Soil microbial beta diversity describes how detected community compositions differ among samples. If you are trying to interpret Bray–Curtis dissimilarity, principal coordinates analysis (PCoA), non-metric multidimensional scaling (NMDS) or a PERMANOVA table in a soil microbiome study, first determine what the distance between samples represents. A close pair of dots can be meaningful only after the distance metric, feature data and plotting method have been identified.
Beta diversity is not the same as the number of microbial species within a sample. Nor is a statistically separated cluster proof that one soil treatment caused a biological change. The result depends on marker or genomic features, library processing, transformations, experimental design, the distance calculation and how much variation a two-dimensional plot retains. Different methods emphasise presence, abundance or evolutionary relatedness.
This guide explains how to read soil microbiome beta-diversity results with practical examples: feature tables, Bray–Curtis and Jaccard distances, PCoA axes, NMDS stress, group centroids, PERMANOVA effect sizes, dispersion tests and cautious interpretation. All numerical illustrations are teaching examples rather than observations from a named field experiment. The purpose is to turn a complex plot and statistical table into statements a reader can verify.
The quick answer: what beta diversity measures
Beta diversity concerns compositional differences between biological samples, based on selected features and a declared mathematical distance or dissimilarity. The data may come from bacterial 16S rRNA amplicons, fungal ITS markers or another defined feature matrix. A distance of zero can indicate identical feature profiles under a specific measure; a larger distance can indicate greater measured dissimilarity. The actual meaning of ‘larger’ depends on the chosen distance formula and data scale.
A two-dimensional ordination is usually a visual summary of many pairwise distances. A PERMANOVA table is a statistical analysis of differences in the multivariate community data under a specified model. These tools are related but not interchangeable. Neither directly reports living microbial cell counts, metabolic fluxes or the causes of compositional differences. Describe detected sequence composition before making broader ecological claims.
Start with the sampling design
A study may compare independent plots under two management regimes, several soil depths or repeated measurements from the same field. These designs require different interpretations. Three libraries made from one pooled soil extract are not equivalent to three independent field plots. The unit of inference should be the sampling unit from which biological replication actually arises, not simply the number of columns in a sequencing table.
Record site identity, depth, season, plant cover, field treatment, extraction batch and sequencing run. Ask whether groups were randomly assigned or observationally selected. If all treated samples come from one location and all controls from another, treatment and location are inseparable in the data. No spatial separation in an ordination can make that confounding disappear. The most important statistical decision often precedes the first distance calculation.
Beta is not alpha diversity
Alpha diversity describes the detected variety within a single sample through metrics such as richness, Shannon or Simpson. Beta diversity describes how whole feature profiles differ across samples. Two soils can share identical alpha diversity yet contain completely different ASVs. Conversely, one soil may contain more observed features than another while still sharing most of the dominant sequence profile.
The distinction should appear clearly in a scientific report. If Shannon diversity remains stable while PCoA separates treatment and control, those observations can be compatible. Community identity and within-sample diversity are different dimensions of the data. Use the alpha-diversity guide to interpret the within-sample question; this article addresses between-sample compositional differences.
The feature table is the starting evidence
An amplicon feature table commonly records how many sequencing reads were assigned to each ASV in each library after quality filtering and denoising. It may be transformed into proportions or presence–absence indicators before distance calculation. A shotgun metagenomic table may represent gene families or inferred taxonomic profiles instead. Distances depend on what counts as a feature, so the table’s identity cannot be omitted.
A presence–absence comparison ignores relative feature counts, while an abundance-weighted comparison responds to their distribution. An ASV and a microbial species are not guaranteed to be equivalent units. If readers need the earlier steps, the soil ASV table guide explains denoising, taxonomy and compositional reads. Do not describe an ordination as ‘the entire microbiome’ when it represents one selected marker and pipeline.
What a distance matrix looks like
Suppose samples A, B and C produce pairwise dissimilarities of 0.15 for A–B, 0.70 for A–C and 0.62 for B–C under a declared metric. A distance matrix stores these pairwise values, normally with zero on the diagonal for comparisons of a sample to itself. A and B are the most similar pair on this measure. That is all the matrix establishes by itself.
Distances are not physical metres between sampling points unless the metric is explicitly geographic. They describe differences in the chosen feature representation. A high value is not a moral judgement that one soil is worse. The correct question is what mathematical ingredients created the dissimilarity, whether technical artefacts might influence it and whether the comparison was planned before the graph was examined.
Bray–Curtis dissimilarity explained
Bray–Curtis dissimilarity is often applied to nonnegative community abundance profiles. One common form compares the sum of absolute differences in feature values between two samples with the sum of their feature values. With comparable nonnegative inputs, it ranges from zero for identical profiles to one for profiles sharing no positive-valued features. Its ecological interpretation depends on whether the inputs are raw counts, proportions or transformed values.
It is sensitive to abundance structure, but not to phylogenetic relatedness unless the features or weights explicitly incorporate such information. A shift among dominant groups can strongly affect Bray–Curtis even if many rare ASVs are shared. When a figure simply says ‘Bray distance,’ check which transformation was applied and how zeros were treated. A dissimilarity name without the preceding table transformation is an incomplete methods description.
A worked Bray–Curtis calculation
Take two illustrative samples with features X and Y. Sample A has counts of 30 and 10, while sample B has 10 and 30. The absolute differences are 20 and 20, totalling 40. The combined counts are 30 plus 10 plus 10 plus 30, or 80. Bray–Curtis dissimilarity is therefore 40 divided by 80, equal to 0.50. This is a comparison of the provided feature vectors, not an estimate that half the organisms differ.
If both samples instead have the same values for each feature, the numerator is zero and the dissimilarity is zero. The calculation illustrates why a method can detect redistribution of feature counts even when total reads are equal. It also reveals a caution: unequal library totals can influence a raw-count calculation. State whether the researcher adjusted depths or transformed counts before applying the formula.
Jaccard distance answers a different question
Jaccard similarity compares shared present features with the union of features observed in either sample. Jaccard distance can be written as one minus that similarity. It is commonly used with presence–absence data, so it does not directly account for the proportion of reads assigned to each feature. A rare detected ASV can contribute to overlap much like a dominant one, subject to the chosen data treatment.
If sample A contains features X, Y and Z while sample B contains Y and Z, there are two shared features among three observed in the union. Similarity is two-thirds and Jaccard distance is one-third. These simple counts are affected by detection sensitivity and filtering. At low sequencing depth, a true rare feature might appear absent, artificially increasing a presence–absence distance. The metric’s biological limitations follow from its definition.
UniFrac uses a phylogenetic tree
UniFrac families compare microbial communities using branch lengths on a phylogenetic tree built for the detected features. Unweighted UniFrac emphasises presence–absence along tree branches; weighted variants incorporate abundance information under their specified formula. This makes phylogenetic context part of the comparison, unlike standard feature-wise Bray–Curtis or binary Jaccard applied without a tree.
Tree construction and marker suitability matter. A poorly supported alignment or branch-length estimate can influence the resulting distances. Do not apply a tree-based label to a dataset that lacks a defensible phylogeny. A greater weighted UniFrac distance is a property of the analysed sequence distribution and tree, not direct evidence that entire soil ecosystems are separated by an equivalent evolutionary or functional distance.
Compositional transformations change the geometry
Microbiome proportions are constrained to sum to a whole. An increase in one group changes the percentages of others, even if some absolute quantities remain fixed. Analysts may use transformations or models designed for compositional data, including log-ratio approaches under suitable assumptions. The resulting distances can differ from raw Bray–Curtis or binary Jaccard because they represent different features of the same observed table.
In log-ratio work, zeros require careful treatment; simply adding an arbitrary tiny constant can materially change results for sparse data. There is no universal transformation that solves every dataset. The reader’s job is to identify what was transformed, why and whether conclusions are robust to reasonable alternatives. Never compare two ordination maps as equivalent merely because each contains colourful sample dots.
Why library-size decisions matter
A library containing 80,000 retained reads and another containing 8,000 reads may differ in rare-feature detection for technical reasons. Some workflows rarefy to a common depth; others use transformations or models that address depth in different ways. Rarefaction can discard reads and exclude low-depth samples, while alternative methods also have assumptions and sensitivity to composition.
Read how the analysis handled depth before interpreting a cluster as ecological. If all control samples have low-quality libraries, method effects can align with treatment labels. A report should disclose thresholds, exclusions and retained sample counts. Ask whether the result holds under appropriate sensitivity checks. A clear ordination drawn from a biased subset may be less persuasive than an imperfect-looking figure with transparent sampling and uncertainty.
PCoA means principal coordinates analysis
Principal coordinates analysis takes a distance matrix and represents samples along axes intended to summarise those distances in a coordinate space. It is often used with Bray–Curtis, Jaccard or UniFrac distances. The horizontal and vertical axes are usually labelled PCoA1 and PCoA2, sometimes with percentages showing how much variation or positive-eigenvalue structure is represented under the chosen procedure.
PCoA axes are not soil pH, time, nutrient concentration or other original field measurements unless a separate procedure explicitly establishes that relationship. They are mathematical coordinates derived from the distance matrix. A sample plotted at a negative coordinate does not have negative biodiversity. Axis signs can flip without changing the underlying pairwise geometry. Focus on relative distances and the amount of structure captured rather than treating axis direction as an ecological cause.
A two-dimensional map does not preserve every distance
Most community tables contain many features, and their relationships need not fit perfectly into two dimensions. PCoA represents selected dimensions of a higher-dimensional distance structure. Points that appear close or far on the first two axes may not reflect their complete original pairwise relationships when substantial variation lies on other axes. Labels giving explained variation help readers judge how much of the structure is visible.
If PCoA1 and PCoA2 capture only a modest proportion, avoid overinterpreting small separations. Examine additional axes, a distance heatmap or the underlying matrix when a claim is important. Some dissimilarities are non-Euclidean and can yield negative eigenvalues, complicating straightforward percentage interpretations. Software may apply mathematical corrections; such choices should be reported rather than concealed.
Reading the percentage beside a PCoA axis
A label such as ‘PCoA1 28%’ is commonly intended to indicate how much of the represented distance structure is associated with that coordinate under the chosen algorithm and conventions. It does not mean that 28% of microbes are explained by the axis, that a treatment caused 28% of a biological effect or that the chart has 28% accuracy.
Ask how percentages were calculated, especially if negative eigenvalues or transformations are involved. The first coordinate captures a prominent mathematical contrast, but its biological meaning requires an association with measured metadata and a suitable design. A coloured treatment grouping is an overlay of labels, not proof that treatment created the axis. Interpret coordinates as summaries first and consider ecology only with additional evidence.
NMDS uses ranks of dissimilarities
Non-metric multidimensional scaling aims to represent the rank order of dissimilarities in a small number of dimensions rather than reproducing exact distances linearly. It often begins with iterative optimisation and can have local solutions, so software settings, repeated starts and random seeds matter. Axes on an NMDS plot do not have inherent ecological units or fixed orientation; maps can rotate or reflect without changing their essential rank relationships.
The most important information is the relative neighbourhood pattern and the reported goodness of fit. Do not compare numerical NMDS1 coordinates across independently fitted plots as if they share a physical scale. It is possible for two studies to have visually similar clusters while using different distance metrics or feature processing. Their resemblance is not evidence that the underlying ecosystems changed in identical ways.
NMDS stress tells you about distortion
NMDS stress summarises disagreement between original dissimilarity ranks and the distances in the reduced ordination. Lower stress generally indicates a better representation under the fitted model, but exact interpretation depends on the stress definition, dimension and data. Informal rules of thumb circulate widely; they should not replace inspection of the method and the actual fit diagnostics.
A highly stressed two-dimensional NMDS may give a misleading impression of who is close to whom. Some analyses add dimensions to improve fit, while others inspect Shepard plots showing observed dissimilarities against fitted ordination distances. State stress and dimension count with the figure. Without them, the reader cannot judge whether attractive cluster shapes faithfully reflect the original dissimilarity structure.
Why NMDS axis direction is not a biological variable
If a group of samples appears on the right of an NMDS plot, that does not automatically mean it has more carbon, nitrogen or a ‘higher’ microbial state. Ordination configurations can rotate, reflect and translate while preserving relevant relationships. Axis labels are coordinates in the fitted representation, not causal gradients established by the analysis alone.
When measured environmental variables are overlaid as vectors, they describe associations with the ordination under the chosen fitting method. They do not convert the axes into those variables. A visually aligned arrow may offer a hypothesis about an environmental gradient, but confounding and sampling design remain relevant. Explain association separately from cause to keep the graph scientifically interpretable.
Group centroids and confidence shapes
Some plots mark a centroid for each treatment group, often representing a central location in ordination space. Ellipses or hulls may illustrate variation, confidence regions or simple geometric ranges. The meanings differ. A convex hull contains the observed points used to draw it, while a confidence ellipse is typically based on a statistical construction with assumptions. A standard-deviation ellipse represents something else again.
Never infer formal significance from overlapping or nonoverlapping shapes without understanding how they were generated. Small sample groups may produce unstable boundaries, and some plotting functions require more observations than are available. Read the caption carefully. An ellipse that looks precise can be the product of a model or graphic convention rather than a direct measure of certainty about community difference.
PERMANOVA asks about multivariate group differences
Permutational multivariate analysis of variance, commonly PERMANOVA, evaluates variation in a distance matrix in relation to group labels or explanatory variables under a specified model. A statistic analogous to an F ratio is evaluated using permutations consistent with the experimental design. Software such as the vegan package’s adonis2 reports pseudo-F, an R-squared-style proportion and a permutation-based p value.
A small p value indicates that the observed association is unusual under the selected null model and permutation procedure. It does not prove causality, biological benefit or a specific mechanism. The mathematical result is conditioned on the distance definition, transformation, covariates and allowable permutations. A responsible description includes these choices, rather than writing simply that ‘the soils were significantly different.’
What PERMANOVA R-squared means
The R-squared value in PERMANOVA describes a proportion of variation in the chosen distance-based representation attributed to model terms under the method. A value of 0.12 might be described as approximately 12% of the analysed multivariate variation in that particular model, but it is not ‘12% of species changed’ or ‘12% of soil health caused by treatment.’
Model order, covariates and whether terms are evaluated sequentially or marginally can affect the interpretation. In multi-factor designs, some reported partial R-squared values do not simply add to a full total in the way a casual reader expects. Read the statistical model and the software’s documentation. A good caption names the metric, predictors and explanatory scope rather than presenting R-squared as a universal ecological percentage.
A significant p value is not a mechanism
Suppose a PERMANOVA gives p equals 0.01 for differences among management groups. This provides statistical evidence of association under the specified test, not proof that management caused the change. Field conditions, sampling site and laboratory batch could be confounded. The size of the effect and group distributions still matter. A tiny effect can be statistically detected with a large dataset without demonstrating ecological importance.
Translate the result as ‘community dissimilarities were associated with management group under the stated model and permutation design.’ If the study was a randomised controlled experiment and other assumptions hold, stronger causal language may be supported. Even then, the particular microbial mechanism requires additional evidence. Treat statistical separation as a starting point for interpretation, not as permission to invent biology.
Dispersion can imitate separation
Groups may differ because their central community compositions are separated, because one group contains more varied samples, or because both are true. PERMANOVA can be sensitive to differences in within-group dispersion. A plot with one tight cluster and one wide scattered group may yield a group-related result that needs careful interpretation. Centroid shifts and dispersion changes are distinct ecological patterns.
The vegan package provides betadisper and related permutation procedures to examine multivariate dispersion, often through distances of samples from group centres. Report this diagnostic when relevant, particularly if PERMANOVA conclusions depend on uneven group variability. A significant dispersion test does not automatically invalidate every PERMANOVA, but it warns against casually interpreting all significance as a difference in group centres.
A worked example with group spread
Imagine two treatments sampled at four independent plots each. Treatment A’s samples cluster tightly in ordination space, while Treatment B’s samples are widely scattered around a similar centre. A figure might make B look less consistent and more heterogeneous. This could represent true within-treatment community variation rather than a clean shift to a different average composition.
A suitable report would examine group centre and dispersion instead of concluding that every B sample moved in one direction. If a permutation test flags differences, inspect dispersion diagnostics and individual pairwise distances. The lesson is conceptual: two groups can differ in the spread of their observations without being separated by an equivalent shift in their typical community profiles. Calling both simply ‘changed composition’ hides an important distinction.
Permutation exchangeability is a design question
Permutation testing works by rearranging labels or residual information under a null model, but not every rearrangement is valid for every study. Repeated samples from the same field, paired plots, blocking and spatial dependence can restrict which observations are exchangeable. Freely permuting all samples may break a study’s dependence structure and produce misleading p values.
Read how permutations respected the design. If sampling dates are repeated within site, a test that ignores site pairing can misstate uncertainty. Small datasets have a limited number of distinct valid permutations, which affects p-value resolution. Rather than reporting a test as ‘nonparametric and therefore assumption-free,’ explain the relevant assumptions about independence, exchangeability and the relationship between the distance measure and the question.
Multiple factors can be entangled
A field study may record soil pH, moisture, organic carbon, management and geographical location. These variables often covary. A PERMANOVA model can include multiple terms, but the interpretation depends on their relationships and the order or testing mode used. Sequential tests can allocate shared variation according to term order, while marginal tests ask different adjusted questions.
Avoid attaching a unique causal explanation to a variable merely because it appears next to a significant p value. Examine collinearity, sampling balance, planned hypotheses and possible confounders. The best biological explanation may involve several linked processes that the design cannot fully separate. A thoughtful report preserves both the observed multivariate association and the limits of individual predictor attribution.
Ordination can be affected by a few dominant features
Bray–Curtis calculated from abundance data may be strongly influenced by common or highly variable features. An unusual sample dominated by one sequence group can pull relationships in a visual display. Presence–absence distances treat rare and dominant features more equally but become more sensitive to detection limits. There is no universally correct choice; each metric reveals a different aspect of observed composition.
Try to understand which biological question each metric answers. For readers, a sensitivity analysis with a second defensible distance can show whether a broad pattern is stable or metric-specific. Disagreement between metrics may be informative, not an error. Do not selectively display only the metric that produces the strongest colour separation without discussing why that measure was chosen.
Why taxonomic aggregation changes the story
An ASV-level table may contain many different variants within the same genus. Collapsing all features to genus or phylum can reduce apparent differences between samples even if their ASV compositions diverge. Conversely, broad grouping can make a dominant phylum difference visually striking while hiding important fine-scale stability or turnover. The aggregation decision belongs in the methods and caption.
An ordination of phylum proportions is not the same analysis as an ordination of unaggregated ASV counts. Species labels may also be uncertain for short markers. Do not call one representation ‘the true community’ without a reason. Readers should compare the analytic resolution with the question: broad composition, detailed lineage changes or functional potential may each require a different feature definition.
Distance-based results are not absolute cell quantities
A microbiome library can change composition even if total microbial biomass is stable. It can also remain compositionally similar while absolute microbial numbers rise or fall. Beta diversity, by design, focuses on comparisons among specified feature profiles. A distance of 0.6 does not mean 60% of soil microorganisms disappeared, nor does it identify grams of carbon or nitrogen exchanged.
If absolute abundance matters, independent assays and calibration are needed. The soil qPCR guide discusses target copies per soil mass, while microbial biomass estimates provide another measurement framework. Use them to formulate complementary questions, not to treat one measure as an exact substitute for another.
Experimental batch effects can create visible clusters
Suppose all samples from the first field were extracted on Monday and all samples from the second field on Friday, using different reagent lots. If the ordination separates fields, the analysis cannot by itself tell whether the cause is true field ecology or batch-associated differences. Sequencing-run differences, storage delays and extraction workflows can all create structure in high-dimensional biological data.
Good design balances important groups across laboratory batches when practical, includes relevant controls and records processing metadata. Analysis may adjust some known effects, but adjustment cannot rescue a completely confounded comparison. When reading a beautiful PCoA figure, check whether technical labels were plotted as well as biological labels. If batch and treatment cluster identically, a cautious conclusion is essential.
Time-series and spatial autocorrelation
Soil communities vary across seasons and distances. Samples collected near each other may be more alike than distant samples for geographical reasons, and repeated observations from a plot share history. Standard permutations or statistical models that assume independence can be inappropriate. A soil depth transect also has ordering that should not be treated as arbitrary labels.
Map the sampling design before interpreting clusters. Ask whether site, spatial position, date and repeated-measure structure were recorded and incorporated appropriately. If the study sampled a gradient, a continuous association may be more informative than dividing points into arbitrary categories. Ecological meaning comes from matching the mathematics to the geography and timing of actual soil samples.
Environmental vectors show associations
Ordination plots sometimes contain arrows labelled pH, moisture or carbon. These may be fitted vectors indicating correlation or association between an environmental variable and the ordination configuration. Their direction and length depend on scaling, method and underlying observations. They do not mean the microbial community literally travelled along the arrow, nor that the variable was experimentally manipulated.
Before describing an arrow, read the fitting method, statistical support and units of the external measurement. If variables are strongly correlated, multiple arrows may align without establishing independent causal pathways. An explanatory sentence should say ‘community composition was associated with soil pH in the fitted analysis’ rather than ‘pH produced all the observed separation.’ Good scientific writing preserves the distinction between an overlay and a measured cause.
Reading a PERMANOVA table in plain language
Start with the response distance matrix and its transformation. Identify every predictor, the reported degrees of freedom and the definition of the test statistic. Check R-squared or partial R-squared and the number and structure of permutations. Then inspect the p value together with group sizes, dispersion and sample plots. A table with strong statistical evidence but tiny R-squared invites a different ecological discussion from one with a large explained share and wide uncertainty.
Do not report only the smallest p value among several exploratory tests. The process of choosing and testing metrics can create selection effects. If many hypotheses were examined, appropriate adjustments or clear exploratory labelling may be needed. The table should support the statement, not serve as a decoration after a preferred narrative has already been written.
A complete reading exercise with fictional data
Imagine twelve independent soil plots, six in each of two treatments, sampled in one season. A Bray–Curtis PCoA shows some overlap but a visible shift in group locations. PERMANOVA reports a model R-squared of 0.14 and p of 0.03 under an appropriate permutation design. A dispersion analysis reports no compelling evidence of different spread, though with limited power. What can the reader say?
A careful answer is that treatment groups are associated with a detectable but partial difference in the analysed sequence-based community composition. The model accounts for about 14% of the chosen multivariate variation under its conventions, leaving much variation unexplained. Without random treatment assignment or other causal support, do not assert treatment alone caused the pattern. Neither test measures soil fertility, metabolic activity or beneficial organisms directly.
Communicating the limits without burying the result
Readers deserve a clear takeaway, not a wall of caveats. Begin with the observed pattern, such as ‘The analysed soil community profiles differed among sampled groups.’ Follow with a single precise statement identifying feature type, distance and statistical result. Then give the most consequential limitation: confounding, sample size, dispersion, library effects or incomplete dimensional representation.
The goal is not to sound uncertain about everything. It is to be specific about what is secure and what is still hypothetical. One sentence can preserve both: ‘The Bray–Curtis analysis found an association between management and detected ASV composition, but the field design does not establish a causal mechanism.’ Good writing carries the scientific boundary into ordinary language without draining the subject of interest.
Common mistakes and practical repairs
Mistake one is calling PCoA1 an ecological variable. Repair it by explaining that it is a mathematical coordinate. Mistake two is treating NMDS axis signs as direction of soil improvement. Repair it by focusing on sample relationships and stress. Mistake three is treating a significant PERMANOVA as proof of group-centre differences without checking dispersion. Repair it by examining within-group variability and the sampling design.
Mistake four is equating Bray–Curtis 0.6 with 60% species loss. Repair it by stating the formula and feature basis. Mistake five is ignoring library depth and extraction batch. Repair it by reviewing preprocessing and controls. Mistake six is treating dependent repeated samples as independent. Repair it with design-respecting models or restricted permutations. Each repair starts by restoring a missing definition rather than by adding stronger language.
A translation glossary for scientists and learners
‘Distance’ here is a numerical comparison of profiles, not necessarily geographic separation. ‘Ordination’ is a compact coordinate representation of multivariate relationships, not an ecological ranking. ‘Stress’ in NMDS is model mismatch, not biological distress. ‘Variance explained’ is a property of a defined statistical model and distance framework. ‘Significant’ is not automatically ‘important,’ and ‘associated’ is not synonymous with ’caused.’
These distinctions matter when reading studies in English or translating them for another language community. The cross-linguistic transfer guide explains why apparently familiar terms can drift across contexts. The translation and meaning guide encourages the reader to preserve the relationship between evidence and claim. Statistical vocabulary should carry its own operational definitions.
Frequently asked questions
Does a PCoA cluster prove the soils came from the same biological community? No; it shows relative proximity under a selected distance and projection. Is NMDS better than PCoA? Neither is universally superior; they optimise different relationships and require different diagnostics. What does low NMDS stress mean? Generally a better rank-distance representation under the reported stress convention. Does PERMANOVA p below 0.05 prove management caused a change? No; causal interpretation depends on design and assumptions.
Can groups differ in beta diversity while Shannon scores remain unchanged? Yes. Can unequal dispersion affect PERMANOVA? Yes, and it warrants diagnosis. Is Bray–Curtis a percentage of lost microorganisms? No. Is the first PCoA axis a direct pH measurement? No. Should plots be interpreted without sample metadata? No; the design, controls and measured environment are needed to make ecological sense of the geometry.
Practice questions with explained answers
Sample A has 30 reads of X and 10 of Y; sample B has 10 of X and 30 of Y. What is Bray–Curtis dissimilarity for those illustrative count vectors? It is 40 divided by 80, equal to 0.5. Does that mean half the organisms died? No; it describes the difference between supplied feature values under the metric. Now suppose an NMDS map is rotated ninety degrees but interpoint relationships remain. Has the biology changed? No; coordinate orientation is arbitrary.
Another exercise: treatment groups produce a small PERMANOVA p value, but one group is much more dispersed. What should be checked before declaring different typical compositions? Examine dispersion diagnostics, individual observations and design-respecting permutations. Finally, two datasets show similar PCoA layouts but use different sequencing markers and distance formulas. Can their axis coordinates be compared directly? No; the feature spaces and derived coordinates are not common measurement units.
Further sources and reading routes
The QIIME 2 diversity documentation explains distance measures and ordination workflows. The vegan adonis2 reference describes distance-based PERMANOVA, model terms and permutation interpretation. The vegan betadisper reference explains multivariate dispersion checks. These are methodological definitions that should be read with the experimental design.
Continue through eduKateSG’s complementary guides: within-sample alpha diversity, amplicon ASV tables and shotgun metagenomics. Each explains a different part of the measurement chain; their roles should be complementary rather than merged into one unqualified claim.
Closing lesson: geometry is evidence, not explanation
Beta diversity reveals structured differences among measured community profiles that would be hard to see by reading thousands of feature counts individually. Distance matrices, PCoA, NMDS and PERMANOVA can expose real patterns and sharpen ecological questions. Their clarity depends on transparent definitions, controlled sampling, valid statistical assumptions and realistic descriptions of uncertainty.
The final task is translation. State what features were analysed, how samples were compared, how reliably groups differed and which alternative explanations remain. That preserves the genuine insight of a complex statistical result without pretending that two axes, a cluster colour or a p value can independently prove microbial activity, ecosystem benefit or causation.