Why Repeated Measurements Matter
A measured omics value combines information about the specimen with variation introduced by measurement. If one specimen has a higher recorded abundance than another, the difference may reflect a real difference between specimens, technical noise, or both. Re-measuring the same specimen provides a way to estimate the technical component: the specimen’s underlying biological state is held fixed while the measurement is repeated.
A technical replicate is a repeated measurement of the same specimen, feature, and intended measurement target. A biological replicate is a different biological specimen. The distinction is essential: treating technical replicates as if they were independent subjects overstates the biological sample size and can make uncertainty look artificially small.
The measurement scale must also be stated. Variance on a log-abundance scale is not variance on the original abundance scale. The models below apply to a single feature at a time, on a chosen analysis scale, and to measurements that are comparable on that scale. They estimate repeatability; by themselves they do not calibrate platforms or establish that platforms measure the same biological quantity without bias.
A Basic Variance-Components Model
Suppose \(m\) specimens are each measured \(r\) times on one platform. A simple random-effects model is
Here \(Y_{ij}\) is the measurement for specimen \(i\) on replicate \(j\), \(\mu\) is the overall mean, \(b_i\) is the deviation for specimen \(i\), and \(\varepsilon_{ij}\) is technical error. Assume specimens are sampled from a population with \(b_i\sim N(0,\sigma_{\text{subject}}^2)\), and replicate errors are independent with \(\varepsilon_{ij}\sim N(0,\sigma_{\text{tech}}^2)\). Assume the subject effects and technical errors are mutually independent, the technical-error variance is common across specimens, and there are no systematic replicate-order or run effects left unmodelled. The same assumptions apply to the transformed measurement scale if a transformation is used.
Under this model, \(\sigma_{\text{subject}}^2\) describes between-subject variation and \(\sigma_{\text{tech}}^2\) describes within-specimen technical variation. The total marginal variance of one measurement is their sum. Because replicate measurements share \(b_i\), they are not independent marginally; their covariance is \(\sigma_{\text{subject}}^2\).
For a balanced design, let \(\bar{Y}_{i\cdot}\) be the mean of specimen \(i\)’s \(r\) measurements, and let \(\bar{Y}_{\cdot\cdot}\) be the grand mean. Define the within- and between-specimen sums of squares and mean squares as follows.
The method-of-moments estimates are \(\widehat{\sigma}_{\text{tech}}^2=MS_W\) and \(\widehat{\sigma}_{\text{subject}}^2=(MS_B-MS_W)/r\). The latter estimate can be negative from sampling variation; a negative variance is not meaningful, so it is commonly reported as zero when a nonnegative variance-component estimate is required. That boundary result signals weak information about between-subject variation, not proof that the true variation is exactly zero.
Why the Mean Squares Separate the Components
The within-specimen deviations remove the shared subject effect: \(Y_{ij}-\bar{Y}_{i\cdot}=\varepsilon_{ij}-\bar{\varepsilon}_{i\cdot}\). Their expected sum of squares is therefore \(m(r-1)\sigma_{\text{tech}}^2\), giving \(E(MS_W)=\sigma_{\text{tech}}^2\).
For the specimen means, \(\bar{Y}_{i\cdot}=\mu+b_i+\bar{\varepsilon}_{i\cdot}\). Their variance is \(\sigma_{\text{subject}}^2+\sigma_{\text{tech}}^2/r\). Multiplying the sum of squared deviations of these means by \(r\) yields \(E(MS_B)=\sigma_{\text{tech}}^2+r\sigma_{\text{subject}}^2\). Subtracting the within mean square and dividing by \(r\) thus isolates the subject component in expectation:
This derivation depends on the balanced, independent-error model. Replicates that share a run-level error, have different error variances, or are systematically shifted by replicate order do not satisfy the simple model without modification.
Worked Example: Separating Subject and Technical Variation
Consider hypothetical measurements of one feature for four specimens, each measured twice on the same platform under a protocol intended to give independent replicate errors. Assume the measurements are on a common scale, with no run or order effect. The pairs are 10 and 12, 14 and 15, 9 and 11, and 17 and 19 units.
The specimen means are 11, 14.5, 10, and 18 units; the grand mean is \(53.5/4=13.375\) units. The within-specimen sum of squares is \(2+0.5+2+2=6.5\) units squared. With \(4(2-1)=4\) degrees of freedom, \(MS_W=6.5/4=1.625\) units squared.
The between-specimen sum of squares is \(2[(11-13.375)^2+(14.5-13.375)^2+(10-13.375)^2+(18-13.375)^2]=79.375\) units squared. Thus \(MS_B=79.375/3=26.458\) units squared. The estimated technical variance is \(1.625\) units squared, and the estimated between-subject variance is \((26.458-1.625)/2=12.417\) units squared.
The corresponding estimated standard deviations are \(\sqrt{1.625}=1.275\) units for technical error and \(\sqrt{12.417}=3.524\) units for between-subject variation. Under the model, this feature varies more between these sampled specimens than across repeated measurements of a specimen. Four specimens provide limited information, so these estimates should not be treated as precise population values.
Paired Differences and Platform-Specific Repeatability
For two technical measurements of each specimen on a given platform, paired differences offer a direct estimate that cancels the specimen’s underlying level. Let \(D_i=Y_{i1}-Y_{i2}\). If each measurement has independent technical error with variance \(\sigma_{\text{tech}}^2\), then \(D_i=\varepsilon_{i1}-\varepsilon_{i2}\) and \(\operatorname{Var}(D_i)=2\sigma_{\text{tech}}^2\). Therefore, if \(s_D^2\) is the sample variance of the differences across \(m\) specimens, \(s_D^2/2\) estimates the technical variance. This uses differences centred around their sample mean, so a consistent offset between replicate positions does not inflate the variance estimate; that offset should nevertheless be investigated as a possible systematic effect.
The estimate is specific to the replicate conditions. If two measurements occur in the same run, a shared run effect cancels from their difference. The estimate then describes within-run repeatability, not run-to-run variation. To estimate run-to-run variation, specimens must be measured across runs in a design that separates specimen effects from run effects; measuring one set of specimens only in one run and a different set only in another confounds those sources.
Worked Example: Comparing Technical Variation Across Platforms
Suppose five hypothetical specimens are each measured twice on each of two platforms. The replicate differences, calculated as measurement 1 minus measurement 2, are \(1,-1,2,-2,0\) units on Platform A and \(4,-2,2,-4,0\) units on Platform B. Assume, separately for each platform, independent replicate errors with equal variance, no specimen-specific replicate effect, and measurements on a common scale within that platform.
For Platform A, the differences have mean 0 units and sample variance \(s_D^2=(1+1+4+4+0)/4=2.5\) units squared. The estimated technical variance is \(2.5/2=1.25\) units squared, giving an estimated technical standard deviation of \(\sqrt{1.25}=1.118\) units.
For Platform B, the differences also have mean 0 units, and \(s_D^2=(16+4+4+16+0)/4=10\) units squared. Its estimated technical variance is \(10/2=5\) units squared, and its estimated technical standard deviation is \(\sqrt{5}=2.236\) units. In this hypothetical comparison, the estimated technical variance on Platform B is four times that on Platform A, and its estimated technical standard deviation is twice as large.
These estimates compare repeatability on the chosen measurement scale; they do not show that either platform is unbiased or that their reported values are directly interchangeable. The calculation also assumes the replicate errors are independent. Shared run effects or specimen-by-platform interactions require a design and model that represent those sources.
Uncertainty, Repeatability, and Diagnostics
A variance estimate is itself uncertain, particularly when the number of specimens or replicate measurements is small. Under independent normal replicate errors, the within-specimen residual sum of squares divided by \(\sigma_{\text{tech}}^2\) follows a chi-square distribution with \(df=m(r-1)\) degrees of freedom. This gives a confidence interval for the technical variance; its endpoints are often widely separated when \(df\) is small.
Worked Example: Quantifying Uncertainty in a Technical Variance
Use the five hypothetical paired differences from Platform A above: \(1,-1,2,-2,0\) units. Their sample variance is \(s_D^2=2.5\) units squared, so the estimated technical variance is \(s_D^2/2=1.25\) units squared. Assume normally distributed, independent replicate errors, with no difference in error variance across specimens.
For the variance of the differences, the 95% chi-square confidence interval is \(\bigl(df\,s_D^2/\chi^2_{0.975,df},\ df\,s_D^2/\chi^2_{0.025,df}\bigr)\). Here \(df=5-1=4\), with chi-square quantiles approximately 11.143 and 0.484. The interval for the difference variance is \((4\times2.5/11.143,\ 4\times2.5/0.484)=(0.897,\ 20.66)\) units squared. Dividing both endpoints by 2 gives an interval for technical variance of \((0.449,\ 10.33)\) units squared, or a technical-standard-deviation interval of approximately \((0.670,\ 3.214)\) units.
The interval is broad: five pairs do not estimate repeatability precisely. Under repeated sampling from the stated normal-error model, 95% of intervals constructed by this procedure would cover the true technical variance. This does not mean there is a 95% probability that this particular interval contains the fixed variance.
A common summary of the basic random-effects model is the intraclass correlation, \(\rho=\sigma_{\text{subject}}^2/(\sigma_{\text{subject}}^2+\sigma_{\text{tech}}^2)\). It is the correlation between two measurements of the same randomly sampled specimen under that model. Using the variance estimates in the first example gives \(\hat{\rho}=12.417/(12.417+1.625)=0.884\). This indicates that much of the single-measurement variance in that hypothetical sample is between specimens. It is not a direct measure of agreement between platforms, and it can be high when between-subject variation is large even if technical error is substantial.
Inspect the design and data before relying on a variance estimate. Plot replicate measurements against each other, plot paired differences against specimen mean and run, and check whether variation grows with the measurement level. Look for a difference mean far from zero, outliers, patterns by replicate order, and unequal spread across specimens or platforms. On a log scale, an approximately constant technical variance may correspond to multiplicative error on the original scale; a constant raw-scale variance should not be presumed.
Design Choices, Alternatives, and Limitations
Replicate design determines what “technical variation” means. Repeated aliquots measured independently within one run estimate a narrow form of repeatability. Replicates distributed across runs can capture run-to-run variation, but only if the same specimens recur across runs sufficiently to separate run from specimen effects. When several sources matter, a model can include their variance components, but the design must contain replication and overlap that identify them. A single observed value per specimen cannot separate specimen effects from residual measurement error using this model.
| Approach | What it estimates or addresses | Key limitation |
|---|---|---|
| Within-specimen ANOVA | Technical and between-subject components for balanced repeated measurements | Relies on a suitable random-effects model and balanced replication for the simple formulas |
| Paired replicate differences | Technical variance from two measurements per specimen | Does not separate multiple technical sources unless the design varies them appropriately |
| Repeatability correlation (ICC) | Relative share of single-measurement variance attributable to subjects | Depends on the population’s between-subject spread and is not cross-platform agreement |
| Replicate means | Can reduce independent measurement error by averaging repeated values | Does not remove systematic bias or correlated technical errors |
The formulas above are for one feature at a time and do not correct platform shifts or establish biological equivalence. On different platforms, variances may be expressed on different scales or may reflect different signal ranges. A high technical standard deviation can also be practically small or large depending on the scale and intended use. Estimating repeatability does not by itself determine whether a biological association is preserved after measurement.
Common mistake: “There are 20 measurements, so the biological sample size is 20.” If those measurements are four specimens each run five times, there are four biological specimens, not 20. The repeats improve information about technical error and possibly the precision of each specimen’s mean; they do not create new independent biological units. Analyse the shared specimen structure rather than counting each repeat as an independent subject.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- In the balanced model, what does \(MS_W\) estimate, and why does it not include the shared subject effect?
- Two measurements of each specimen occur in the same run. What component is cancelled by paired differencing, and what broader component may therefore remain unestimated?
- For paired differences with sample variance 6 units squared, what is the estimated technical variance under independent equal-variance errors?
- Why is a high intraclass correlation not sufficient evidence that two platforms agree?
Solutions. (1) \(MS_W\) estimates \(\sigma_{\text{tech}}^2\); subtracting each specimen’s replicate mean removes its shared \(b_i\). (2) A run effect shared by the two measurements cancels; run-to-run variation may remain unestimated. (3) The estimate is \(6/2=3\) units squared. (4) The ICC describes relative between-subject variation under one model and depends on the sampled population’s spread; it does not test platform bias, scale differences, or agreement.