 10.1037/1082-989X.2.3.292 skewness and kurtosis: Higher BCs result ity. A probability density function for the Freeman, J. B., and Dale, R. (2013). Assessing from high absolute values of skewness BC, however, cannot be derived (Knapp, bimodality to detect the presence of a dual cog- and low or negative values of kurtosis. 2007). This is a major drawback because it nitive process. Behav. Res. Methods 45, 83–97. doi: 10.3758/s13428-012-0225-x Especially the influence of skewness can precludes a thorough null-hypothesis sig- Hartigan, J. A., and Hartigan, P. M. (1985). The dip result in undesired behavior of the BC. As nificance test. test of unimodality. Ann. Stat. 13, 70–84. doi: an illustration, four hypothetical distribu- A suitable alternative test for bimodal- 10.1214/aos/1176346577 tions of 100 values each (range 1–11) are ity is the dip test (Hartigan and Hartigan, Hartigan, P. M. (1985). Computation of the dip plotted in Figure 1, including their skew- 1985) that probes for deviations from statistic to test for unimodality. J. R. Stat. Soc. Ser. C (Applied Statistics) 34, 320–325. doi: ness, kurtosis, and the resulting BC (see unimodality (see also Freeman and Dale, 10.2307/2347485 Appendix for the raw data). 2013, for a more detailed description). Joanes, D. N., and Gill, C. A. (1998). Comparing mea- Comparing distribution A and B reveals An algorithm for this test was proposed sures of sample skewness and kurtosis. Statistician theexpectedbehavioroftheBC:The after its publication (Hartigan, 1985) 47, 183–189. doi: 10.1111/1467-9884.00122 two obvious modes in distribution B and this algorithm has meanwhile been Knapp, T. R. (2007). Bimodality revisited. J. Mod. Appl. Stat. Methods 6, 8–20. decrease kurtosis and increase the BC. adopted for MATLAB (Mechler, 2002). Larkin, R. P. (1979). An algorithm for assessing Distribution C, however, is clearly uni- Additionally, an up-to-date, bug-corrected bimodality vs. unimodality in a univariate distri- modal when inspected by eye but its heavy versionwasrecentlypublishedasanR bution. Behav. Res. Methods Instrum. 11, 467–468. skew also leads to a large BC. In terms package (diptest, Maechler, 2012). doi: 10.3758/BF03205709 of the BC, distribution C is even more A direct comparison of the BC and Maechler, M. (2012). diptest: Hartigan’s dip test statis- tic for unimodality – corrected code.Rpackagever- bimodal than distribution D even though the dip test (Freeman and Dale, 2013) sion 0.75-74. Available online at: http://CRAN.R- distribution D clearly has two modes, revealed that both measures have merit project.org/package=diptest. [Retrieved: January but otherwise both are very similar.