Bridge Samples to Link Cohorts

Why Bridge Specimens Link Batches

When two cohorts are measured in separate batches, their observed values combine biological differences with measurement differences. A specimen measured in both batches provides a comparison in which the specimen’s underlying abundance is held fixed. That comparison can estimate a relative batch shift, which can then help place measurements from the cohorts on a common scale.

A bridge specimen is the same biological material measured in two or more batches. It may be a stored aliquot from a cohort, or another specimen that is deliberately included across batches. Its key contribution is not that it represents every cohort perfectly; it is that its underlying value does not change between the batch measurements. Repeated measurements of one bridge specimen are not, by themselves, evidence that a batch shift is constant across specimens. That constancy is a modeling assumption to assess.

Definition: For feature $g$, a bridge measurement is an observation of the same specimen $i$ in multiple batches. Under an additive batch model, write $$ y_{igb}=\theta_{ig}+\beta_{gb}+\epsilon_{igb}, $$ where $y_{igb}$ is the measured value for specimen $i$, feature $g$, and batch $b$; $\theta_{ig}$ is the specimen’s underlying feature value; $\beta_{gb}$ is the batch-specific shift for that feature; and $\epsilon_{igb}$ is measurement error. The target is usually a contrast $\beta_{gb}-\beta_{gb'}$, not an absolute batch effect.

The model is additive on the scale used for $y$. For example, an additive shift on a log2 scale represents a multiplicative change on the original scale. The model assumes that the shift for a feature is stable across the specimens to which it will be applied, and that bridge specimens are measured comparably in each batch. For standard paired uncertainty calculations, assume bridge-specimen differences are independent across specimens, have mean zero error, and are approximately normally distributed; the usual small-sample $t$ interval is exact if those differences are normally distributed. These assumptions are specific to the contrast calculation, not guaranteed by the fact that specimens were shared.

Estimating a Direct Batch Contrast

For a bridge specimen measured in batches $b$ and $b'$, subtract the two observations. Since the same $\theta_{ig}$ appears in both, it cancels:

y_{igb}-y_{igb'}=(\beta_{gb}-\beta_{gb'})+ (\epsilon_{igb}-\epsilon_{igb'}).

With $m$ bridge specimens, the average of these paired differences estimates the contrast. The pairing matters: it removes between-specimen variation from the contrast, while retaining the measurement-error variation in the within-specimen differences. Define $d_i=y_{igb}-y_{igb'}$, $\bar d=\sum_{i=1}^{m}d_i/m$, and let $s_d$ be the sample standard deviation of the $d_i$. The estimated standard error is $s_d/\sqrt{m}$. A two-sided $100(1-\alpha)\%$ confidence interval is $\bar d\pm t_{1-\alpha/2,m-1}s_d/\sqrt{m}$.

Worked Example: A Paired Estimate for Two Batches

Suppose four hypothetical bridge specimens are measured for one feature on a log2 scale. Batch 2 minus Batch 1 gives the following paired differences: $0.4,\ 0.6,\ 0.3,\ 0.7$ log2 units. Assume the four specimens are independent, the differences are approximately normal, and the additive shift is shared across the specimens. These conditions make a paired $t$ interval appropriate for this illustration.

The estimated contrast is $\bar d=(0.4+0.6+0.3+0.7)/4=2.0/4=0.5000$ log2 units. The deviations from the mean are $-0.1,\ 0.1,\ -0.2,\ 0.2$, so $s_d^2=[0.01+0.01+0.04+0.04]/(4-1)=0.03333$, and $s_d=\sqrt{0.03333}=0.1826$ log2 units. Thus $SE=0.1826/\sqrt{4}=0.09129$ log2 units.

For a 95% interval, $df=3$ and $t_{0.975,3}=3.182$. Therefore, $0.5000\pm3.182(0.09129)=0.5000\pm0.2905$, giving an interval of $[0.2095,\ 0.7905]$ log2 units. The estimated Batch 2 value is higher by 0.5000 log2 units for this feature. Under repeated sampling from the stated model, the interval procedure has 95% coverage; this interval does not mean there is a 95% probability that the fixed contrast lies inside this particular interval.

This estimate is for the direction Batch 2 minus Batch 1. Reversing the subtraction changes the sign of the estimate and both interval endpoints. Keeping that direction explicit prevents a sign error when applying a correction.

Combining Bridges Across a Batch Network

Not every bridge specimen needs to be measured in every batch. Some specimens may link batches 1 and 2, while others link batches 2 and 3. Represent batches as nodes and measured pairwise contrasts as edges. An edge from $b'$ to $b$ records an estimate of $\beta_{gb}-\beta_{gb'}$. To estimate offsets across the network, choose a reference batch, such as $\beta_{g1}=0$, and combine the edge estimates.

Let $\hat{\boldsymbol d}$ collect the observed edge contrasts, and let $A$ be the edge-by-batch contrast matrix after the reference batch is removed. Each row has a $+1$ for the destination batch and a $-1$ for the source batch, omitting the reference column. If edge errors are independent with known variances, let $W$ be diagonal with entries equal to their inverse variances. The weighted least-squares estimate minimizes the weighted squared disagreement between edge observations and the offsets it implies:

\hat{\boldsymbol\beta}_g =\arg\min_{\boldsymbol\beta_g} (\hat{\boldsymbol d}-A\boldsymbol\beta_g)^\mathsf{T} W(\hat{\boldsymbol d}-A\boldsymbol\beta_g).

Differentiating this quadratic objective and setting the derivative to zero gives the normal equations $A^\mathsf{T}WA\hat{\boldsymbol\beta}_g=A^\mathsf{T}W\hat{\boldsymbol d}$. When the batch network is connected and the reference is fixed, the solution is $(A^\mathsf{T}WA)^{-1}A^\mathsf{T}W\hat{\boldsymbol d}$. If the estimates share bridge specimens, their errors may be correlated; then use the full inverse covariance matrix in place of $W$, rather than treating the edges as independent. Connectivity is essential: if no path links a batch to the reference, its relative offset cannot be determined from these contrasts.

Worked Example: Combining Three Noisy Batch Contrasts

Consider hypothetical batches A, B, and C for one feature, with A chosen as the reference ($\beta_A=0$). Three independently estimated bridge contrasts are: B minus A, $0.4$ with $SE=0.1$; C minus B, $0.3$ with $SE=0.2$; and C minus A, $0.8$ with $SE=0.1$, all in log2 units. Assume the edge estimates are unbiased, their stated standard errors are valid, and their errors are independent. The direct A-to-C contrast does not exactly equal the sum of the other two, so weighted least squares reconciles the discrepancy.

Write $b=\beta_B$ and $c=\beta_C$. Inverse-variance weights are $100,\ 25,\ 100$. The objective is $100(b-0.4)^2+25(c-b-0.3)^2+100(c-0.8)^2$. Differentiating with respect to $b$ and $c$ yields $125b-25c=32.5$ and $-25b+125c=87.5$. Dividing by 25 gives $5b-c=1.3$ and $-b+5c=3.5$. Substituting $c=5b-1.3$ into the second equation gives $24b=10$, so $b=0.4167$ and $c=0.7833$ log2 units.

The fitted contrasts are therefore B minus A $=0.4167$, C minus B $=0.7833-0.4167=0.3667$, and C minus A $=0.7833$ log2 units. The precise A-to-C edge receives more weight than the less precise B-to-C edge, but all three observations contribute. These fitted values describe relative offsets, not absolute measurements or biological differences between cohorts.

Applying a Bridge Estimate to Separate Cohorts

A bridge can support a comparison of cohorts when it is measured in the batches containing those cohorts. For example, suppose three hypothetical bridge specimens have Batch 2 minus Batch 1 differences $0.4,\ 0.5,\ 0.6$ log2 units. The mean contrast is $(0.4+0.5+0.6)/3=0.5000$. If the mean for a cohort measured in Batch 1 is $8.0$ log2 units and the mean for a separate cohort measured in Batch 2 is $9.2$, subtracting the estimated Batch 2 offset gives an adjusted Batch 2 mean of $9.2-0.5=8.7$. The adjusted difference is $8.7-8.0=0.7$ log2 units.

This arithmetic is justified only if the bridge-derived offset applies to those cohort specimens for this feature. It does not prove that the remaining difference is a causal cohort effect, and it does not remove other sources of cohort differences. In particular, a shared bridge makes the batch contrast estimable under the model; it does not make an additive, specimen-invariant batch effect true by design.

Specify the measurement scale and contrast direction.
State the feature, the two batches being compared, and which batch is subtracted from which. The offset can depend on feature and scale.

Verify genuine specimen sharing.
Confirm that paired measurements are from the same biological material, with compatible processing and feature definitions. Similar but different specimens do not cancel the specimen value.

Estimate contrasts and uncertainty.
Use paired differences for direct links. For multiple links, account for their precision and any covariance from shared bridge specimens.

Check whether the bridge network connects the batches.
A path permits relative offsets to be compared; an isolated batch remains unlinked. Examine whether estimates depend heavily on one bridge or one edge.

Assess transfer to cohort specimens.
Look for feature-specific inconsistencies and unusually variable paired differences. A bridge estimate should not be applied automatically if the additive shift appears specimen-dependent.

Diagnostics, Alternatives, and Limitations

A useful diagnostic is to inspect paired differences by bridge specimen and, where possible, by bridge type or source cohort. Large variability, outliers, or systematic patterns suggest that a single additive contrast may be inadequate. With multiple paths between batches, compare the observed edge contrasts around each path: substantial disagreement may reflect noise, an incorrectly recorded pairing, or shifts that are not stable across specimens. A connected design can still be weakly informative if its links are based on very few or unusually variable bridge specimens.

The main alternatives answer different design problems. A technical replicate estimates repeatability under the conditions where it is repeated; it links batches only when the replicated material actually crosses those batches. A reference material can provide a standardized shared input, but its behavior and comparability to cohort specimens must be evaluated separately. Modeling batch as a fixed or random effect can use a broader design, but it cannot recover a batch contrast that the measurements do not identify. Bridge measurements add direct overlap; they do not replace design checks or justify adjustment beyond the model.

Approach	What it contributes	Main caution
Paired bridge contrast	Direct relative shift from the same specimens in two batches	Requires shared material and a transferable additive shift
Network of bridge contrasts	Relative offsets for connected batches, combining links by precision	Disconnected batches are not linked; shared bridges can induce correlated edge errors
Within-batch technical replicates	Information about repeatability under the replicate conditions	Do not estimate between-batch shifts unless they span batches
Model-based batch adjustment	Can represent batch effects alongside other modeled structure	Cannot resolve a contrast absent from the design or its assumptions

A bridge estimate is feature-specific: one feature’s offset does not establish another’s. It also has sampling uncertainty, which should be considered when interpreting adjusted cohort differences. If the bridge material is atypical, degraded, processed differently, or affected by a batch-by-specimen interaction, its contrast may fail to transfer. Finally, choosing a reference batch only fixes the arbitrary origin of the offsets; it does not change the estimated differences between batches.

Common Mistake

Wrong reasoning: “The bridge values in Batch 2 are 0.5 units higher, so every Batch 2 cohort value should be reduced by exactly 0.5, and the corrected cohorts are directly comparable.” The first clause may describe a point estimate, but it ignores its uncertainty and assumes the shift transfers to every relevant specimen. It also treats adjustment as proof that all remaining differences are biological.

Correction: Report the direction, scale, uncertainty, and assumptions of the bridge contrast. Apply it only when an additive, sufficiently stable shift is plausible for the feature and target specimens. Then describe the resulting comparison as adjusted under that model, not as a causal or assumption-free cohort difference.

Key takeaway: Shared specimens identify relative batch shifts because paired subtraction cancels the specimen’s underlying value. Reliable cohort linkage additionally requires a connected bridge design, valid uncertainty calculations, and a defensible assumption that each estimated shift transfers to the specimens being compared.

Check Your Understanding

Answer each question, then compare with the brief solutions below.

Three bridge differences, defined as Batch B minus Batch A, are $0.2,\ 0.4,\ 0.6$ log2 units. What is the estimated contrast?
If the Batch B minus Batch A estimate is $0.4$, what is the estimate for Batch A minus Batch B?
Batch A is linked to B and B is linked to C, but A and C have no direct bridge. Can their relative offsets be estimated under the additive model?
Why is it not valid to use measurements from different, merely similar specimens as paired bridge differences?
Two edge estimates use the same bridge specimens. What issue should be considered when combining them?

Solutions. (1) The mean is $(0.2+0.4+0.6)/3=0.4000$ log2 units. (2) Reversing the subtraction changes the sign, giving $-0.4$ log2 units. (3) Yes, if the A-to-B and B-to-C links are valid: the path connects all three, and the A-to-C contrast is the sum of the two fitted contrasts. (4) Different specimens generally have different underlying values, so subtraction would not cancel $\theta_{ig}$ and would mix specimen variation with the batch contrast. (5) The edge errors may be correlated; combining them as independent can misstate uncertainty, so their covariance should be included when available.

What You'll Learn