Near Confounding and Variance Inflation

Why Uneven Representation Costs Precision

When both biological conditions appear in each batch, a condition effect can be estimated by comparing conditions within batches and combining those comparisons. But the amount of information depends on how many observations of each condition are available in each batch. A batch with 19 control observations and 1 treated observation offers a much less precise within-batch comparison than a batch with 10 of each. If condition composition is nearly determined by batch, adjustment must rely on a small amount of within-batch information.

This loss of precision is called variance inflation. It does not mean that adjustment necessarily makes the estimate biased. Under a suitable model, adjustment removes batch differences from the condition comparison, while uneven representation increases the estimate’s sampling variance. The distinction matters: omitting batch may produce a numerically more precise but biased condition comparison when batch effects are associated with condition.

Model and Assumptions

Definition: Consider one gene’s expression response $y_i$ for observation $i$, binary condition indicator $z_i$ (0 for control and 1 for treated), and batch $b$. The additive fixed-batch model is $$ y_i=\mu+\tau z_i+\sum_{b=2}^{B}\gamma_b I(\text{observation }i\text{ is in batch }b)+\varepsilon_i. $$ Here $\mu$ is the control mean in the reference batch, $\tau$ is the common treated-minus-control condition effect, $\gamma_b$ is the difference between batch $b$ and the reference batch, and $I(\cdot)$ equals 1 when its statement is true and 0 otherwise. The usual variance calculation assumes independent errors with mean 0 and common variance $\sigma^2$, and assumes the additive model adequately describes the mean response.

For the formulas below, suppose each of the $B$ batches contains observations from both conditions. Let $n_{1b}$ and $n_{0b}$ be the treated and control counts in batch $b$, and let $n_b=n_{1b}+n_{0b}$. The total counts are $n_1=\sum_b n_{1b}$, $n_0=\sum_b n_{0b}$, and $N=n_1+n_0$. These are counts of the observations to which the model is applied. In single-cell data, treating every cell as an independent observation may be inappropriate when cells from the same biological sample are correlated; the independence assumption must match the analysis unit and design.

The key quantity is the within-batch information about condition:

S_{zz}=\sum_{b=1}^{B}\frac{n_{1b}n_{0b}}{n_b}.

A batch contributes more when both conditions are well represented. If one condition is rare in a batch, its contribution is small. For fixed $n_b$, the contribution is largest when the two condition counts are as equal as possible.

Deriving the Adjusted Variance

Within a batch, the treated-minus-control sample mean difference is $d_b=\bar{y}_{1b}-\bar{y}_{0b}$. Under the additive model, its expected value is $\tau$: the shared batch effect cancels in the subtraction. Under independent errors of variance $\sigma^2$, the variance of that difference is

\operatorname{Var}(d_b)=\sigma^2\left(\frac{1}{n_{1b}}+\frac{1}{n_{0b}}\right) =\frac{\sigma^2}{S_b}, \qquad S_b=\frac{n_{1b}n_{0b}}{n_b}.

Thus a within-batch difference with a larger $S_b$ has lower variance. Combining the batch differences with weights proportional to their information gives the adjusted condition estimate:

\hat{\tau} =\frac{\sum_{b=1}^{B}S_b d_b}{\sum_{b=1}^{B}S_b}, \qquad \operatorname{Var}(\hat{\tau})=\frac{\sigma^2}{\sum_{b=1}^{B}S_b} =\frac{\sigma^2}{S_{zz}}.

The variance follows directly because the batches’ errors are independent under the model. The variance of a weighted average with weights $S_b/S_{zz}$ is $\sum_b (S_b/S_{zz})^2(\sigma^2/S_b)=\sigma^2/S_{zz}$. If $\sigma^2$ is unknown, it is commonly replaced by the residual mean square from the fitted model. With all $B$ batches represented and both conditions present in each, the model has $B+1$ coefficients, so the residual degrees of freedom are $N-B-1$, and the estimated standard error is $\sqrt{\mathrm{MSE}/S_{zz}}$.

Worked Example: Combining Two Uneven Within-Batch Comparisons

Suppose hypothetical normalized expression measurements follow the stated additive model in two batches. Batch A has 36 controls and 4 treated observations, with sample means 2.0 and 2.4 expression units, respectively. Batch B has 4 controls and 36 treated observations, with means 3.0 and 3.2 units. Assume independent, equal-variance errors and, for this calculation, a known residual standard deviation $\sigma=0.5$ expression units.

The batch-specific information values are $S_A=(36)(4)/40=3.6$ and $S_B=(4)(36)/40=3.6$. The within-batch differences are $d_A=2.4-2.0=0.4$ units and $d_B=3.2-3.0=0.2$ units. Therefore,

\hat{\tau} =\frac{(3.6)(0.4)+(3.6)(0.2)}{3.6+3.6} =\frac{2.16}{7.2} =0.3000\text{ expression units}.

The estimate is a weighted average of the within-batch differences; here the weights are equal. Its standard error is $\sigma/\sqrt{S_{zz}}=0.5/\sqrt{7.2}=0.1863$ expression units. This describes sampling uncertainty under the assumed model and error variance. It does not establish that the observed condition difference is causal.

Quantifying Variance Inflation

A useful reference is the condition-only model, which includes an intercept and condition but omits batch. Its condition coefficient has variance $\sigma^2/(n_1n_0/N)$ under the same independent, equal-variance error calculation. That variance is a mathematical benchmark, not a recommendation to omit batch: if batch means differ and condition composition varies by batch, the condition-only estimate can be biased.

The variance inflation factor (VIF) compares the adjusted estimate’s variance with this condition-only variance benchmark:

\mathrm{VIF} =\frac{\sigma^2/S_{zz}}{\sigma^2/(n_1n_0/N)} =\frac{n_1n_0/N}{S_{zz}}.

The numerator $n_1n_0/N$ is the total condition information before accounting for batch; the denominator is the information that remains in within-batch comparisons. Equivalently, if $R^2$ is the proportion of variation in the binary condition indicator $z_i$ explained by batch indicators and an intercept, then

\mathrm{VIF}=\frac{1}{1-R^2}.

To see this, the total centered sum of squares of the binary condition indicator is $z^\mathsf{T}(I-P_1)z=n_1n_0/N$, where $P_1$ projects onto the intercept. Removing the variation explained by batch leaves $S_{zz}=z^\mathsf{T}(I-P_B)z$, where $P_B$ projects onto the batch indicators. Hence $S_{zz}=(1-R^2)n_1n_0/N$, giving the stated ratio. The standard error is multiplied by $\sqrt{\mathrm{VIF}}$, not by VIF itself.

Worked Example: Calculating the Precision Penalty

Consider two hypothetical batches of 40 observations each. In batch A there are 24 controls and 16 treated observations; in batch B there are 16 controls and 24 treated observations. Assume the additive model and independent errors with a common variance.

Each batch contributes $S_b=(24)(16)/40=9.6$, so $S_{zz}=19.2$. Across the study there are 40 controls and 40 treated observations, giving the condition-only benchmark $n_1n_0/N=(40)(40)/80=20$. Thus

\mathrm{VIF}=\frac{20}{19.2}=1.0417, \qquad \sqrt{\mathrm{VIF}}=1.0206.

Adjusting for batch therefore increases the variance by about 4.17% relative to the condition-only benchmark, and increases the standard error by about 2.06%. If the known residual standard deviation were 0.4 expression units, the adjusted standard error would be $0.4/\sqrt{19.2}=0.09129$ units, compared with $0.4/\sqrt{20}=0.08944$ units for the benchmark. The adjustment’s precision cost is modest because both batches have substantial representation of both conditions.

Worked Example: Improving Overlap Without Changing Total Condition Counts

Suppose a hypothetical study has 80 observations across two batches of 40. Initially, batch A has 38 controls and 2 treated observations, while batch B has 2 controls and 38 treated observations. There are 40 observations per condition overall, but each batch provides little within-batch information: $S_{zz}=2(38)(2)/40=3.8$. The condition-only benchmark is $40(40)/80=20$, so the VIF is $20/3.8=5.263$, and the standard error is multiplied by $\sqrt{5.263}=2.294$.

Now consider a hypothetical allocation with the same total of 80 observations and the same 40 observations per condition: batch A has 28 treated and 12 controls, while batch B has 12 treated and 28 controls. Each batch contributes $(28)(12)/40=8.4$, giving $S_{zz}=16.8$. The VIF becomes $20/16.8=1.190$, and the standard-error multiplier becomes $\sqrt{1.190}=1.091$. If $\sigma=0.4$ expression units, the initial adjusted standard error is $0.4/\sqrt{3.8}=0.2052$ units, compared with $0.4/\sqrt{16.8}=0.09759$ units after the more overlapping allocation.

The overall condition totals did not change; the distribution across batches did. Under the model, the second allocation provides substantially more information for separating the condition difference from batch differences.

Diagnostics, Alternatives, and Limitations

The main design diagnostic for this calculation is the condition-by-batch count table. Calculate $S_{zz}$ and the VIF, and inspect the individual contributions $S_b$: a small total may be driven by one or more batches with very rare representation of one condition. A large VIF signals that the adjusted estimate will be imprecise relative to the condition-only benchmark. It does not by itself diagnose whether the additive model is correct, whether observations are independent, or whether the measurement scale is appropriate.

The weighted within-batch calculation and the fixed-batch regression are two ways to express the same additive comparison when the assumptions above hold. The regression formulation is convenient when there are additional modeled covariates; the within-batch form makes clear which comparisons supply information. In contrast, omitting batch can yield a smaller variance on paper, but it does not account for batch differences. Its condition estimate need not represent a batch-adjusted condition effect. A comparison restricted to one well-represented batch can avoid combining across batches, but uses less of the available data and targets the comparison within that batch rather than the combined additive estimate.

The variance formula depends on the stated model. Unequal error variances, correlated observations, or a condition effect that differs across batches can change the uncertainty calculation or the meaning of a single common $\tau$. In single-cell studies, many cells may come from the same biological sample; regarding them as independent can make standard errors too small. The formula quantifies the precision available under its assumptions, not a guarantee that the assumptions hold.

Common Mistakes

Using total counts alone. Equal numbers of treated and control observations overall do not ensure good adjusted precision. The allocation within each batch determines $S_{zz}$.
Multiplying the standard error by the VIF. The variance is multiplied by VIF; the standard error is multiplied by $\sqrt{\mathrm{VIF}}$.
Calling the condition-only benchmark an adjusted estimate. It omits batch and can be biased when condition and batch are associated. The VIF comparison concerns sampling variance, not validity.
Interpreting low precision as proof of no condition effect. A large standard error indicates limited information under the model; it does not establish that the true effect is zero.

Key takeaway: Under an additive batch model with independent, equal-variance errors, the adjusted condition estimate has variance $\sigma^2/S_{zz}$, where $S_{zz}=\sum_b n_{1b}n_{0b}/n_b$. Uneven condition representation reduces this within-batch information and increases uncertainty.

Check Your Understanding

Try the calculations first, then compare your reasoning with the solutions.

Two batches each contain 20 observations. Batch A has 15 controls and 5 treated observations; batch B has 5 controls and 15 treated observations. Calculate $S_{zz}$.
For the design in question 1, there are 20 observations per condition overall. Calculate the VIF and the standard-error multiplier.
In one batch, $n_{1b}=8$ and $n_{0b}=32$. Calculate that batch’s contribution to $S_{zz}$.
Why can omitting batch produce a smaller calculated standard error without providing a valid batch-adjusted condition estimate?

Solutions. (1) Each batch contributes $(15)(5)/20=3.75$, so $S_{zz}=3.75+3.75=7.50$. (2) The condition-only benchmark is $(20)(20)/40=10$; therefore VIF $=10/7.50=1.333$, and the standard-error multiplier is $\sqrt{1.333}=1.155$. (3) The contribution is $(8)(32)/(8+32)=256/40=6.400$. (4) The condition-only calculation does not spend precision estimating batch differences, but it also ignores them. If batch means differ and condition composition varies across batches, the resulting condition estimate can be biased; its smaller standard error does not make it equivalent to an adjusted estimate.