One Overdispersed Pattern, More Than One Cause
A gene’s counts can vary more across cells than a Poisson model with one fixed mean predicts. That extra variation is called overdispersion relative to the Poisson mean-variance relationship. It is a description of the observed distribution, not a diagnosis of its cause. Variation in technical capture can make cells yield different numbers of observed molecules, while biological heterogeneity can make cells differ in their underlying expression. Either mechanism can produce a count variance larger than the count mean.
The distinction matters because a statistical description of the counts does not by itself reveal the data-generating process. A negative-binomial distribution, for example, may describe counts generated by a mixture of Poisson rates, but the latent rate variation could represent technical differences, biological differences, or a combination. The calculations below make this ambiguity explicit.
The interpretation of \(\Lambda\) depends on the mechanism being considered. In a technical model, it can be a biological mean multiplied by a cell-specific capture factor. In a biological model, it can be the cell’s own expression rate. These are different claims, even when the resulting count distributions have the same mean and variance.
Why a Random Rate Adds Variance
The key calculation uses the law of total variance: the variance of \(Y\) is the average conditional variance plus the variance of the conditional mean. For \(Y\mid\Lambda\sim\operatorname{Poisson}(\Lambda)\), both the conditional mean and the conditional variance equal \(\Lambda\). Therefore:
This derivation identifies two components. The term \(E(\Lambda)\) is the Poisson sampling variation remaining when the rate is specified. The term \(\operatorname{Var}(\Lambda)\) is added variation because rates differ across cells. The equation does not say what caused those rate differences; that interpretation requires assumptions and evidence beyond the marginal variance.
Worked Example: Variable Technical Capture
Worked Example: Variable Technical Capture
Suppose, hypothetically, that a gene has the same biological mean of 8 molecules per cell in a population of cells. Each cell has a dimensionless technical capture multiplier \(C\), independent of that gene’s biological expression, with \(E(C)=1\) and \(\operatorname{Var}(C)=0.25\). Conditional on capture, let the observed count satisfy \(Y\mid C\sim\operatorname{Poisson}(8C)\). Calculate its marginal mean and variance.
The conditional mean is \(E(Y\mid C)=8C\), and the conditional variance is also \(8C\). Averaging the conditional mean gives:
For the variance, the law of total variance gives:
Thus the expected count is 8 and its variance is 24 counts squared, three times the mean. Under this model the excess variance arises from cell-to-cell variation in technical capture, not from differences in the gene’s underlying biological mean. This conclusion depends on the stated assumptions; the counts alone do not establish that those assumptions are true.
The model’s assumptions are worth making explicit. The multiplier is nonnegative, has mean 1, and acts multiplicatively on the rate. Conditional on that multiplier, counts follow a Poisson distribution. The example treats the biological mean as fixed across cells and attributes the modeled rate variation to capture. In real data, those simplifying assumptions may not all hold.
Worked Example: Heterogeneous Biological Expression
Worked Example: Heterogeneous Biological Expression
Now suppose, hypothetically, that capture is constant but cells have different biological expression rates \(\Lambda\). Let \(Y\mid\Lambda\sim\operatorname{Poisson}(\Lambda)\), with \(E(\Lambda)=8\) and \(\operatorname{Var}(\Lambda)=16\). These assumptions describe nonnegative cell-specific rates followed by Poisson counting. Calculate the marginal mean and variance.
The expected count is the average biological rate:
The conditional Poisson variance contributes 8, and variation in biological rates contributes 16:
The mean and variance match the technical-capture example: 8 and 24. Here, however, the excess variance comes from biological differences in the cell-specific expression rates. If \(\Lambda\) has a gamma distribution with shape 4 and scale 2, its mean is \(4(2)=8\) and its variance is \(4(2^2)=16\). The resulting Poisson-gamma mixture is negative binomial with mean 8 and shape 4, illustrating a standard way biological rate heterogeneity can yield overdispersed counts.
The gamma distribution is an optional, more specific assumption in this example; the mean-variance calculation only requires the stated first two moments of \(\Lambda\). Neither the gamma assumption nor the variance calculation identifies whether a latent rate is biological in a real data set.
Matching Marginal Distributions Do Not Identify Mechanisms
The preceding examples have the same marginal mean and variance because both can be written as a Poisson count with a variable rate. The technical rate is \(8C\); the biological rate is \(\Lambda\). If \(C\) is gamma distributed with shape 4 and scale 0.25, then \(E(C)=1\) and \(\operatorname{Var}(C)=4(0.25^2)=0.25\). Consequently, \(8C\) has a gamma distribution with shape 4 and scale 2, exactly the rate distribution used in the biological example. Under these assumptions, the two models produce the same marginal count distribution, not merely the same first two moments.
This is a form of observational non-identifiability: one observed count distribution can be consistent with distinct accounts of how the rate varies. A good analysis therefore separates two questions: does a fixed-rate Poisson model fit adequately, and what evidence supports a particular source of rate variation? Overdispersion can support the first conclusion’s rejection, but does not answer the second question by itself.
Worked Example: Same Marginal Moments, Different Cross-Gene Pattern
Consider two genes with fixed biological means 4 and 6, respectively. Suppose one technical multiplier \(C\) is shared by both genes in each cell, with \(E(C)=1\) and \(\operatorname{Var}(C)=0.25\). Given \(C\), their counts \(Y_A\) and \(Y_B\) are independent Poisson variables with rates \(4C\) and \(6C\). Find their marginal variances and covariance.
For gene A, the mean is \(4E(C)=4\), and its variance is:
For gene B, the mean is \(6E(C)=6\), and its variance is:
Conditional independence does not imply marginal independence when both counts depend on the same random multiplier. The covariance is:
The shared capture variation predicts positive covariance between these genes under this model. By contrast, suppose the two genes have independent biological latent rates, each with mean and variance chosen to give marginal count means 4 and 6 and variances 8 and 15. Their counts can have the same marginal moments as above, but their covariance is zero under conditional Poisson sampling and independent rates. Thus cross-gene patterns may offer clues about a shared technical factor, but this contrast relies on assumptions about which rates vary together.
How to Compare Explanations in Practice
The following checks help organize interpretation without treating any one pattern as definitive proof of a mechanism:
- Check the marginal mean-variance pattern. Compare observed variability with the Poisson benchmark, while accounting for sampling uncertainty and the way cells and genes were selected. Excess variation is a signal to investigate, not a label for its cause.
- Ask whether a shared technical factor is plausible. A cell-wide capture multiplier can induce positive covariance across many genes. A pattern spanning genes with different biological roles may be consistent with such a factor, but shared biological states can also induce coordinated expression.
- Look for measured technical structure. If rate variation tracks independently recorded technical features, that supports investigating a technical explanation. Such associations are not automatically causal: technical features may be related to biological composition or other cell properties.
- Consider which features should co-vary under each proposed model. A model that gives every gene an independent biological rate has different joint predictions from one with a common cell-level multiplier. A model allowing coordinated biological programs can, in turn, produce cross-gene covariance.
The introductory material on using replicates to separate technical and biological noise motivates comparing independent sources of information. Here the central point is narrower: matching count moments does not settle the mechanism. Evidence about shared patterns, measured technical variables, and experimental design can make explanations more or less credible, but the conclusion remains conditional on the model.
| Feature | Variable technical capture | Heterogeneous biological expression |
|---|---|---|
| What varies between cells? | A technical multiplier on the observed rate | The underlying biological expression rate |
| Conditional count model in the examples | Poisson given the capture multiplier | Poisson given the biological rate |
| How does extra marginal variance arise? | Variation in the multiplier changes the conditional mean | Variation in the cell-specific expression rate changes the conditional mean |
| Possible joint pattern across genes | A shared multiplier can induce covariance | Independent rates need not induce covariance; coordinated rates can |
| What do overdispersed marginal counts establish? | They indicate more variation than the fixed-mean Poisson benchmark, but alone do not distinguish these explanations. | |
Common Mistake and Limitations
Wrong reasoning: “The variance is three times the mean, so the gene must be biologically heterogeneous.” Correction: The variance ratio shows overdispersion relative to a fixed-rate Poisson model, but the same mean and variance can arise from variable capture. A mechanism claim requires additional assumptions or evidence.
A second pitfall is to treat a negative-binomial fit as proof of a particular biological process. It is a useful count distribution for representing overdispersion, but the fit describes the observed counts under its assumptions; it does not by itself specify why latent rates differ. Conversely, a simple technical multiplier is not guaranteed to explain all extra variability. Technical and biological variation may coexist, and cell-level factors can affect multiple genes in related ways.
Moment comparisons also have limitations. With a finite number of cells, sample variance is uncertain, and a few unusual observations can strongly affect it. A variance greater than the sample mean is not, without an uncertainty assessment and an appropriate design, a formal proof that the population is overdispersed. Even a well-supported rejection of the fixed-mean Poisson model does not identify the source of the departure.
Check Your Understanding
Try each question before reading its solution.
- A hypothetical Poisson-mixture count has \(E(\Lambda)=5\) and \(\operatorname{Var}(\Lambda)=3\). What are the marginal mean and variance, and is it overdispersed relative to a fixed-rate Poisson count with the same mean?
- A hypothetical technical multiplier has mean 1 and variance 0.1, and the fixed biological mean is 10. Calculate the marginal count mean and variance.
- Two proposed models have the same marginal count mean and variance. Does that establish that they have the same data-generating mechanism? Explain.
- A shared random capture multiplier affects two genes with fixed means \(\mu_A\) and \(\mu_B\). Under conditional independence given that multiplier, what expression gives the marginal covariance?
Solutions. For 1, \(E(Y)=E(\Lambda)=5\), and \(\operatorname{Var}(Y)=5+3=8\). Since 8 exceeds 5, the mixture is overdispersed relative to the fixed-rate Poisson benchmark. For 2, \(E(Y)=10(1)=10\), and \(\operatorname{Var}(Y)=10+10^2(0.1)=20\). For 3, no. Different latent mechanisms can produce identical marginal moments, and even an identical marginal distribution, so the moments alone do not identify the mechanism. For 4, if the multiplier is \(C\), the covariance is \(\mu_A\mu_B\operatorname{Var}(C)\): conditional covariance is zero, while covariance of the conditional means \(\mu_A C\) and \(\mu_B C\) equals that expression.