# Bimodal Data: Distribution Examples A bimodal distribution is a frequency distribution that has two distinct peaks, or modes, when the data are plotted as a histogram or a smooth density curve. Bimodal data therefore describe a single variable whose values cluster around two different centers rather than one. That single sentence carries more weight than it first appears. Most [statistical tests](/blog/guides/statistical-tests-choosing-the-right-one-for-your-data) taught in introductory courses assume one peak. When a variable actually has two, the mean sits in the valley between them, the standard deviation inflates, and any test that treats the sample as homogeneous can return a confident but wrong answer. Recognizing bimodality is often the first clue that a hidden grouping variable exists, such as sex, disease subtype, or a switch-like biological process. This guide covers what bimodality is, how to see it, how to test it, and how to avoid the two errors that plague student reports: calling a skewed curve bimodal, and calling a real second peak noise. ## What Bimodality Actually Means A distribution is bimodal when its probability density function has two local maxima separated by a local minimum. In a histogram, that appears as two humps with a dip between them. In a kernel density estimate, it appears as two smooth peaks. The key word is local. The second peak must rise above the surrounding density, not merely appear as a shoulder or a long tail. A common confusion is treating any two clusters of points as bimodality. Clusters are a property of the data points in a scatterplot or a clustering algorithm. Modes are a property of the density of a single variable. A dataset can form two tight clusters in two dimensions while the marginal distribution of each variable is perfectly unimodal. Conversely, a genuinely bimodal variable can look like one smear if the peaks overlap heavily. Bimodality is about the shape of the density along one axis, and it must be demonstrated on that axis. Two other terms get mixed into the discussion. A multimodal distribution has three or more peaks. A mixture distribution is a model that describes data as a weighted sum of two or more component distributions. A mixture of two normal distributions is often bimodal, but not always. If the two component means are close relative to their standard deviations, the sum is unimodal with a wide, flat top. Bimodality is an observable property of the density. A mixture is a generative model. They are related but not identical. | Term | What it describes | How you detect it | |--|--|--| | Bimodal distribution | Two local maxima in a density | Histogram, kernel density estimate, dip test | | Multimodal distribution | Three or more local maxima | Same tools, plus mode counting | | Mixture distribution | Data generated by two or more components | Gaussian mixture modeling, information criteria | | Skewed unimodal | One peak with a long tail | Density plot, skewness, dip test not significant | | Two clusters | Grouping of points in multivariate space | Clustering algorithms, not density shape | ## Why Bimodality Matters in Biology Biology produces bimodal data for a simple reason: many biological variables are controlled by switches, thresholds, or discrete categories. Sex is the clearest example. A sexually dimorphic species has males and females with different mean body masses, so a pooled sample of adult body mass is often bimodally distributed. The two peaks correspond to the two sexes, and the dip between them reflects the rarity of intermediate sizes. [Gene expression](/blog/guides/gene-expression) provides a second classic case. Many genes are effectively off or on. A population of single cells measured for one gene's transcript count often shows a large peak near zero and a second peak at a higher expression level. This on/off pattern is a bimodal distribution driven by transcriptional bursting and regulatory switching. The same logic applies to epigenetic marks, where a locus is either methylated or unmethylated in most cells. Response times and latency measures form a third family. In cognitive and educational testing, engaged responses and rapid-guessing responses have overlapping but distinct time distributions. When a substantial fraction of test takers guess quickly, the overall response time distribution becomes bimodal, with a fast mode and a slow mode [1]. A simulation study of rapid-guessing detection found that mixture lognormal methods were most robust precisely when bimodality was hardest to detect, which shows how much practical work depends on getting the shape right [1]. Clinical and epidemiological data add more examples. In a study of 659 consecutive pediatric orchiopexy patients, age at surgery was bimodal, with an early component centered near 1.31 years and a late component near 6.28 years [2]. The authors interpreted this as consistent with congenital and acquired subtypes of cryptorchidism, though they were careful to note that bimodality alone does not prove two distinct biological entities [2]. In a community sample, adjusted aldosterone levels in people with low-renin hypertension were bimodal, while the same measure was unimodal in normal-renin hypertension and in normotensive controls [3]. That contrast is a good illustration of how bimodality can mark a subgroup rather than a whole population. Bimodality also appears in environmental and molecular measurements. Ambient volatile organic compounds in Seoul showed a bimodal diurnal pattern, with concentration peaks during morning and evening rush hours, reflecting traffic emission cycles [4]. Peptide collision cross sections measured by ion mobility spectrometry are often bimodal, corresponding to extended and compact conformations in the gas phase [5]. In each case, the two peaks carry a mechanistic interpretation that a single summary statistic would hide. ## How to See Bimodality: Histograms and Density Plots The first tool is the histogram. Bin the variable, count observations per bin, and plot the counts. Two peaks with a clear dip between them is the visual signature. The bin width controls what you see. Bins that are too wide merge the peaks into one hump. Bins that are too narrow produce a spiky plot where every random fluctuation looks like a mode. There is no universal correct bin width, but a useful habit is to plot several widths and check whether the two-peak structure persists across a reasonable range. If it appears only at one arbitrary width, treat it as an artifact until proven otherwise. The second tool is the kernel density estimate. This replaces the histogram's blocky steps with a smooth curve by placing a small bump at each data point and summing the bumps. The width of that bump is the bandwidth, and it plays the same role as bin width. A small bandwidth produces a wiggly curve with spurious modes. A large bandwidth oversmooths and erases real structure. Any report of a bimodal curve from a kernel density estimate should state the bandwidth and the kernel used, because the conclusion depends on both. Sample size matters as much as smoothing. With 20 observations, a histogram is mostly noise, and two apparent peaks can easily arise by chance. With several hundred observations, a stable dip between two modes is meaningful. A practical rule is that visual assessment of modality becomes reasonably reliable in the low hundreds of observations for well-separated modes, and requires more data when the modes overlap. The pediatric orchiopexy study used 659 patients and confirmed the structure with bootstrap mode stability analysis, which is a good model for how much evidence a strong claim needs [2]. A third visual aid is the rug plot or the individual data points drawn beneath the density. It shows where the observations actually fall and prevents the illusion that a smooth curve contains more information than the data support. When the dip between two peaks contains very few points, the bimodality is well supported. When the dip is shallow and populated, the two modes may be an artifact of smoothing. ## Formal Tests: The Dip Test and the Silverman Test Eyeballing a histogram is not a statistical test. Two formal approaches dominate practice. Hartigan's dip test measures the maximum difference between the empirical distribution and the closest unimodal distribution. A small dip statistic means the data are consistent with one mode. A large dip statistic, with a small p value, means unimodality is rejected. The dip test is widely used because it makes no assumption about the shape of the components and works on the raw data rather than a smoothed curve. In the orchiopexy study, the dip statistic was 0.134 with p < 0.001, a decisive rejection of unimodality [2]. In the aldosterone study, the dip test gave p = 0.008 for the low-renin hypertension group, while the normal-renin and normotensive groups had p values near 0.9, showing unimodality was not rejected there [3]. A meta-analysis of immune parameters in psychosis used the same dip test to assess whether cytokine distributions were unimodal or not [6]. The Silverman test, sometimes called the bandwidth test, takes a different route. It asks how much you would have to smooth the data before the density becomes unimodal. If a small amount of smoothing removes the second peak, that peak was probably noise. If a large amount of smoothing is needed, the second peak is robust. The test produces a p value for the null hypothesis of one mode. A comparison of dip and bandwidth tests in flow cytometry found that both could distinguish unimodal from bimodal densities with low error rates, while criteria previously used in automated gating could not [7]. That study also noted that the significance level can be adjusted to trade off the risk of missing a second mode against the risk of seeing one that is not there [7]. | Method | What it tests | Main limitation | |--|--|--| | Histogram inspection | Visual presence of two peaks | Sensitive to bin width, subjective | | Kernel density estimate | Smooth estimate of density shape | Sensitive to bandwidth, can create or hide modes | | Hartigan dip test | Null hypothesis of unimodality against any alternative | Low power with small samples or heavily overlapping modes | | Silverman bandwidth test | Whether smoothing is needed to reach unimodality | Result depends on the kernel and critical bandwidth choice | | Gaussian mixture modeling | Whether a two-component model fits better than one | Model-based, can fit two components to skewed unimodal data | | Bootstrap mode stability | Whether the second mode persists across resamples | Computationally heavier, still descriptive | ## Distinguishing True Bimodality from Look-Alikes The most common error in student work is calling a right-skewed unimodal distribution bimodal. A lognormal or exponential distribution has a single peak near the low end and a long tail to the right. If the tail is long enough, random sampling can create a small secondary bump far out, and a narrow bin width can make that bump look like a second mode. The dip test is designed to catch this. A skewed unimodal distribution will usually fail to reject unimodality, while a true two-mode distribution will reject it. The second look-alike is a mixture of two normal distributions with different variances and similar means. If one component is tight and the other is broad, and their means are close, the sum can be unimodal even though the data were generated by two processes. This is the reverse error: assuming that a mixture model with two components implies a bimodal curve. It does not. A two-component Gaussian mixture can fit a skewed unimodal distribution perfectly well, which is why model selection criteria such as BIC should be paired with a density plot and a dip test rather than used alone. The orchiopexy study used all three approaches together: dip test, kernel density with bootstrap stability, and Gaussian mixture modeling with information criteria [2]. That combination is the standard to aim for. A third look-alike is a censored or truncated variable. If measurements below a detection limit are all recorded as zero, a spike at zero can create an apparent second mode that reflects the assay, not biology. Similarly, a variable with a hard floor or ceiling can produce an artificial peak at the boundary. Always check whether a mode sits exactly at a known detection limit, saturation point, or rounding value. A fourth issue is pooling. If you combine two groups that differ in mean, you can manufacture bimodality that does not exist within either group. This is Simpson's paradox in density form. Before interpreting a bimodal curve as evidence of two subtypes, check whether a known grouping variable, such as batch, site, sex, or treatment arm, explains the two peaks. The aldosterone finding is instructive here: bimodality appeared only in the low-renin hypertension subgroup and not in the other groups, which argues against a simple pooling artifact [3]. ## Step-by-Step: A Practical Workflow 1. Plot the raw data. Start with a histogram at several bin widths and a kernel density estimate at several bandwidths. Note whether the two-peak structure is stable. 2. Check the sample size. If you have fewer than roughly 100 observations, treat visual bimodality as a hypothesis, not a finding. 3. Look for known grouping variables. Color the histogram by sex, batch, site, or treatment. If the two peaks separate cleanly by group, you have a mixture of groups rather than a mysterious bimodal variable. 4. Run the dip test. Report the dip statistic and the p value. A non-significant result means you cannot reject unimodality. 5. Run the Silverman test or a bootstrap mode stability analysis. Report the bandwidth or the resampling scheme. 6. If you fit a mixture model, report the number of components, the selection criterion, and the component parameters. Do not present a mixture model as proof of bimodality without a density plot. 7. State the bandwidth and kernel for any kernel density estimate you publish. State the bin width for any histogram. 8. Interpret the two modes biologically. A bimodal distribution is a clue about mechanism, not a mechanism itself. ## Biological Examples in Detail Sexual dimorphism in body mass is the textbook case. In a species where adult males and females differ substantially in size, a pooled sample of body mass from adults will often be bimodal, with one peak for each sex. The dip between the peaks is populated by juveniles, subadults, or unusually small or large individuals, depending on how the sample was collected. If the sample includes only adults and the dimorphism is strong, the dip can be nearly empty. This is why body mass in dimorphic species is a standard teaching example of a bimodal distribution. Gene expression on/off states form the second major example. Single-cell RNA sequencing measures transcript counts per cell. For a gene that is actively regulated in a switch-like manner, the count distribution across cells often has a large peak at or near zero and a second peak at a higher level. The two modes correspond to the off and on states of the gene in individual cells. This pattern is common enough that it shapes how analysts model single-cell data, and it is one reason simple mean expression values can be misleading for switch-like genes. Latency and response time data form the third example. In timed tasks, a fast mode can reflect automatic or guessing responses, while a slow mode reflects engaged processing. The two distributions overlap, which is exactly the situation the rapid-guessing simulation study addressed [1]. The practical consequence is that a single mean response time can describe neither the fast nor the slow process well. Clinical biomarkers add more cases. Adjusted aldosterone in low-renin hypertension was bimodal in a community sample, suggesting a subgroup with higher aldosterone consistent with primary aldosteronism and a subgroup with lower levels [3]. Age at orchiopexy was bimodal, with early and late components that may correspond to congenital and acquired presentations [2]. In both cases, the bimodality is a hypothesis-generating observation about subtypes, and the authors were explicit that it does not by itself prove distinct disease entities [2]. Molecular and environmental measurements round out the list. Peptide collision cross sections are often bimodal because peptides adopt extended and compact conformations in the gas phase [5]. Ambient volatile organic compounds in Seoul showed a bimodal diurnal pattern tied to morning and evening traffic [4]. Neural fragility scores in simulated intracranial EEG data were bimodally distributed, with a threshold separating epileptogenic from non-epileptogenic nodes [8]. These examples show that bimodality is not confined to organismal traits. It appears wherever a system has two stable states, two dominant sources, or two subpopulations. ## Common Mistakes and Limitations The first mistake is reporting a bimodal distribution from a single histogram at one bin width. Always check stability across bin widths and, ideally, across a kernel density estimate with a stated bandwidth. The second mistake is confusing clusters with modes. Clustering algorithms partition points. Modes are features of a density along one variable. A two-cluster solution does not imply a bimodal marginal distribution, and a bimodal marginal does not imply two clean clusters in multivariate space. The third mistake is treating a two-component mixture model as proof of bimodality. A mixture model can fit a skewed unimodal distribution. Pair it with a dip test and a density plot. The fourth mistake is ignoring sample size. With small samples, the dip test has low power, and visual inspection is unreliable. With very large samples, trivial departures from unimodality can become statistically significant while being biologically meaningless. Report effect sizes and plots alongside p values. The fifth mistake is overinterpreting the dip. A statistically significant dip test tells you the data are unlikely under a unimodal model. It does not tell you the two modes correspond to two biological entities. That requires independent evidence, such as a known grouping variable or a mechanistic experiment. A genuine limitation of all these methods is that they assess the shape of the observed sample, not the shape of the population. Sampling bias, measurement error, and rounding can all create or erase modes. The dip test and the Silverman test are tools for quantifying uncertainty about shape, not instruments that reveal truth. Individual datasets always deserve a careful look at how the data were collected before the shape is interpreted. ## Quick Review - Bimodality means two local maxima in the density of one variable, not two clusters of points. - Histograms and kernel density estimates are the primary visual tools. Bin width and bandwidth must be reported. - The Hartigan dip test and the Silverman bandwidth test are the standard formal tests for unimodality. - A skewed unimodal distribution and a mixture of two normals with different variances are the two most common look-alikes. - Biological examples include body mass in sexually dimorphic species, on/off gene expression, and fast versus slow response times. - A significant dip test supports bimodality but does not prove two biological subtypes. - Always check whether a known grouping variable explains the two peaks before proposing a new mechanism. ## Frequently Asked Questions ### What is the difference between bimodal and bimodal data distribution? They mean the same thing. Bimodal describes a distribution with two peaks, and bimodal data distribution is simply the full phrase for that pattern. The shorter form is standard in statistics writing. ### Can a bimodal distribution have unequal peaks? Yes. The two modes can have very different heights and widths. Unequal peaks are common when one subgroup is larger than the other, such as a sample with more females than males in a dimorphic species. ### Is a bimodal distribution always a mixture of two normal distributions? No. Bimodality can arise from many shapes, including two lognormal components, two skewed components, or a single process with a threshold. Fitting two normal components is a modeling choice, not a requirement. ### How many observations do I need to detect bimodality? There is no fixed number, but visual assessment becomes reasonably reliable in the low hundreds of observations when the modes are well separated. Overlapping modes require more data and a formal test. ### What does a non-significant dip test mean? It means you cannot reject the hypothesis that the data come from a unimodal distribution. It does not prove unimodality, especially with small samples where the test has low power. ### Can bimodality appear in time series data? Yes. Ambient volatile organic compounds in Seoul showed a bimodal diurnal pattern with morning and evening peaks, reflecting traffic cycles [4]. In time series, the two modes often correspond to recurring daily or seasonal events. ## Related Articles - [How to Choose the Right Probability Distribution for Your Biological Data](/knowledge/bioinformatics/how-to-choose-the-right-probability-distribution-for-your-biological-data-a-decision-framework) - [Growth Curve Modeling for Biological Data](/knowledge/bioinformatics/growth-curve-modeling-for-biological-data-from-linear-to-nonlinear-trajectories) - [How to Perform a Bacterial Growth Curve: Protocol and Data Analysis](/knowledge/diagnostics/microbiology/bacterial-growth-curve-protocol) - [How to Build a Data Extraction Table for a Systematic Review](/blog/research-skills/how-to-build-a-data-extraction-table-for-a-systematic-review-a-template-with-life-science-examples) - [Qualitative Data Examples: How to Recognize and Use Non-Numerical Evidence](/blog/guides/qualitative-data-examples-how-to-recognize-and-use-non-numerical-evidence) - [RNA Location: Types, Functions, and Cellular Distribution](/knowledge/molecular-biology/rna-location) - [Bell Curve: Normal Distribution Explained](/blog/research-skills/bell-curve-normal-distribution-explained) - [Tabular Data: What It Is and How to Analyze It](/blog/research-skills/tabular-data-what-it-is-and-how-to-analyze-it) - [Poisson Distribution: Formula and Examples](/blog/research-skills/poisson-distribution-formula-and-examples) ## Sources 1. [Misclassification Produced by Rapid-Guessing Identification Methods and Their Suitability Under Various Conditions.](https://pubmed.ncbi.nlm.nih.gov/41743843/) 2. [Statistical Evidence for Bimodal Age Distribution in Pediatric Orchiopexy: Support for Congenital and Acquired Cryptorchidism Subtypes.](https://pubmed.ncbi.nlm.nih.gov/41510419/) 3. [Bimodal aldosterone distribution in low-renin hypertension.](https://pubmed.ncbi.nlm.nih.gov/23757402/) 4. [Spatial and Temporal Distribution Characteristics of VOCs in Seoul Ambient Air and Identification of Potential Pollution Sources Using Principal Component Analysis.](https://pubmed.ncbi.nlm.nih.gov/42515119/) 5. [Bimodal Peptide Collision Cross Section Distribution Reflects Two Stable Conformations in the Gas Phase.](https://pubmed.ncbi.nlm.nih.gov/42053382/) 6. [A Meta-analysis of Immune Parameters, Variability, and Assessment of Modal Distribution in Psychosis and Test of the Immune Subgroup Hypothesis.](https://pubmed.ncbi.nlm.nih.gov/30407606/) 7. [What is a "unimodal" cell population? Using statistical tests as criteria for unimodality in automated gating and quality control.](https://pubmed.ncbi.nlm.nih.gov/28759711/) 8. [Interictal Epileptogenic Zone Localization using Neural Fragility in Simulated Electroencephalogram Data().](https://pubmed.ncbi.nlm.nih.gov/41336485/)