# Subsampling vs. True Replication ## Quick Answer - True replication means each independent experimental unit receives its own treatment assignment, while subsampling takes multiple measurements from the same unit and treats them as if they were independent. - Treating subsamples as true replicates inflates statistical power and produces false precision, so researchers must identify the experimental unit before designing the study. - The correction requires analyzing subsample data with nested or mixed-effects models that account for the hierarchical structure of the data. ## The Experimental Unit Problem in Biological Research The distinction between subsampling and true replication is one of the most consequential decisions in experimental design, yet it is frequently misunderstood in biology laboratories. When a researcher measures multiple cells from the same culture dish, multiple wells from the same plate, or multiple tissue sections from the same animal, those measurements share a common source of variation. They are not independent observations in the statistical sense because they all originate from a single experimental unit. The experimental unit is the smallest entity that is independently assigned to a treatment. In a study where three different culture conditions are applied to three separate flasks, the flask is the experimental unit. If the researcher then takes five samples from each flask and measures each sample, those five measurements are subsamples. They provide information about measurement error and within-flask variation, but they do not increase the number of independent observations available for testing the treatment effect. This confusion leads to a specific and predictable error. Researchers count every measurement as a replicate, report a sample size of 15 when only three independent units exist, and obtain p-values that appear highly significant. The apparent precision is an artifact of pseudoreplication, a term introduced to describe the analysis of subsamples as if they were independent replicates. The consequences extend beyond statistical correction. Studies designed with pseudoreplication cannot support the conclusions they claim because the true sample size is smaller than reported. This undermines reproducibility and can lead other laboratories to build on findings that do not hold up under proper analysis. The [National Library of Medicine Research Methods Resources](https://www.ncbi.nlm.nih.gov/books) provides access to biomedical research methodology references that address these design principles in detail. ## Defining True Replication True replication requires that each experimental unit receives its treatment independently and that the units vary in ways that reflect the population to which the researcher intends to generalize. In a cell culture experiment, true replication means using multiple separate cultures, ideally prepared on different days or from different cell passages, each assigned to a treatment condition. In an animal study, true replication means using multiple animals per treatment group, with each animal representing an independent biological entity. The key feature of true replication is independence. Each replicate provides an independent estimate of the treatment effect because the units are subject to independent sources of biological and environmental variation. This independence is what allows the researcher to draw inferences about the broader population from which the units were drawn. Consider a study of gene expression in response to a drug. If the researcher treats three separate animals with the drug and three separate animals with a vehicle control, then measures gene expression in multiple tissue samples from each animal, the animal is the experimental unit. The tissue samples are subsamples. The study has three true replicates per group, not 15 or 30, regardless of how many tissue samples are measured. The distinction matters because the statistical analysis must reflect the design. Analyzing subsamples as independent replicates violates the assumption of independence that underlies standard statistical tests. The result is an inflated type I error rate, meaning the researcher is more likely to declare a treatment effect when none exists. ## Subsampling Defined Subsampling is the practice of taking multiple measurements from a single experimental unit. These measurements are sometimes called technical replicates, repeated measures, or within-unit observations. Subsampling serves legitimate purposes in experimental design, but those purposes are different from the purpose of true replication. Subsampling provides information about measurement precision. If a researcher measures the same sample multiple times on the same instrument, the variation among those measurements reflects instrument noise and technical error. If the researcher takes multiple samples from the same culture and measures each one, the variation reflects both sampling error and measurement error, but not biological variation among independent units. Subsampling can also improve the accuracy of the estimate for a particular unit. Averaging multiple measurements from the same unit reduces the impact of measurement error and provides a more reliable estimate of that unit's true value. This is valuable when the measurement technique is noisy or when the sample is heterogeneous. The critical point is that subsampling does not increase the number of independent observations. The researcher gains precision about each unit but does not gain additional units. Statistical tests that treat subsamples as independent replicates are analyzing the data incorrectly. The [EQUATOR Network](https://www.equator-network.org/) provides reporting guidelines that help researchers describe their experimental designs accurately. Transparent reporting of how many units were assigned to treatment and how many measurements were taken from each unit allows reviewers and readers to assess whether the analysis matches the design. ## The Statistical Consequences of Pseudoreplication Pseudoreplication produces several predictable statistical problems. The most serious is the inflation of the degrees of freedom in statistical tests. When subsamples are treated as independent replicates, the error term in the analysis is based on within-unit variation instead of between-unit variation. This makes the error term smaller than it should be, which makes the test statistic larger and the p-value smaller. The result is false precision. A study with three animals per group and ten measurements per animal might report a p-value of 0.001 when the correct analysis, based on three animals per group, would produce a p-value of 0.20. The researcher concludes that the treatment has an effect when the data do not support that conclusion. The magnitude of the problem depends on the ratio of between-unit to within-unit variation. If the variation among subsamples within a unit is small relative to the variation among units, the inflation is severe. If the within-unit variation is large, the inflation is less severe but still present. Another consequence is that confidence intervals are too narrow. The reported confidence interval suggests a level of precision that the design does not support. This can mislead other researchers who use the reported effect size and confidence interval to plan their own studies. The problem is not limited to hypothesis testing. Effect size estimates can also be biased when subsamples are treated as independent replicates, particularly when the number of subsamples per unit varies. Studies with unequal subsampling can produce estimates that are weighted toward units with more measurements. ## Identifying the Experimental Unit The first step in avoiding pseudoreplication is to identify the experimental unit before the study begins. This requires asking a specific question: what entity was independently assigned to the treatment condition? In cell culture experiments, the experimental unit is typically the culture vessel that received the treatment. If a researcher prepares six dishes and assigns three to treatment A and three to treatment B, the dish is the experimental unit. Multiple measurements from each dish are subsamples. In animal studies, the experimental unit is typically the individual animal. If a researcher assigns five mice to each of two treatment groups, the mouse is the experimental unit. Multiple tissue samples from each mouse are subsamples. In some designs, the experimental unit is not the individual animal but a group of animals. This occurs in behavioral studies where animals are housed together and the treatment is applied to the cage. If two cages receive treatment A and two cages receive treatment B, the cage is the experimental unit, even though each cage contains multiple animals. In field studies, the experimental unit might be a plot, a field, or a pond. The treatment is applied to the entire unit, and measurements are taken from multiple locations within the unit. The identification of the experimental unit should be documented in the study protocol before data collection begins. This documentation prevents the temptation to reclassify the experimental unit after seeing the results. The [NIH Grants and Funding](https://grants.nih.gov/) resources describe the expectations for rigorous experimental design in funded research, including the need to specify the experimental unit in the application. ## Common Experimental Designs and Their Units Different experimental designs present different challenges for identifying the experimental unit. Understanding the structure of common designs helps researchers recognize when they are at risk of pseudoreplication. ### Cell Culture Experiments In cell culture, the experimental unit is the culture vessel that receives the treatment. This could be a flask, a dish, a well in a multiwell plate, or a bioreactor. The key question is whether the treatment was applied to the vessel independently. A common error occurs with multiwell plates. If a researcher adds a drug to all wells in one column of a 96-well plate and a vehicle to all wells in another column, the plate is the experimental unit for the comparison, not the individual wells. The wells within a column share the same treatment application and are subject to plate-position effects. A proper design would randomize the treatment assignments across wells, with multiple plates serving as replicates. Another common error occurs when cells from a single culture are split into multiple vessels and each vessel is treated differently. The vessels are subsamples of the original culture, not independent replicates, because they share the same source population and passage history. True replication requires starting with separate cultures, ideally from different passages or different freezing stocks. ### Animal Studies In animal studies, the experimental unit is usually the individual animal. However, complications arise when animals are housed together or when treatments are applied to groups. If animals are housed in cages and the treatment is administered through the feed or water, the cage is the experimental unit because all animals in the cage receive the same treated feed. Individual animals within the cage are subsamples. This design requires multiple cages per treatment group to achieve true replication. If the treatment is administered to each animal individually, such as by injection or oral gavage, the animal is the experimental unit, even if the animals are housed together. The treatment assignment is independent for each animal. The distinction between housing and treatment application is critical. Researchers must document whether the treatment was applied at the level of the individual or the group. ### Tissue and Organ Studies When tissues are harvested from animals and studied ex vivo, the experimental unit depends on the research question. If the question concerns the effect of a treatment applied to the animal, the animal is the experimental unit. Multiple tissue samples from the same animal are subsamples. If the question concerns the effect of a treatment applied to the tissue after harvest, the tissue sample is the experimental unit. For example, if a researcher takes brain slices from several animals and applies a drug to some slices and a vehicle to others, the slice is the experimental unit, provided that slices from the same animal are randomly assigned to treatments. The distinction is subtle but important. The experimental unit is determined by the level at which the treatment is applied, not by the level at which the measurement is taken. ### Field and Environmental Studies In field studies, the experimental unit is typically the plot or the field to which the treatment is applied. Multiple soil samples, plant measurements, or water samples from the same plot are subsamples. The spatial arrangement of plots introduces additional considerations. Plots that are close together may be more similar than plots that are far apart, violating the assumption of independence. Randomized block designs and spatial analysis methods address this issue. ## Statistical Analysis of Subsampled Data When subsampling is part of the design, the statistical analysis must account for the hierarchical structure of the data. Several approaches are available, and the choice depends on the design and the research question. ### Averaging Subsamples The simplest approach is to average the subsamples for each experimental unit and analyze the unit-level averages. This approach is appropriate when the research question concerns the treatment effect on the unit, not the variation among subsamples. Averaging reduces the data to one observation per experimental unit, which can then be analyzed with standard methods such as t-tests or ANOVA. The disadvantage is that information about within-unit variation is lost, and the analysis does not account for differences in the number of subsamples per unit. ### Nested Analysis of Variance Nested ANOVA, also called hierarchical ANOVA, partitions the variation into between-unit and within-unit components. This approach is appropriate when the researcher wants to estimate both the treatment effect and the components of variance. In a nested design, the subsamples are nested within experimental units, and the experimental units are nested within treatment groups. The analysis tests the treatment effect against the between-unit variation, not the within-unit variation. This produces the correct degrees of freedom and avoids the inflation associated with pseudoreplication. ### Mixed-Effects Models Mixed-effects models, also called multilevel models or hierarchical linear models, provide a flexible framework for analyzing subsampled data. These models include both fixed effects, such as treatment, and random effects, such as experimental unit. The random effect for experimental unit accounts for the correlation among subsamples from the same unit. This correlation is the statistical signature of subsampling. Ignoring it, by treating subsamples as independent, is the essence of pseudoreplication. Mixed-effects models can accommodate unbalanced designs, where the number of subsamples varies among units, and can include additional levels of nesting, such as subsamples within tissue samples within animals. The [National Library of Medicine Research Methods Resources](https://www.ncbi.nlm.nih.gov/books) provides access to statistical methodology references that describe these models in detail. ### Choosing the Analysis The choice of analysis depends on the research question and the design. If the research question concerns the treatment effect and the design is balanced, averaging and nested ANOVA produce similar results. If the design is unbalanced or the researcher wants to estimate variance components, mixed-effects models are preferred. The analysis should be specified in the study protocol before data collection. This prevents the researcher from choosing the analysis that produces the most favorable results after seeing the data. ## Designing Studies with Appropriate Replication The design phase is where pseudoreplication is prevented. Several practices help researchers achieve appropriate replication. ### Determine the Experimental Unit First The experimental unit should be identified before the study begins. This determination should be based on the level at which the treatment is applied, not the level at which measurements are taken. The identification should be documented in the protocol. ### Estimate the Required Number of Units The number of experimental units needed depends on the expected effect size, the variation among units, and the desired statistical power. Power calculations should be based on the number of experimental units, not the number of measurements. Researchers often underestimate the number of units needed because they confuse subsamples with replicates. A power calculation based on the number of subsamples will produce a sample size that is too small, leading to an underpowered study. ### Randomize Treatment Assignment Treatments should be assigned to experimental units randomly. Randomization prevents systematic bias and ensures that the treatment groups are comparable. In multiwell plates, treatments should be randomized across wells and plates. In animal studies, animals should be randomly assigned to treatment groups. ### Include Sufficient Subsamples Subsampling is valuable for reducing measurement error and characterizing within-unit variation. The number of subsamples per unit should be based on the measurement error and the within-unit variation. More subsamples are needed when the measurement is noisy or the unit is heterogeneous. The optimal number of subsamples depends on the cost of additional units relative to the cost of additional measurements. If experimental units are expensive, such as large animals, more subsamples per unit may be efficient. If measurements are expensive, fewer subsamples per unit may be appropriate. ### Document the Design The design should be documented in enough detail that another researcher can determine the experimental unit and the number of true replicates. This documentation is essential for reproducibility and for reviewers to assess the validity of the analysis. The [EQUATOR Network](https://www.equator-network.org/) provides reporting guidelines that specify the design information that should be included in publications. Following these guidelines ensures that the experimental unit and the replication strategy are transparent to readers. ## At a Glance | Design Element | True Replication | Subsampling | Pseudoreplication Error | |---|---|---|---| | Definition | Independent experimental units assigned to treatment | Multiple measurements from the same unit | Treating subsamples as independent replicates | | Example | Three separate animals per treatment group | Five tissue samples from each animal | Reporting 15 replicates when only 3 animals were used | | Statistical role | Provides the error term for hypothesis testing | Reduces measurement error, characterizes within-unit variation | Inflates degrees of freedom, produces false precision | | Analysis | Standard tests on unit-level data | Nested ANOVA or mixed-effects models | Standard tests on all measurements | | Sample size | Number of units assigned to treatment | Number of measurements per unit | Reported as total measurements, not units | | Common error | Too few units for the desired power | Too few subsamples to characterize the unit | Counting wells, samples, or measurements as replicates | ## Analysis Methods Comparison | Method | When to Use | Key Advantage | Key Limitation | |---|---|---|---| | Unit-level averaging | Balanced designs, simple treatment comparisons | Straightforward, uses standard tests | Loses within-unit variation information | | Nested ANOVA | Estimating variance components, balanced designs | Partitions between-unit and within-unit variation | Requires balanced data, less flexible | | Mixed-effects models | Unbalanced designs, multiple nesting levels | Handles complex structures, accounts for correlation | Requires specialized software and expertise | ## Practical Steps for Implementing Correct Replication Implementing correct replication requires a systematic approach that begins before data collection and continues through analysis and reporting. ### Step 1: Define the Research Question and the Experimental Unit Write the research question in a way that identifies the level of inference. If the question asks whether a treatment affects cells, the experimental unit is the culture vessel. If the question asks whether a treatment affects animals, the experimental unit is the animal. Document this decision in the protocol. ### Step 2: Calculate the Required Number of Experimental Units Use a power calculation based on the expected effect size and the variation among experimental units. The variation among units should be estimated from previous studies or pilot data. If no estimate is available, use a conservative estimate and plan to verify the power after data collection. ### Step 3: Plan the Subsampling Strategy Determine how many measurements will be taken from each experimental unit. The number of subsamples should be based on the measurement error and the within-unit variation. More subsamples reduce the impact of measurement error but do not increase the number of independent observations. ### Step 4: Randomize Treatment Assignment Assign treatments to experimental units using a randomization procedure. For multiwell plates, randomize the treatment assignments across wells and use multiple plates. For animal studies, randomize animals to treatment groups and consider blocking by litter, weight, or other relevant factors. ### Step 5: Collect Data with Clear Unit Identification Label each measurement with the experimental unit from which it was taken. This labeling is essential for the analysis and for verifying that the design was implemented correctly. The data file should include a column for the experimental unit identifier and a column for the subsample identifier. ### Step 6: Analyze the Data with the Correct Model Analyze the data using a method that accounts for the hierarchical structure. If the design is balanced and the research question concerns the treatment effect, averaging the subsamples and analyzing the unit-level means is appropriate. If the design is unbalanced or the researcher wants to estimate variance components, use a mixed-effects model with the experimental unit as a random effect. ### Step 7: Report the Design Transparently Report the number of experimental units per treatment group and the number of subsamples per unit. Describe the experimental unit in the methods section and specify the statistical model used for the analysis. This transparency allows readers to assess whether the analysis matches the design. ## Records and Measurements Accurate records are essential for distinguishing subsamples from true replicates and for defending the analysis in peer review. ### Experimental Unit Log Maintain a log that records each experimental unit, its treatment assignment, and its identifying information. For cell culture experiments, record the culture vessel identifier, the cell passage number, and the date the culture was established. For animal studies, record the animal identifier, the housing assignment, and the treatment administration details. ### Subsample Log Maintain a log that records each subsample, the experimental unit from which it was taken, and the measurement details. This log should include the date and time of the measurement, the instrument used, and any observations about the sample quality. ### Data File Structure Structure the data file with one row per subsample and columns for the experimental unit identifier, the treatment assignment, and the measured value. This structure allows the analysis software to account for the nesting of subsamples within units. ### Analysis Documentation Document the statistical analysis, including the model specification, the software used, and the rationale for the chosen approach. This documentation is important for reproducibility and for responding to reviewer questions about the analysis. The [NIH Data Management and Sharing Policy](https://sharing.nih.gov/data-management-and-sharing-policy) describes expectations for data management and sharing in NIH-funded research. These expectations include documenting the data collection and analysis procedures in enough detail that another researcher can understand and reproduce the work. ## Common Failure Patterns Several recurring errors lead to pseudoreplication in biological research. Recognizing these patterns helps researchers avoid them. ### Counting Wells as Replicates The most common error in cell culture research is counting individual wells in a multiwell plate as independent replicates. Wells on the same plate share the same plate environment, the same cell preparation, and often the same pipetting session. They are subsamples, not replicates. The correct approach is to treat the plate as the experimental unit or to use a design where each well is independently prepared and randomized. If the researcher wants to treat wells as replicates, each well must be an independent culture prepared separately. ### Counting Samples from the Same Animal as Replicates In animal studies, multiple tissue samples from the same animal are often counted as replicates. This error is common in histology, where multiple sections from the same tissue block are measured, and in biochemistry, where multiple aliquots from the same homogenate are assayed. The animal is the experimental unit. The tissue sections and aliquots are subsamples. The correct analysis uses the animal as the unit of analysis, with the subsamples providing a more reliable estimate of each animal's value. ### Ignoring the Source of the Cells or Animals Cells from a single culture that are split into multiple vessels are not independent replicates because they share the same source population. Animals from the same litter are more similar than animals from different litters because they share genetic and environmental factors. The correct approach is to use multiple independent cultures or multiple litters and to account for the source in the analysis. Blocking by source, such as using litter as a blocking factor, can improve the power of the study while maintaining the correct experimental unit. ### Analyzing Repeated Measures as Independent Observations Longitudinal studies, where the same unit is measured at multiple time points, present a related problem. The measurements from the same unit over time are correlated and cannot be treated as independent observations. The correct approach is to use repeated-measures ANOVA or a mixed-effects model with time as a within-unit factor. These methods account for the correlation among measurements from the same unit. ### Reporting the Total Number of Measurements as the Sample Size Publications that report the sample size as the total number of measurements, instead of the number of experimental units, mislead readers about the precision of the study. This reporting error is common in abstracts and methods sections. The correct approach is to report both the number of experimental units and the number of subsamples per unit. For example, a study might report that 5 animals per group were used, with 10 tissue sections measured per animal. ## Welfare and Safety Context The distinction between subsampling and true replication has implications for animal welfare and research ethics. Using too few animals because the researcher plans to count subsamples as replicates can lead to underpowered studies that fail to detect treatment effects. This wastes animal lives and resources. Conversely, using more animals than necessary because the researcher does not understand the distinction between subsamples and replicates is also problematic. The [Committee on Publication Ethics Core Practices](https://publicationethics.org/core-practices) address the responsible conduct of research, including the ethical use of animals and the accurate reporting of experimental designs. The principle of reduction in animal research, one of the three Rs, calls for minimizing the number of animals used while maintaining the scientific validity of the study. Correctly identifying the experimental unit and using appropriate power calculations ensures that the number of animals is neither too high nor too low. Laboratory safety is also relevant to the design of experiments with subsampling. The [WHO Laboratory Biosafety Manual](https://www.who.int/publications/i/item/9789240011311) provides guidance on safe handling of biological materials. When multiple subsamples are taken from a single unit, each subsampling operation carries a risk of contamination or exposure. The number of subsamples should balance the need for precision with the safety considerations of handling biological materials. ## Limitations of Subsampling Subsampling has inherent limitations that researchers must recognize. These limitations affect the conclusions that can be drawn from subsampled data. ### Subsampling Cannot Replace Replication The most important limitation is that subsampling cannot compensate for insufficient replication. If a study has too few experimental units, no amount of subsampling can provide the statistical power needed to detect a treatment effect. The researcher must increase the number of experimental units. ### Subsampling Does Not Increase Generalizability Subsampling provides information about the specific experimental units in the study. It does not increase the generalizability of the findings to the broader population. Generalizability depends on the number and representativeness of the experimental units. ### Subsampling Can Be Misleading Subsampling can be misleading if the within-unit variation is not representative of the between-unit variation. If the subsamples from a single unit are more similar than units are to each other, the subsamples will underestimate the true variation in the population. ### Subsampling Adds Complexity to the Analysis Subsampled data require more complex statistical analysis than data from a simple design. The analysis must account for the hierarchical structure, which requires specialized software and expertise. Researchers who are not familiar with nested designs or mixed-effects models may struggle to analyze their data correctly. ## Professional Escalation Criteria Researchers should seek additional expertise when they encounter situations that exceed their statistical or design knowledge. ### Consult a Biostatistician Before Data Collection If the researcher is uncertain about the experimental unit or the appropriate sample size, a biostatistician should be consulted before data collection begins. The cost of consulting a statistician is small compared to the cost of collecting data that cannot be analyzed correctly. ### Seek Guidance for Complex Designs Complex designs, such as those with multiple levels of nesting, repeated measures, or spatial structure, require specialized statistical expertise. Researchers should seek guidance from a statistician or a collaborator with relevant experience. ### Escalate When Reviewers Question the Design If reviewers question the experimental unit or the analysis, the researcher should take the concern seriously and seek expert advice. Reviewer concerns about pseudoreplication are often valid and indicate a problem with the design or analysis. ### Use Reporting Guidelines The [EQUATOR Network](https://www.equator-network.org/) provides reporting guidelines for many types of studies. These guidelines specify the design information that should be reported and can help researchers identify gaps in their reporting. Following the relevant guideline for the study type is a professional expectation for publication. ### Maintain Researcher Identity and Records The [ORCID for Researchers](https://info.orcid.org/researchers) resource describes how researchers can maintain a persistent identifier that links their publications and research activities. Maintaining accurate records of experimental designs and analyses is part of responsible research conduct and supports the reproducibility of the research. ## A Practical Decision Framework for Subsampling and Replication Researchers often struggle to apply the subsampling versus replication distinction in real time, especially when designing studies under time pressure or adapting protocols mid-experiment. A structured decision framework helps standardize these choices and reduces the risk of pseudoreplication creeping into the design. ### The Five-Question Design Audit Before finalizing any experimental protocol, work through these five questions in order. Document the answers in the study protocol so the reasoning is transparent to collaborators and reviewers. **Question 1: What entity received the treatment?** Identify the physical object or organism to which the treatment was applied. This is the candidate experimental unit. Write it down as a specific noun, such as "culture flask," "mouse," "cage of five mice," or "field plot." Avoid vague terms like "sample" or "specimen" because these do not specify the level of treatment application. **Question 2: Could this entity have received a different treatment?** The experimental unit must be capable of receiving any of the treatments in the study. If the entity could only have received the treatment it actually got, it is likely the experimental unit. If the entity could not have received a different treatment because it is part of a larger unit that received the treatment, then it is a subsample. **Question 3: Are the entities independent?** Independence means the treatment assignment and outcome for one entity do not influence another entity. Cells from the same culture split into multiple dishes are not independent because they share a common source. Animals from the same litter housed together are not fully independent because they share genetics and environment. If entities share a common source or are physically connected, they are subsamples of a larger unit. **Question 4: What is the smallest unit I can randomize?** Randomization is the operational test of the experimental unit. The experimental unit is the smallest entity that can be randomly assigned to a treatment. If you can randomize individual wells in a plate to different treatments, the well is the experimental unit. If you can only randomize whole plates, the plate is the experimental unit. **Question 5: To what population do I want to generalize?** The answer to this question determines the level of replication needed. If you want to generalize to all animals of a certain strain, individual animals are the experimental units. If you want to generalize to all cultures prepared from a particular cell line, independent culture preparations are the experimental units. Subsampling cannot extend the inference beyond the units actually studied. ### Decision Tree for Analysis Selection Once the experimental unit is identified, use the following decision tree to select the appropriate analysis. This tree assumes the researcher has already confirmed the experimental unit and collected data with clear unit identifiers. **Branch 1: Balanced design with equal subsamples per unit** If every experimental unit has the same number of subsamples and the research question concerns the treatment effect, average the subsamples for each unit and analyze the unit-level means with a standard t-test or ANOVA. This approach is simple, transparent, and produces correct inference. **Branch 2: Unbalanced design with unequal subsamples per unit** If the number of subsamples varies among units, averaging gives equal weight to units with few and many measurements, which can bias the analysis. Use a mixed-effects model with the experimental unit as a random effect. This model appropriately weights units by their precision while maintaining the correct degrees of freedom for the treatment effect. **Branch 3: Multiple levels of nesting** If the data have more than one level of nesting, such as subsamples within tissue sections within animals, use a mixed-effects model with random effects for each level. Specify the nesting structure explicitly in the model formula so the software accounts for the hierarchy correctly. **Branch 4: Variance components are the research focus** If the research question concerns how much variation exists at each level, such as measurement error versus biological variation, use nested ANOVA or a variance-components model. This analysis partitions the total variation into between-unit and within-unit components. ### The Unit-Mismatch Troubleshooting Method When a study produces unexpected results or reviewers question the analysis, use this troubleshooting method to identify whether pseudoreplication is the cause. **Step 1: Reconstruct the design from the data file** Open the raw data file and identify the columns that represent experimental units and subsamples. If the data file does not contain a column for the experimental unit, the design was likely not documented properly. Reconstruct the design from laboratory notebooks or original records. **Step 2: Count the true replicates** Count the number of distinct experimental units per treatment group. Compare this number to the sample size reported in the analysis. If the reported sample size is larger than the number of experimental units, pseudoreplication is present. **Step 3: Compare the analysis to the design** Check whether the statistical model accounts for the nesting of subsamples within units. If the analysis treats every measurement as an independent observation, the degrees of freedom are inflated and the p-values are too small. **Step 4: Reanalyze with the correct structure** Reanalyze the data using the experimental unit as the level of analysis. If the conclusions change, the original analysis was affected by pseudoreplication. The corrected analysis provides the valid inference. **Step 5: Document the correction** If pseudoreplication is discovered after data collection, document the problem and the correction in the study records. Report the corrected analysis in any publication and explain the discrepancy transparently. The [Committee on Publication Ethics Core Practices](https://publicationethics.org/core-practices) address the responsible handling of data errors and corrections in published research. ### A Worked Example Using the Framework Consider a researcher studying the effect of a novel compound on cytokine production in cultured macrophages. The researcher prepares six flasks of cells from the same source culture, treats three flasks with the compound and three with vehicle, and then measures cytokine levels in four aliquots from each flask. Applying the five-question audit: the flask received the treatment, each flask could have received either treatment, the flasks share a common cell source so they are not fully independent, the smallest randomizable unit is the flask, and the researcher wants to generalize to the cell line. The flask is the experimental unit, and the four aliquots are subsamples. The design is balanced with four subsamples per unit. According to the decision tree, the researcher should average the four aliquots for each flask and analyze the three flask-level means per group with a t-test. The study has three true replicates per group, not 12 or 24. If the researcher instead treats all 12 measurements per group as independent replicates, the analysis will report inflated significance. The unit-mismatch troubleshooting method would reveal this by comparing the reported sample size of 12 to the actual number of experimental units, which is three. ### Common Failure Patterns in Applying the Framework Several recurring mistakes undermine the decision framework in practice. **Skipping the audit for familiar protocols** Researchers who have used the same protocol for years often assume the experimental unit is already established. Protocols can change subtly, such as switching from individual treatment application to pooled treatment application, without the researcher noticing the shift in experimental unit. Run the audit every time the protocol changes. **Confusing the measurement unit with the experimental unit** Instruments often define the measurement unit, such as a well in a plate or a field of view on a microscope slide. The measurement unit is not the experimental unit unless it received the treatment independently. The audit question about treatment application resolves this confusion. **Applying the framework after data collection** The framework is most useful when applied before data collection. Applying it after the fact can identify problems but cannot recover lost replication. If the audit reveals that the experimental unit is larger than assumed, the study may be underpowered and may need to be repeated with more units. **Ignoring the source of biological material** Cells from the same source culture and animals from the same litter share common variation that makes them more similar than independent units. The audit question about independence captures this issue. Researchers should document the source of all biological material and account for it in the design. ### Integration with Reporting Guidelines The decision framework produces the information needed for transparent reporting. The [EQUATOR Network](https://www.equator-network.org/) provides reporting guidelines that specify what design information should be included in publications. The answers to the five audit questions map directly to the reporting requirements: the experimental unit, the number of units per group, the number of subsamples per unit, and the statistical model. The [NIH Grants and Funding](https://grants.nih.gov/) resources describe the expectation that grant applications specify the experimental unit and the replication strategy. Using the decision framework during the planning phase produces the documentation needed for grant applications and for the [NIH Data Management and Sharing Policy](https://sharing.nih.gov/data-management-and-sharing-policy) requirements for reproducible research records. The [National Library of Medicine Research Methods Resources](https://www.ncbi.nlm.nih.gov/books) provides access to methodology references that describe the statistical foundations of the decision framework, including the assumptions underlying nested designs and mixed-effects models. ## Frequently Asked Questions ### What is the difference between a subsample and a true replicate? A true replicate is an independent experimental unit that receives its own treatment assignment. A subsample is a measurement taken from an experimental unit. Multiple subsamples from the same unit are correlated because they share the same source, so they cannot be treated as independent observations in the statistical analysis. ### How do I identify the experimental unit in my study? The experimental unit is the smallest entity that is independently assigned to a treatment. Ask what entity received the treatment. If you applied the treatment to a culture dish, the dish is the experimental unit. If you applied the treatment to an animal, the animal is the experimental unit. Multiple measurements from the same dish or animal are subsamples. ### Why does pseudoreplication produce false precision? Pseudoreplication treats correlated measurements as independent observations. This inflates the degrees of freedom in the statistical test, making the error term smaller than it should be. The smaller error term produces larger test statistics and smaller p-values, which makes the results appear more precise than the design supports. ### Can I average my subsamples and analyze the averages? Yes, averaging subsamples for each experimental unit and analyzing the unit-level averages is a valid approach. This approach is appropriate when the research question concerns the treatment effect on the unit. The analysis should use the number of experimental units as the sample size, not the total number of measurements. ### What is a nested analysis of variance? Nested ANOVA, also called hierarchical ANOVA, partitions the variation into between-unit and within-unit components. The treatment effect is tested against the between-unit variation, which provides the correct degrees of freedom. This approach is appropriate when the researcher wants to estimate both the treatment effect and the components of variance. ### When should I use a mixed-effects model for subsampled data? Mixed-effects models are appropriate when the design is unbalanced, when there are multiple levels of nesting, or when the researcher wants to estimate variance components. The experimental unit is included as a random effect, which accounts for the correlation among subsamples from the same unit. ### How many subsamples should I take from each experimental unit? The number of subsamples should be based on the measurement error and the within-unit variation. More subsamples reduce the impact of measurement error but do not increase the number of independent observations. The optimal number balances the cost of additional measurements against the improvement in precision. ### What should I report in my publication about replication and subsampling? Report the number of experimental units per treatment group and the number of subsamples per unit. Describe the experimental unit in the methods section and specify the statistical model used for the analysis. This transparency allows readers to assess whether the analysis matches the design. ## Related Diagnostic Guides - [Veterinary Radiography and Contrast Studies](/knowledge/diagnostics/imaging/veterinary-radiography) - [RNA Sequencing Protocol: From Sample to Data](/knowledge/diagnostics/molecular/rna-sequencing-protocol-from-sample-to-data) - [Animal Biology](/blog/guides/animal-biology) ## References and Further Reading - [Research Methods Resources](https://www.ncbi.nlm.nih.gov/books). National Library of Medicine. - [EQUATOR Network](https://www.equator-network.org/). EQUATOR Network. - [Core Practices](https://publicationethics.org/core-practices). Committee on Publication Ethics. - [NIH Grants and Funding](https://grants.nih.gov/). National Institutes of Health. - [ORCID for Researchers](https://info.orcid.org/researchers). ORCID. - [Data Management and Sharing Policy](https://sharing.nih.gov/data-management-and-sharing-policy). National Institutes of Health. - [Laboratory Biosafety Manual](https://www.who.int/publications/i/item/9789240011311). World Health Organization. - [NCBI Literature Resources](https://www.ncbi.nlm.nih.gov/). National Center for Biotechnology Information. - [Association of Phosphorylated Tau Biomarkers With Amyloid Positron Emission Tomography vs Tau Positron Emission Tomography.](https://pubmed.ncbi.nlm.nih.gov/36508198). JAMA neurology, 2023. - [25th Annual Computational Neuroscience Meeting: CNS-2016.](https://pubmed.ncbi.nlm.nih.gov/27534393). BMC neuroscience, 2016. - [Genetic signatures of suicide attempt behavior: insights and applications.](https://pubmed.ncbi.nlm.nih.gov/38315076). Expert review of proteomics, 2024. > This article is educational and does not replace validated laboratory procedures, institutional biosafety review, manufacturer instructions, or professional interpretation.