When More Cells Stop Adding Much Information

Why Cell Counts Do Not Tell the Whole Story

A single-cell experiment may measure thousands of cells, but cells from the same donor are not generally independent observations of a population-level biological quantity. Cells from one donor share biological and technical features, so their measurements tend to be correlated. As a result, increasing the number of cells per donor improves the estimate of that donor’s cell-level mean, but does not provide the same amount of new information as sampling more independent donors.

The design effect makes this distinction quantitative. Under a balanced sampling model, it shows how much the variance of a grand mean is inflated by within-donor correlation relative to a hypothetical sample with the same total number of independent cells. The associated effective sample size is a variance-equivalent summary, not a replacement count of biological replicates.

Variance Decomposition and the Design Effect

Definition: Suppose \(D\) independent donors each contribute \(m\) cells, and let \(Y_{ij}\) be a measurement on cell \(j\) from donor \(i\). In the random-intercept model \(Y_{ij}=\mu+b_i+e_{ij}\), \(b_i\) is a donor effect with variance \(\tau^2\), and \(e_{ij}\) is a cell-level deviation with variance \(\sigma_e^2\). Assume the donor effects are independent across donors, the cell-level deviations are independent across cells and donors, and the two components are independent. The marginal variance is \(\sigma^2=\tau^2+\sigma_e^2\), and the within-donor intraclass correlation is \(\rho=\tau^2/\sigma^2\).

The target considered here is the mean \(\mu\) over the donor and cell populations represented by this model. Each donor contributes the same number of cells, and the grand mean is \(\bar{Y}=(1/D)\sum_{i=1}^{D}\bar{Y}_i\), where \(\bar{Y}_i=(1/m)\sum_{j=1}^{m}Y_{ij}\) is donor \(i\)’s cell mean. The calculations below treat the variance components as known; in practice, they are estimated, and uncertainty in those estimates can matter.

For one donor, averaging \(m\) cells reduces the cell-level component of variance, but not the shared donor component:

\operatorname{Var}(\bar{Y}_i)=\tau^2+\frac{\sigma_e^2}{m}.

Because the \(D\) donors are independent and equally weighted, the grand mean has variance \(\operatorname{Var}(\bar{Y})=\operatorname{Var}(\bar{Y}_i)/D\). Substituting the variance components and expressing them through \(\rho\) gives the variance decomposition:

\operatorname{Var}(\bar{Y}) =\frac{1}{D}\left(\tau^2+\frac{\sigma_e^2}{m}\right) =\frac{\sigma^2}{D}\left(\rho+\frac{1-\rho}{m}\right) =\frac{\sigma^2}{Dm}\left[1+(m-1)\rho\right].

The total number of cells is \(N=Dm\). If these \(N\) measurements were independent with marginal variance \(\sigma^2\), the variance of their mean would be \(\sigma^2/N\). The ratio of the actual variance to this independent-cell variance is the design effect.

Definition: For equal cluster size \(m\) and within-donor intraclass correlation \(\rho\), the design effect is \(DE=1+(m-1)\rho\). The variance-equivalent effective sample size is \(n_{\mathrm{eff}}=Dm/DE\), because \(\operatorname{Var}(\bar{Y})=\sigma^2/n_{\mathrm{eff}}\) under the model.

The effective sample size answers a limited, useful question: how many independent measurements with marginal variance \(\sigma^2\) would give the same variance for this grand mean? It does not mean that the experiment contains \(n_{\mathrm{eff}}\) independent donors. For inference about donor-to-donor variation, the number of independent donors remains \(D\).

Why the Precision Gain Diminishes

Holding \(D\), \(\tau^2\), and \(\sigma_e^2\) fixed, the variance can be written as a sum of a donor-level component and a cell-level component:

\operatorname{Var}(\bar{Y})=\frac{\tau^2}{D}+\frac{\sigma_e^2}{Dm}.

As \(m\) grows, the second term shrinks, while the first term remains. If \(\tau^2>0\), the limiting variance is \(\tau^2/D\), and the limiting standard error is \(\tau/\sqrt{D}\). More cells can approach this limit but cannot remove donor-to-donor variability when the number of donors is fixed.

The exact reduction in variance from increasing the per-donor count from \(m\) to \(m+1\) is:

\operatorname{Var}_m(\bar{Y})-\operatorname{Var}_{m+1}(\bar{Y}) =\frac{\sigma_e^2}{D}\left(\frac{1}{m}-\frac{1}{m+1}\right) =\frac{\sigma_e^2}{D\,m(m+1)}.

This reduction gets smaller as \(m\) increases. In the same model, when \(\rho>0\), the effective sample size tends to \(D/\rho\) as \(m\) tends to infinity. That finite limit expresses saturation in variance-equivalent cell information; it is not a limit on how many cells can be observed or a claim that further cells have no value.

Worked Example: Calculating the Design Effect

Suppose a hypothetical study samples \(D=20\) donors, with \(m=100\) cells per donor. Assume the within-donor correlation for the measurement of interest is \(\rho=0.10\). Calculate the design effect and effective sample size.

There are \(N=Dm=20(100)=2{,}000\) measured cells. Under the equal-size model, the design effect is:

DE=1+(m-1)\rho=1+(100-1)(0.10)=10.9.

Therefore the effective sample size is:

n_{\mathrm{eff}}=\frac{Dm}{DE}=\frac{2{,}000}{10.9}=183.5.

Under the stated model, the grand-mean variance is the same as it would be for about 183.5 independent measurements with the same marginal variance, rather than 2,000 independent measurements. The study still has 20 independent donors. The calculation does not make 183.5 donors available for donor-level inference.

Worked Example: Comparing Precision at Different Cell Counts

Consider a hypothetical study with \(D=12\) donors. For a measurement in standardized units, suppose the marginal variance is \(\sigma^2=1\) and the within-donor correlation is \(\rho=0.60\). Compare the grand-mean standard errors when each donor contributes \(m=5\), \(m=20\), or \(m=80\) cells. Assume the model and equal donor counts apply at all three sampling depths.

The variance formula is \(\operatorname{Var}(\bar{Y})=(\sigma^2/D)[\rho+(1-\rho)/m]\). At \(m=5\):

\operatorname{Var}(\bar{Y}) =\frac{1}{12}\left(0.60+\frac{0.40}{5}\right) =\frac{0.68}{12}=0.05667, \qquad SE=\sqrt{0.05667}=0.2380.

At \(m=20\):

\operatorname{Var}(\bar{Y}) =\frac{1}{12}\left(0.60+\frac{0.40}{20}\right) =\frac{0.62}{12}=0.05167, \qquad SE=\sqrt{0.05167}=0.2273.

At \(m=80\):

\operatorname{Var}(\bar{Y}) =\frac{1}{12}\left(0.60+\frac{0.40}{80}\right) =\frac{0.605}{12}=0.05042, \qquad SE=\sqrt{0.05042}=0.2245.

Increasing from 5 to 20 cells per donor reduces the standard error by \(0.2380-0.2273=0.0107\) standardized units. Increasing from 20 to 80 reduces it by only \(0.2273-0.2245=0.0028\) standardized units, despite adding three times as many cells as in the first increase. The variance floor is \(\tau^2/D=\rho\sigma^2/D=0.60/12=0.0500\), consistent with both computed variances approaching, but remaining above, that floor.

Worked Example: Same Total Cells, Different Donor Counts

Two hypothetical designs each measure 200 cells and have the same marginal variance \(\sigma^2=1\) and within-donor correlation \(\rho=0.30\). Design A samples \(D=10\) donors with \(m=20\) cells each; Design B samples \(D=20\) donors with \(m=10\) cells each. Assume donors are independent, both designs target the same population mean, and the model is valid.

For Design A, \(DE=1+(20-1)(0.30)=6.70\), so \(n_{\mathrm{eff}}=200/6.70=29.85\). Its grand-mean variance is \(1/29.85=0.03350\), and its standard error is \(\sqrt{0.03350}=0.1830\).

For Design B, \(DE=1+(10-1)(0.30)=3.70\), so \(n_{\mathrm{eff}}=200/3.70=54.05\). Its grand-mean variance is \(1/54.05=0.01850\), and its standard error is \(\sqrt{0.01850}=0.1360\).

In this example, distributing the same total number of cells across more independent donors yields a smaller grand-mean variance. This comparison concerns the stated balanced designs and model; it does not establish that one allocation is always preferable when recruitment costs, cell yields, or the scientific target differ.

Assumptions, Diagnostics, and Alternatives

The simple design-effect calculation is most directly applicable when each of \(D\) independent donors contributes the same number \(m\) of cells, cells share a common marginal variance, and their dependence is captured by a common within-donor correlation. It also assumes that the target and measurement are comparable across donors and that the random-intercept decomposition is a reasonable representation of the relevant variation. If correlation varies substantially across donors or cells, or if donor-specific responses differ in ways not represented by a shared intercept, this single-\(\rho\) formula may not describe the design accurately.

Useful checks include comparing observed within-donor similarity with the assumed correlation, examining whether estimated donor and cell-level variance components are plausible, and checking whether the sampling plan is actually balanced. A correlation estimate near zero makes the design effect close to one; a very large estimate makes additional within-donor cells contribute relatively little. Because correlation estimates can be uncertain, especially with few donors, it is prudent to assess how conclusions change over a scientifically plausible range of \(\rho\), rather than treating one estimate as exact.

The formula also has an important boundary: unequal cell counts across donors do not generally permit replacing \(m\) with the simple average count and applying the balanced formula without qualification. The grand mean’s weighting and variance then depend on the actual donor-specific counts and the analysis target. The calculations here compare balanced designs; they are not a procedure for correcting an imbalanced design.

The design effect is a compact variance comparison, not an analysis method that replaces donor-aware estimation. A donor-level analysis works with donor summaries, while a random-intercept model represents donor and cell-level variation directly. Both can respect the dependence structure when used with suitable assumptions. Treating every cell as independent is the contrasting approach: it uses the nominal cell count as though \(\rho=0\), which can underestimate variance when within-donor correlation is positive. The design-effect calculation helps explain the size and direction of that problem, but does not by itself handle covariates, unequal weights, or complex condition contrasts.

Common Mistake

Common mistake: “We measured 2,000 cells, so the sample size for biological inference is 2,000.” This reasoning treats correlated cells from a donor as independent biological replicates. Under the balanced random-intercept model, positive \(\rho\) inflates variance by \(1+(m-1)\rho\) relative to independent cells. Use the design effect to describe variance-equivalent information, and keep the distinction between cell count and independent donor count explicit.

A second overinterpretation is to say that once the variance approaches its floor, more cells are useless. The cell-level component continues to decline with added cells, and deeper sampling can still improve donor-specific estimates or support other measurement goals. The narrower conclusion is that, for this grand mean under this model and fixed \(D\), precision gains diminish and cannot remove the between-donor component.

Check Your Understanding

Use the balanced random-intercept model and show the calculation for each answer.

With \(m=10\) and \(\rho=0.20\), calculate the design effect.
For \(D=8\), \(m=25\), and \(\rho=0.08\), calculate the effective sample size to three significant figures.
For fixed \(D\), what is the limiting grand-mean variance as \(m\) grows, and what component of the model determines it?
Why is the effective sample size not the number of independent biological replicates?
If \(\rho=0\), what happens to the design effect and to the relationship between effective sample size and total cell count?

Solutions. (1) \(DE=1+(10-1)(0.20)=2.80\). (2) The total is \(Dm=8(25)=200\), and \(DE=1+(25-1)(0.08)=2.92\), so \(n_{\mathrm{eff}}=200/2.92=68.5\). (3) From \(\operatorname{Var}(\bar{Y})=\tau^2/D+\sigma_e^2/(Dm)\), the limit is \(\tau^2/D\), determined by the between-donor variance component. (4) It is defined by matching the variance of the grand mean to that from independent measurements; it does not change the fact that only \(D\) donors were independently sampled. (5) \(DE=1\), so \(n_{\mathrm{eff}}=Dm\): with zero within-donor correlation, the model gives the same grand-mean variance as \(Dm\) independent measurements.