Soil indicator species analysis asks which detected organisms or sequence features are consistently associated with a particular habitat, treatment or group of sampling sites. If you are reading an IndVal result, an IndVal.g score, an indicspecies multipatt table, specificity, fidelity or a permutation p value, begin by identifying the groups, the features and the exact association index rather than treating every highlighted taxon as an ecological indicator.
An indicator ASV is not automatically a beneficial microbe, a diagnostic species or proof of a causal soil process. The statistical analysis looks for an association between a detected feature and one or more predefined groups of samples. Such an association can arise because environmental conditions differ, because sites share a history or because sampling and laboratory procedures differ. A label such as ‘indicator of restored soil’ is therefore a statement about a study-defined classification, not a certificate that the feature creates restoration.
This guide explains soil indicator-taxon interpretation from the sample table to the final biological claim. You will learn how specificity and fidelity work, how IndVal and IndVal.g are related, why group combinations matter, how permutation tests obtain p values, why multiple testing requires caution and how to communicate the findings without confusing sequence detection with a functional mechanism. Every numerical example is illustrative and designed to be independently checked.
The immediate answer: what counts as an indicator?
An indicator feature is a measured entity whose pattern of occurrence or abundance is strongly associated with a chosen group of sites according to an explicit method. In microbiome studies, the feature could be an amplicon sequence variant, an operational taxonomic unit, a genus-level aggregation or an independently measured organism. The study design determines what ‘group’ means: soil depth, land use, management, season, pH class or another declared classification.
A useful interpretation names both sides of the association. ‘ASV-14 was more specifically and consistently detected in the wetter soil group under this study’s grouping and index’ is defensible if supported. ‘ASV-14 causes wet soil’ is a causal claim that the index does not establish. The difference is the centre of the topic. Indicator analysis describes pattern reliability with respect to chosen groups; ecology explains the circumstances of that pattern through additional evidence.
First understand the sampling groups
Indicator statistics do not decide which labels are scientifically appropriate. A researcher supplies a grouping of field plots or samples, and the analysis looks for taxa associated with those groups. If samples are labelled ‘healthy’ and ‘unhealthy’ without independently validated criteria, the indicator result cannot repair that definition. It will simply identify associations with the labels that the investigator supplied.
Read how groups were created and whether the categories existed before examining sequencing results. A grouping based on observed microbial composition can make an indicator analysis partly circular if it is then presented as an independent test of the same structure. Record sample size per group, biological independence, collection dates, depths, environmental covariates and whether group categories overlap in real life. Good interpretation starts with clear, defensible group definitions.
An ASV is not the same as a named species
Many soil indicator analyses begin with an amplicon sequence variant table derived from 16S rRNA or fungal ITS sequencing. An ASV is a computationally inferred marker-sequence feature. It may correspond to more than one closely related species, while a species can sometimes contain multiple marker variants. Taxonomic classifiers assign labels under the information available, not by cultivating and observing every represented organism.
When a paper calls an ASV an ‘indicator species’, check whether the species identification was independently established. If the sequence supports only a family or genus classification, preserve that limitation. The eduKateSG soil ASV interpretation guide describes the upstream read counts, taxonomic uncertainty and compositional denominator. Those conditions remain important even after an association score is calculated.
Relative abundance is a constrained measurement
Amplicon read counts do not directly count living cells. Their relative proportions are affected by sequencing depth, primer amplification, extraction, marker-copy number and processing. A feature that represents 12% of reads in one group and 5% in another can be an interesting compositional signal, but that contrast does not prove a corresponding change in cell density. Indicator metrics based on abundance inherit these limits.
Presence–absence indicators avoid some proportional comparisons but have their own sensitivity problem: a feature may go undetected when read depth is low, even if it is present in the original soil. Before running or interpreting the indicator analysis, ask how low-depth samples and rare features were handled. A strong statistical association on a badly biased input table will still be a distorted description of the underlying ecology.
What the original Indicator Value idea combines
The classic Indicator Value idea combines two intuitive ingredients. Specificity describes how strongly the species’ abundance is concentrated in the target group rather than other groups. Fidelity describes how consistently it occurs among sampling units inside that group. An ideal indicator is concentrated in the target category and found in most or all of that category’s sites, rather than appearing sporadically.
These concepts complement each other. A feature restricted to one soil treatment but detected in only one of its twenty plots may be highly specific but have low fidelity. Another feature found in every treatment plot and every control plot may have high fidelity within treatment but poor specificity. Both ideas are needed before calling a feature a useful identifier of the group.
Specificity asks whether the group is distinctive
In one common formulation of IndVal, specificity is related to the mean abundance of a feature within a target group divided by its mean abundance across the groups being compared, under a clearly defined calculation. If the feature is common in every group, specificity to any one group is low. If it is concentrated mainly in the target group, specificity approaches its upper bound.
This summary assumes the same group conventions and implementation rules as the method being used. Extensions can adjust how group size is handled. Do not reverse-engineer a printed score without reading the software’s definition. In a three-group experiment, values depend on all groups included; adding a new habitat category can change an association even when the original samples remain identical.
Fidelity asks how regularly a feature appears
Fidelity is commonly calculated as the fraction of sites in the target group where a feature is detected under the analysis’s presence rule. If a bacterial ASV is observed in eight of ten sampled treatment plots, its fidelity is 0.8. That is a frequency of detection among sampled units, not a claim that the ASV occupies 80% of the soil volume or 80% of the living microbial cells.
Read the detection threshold carefully. A zero may mean no reads after filtering rather than verified biological absence. If the sequencing depth differs among groups, fidelity estimates may be distorted through unequal opportunities to detect rare features. A report should distinguish observed fidelity under the assay from a universal probability that a taxon exists in a habitat.
An easy specificity-and-fidelity example
Imagine two groups, A and B, each containing ten independent soil plots. A target ASV is detected in all ten plots of A and in none of B, under the same assay. In an idealised setting its fidelity for A is one and its specificity to A is also maximal. It therefore appears to be an excellent indicator of the group distinction in that dataset.
Now imagine a second ASV appears only in one of the ten A plots and in no B plots. Its strong exclusivity may yield high specificity, but its low occurrence frequency makes fidelity only 0.1. It could be useful for understanding a rare ecological niche, yet it would be a poor universal screening marker for identifying an arbitrary plot as A. The contrast explains the two components better than a single score does.
IndVal and IndVal.g require attention to definitions
The original IndVal formulation and its extensions are related but not identical in every detail. The indicspecies R package’s multipatt function commonly uses IndVal.g by default. Its documentation explains that the returned association statistic for this setting is the square root of the underlying IndVal-style product. A score printed by multipatt should therefore be interpreted under that documented transformation, not silently treated as the untransformed A times B.
This distinction matters in arithmetic. If specificity A equals 0.81 and fidelity B equals 0.64, their product is 0.5184 and its square root is 0.72. A displayed value of 0.72 in a square-root convention is not evidence that A multiplied by B equals 0.72. The underlying interpretation remains a combination of specificity and fidelity, but the numeric scale must be named.
A checked numerical IndVal illustration
Suppose for a particular target group a feature has calculated specificity A of 0.75 and fidelity B of 0.64. The underlying product A times B is 0.48. Its square root is approximately 0.693. If a study reports the square-root form, a value near 0.69 would be compatible with this illustrative pair of components. If it reports the product instead, the corresponding value is 0.48.
The example is a mathematical demonstration, not a reconstruction of the exact internal calculation for every indicspecies option. Values of A and B are themselves determined by the selected groups, observations and method. Always check whether the output uses IndVal, IndVal.g, r or r.g before applying a numeric interpretation. Different association functions should not be compared as interchangeable percentages.
Why a strong association need not be frequent everywhere
An indicator is defined relative to the target group in the study. A taxon might be common in one soil horizon in a particular region yet rare in a similar horizon elsewhere. Its indicator value can depend on climate, vegetation, mineralogy, season, collection depth or methods. The label does not travel automatically across ecosystems.
A strong indicator in a local study is useful as a local pattern. Transferring it requires testing the association on independent samples representing the new conditions. Do not turn ‘indicator of this sampled group’ into ‘global marker of fertile soils’ without evidence. External validation is not just a desirable extra; it defines whether the indicator is useful outside the original experiment.
Group combinations can reveal shared associations
Multilevel pattern analysis can test associations with individual site groups or combinations of groups. A feature may occur strongly in both moderately wet and very wet categories while being rare in dry sites. Under a suitable analysis, a combination of the two wetter groups might be the strongest indicator target. This can be ecologically informative, but it complicates quick summaries.
Read the displayed group-combination coding. A result assigned to groups 2 and 3 is not necessarily an indicator of group 2 alone. Some multipatt options restrict the maximum combination size or analyse only single groups. Different configuration choices change the search space and potentially the selected best association. An honest caption spells out the actual group combination represented by each indicator.
Combinations create a selection problem
When the software searches many candidate combinations, it may pick the strongest apparent association for each feature. A high selected score can partly reflect the fact that many targets were considered. The permutation procedure and its null hypothesis should reflect the way the association was chosen. Readers should check which group combinations were allowed and whether that decision was made before inspecting the outputs.
If a researcher tests every possible soil grouping and reports only the most exciting one, the result is exploratory unless selection is adequately accounted for. A restricted, biologically motivated set of groups can sometimes produce more interpretable outcomes. The purpose is not to forbid exploration, but to describe what was searched so the statistical evidence is not overstated.
What a permutation test asks
A permutation test evaluates how unusual the observed feature–group association would be under a null arrangement consistent with the experimental design. It generally rearranges group labels or uses another permitted permutation scheme and recalculates the association. The proportion of randomised outcomes at least as extreme as the observed one helps estimate a p value under the selected testing convention.
A small p value does not certify that the taxon causes the habitat condition. It indicates evidence against a specified lack-of-association pattern, conditional on the sampling and permutation assumptions. If sites are spatially clustered, paired or repeatedly observed, arbitrary shuffling may break the design. A valid analysis must respect which samples are exchangeable under its null hypothesis.
Permutation count limits p-value resolution
If a test uses 999 random permutations and a conventional plus-one p-value calculation, the smallest attainable estimated p value may be 1 divided by 1,000, or 0.001, if no permuted result is at least as extreme as the observed statistic. Software conventions and exhaustive versus sampled permutations can differ, so the exact calculation should be read from the method rather than assumed.
Reporting a result as p equals zero is usually inappropriate for such a finite sampled permutation procedure. It merely means none of the tested random arrangements was as extreme under that implementation. More permutations can improve the resolution of very small p estimates, but they do not cure confounding, dependence or inappropriate experimental units. Design quality matters before permutation quantity.
Biological replicates are not sequencing reads
A sequencing experiment might contain 100,000 reads per sample but only three independent field plots in each soil group. The statistical sample size for a treatment comparison is related to the independent plots, not the number of reads. Three technical libraries from the same soil core cannot automatically replace biological replication across sites. Miscounting the unit inflates confidence in both IndVal components and the permutation test.
When reading results, look for plot identifiers, pooling, sampling dates, nested fields and repeated measures. If the design includes paired upstream and downstream sites, permutations may need to preserve pairs. If sites are nested within regions, unrestricted label shuffling can attribute regional differences to treatment. The correct permutation structure follows the physical sampling arrangement, not the convenience of the software interface.
Unequal group sizes need explicit thought
Suppose a study has thirty samples from one soil type and five from another. The opportunity to detect rare features and estimate fidelity differs across groups. Some association index variants are designed with group-size adjustments, while other choices weight the available observations differently. Such differences can affect which taxa rank highly and how confidence is interpreted.
Report the sample counts and the exact IndVal variant. Do not assume that a high-score taxon found in a tiny group has the same evidence base as one detected consistently across dozens of independent plots. Unequal designs can still be analysed, but the sensitivity and precision should be considered. If possible, show both the association components and underlying detection counts so the reader can judge how much information supports the score.
Multiple taxa mean multiple hypothesis tests
Soil microbiome tables may contain hundreds or thousands of ASVs. Testing each against group membership raises a multiple-testing problem. Even when no true associations exist, some features can appear significant by chance at a conventional unadjusted threshold. The indicspecies guidance notes that community-level conclusions involving many species tests need particular care.
Ask whether p values were adjusted for the declared family of comparisons, whether hypotheses were prespecified and whether the analysis filtered features before testing. Different decisions can change the number of reported indicators. Do not report ‘forty confirmed indicator species’ from forty unadjusted p values without discussing the testing burden and taxonomic uncertainty. A careful paper distinguishes candidate indicators from validated markers.
False discovery rate and effect size answer different questions
A false-discovery-rate procedure aims to control a particular expected proportion of incorrect discoveries among rejected hypotheses under its assumptions. It does not tell you the probability that one highlighted taxon is real, nor does it measure the indicator’s ecological effect size. The IndVal-related score expresses association strength under a model; a p value addresses statistical evidence against a chosen null.
Report both, together with specificity, fidelity and replicate counts when available. A weak association can be statistically detectable with a large study, while a potentially useful high-score indicator may have wide uncertainty in a small one. Neither number alone justifies a management decision. The reader needs magnitude, reliability, context and external validation.
IndVal is not the same as log fold change
Differential-abundance analysis asks whether the abundance of a feature differs between groups under a statistical model, often on a logarithmic or compositional scale. Indicator analysis asks how specifically and consistently a feature is associated with predefined groups. A taxon could show a large estimated abundance shift but remain absent from most treatment plots, weakening its usefulness as a practical indicator.
The soil microbiome differential-abundance guide focuses on log fold change, multiple testing and denominator effects. Treat it as a complementary method. A large effect under one analysis and a low indicator score under another can be informative because the questions differ, not necessarily because either method is wrong.
IndVal is not the same as network centrality
A co-occurrence network identifies inferred statistical relationships between features across samples. A highly connected network node is not automatically a strong habitat indicator. It may be broadly common across many habitats while exhibiting numerous correlations. Conversely, a feature strongly restricted to one group could have few inferred network associations. The methods use different definitions of relevance.
The soil microbial network guide explains edges, hub taxa and compositional pitfalls. Do not casually call an IndVal-associated ASV a ‘keystone microbe’ or claim that its network position drives the habitat. Such causal and functional descriptions require separate experimental evidence.
Differential detectability is a practical confounder
If treatment samples were sequenced at much greater depth, their rare features may be detected more often. Their measured fidelity could rise simply because more reads were collected, not because organisms occurred in more treatment plots. Primer mismatch and extraction bias can likewise vary with soil chemistry. The presence rule used by the indicator analysis is downstream of those procedures.
Before interpreting an indicator as a biological marker, review sample read depths, quality filtering, negative controls and laboratory batches. Try defensible sensitivity analyses with alternative thresholds or depth handling. A candidate that disappears under small reasonable processing changes should be labelled exploratory. Robustness is especially important if the result might influence environmental monitoring or restoration assessments.
A false-negative detection is not biological absence
A feature with zero observed reads in a plot could be genuinely absent, but the assay might also have missed it. Fidelity therefore reflects detection under conditions rather than a perfect ecological occupancy census. If an indicator is used to identify a soil condition, false negatives matter: a genuinely related group might be classified incorrectly because the relevant ASV went undetected.
A monitoring programme needs a detection protocol, independent calibration and an understanding of expected error rates. IndVal alone does not provide a complete diagnostic sensitivity and specificity evaluation on external data. Its component named specificity has a statistical ecological definition; it should not be confused automatically with the clinical-style diagnostic specificity of a deployed classifier.
Training and validation samples answer different questions
Selecting indicator taxa and testing them on the same dataset provides evidence of association within the sampled sites. It does not establish performance on independent future sites. If the goal is practical classification, reserve an appropriately independent validation set or conduct a genuinely external study. Then measure how well the candidate features identify new soil conditions under the intended monitoring procedures.
A simple random train–test split can be inadequate when nearby plots or repeated dates share conditions. Spatial or temporal blocking may better represent the real generalisation task. A useful validation report distinguishes discovery scores, classification performance and laboratory reproducibility. Calling an indicator ‘validated’ requires evidence beyond a low p value in the discovery dataset.
Environmental gradients can make groups artificial
Soil moisture, temperature, pH and organic carbon often vary continuously. Dividing a continuum into ‘low’ and ‘high’ can create an indicator contrast dependent on the chosen cutoff. A taxon that responds smoothly to pH may be labelled a high-pH indicator because the analyst created two bins. The label can be useful operationally but should not be mistaken for a sharp biological threshold.
Examine the original environmental measurements where possible. A regression along the continuous gradient may reveal a more realistic response than a two-group indicator test. If threshold categories are needed for management, justify the boundaries and test whether the result changes under nearby reasonable thresholds. The distinction between an administrative category and a biological discontinuity is important.
A three-group worked scenario
Imagine soil groups labelled dry, intermediate and wet, each with independent plots. ASV-A appears in nearly every wet plot and rarely elsewhere. ASV-B appears in both intermediate and wet plots, but not dry plots. A single-group indicator approach may rank ASV-A strongly for wet soil. A multilevel combination analysis may find that ASV-B best indicates the shared intermediate–wet category.
This does not mean that ASV-B ‘prefers’ two climates in a physiological sense. The result describes where its reads were detected under this design. The question for further study is whether environmental moisture, correlated soil chemistry or some other variable explains the pattern. Such a comparison teaches why group combinations must be named exactly when reporting indicators.
Indicator taxa and the meaning of ‘healthy soil’
An environmental report might present taxa associated with sites classified as high quality. But soil health or quality must be defined by independently relevant outcomes, such as fertility, structure, ecological function or a validated monitoring standard. An indicator result cannot establish that a taxon itself caused those outcomes or that its presence is always desirable in every soil.
A sound conclusion identifies the observation: ‘This taxon was associated with plots classified as higher-quality under the study’s specified criteria.’ If useful, it then proposes prospective validation across independent sites. Avoid universal promises, particularly when soil types, crops and management goals differ. A marker may be locally useful without becoming a general biological standard.
Distinguish indicator from pathogen diagnosis
Some sequence names may be associated with organisms that include pathogenic members, yet a short marker region can lack strain or species resolution. An IndVal association does not diagnose infection, toxin production or environmental exposure. If a result could influence health or safety decisions, a validated targeted assay and appropriate domain expertise are needed.
Use cautious wording such as ‘sequence feature assigned to this taxonomic group’ when identification is limited. The interpretation should not jump from a familiar genus name to an unsupported strain-level hazard. Different strains can have different functions, and DNA may be present without viable organisms. Responsible translation keeps ecological indicators and diagnostic clinical tests separate.
How to read a typical multipatt output
A multipatt result may list a feature name, the associated group or group combination, a statistic and a permutation-derived p value. Some summaries additionally reveal specificity and fidelity components. The group columns can encode membership in an association combination rather than one category. Read the package documentation to establish whether the reported statistic is IndVal.g, IndVal, r or another supported method.
A reader’s summary should include the feature definition, selected group, index variant, association statistic, detection frequency, statistical evidence and any adjustments. If data were transformed before analysis, include that detail. Do not reconstruct missing metadata from the printed output alone. The software table is one layer; sample identity, sequencing quality and biological interpretation come from the wider study.
Create a useful plain-language report
A precise conclusion might read: ‘Within the sampled plots, ASV-27 showed a relatively strong association with the upper soil horizon according to the stated IndVal.g model, with higher specificity and frequent detection in that group. The result is an association of marker-sequence observations and does not identify a causal ecological function.’ If multiple comparisons were performed, add how they were handled.
The sentence tells the reader what the method can defend and why the result matters. For a general audience, translate fidelity as ‘found consistently across sampled plots’ and specificity as ‘concentrated in the chosen group compared with other groups.’ Avoid replacing the technical meanings with catchy but wrong claims such as ‘this bacterium fixes the soil’ or ‘this group determines soil quality.’
An independent comparison across seasons
Suppose a candidate indicator scores strongly for one soil group during spring sampling but weakly when the same plots are visited in autumn. This is not necessarily a failure of arithmetic. The original association may have depended on seasonal moisture, plant cover, detectable abundance or environmental conditions that shifted between observations. A time-specific indicator can still be useful when its scope is described accurately.
To test durability, compare the same feature, assay and group definition across seasons while accounting for repeated samples from individual plots. Ask whether sequencing depth, DNA extraction and storage procedures remained comparable. If the same association recurs despite changed conditions, that strengthens generalisation within the tested range. If it vanishes, the result suggests a more conditional relationship rather than a universal biological marker.
Checking classification performance without circularity
Consider using a discovered indicator ASV as a practical test that classifies new soil samples into two groups. The discovery dataset cannot by itself provide an unbiased estimate of how frequently that classification will succeed. One needs new observations with independently defined group labels and a decision threshold chosen without peeking at the validation answers.
If a marker is counted as present whenever one read is detected, noise and cross-contamination may create false positive classifications. A stricter threshold can reduce some false positives while increasing missed detections. Report the threshold, numbers of true and false classifications, and the target use case. In an environmental survey, an apparently impressive IndVal association is a candidate for validation, not a replacement for a complete screening-performance study.
Errors worth checking before publication
A common mistake is to treat the strongest p value as the strongest biological indicator. Repair it by reading association magnitude and detection frequency. Another mistake is to interpret IndVal.g as the direct product of its components; check the square-root convention. A third is to assume a label means species-level identification. Keep the ASV or supported rank visible.
Other failures include confusing technical sequencing replicates with independent plots, ignoring spatial dependence in permutations, failing to adjust for multiple testing, accepting laboratory batch effects as ecology, and claiming indicator status transfers globally. Each repair begins with an explicit input definition or design check. Better reporting is achieved by inspecting the evidence, not by expanding the confidence of the prose.
Frequently asked questions
Is an indicator taxon automatically beneficial? No. Its label reflects association with a defined group. Can a taxon be an indicator of more than one group? Yes, multilevel methods can test combinations. Does fidelity measure a percentage of soil occupied? No, it describes detection across sampled sites under the method. Does specificity mean the same as diagnostic specificity? Not necessarily; the ecological association component has its own definition.
Why may a high IndVal.g score have a different raw product? The function can return the square root of the IndVal-style product. Is a permutation p value proof of causality? No. Why might an indicator vanish after sequencing-depth adjustment? Detection bias may have influenced its frequency. Can a single study validate a global soil marker? No; independent transfer testing is required. Should dozens of unadjusted significant ASVs be called confirmed discoveries? No, the multiplicity and study design require scrutiny.
Independent practice with answers
A feature is detected in nine of ten target plots and one of ten control plots. What is its observed fidelity for the target group? Nine divided by ten, or 0.9. Does that prove it exists in 90% of the soil volume? No, the numerator and denominator are sampled plots. Another feature appears in only one target plot and nowhere else. Can it be very specific yet unreliable for routine classification? Yes. Its rare occurrence limits fidelity.
Suppose specificity A equals 0.49 and fidelity B equals 0.81. Their product is 0.3969 and the square root is 0.63. If the reported statistic is the square-root variant, expect about 0.63 rather than 0.3969. Finally, if a feature’s association holds in discovery sites but fails in an independent landscape, does the original calculation become arithmetically wrong? No. Its external generalisation was limited, which should shape the wider claim.
Authoritative methods and further learning
The developers’ indicspecies Indicator Species Analysis guide explains IndVal, IndVal.g, fidelity, specificity and multipatt. The indicspecies package documentation and reference manual describe supported association functions and permutation options. Use the version actually applied by a study. Method names and output conventions should never be inferred from a chart label alone.
To understand sequence sources, read soil amplicon tables. To compare distinct statistical jobs, use soil differential abundance and soil microbial networks. For communicating precise concepts, cross-linguistic transfer and context-aware translation explain why familiar words can conceal different definitions.
The final translation: association is an opportunity to investigate
Indicator analysis identifies sequence features that are unusually concentrated or consistently detected in selected groups. Its strength comes from making the association measurable and testable, not from proving which microbes cause the observed habitat. The formula, sample groups, index variant, detection thresholds, permutations and multiple comparisons all affect what a score means.
A trustworthy conclusion names the target group, the feature, the association evidence and the limitations. Independent validation then asks whether the pattern survives new sites, seasons and methods. That is how a statistical indicator becomes useful knowledge: by carrying its definition and uncertainty forward rather than converting an interesting association into a universal claim of ecological function.