Quality is often more important than quantity.
Science relies heavily on replicate measurements. Additional replicates generally yield more accurate and reliable summary statistics in experimental work. But the straightforward question, 'how many and what kind of replicates should I run?' belies a deep set of distinctions and tradeoffs that affect statistical testing. We illustrate different types of replication in multilevel ('nested') experimental designs and clarify basic concepts of efficient allocation of replicates.
Replicates can be used to assess and isolate sources of variation in measurements and limit the effect of spurious variation on hypothesis testing and parameter estimation. Biological replicates are parallel measurements of biologically distinct samples that capture random biological variation, which may itself be a subject of study or a noise source. Technical replicates are repeated measurements of the same sample that represent independent measures of the random noise associated with protocols or equipment. For biologically distinct conditions, averaging technical replicates can limit the impact of measurement error, but taking additional biological replicates is often preferable for improving the efficiency of statistical testing.
Nested study designs can be quite complex and include many levels of biological and technical replication (Table 1). The distinction between biological and technical replicates depends on which sources of variation are being studied or, alternatively, viewed as noise sources.
An illustrative example is genome sequencing, where base calls (a statistical estimate of the most likely base at a given sequence position) are made from multiple DNA reads of the same genetic locus. These reads are technical replicates that sample the uncertainty in the sequencer readout but will never reveal errors present in the library itself. Errors in library construction can be mitigated by constructing technical replicate libraries from the same sample. If additional resources are available, one could potentially return to the source tissue and collect multiple samples to repeat the entire sequencing workflow. Such replicates would be technical if the samples were considered to be from the same aliquot or biological if considered to be from different aliquots of biologically distinct material1. Owing to historically high costs per assay, the field of genome sequencing has not demanded such replication. As the need for accuracy increases and the cost of sequencing falls, this is likely to change.
How does one determine the types, levels and number of replicates to include in a study, and the extent to which they contribute information about important sources of variation? We illustrate the approach to answering these questions with a single-cell sequencing scenario in which we measure the expression of a specific gene in liver cells in mice. We simulated three levels of replication: animals, cells and measurements (Fig. 1a). Each level has a different variance, with animals (σA2 = 1) and cells (σC2 = 2) contributing to a total biological variance of σB2 = 3. When technical variance from the assay (σM2= 0.5) is included, these distributions compound the uncertainty in the measurement for a total variance of σTOT2 = 3.5. We next simulated 48 measurements, allocated variously between biological replicates (the number of animals, nA and number of cells sampled per animal, nC) and technical replicates (number of measurements taken per cell, nM) for a total number of measurements nAnCnM = 48. Although we will always make 48 measurements, the effective sample size, n, will vary from about 2 to 48, depending on how the measurements are allocated. Let us look at how this comes about.
Our ability to make accurate inferences will depend on our estimate of the variance in the system, Var(X). Different choices of nA, nC and nM impact this value differently. If we sample nC = 48 cells from a single animal (nA = 1) and measure each nM = 1 times, our estimate of the total variance σTOT2 will be Var(X) = 2.5 (Fig. 1b). This reflects cell and measurement variances (σC2 + σM2) but not animal variation; with only one animal sampled we have no way of knowing what the animal variance is. Thus Var(X) certainly underestimates σTOT2, but we would not know by how much. Moreover, the uncertainty in Var(X) (error bar at nA = 1; Fig. 1b) is the error in σC2 + σM2 and not σTOT2. At another extreme, if all our measurements are technical replicates (nA = nC = 1, nM = 48) we would find Var(X) = 0.5 (not represented in Fig. 1). This is only the technical variance; if we misinterpreted this as biological variation and used it for biological inference, we would have an excess of false positives. Be on the lookout: unusually small error bars on biological measurements may merely reflect measurement error, not biological variation. To obtain the best estimate of σTOT2 we should sample nC = 1 cells from nA = 48 animals because each of the 48 measurements will independently sample each of the distributions in Figure 1a.
Our choice of the number of replicates also influences Var(), the precision in the expression mean. The optimal way to minimize this value is to collect data from as many animals as possible (nA = 48, nC = nM = 1), regardless of the ratios of variances in the system. This comes from the fact that nA contributes to decreasing each contribution to Var(), which is given by σA2/nA + σC2/nAnC + σM2/nAnCnM. Although technical replicates allow us to determine σM2, unless this is a quantity of interest, we should omit technical replicates and maximize nA. Of course, good blocking practice suggests that samples from the different animals and cells should be mixed across the sequencing runs to minimize the effect of any systematic run-to-run variability (not present in simulated data here).
The value in additional measurements can be estimated by the prospective improvement in effective sample size. We have seen before that the variance in the mean of a random variable is related to its variance by Var(X) = nVar(). The ratio of Var(X) to Var() can therefore be used as a measure of the equivalent number of independent samples. From Figure 1b, we can see that n = 48 only for nA = 48 and drops to n = 25 for nA, nC = 12, 4 and is as low as about 2 for nA= 1. In other words, even though we may be collecting additional measurements they do not all contribute equally to an increase in the precision of the mean. This is because additional cell and technical replicates do not correspond to statistically independent values: technical replicates are derived from the same cell and the cell replicates from the same animal. If it is necessary to summarize expression variability at the level of the animals, then cells from a given animal are pseudoreplicates—statistically correlated in a way that is unique to that animal and not representative of the population under study. Not all replicates yield statistically independent measures, and treating them as if they do can erroneously lower the apparent uncertainty of a result.
The number of replicates has a practical effect on inference errors in analysis of differences of means or variances. We illustrate this by enumerating inference errors in 10,000 simulated drug-treatment experiments in which we vary the number of animals and cells (Fig. 2). We assume a 10% effect chance for two scenarios: a twofold increase in variance, σC2, or a 10% increase in mean, μA, using the same values for other variances and 48 total measurements as in Figure 1. Applying the t-test, we show false discovery rate (FDR) and power for detecting these differences (Fig. 2). If we want to detect a difference in variation across cells, it is best to choose nA ≈ nC in our range. On the other hand, when we are interested in changes in mean expression across mice, it is better to sample as many mice as possible. In either case, increasing the number of measurements from 48 to 144 by taking three technical replicates (nM = 3) improves inference only slightly.
Biological replicates are preferable to technical replicates for inference about the mean and variance of a biological population (Fig. 2). For example, changing nA,nC,nM from 8,6,3 (where power is highest) to 12,12,1 doubles the power (0.43 to 0.88) in detecting a twofold change in variance. In the case of detecting a 10% difference in means, changing nA,nC,nM from 24,2,3 to 72,2,1 increases power by about 50% from 0.66 to 0.98. Practically, the cost difference between biological and technical replicates should be considered; this will affect the cost-benefit tradeoff of collecting additional replicates of one type versus the other. For example, if the cost units of animals to cells to measurements is 10:1:0.1 (biological replicates are likely more expensive than technical ones) then an experiment with nA,nC,nM of 12,12,1 is about twice as expensive as that with 8,6,3 (278 versus 142 cost units). However, power in detecting a change in variance is doubled as well, so the cost increase is commensurate with increase in efficiency. In the case of detecting differences in means, 72,2,1 is about three times as expensive as 24,2,3 (878 versus 302 cost units) but increases power only by 50%, making this a lower-value proposition.
Typically, biological variability is substantially greater than technical variability, so it is to our advantage to commit resources to sampling biologically relevant variables unless measures of technical variability are themselves of interest, in which case increasing the number of measurements per cell, nM, is valuable.
Good experimental design practice includes planning for replication. First, identify the questions the experiment aims to answer. Next, determine the proportion of variability induced by each step to distribute the capacity for replication of the experiment across steps. Be aware of the potential for pseudoreplication and aim to design statistically independent replicates.
As our capacity for higher-throughput assays increases, we should not be misled into thinking that more is always better. Clear thinking about experimental questions and sources of variability is still crucial to produce efficient study designs and valid statistical analyses.
Robasky, K., Lewis, N.E. & Church, G.M. Nat. Rev. Genet. 15, 56–62 (2014).
The authors declare no competing financial interests.
About this article
Cite this article
Blainey, P., Krzywinski, M. & Altman, N. Replication. Nat Methods 11, 879–880 (2014). https://doi.org/10.1038/nmeth.3091
This article is cited by
Cellular & Molecular Biology Letters (2023)
Empirical methods for the validation of time-to-event mathematical models taking into account uncertainty and variability: application to EGFR + lung adenocarcinoma
BMC Bioinformatics (2023)
Temporal transcriptome profiling of floating apical out chicken enteroids suggest stability and reproducibility
Veterinary Research (2023)
Sensitivity of endogenous autofluorescence in HeLa cells to the application of external magnetic fields
Scientific Reports (2023)
Sumoylated SnoN interacts with HDAC1 and p300/CBP to regulate EMT-associated phenotypes in mammary organoids
Cell Death & Disease (2023)