--- title: "Rarefaction and Standardization" author: "Gilles Colling" date: "`r Sys.Date()`" output: rmarkdown::html_vignette vignette: > %\VignetteIndexEntry{Rarefaction and Standardization} %\VignetteEngine{knitr::rmarkdown} %\VignetteEncoding{UTF-8} --- ```{r setup, include = FALSE} knitr::opts_chunk$set( collapse = TRUE, comment = "#>", fig.width = 7, fig.height = 5, dev = "svglite", fig.ext = "svg" ) ``` ```{r theme-helper, include = FALSE} if (requireNamespace("ggplot2", quietly = TRUE)) { transparent <- ggplot2::theme( panel.background = ggplot2::element_rect(fill = "transparent", color = NA), plot.background = ggplot2::element_rect(fill = "transparent", color = NA) ) } ``` ## Introduction Species richness counted from raw data confounds two things: how diverse a community is and how hard it was sampled. A plot visited for an afternoon yields fewer species than the same plot visited for a week, even though the underlying community has not changed. The same effect appears across sites. A site with 500 recorded individuals will, on average, show more species than a neighbour with 50, because more individuals expose more of the rare tail of the community. Any comparison of raw counts therefore partly measures effort rather than ecology. The bias is large. Richness rises steeply with the first few dozen individuals and only flattens once the common species are saturated and new records start repeating species already seen. Two sites at different points on that rising curve are not comparable. Reporting that site A has 18 species and site B has 11 says little until both numbers refer to the same amount of sampling. Standardization removes the effort axis so that the diversity axis can be read on its own. Two families of methods do this. Rarefaction fixes a common sample size and asks how many species would be expected at that size. Coverage-based methods fix a common completeness, the fraction of total community abundance that the observed species account for, and compare richness at that shared completeness. This vignette develops both, shows the estimators **spacc** provides for each, and ends with concrete rules for choosing between them. The running theme is that the standardization target, equal size or equal coverage, can change which site looks more diverse. The distinction matters because size and coverage are not interchangeable. A fixed number of individuals corresponds to very different completeness depending on how even the community is. In an even community a few hundred individuals already capture almost every species, so coverage is high. In a community with a handful of dominants and a long tail of rare species, the same few hundred individuals leave much of the tail undetected, so coverage stays low. Comparing two such communities at equal sample size therefore compares them at unequal completeness, which reintroduces a version of the bias that rarefaction was meant to remove. Coverage-based standardization closes that gap by making completeness itself the common currency. The cost of coverage standardization is that it needs counts of individuals. The singleton and doubleton frequencies that drive the coverage estimate exist only in abundance data, not in presence/absence records. When the data are binary, or when effort varies through the number of plots rather than the number of individuals within a plot, the sample-based rarefaction family is the right tool instead. **spacc** carries all three families, and the sections below show how to move between them on a single simulated dataset where the true community is known. ## Methods and theory ### Individual-based rarefaction Individual-based rarefaction asks a counterfactual question. Given a pooled sample of $N$ individuals containing $S_{obs}$ species, how many species would be expected in a random subsample of $n \le N$ individuals drawn without replacement. Hurlbert (1971) gives the exact expectation. Let $N_i$ be the total abundance of species $i$, so that $N = \sum_i N_i$. The probability that species $i$ appears at least once in a subsample of $n$ is one minus the probability of drawing $n$ individuals that all avoid species $i$: $$ P(\text{species } i \in \text{subsample}) = 1 - \frac{\binom{N - N_i}{n}}{\binom{N}{n}}. $$ Summing detection probabilities over all observed species gives the expected rarefied richness, since the expectation of a sum of indicators is the sum of their probabilities: $$ E[S_n] = \sum_{i=1}^{S_{obs}} \left[ 1 - \frac{\binom{N - N_i}{n}}{\binom{N}{n}} \right]. $$ Here $E[S_n]$ is expected richness at subsample size $n$, $N$ is the total number of individuals pooled across the sites being compared, $N_i$ is the abundance of species $i$, and $S_{obs}$ is the number of species observed in the full pooled sample. The curve $E[S_n]$ starts at 1 when $n = 1$, rises monotonically, and reaches $S_{obs}$ exactly when $n = N$. The formula is computed on a log scale in **spacc** to avoid overflow in the binomial coefficients. The expectation is exact, not a simulation average. Drawing many random subsamples of size $n$ and averaging their richness would converge to $E[S_n]$, but the Hurlbert formula gives that limit in closed form, which is both faster and free of Monte Carlo noise. The shape of the curve carries information of its own. A curve that has flattened by $n = N/2$ indicates a community whose common species are saturated, so doubling the remaining effort would add few species. A curve still rising steeply at $n = N$ indicates an undersampled community where the species total is an underestimate. **spacc** attaches a bootstrap interval to the curve by resampling individuals with replacement and recomputing the expectation, which captures sampling variation in the underlying frequencies. The same logic extends to Hill numbers of order $q$, which weight species by abundance. Order $q = 0$ is richness, where all species count equally and the rare tail dominates. Order $q = 1$ is the exponential of Shannon entropy, which weights species by their frequency. Order $q = 2$ is the inverse Simpson concentration, dominated by the common species. The Hill number of order $q$ of a community with relative abundances $p_i$ is $$ {}^{q}D = \left( \sum_{i=1}^{S} p_i^{q} \right)^{1/(1-q)}, \qquad q \ne 1, $$ with the $q = 1$ case taken as the limit ${}^{1}D = \exp(-\sum_i p_i \log p_i)$. Each ${}^{q}D$ is an effective number of species, the count of equally abundant species that would give the same diversity value. Higher $q$ down-weights rare species, so the rarefaction curve for $q = 2$ saturates faster and is far less sensitive to subsample size than the curve for $q = 0$. The choice of $q$ encodes a research question rather than a technical preference. Richness ($q = 0$) treats a species seen once and a species seen a thousand times as equal contributions, which makes it the most sensitive to sampling effort and to the rare tail. Shannon diversity ($q = 1$) weights each species by its abundance, giving a balanced summary that is less affected by the few rarest records. Inverse Simpson ($q = 2$) is governed almost entirely by the common species and barely moves as effort grows, which makes it a stable comparison when the rare tail is suspected to be noise. Scanning $q$ from 0 upward across the same data traces a diversity profile, and the steepness of that profile is itself a measure of how uneven the community is. ### Sample coverage Sample size is a poor common currency when communities differ in evenness. A fixed number of individuals can represent a near-complete census of an even community yet barely scratch a community with a long rare tail. Sample coverage measures completeness directly. Coverage $C$ is the fraction of the total community abundance belonging to the species already observed, so $1 - C$ is the probability that the next individual sampled belongs to a species not yet recorded. Chao and Jost (2012) estimate it from the frequency counts of the sample: $$ C = 1 - \frac{f_1}{N} \cdot \frac{(N-1) f_1}{(N-1) f_1 + 2 f_2}. $$ Here $f_1$ is the number of singletons, species represented by exactly one individual in the pooled sample, $f_2$ is the number of doubletons, species represented by exactly two individuals, and $N$ is the total number of individuals. The estimator rests on the Good-Turing insight that singletons carry the signal about undetected species: many singletons relative to doubletons means the rare tail is poorly sampled and coverage is low. When $f_1 = 0$ every observed species has been seen more than once, the correction vanishes, and $C = 1$. Coverage-based rarefaction interpolates richness or any ${}^{q}D$ to a target coverage rather than a target sample size, so two communities are compared at equal proportional representation. The deficit $1 - C$ has a direct reading as a probability. It estimates the chance that the next individual sampled belongs to a species not yet on the list, so it answers the practical question of how much new information one more sample would bring. A deficit of 0.05 means coverage of 0.95, and the next individual has roughly a one-in-twenty chance of being a new species. This interpretation is what makes coverage a fair common currency: two samples at the same deficit have the same residual probability of discovery, regardless of how many individuals each took to reach that point. For plot-based spatial data, where the sampling unit is a site rather than an individual, the Chiu (2023) variant uses incidence frequencies, the counts of species found in exactly one or two sites, and is the more appropriate estimator. `spaccCoverage()` exposes both through its `coverage` argument. ## Simulating uneven effort The simulation idiom from the package, a site-by-species matrix of Poisson abundances, makes the effort bias concrete because the ground truth is known. Here every site is drawn from the same 20-species pool, so the true per-site richness ceiling is identical across all sites. What differs is the Poisson rate, which sets per-site total abundance. Low-rate sites are sampled shallowly and will miss species purely because few individuals were drawn. Building the matrix this way separates the two things real data conflate. The biological signal, the species pool, is held fixed at 20 species for every site. The sampling signal, the per-site rate $\lambda$, is varied deliberately across three tiers. Any difference in observed richness between tiers is therefore attributable to effort alone, because the generating community is identical. This is the controlled setting in which a standardization method can be checked: a method that works should erase the between-tier difference, recovering the shared 20-species ceiling, while a method that fails will leave the effort gradient intact. Real surveys never come with this guarantee, which is why a simulation with known truth is the right place to build intuition before applying the same tools to field data. ```{r data} library(spacc) set.seed(42) n_sites <- 60 coords <- data.frame(x = runif(n_sites), y = runif(n_sites)) # Variable total abundances across sites (realistic uneven sampling) lambdas <- rep(c(1, 3, 5), each = 20) species <- matrix(0, nrow = n_sites, ncol = 20) for (i in seq_len(n_sites)) { species[i, ] <- rpois(20, lambda = lambdas[i]) } colnames(species) <- paste0("sp", 1:20) ``` The three effort tiers, Poisson rates 1, 3, and 5, span a fivefold range in expected individuals per site. The induced bias is visible directly in the per-site counts: low-effort sites record fewer individuals and fewer species, even though all sites share the same underlying community. ```{r effort-bias} tier <- factor(lambdas) indiv <- rowSums(species) rich <- rowSums(species > 0) aggregate(cbind(individuals = indiv, richness = rich), list(tier = tier), mean) ``` Mean observed richness climbs from the low-effort tier to the high-effort tier despite an identical 20-species ceiling. That gap is the artefact standardization is built to remove. A naive ranking of these sites by raw richness would simply recover the sampling-effort gradient. The mean individuals per site track the Poisson rates closely, around one, three, and five times 20 species, since each cell of the matrix is an independent Poisson draw. Mean richness rises with them, but more slowly, because the high tiers are already close to the 20-species ceiling and have little room left to climb. The low tier sits far below the ceiling, missing several species per site purely because few individuals were drawn. This is the saturating-curve effect seen at the level of a single site: the first individuals reveal common species quickly, and each additional individual is progressively more likely to repeat a species already recorded. ## Individual-based rarefaction The `rarefy()` function takes the abundance matrix, pools it across sites, and returns the Hurlbert expectation $E[S_n]$ over a grid of subsample sizes. The returned `spacc_rare` object carries the curve, a bootstrap standard deviation, and 95 percent bounds. ```{r rarefy, eval = requireNamespace("ggplot2", quietly = TRUE)} # Rarefy to the pooled abundance grid (q = 0, richness) rare <- rarefy(species) print(rare) ``` The numeric curve is recovered with `as.data.frame()`, which exposes the sample sizes, the expected diversity, and the bootstrap interval. The expectation flattens well before the full sample size, the signature of a saturating community. ```{r rarefy-extract} rare_df <- as.data.frame(rare) head(rare_df, 4) tail(rare_df, 2) ``` ```{r rarefy-plot, eval = requireNamespace("ggplot2", quietly = TRUE)} plot(rare) + transparent ``` Rarefaction at higher orders down-weights the rare species. Passing `q = 1` and `q = 2` returns the Shannon and Simpson effective numbers, which saturate faster because they are governed by the common species rather than the long tail. `rarefy()` accepts any non-negative order; orders 0, 1, and 2 use exact estimators while other orders report the Hill number of order `q` of the sampled composition. ```{r rarefy-hill} rare_q1 <- rarefy(species, q = 1) rare_q2 <- rarefy(species, q = 2) c(q0 = max(rare$expected), q1 = max(rare_q1$expected), q2 = max(rare_q2$expected)) ``` ```{r rarefy-anyq} rare_q3 <- rarefy(species, q = 3) rare_half <- rarefy(species, q = 0.5) c(q0.5 = max(rare_half$expected), q3 = max(rare_q3$expected)) ``` ### Analytical sample-based alternatives Rarefaction can also be sample-based: instead of subsampling individuals, the unit subsampled is the site. **spacc** provides three analytical curves over sites that need no simulation. `mao_tau()` is the exact expected sample-based rarefaction of Colwell et al. (2004); it binarizes the matrix and returns a data frame with `sites`, `expected`, `sd`, and analytical 95 percent `lower`/`upper` bounds. `coleman()` returns the Coleman et al. (1982) expectation with columns `sites`, `expected`, `sd`. `collector()` returns the species count in data order, with columns `sites`, `species`, which traces how the actual survey accumulated species rather than a randomized expectation. ```{r analytical} mt <- mao_tau(species) cm <- coleman(species) cl <- collector(species) head(mt, 3) ``` Plotting the three together shows that the randomized expectations from `mao_tau()` and `coleman()` lie close to each other, while the collector's curve fluctuates around them because it follows the arbitrary order of the rows. The Mao Tau band gives the analytical uncertainty without any bootstrap. ```{r analytical-plot, eval = requireNamespace("ggplot2", quietly = TRUE)} df <- rbind( data.frame(sites = mt$sites, S = mt$expected, kind = "Mao Tau"), data.frame(sites = cm$sites, S = cm$expected, kind = "Coleman"), data.frame(sites = cl$sites, S = cl$species, kind = "Collector") ) ggplot2::ggplot(df, ggplot2::aes(sites, S, color = kind)) + ggplot2::geom_line(linewidth = 0.9) + ggplot2::labs(x = "Sites", y = "Species", color = NULL) + transparent ``` The Mao Tau and Coleman expectations are the sample-based analogue of Hurlbert's individual-based curve. They standardize on number of sites; `rarefy()` standardizes on number of individuals. Which axis matters depends on whether effort varies through plot count or through depth of sampling within plots. ## Coverage-based standardization `spaccCoverage()` accumulates sites spatially while tracking the Chao-Jost coverage at each step, across many random starting seeds. The default estimator is `"chao"`; a `"chiu"` option uses incidence frequencies and is recommended when the sampling units are plots rather than independent individuals. The result is a `spacc_coverage` object holding `n_seeds` by `n_sites` matrices of richness, individuals, and coverage. ```{r coverage, eval = requireNamespace("ggplot2", quietly = TRUE)} cov_result <- spaccCoverage(species, coords, n_seeds = 30, progress = FALSE) print(cov_result) ``` ```{r coverage-plot, eval = requireNamespace("ggplot2", quietly = TRUE)} plot(cov_result) + transparent ``` ### Interpolation at fixed coverage `interpolateCoverage()` reads richness off each seed's curve at the requested coverage levels by linear interpolation. The return is a data frame with one row per seed and one column per target, so the spread across rows is the across-seed uncertainty introduced by the choice of starting point. ```{r interpolate} interp <- interpolateCoverage(cov_result, target = c(0.90, 0.95, 0.99)) round(colMeans(interp), 2) round(apply(interp, 2, sd), 3) ``` The across-seed standard deviation widens toward higher coverage targets, because those targets sit near or beyond the end of some seeds' curves where interpolation has less support. The mean richness at 90 percent coverage is the fair point estimate; the column standard deviation is its seed-induced uncertainty. Each seed is a different spatial path through the sites, starting from a different point and accumulating neighbours in a different order. Because the sites are spatially structured, the order in which they are added changes the coverage reached at each step, and so two seeds can hit 90 percent coverage at different richness values. Averaging over seeds marginalizes out the arbitrary choice of starting point, and the spread of the per-seed values is the honest uncertainty that choice introduces. Reporting only the mean would hide that spread; reporting per-seed values exposes it, which is why `interpolateCoverage()` returns the full matrix rather than a collapsed summary. ## Extrapolation beyond observed When a target coverage exceeds the maximum any seed reached, interpolation is impossible and `extrapolateCoverage()` switches to the Chao et al. (2014) asymptotic estimators. For $q = 0$ it uses the Chao1 coverage deficit; for $q = 1$ and $q = 2$ it applies the corresponding Shannon and Simpson asymptotics. The returned `spacc_coverage_ext` object reports the observed coverage and richness alongside the extrapolated values. ```{r extrapolate} extrap <- extrapolateCoverage(cov_result, target_coverage = c(0.95, 0.99), q = 0) print(extrap) summary(extrap) ``` ```{r extrapolate-plot, eval = requireNamespace("ggplot2", quietly = TRUE)} plot(extrap) + transparent ``` Extrapolation is reliable only over a limited range past the observed coverage. The dashed reference line marks the observed coverage; the orange points beyond it are model-based projections whose uncertainty grows with distance from the data. Targets far past the observed maximum should be read as informed extrapolation, not measurement. ## Hill numbers with coverage `spaccHillCoverage()` tracks Hill numbers for several orders and coverage at the same time, so the three diversity orders can be standardized to one completeness. Supplying `target_coverage` adds a `standardized` component holding each order's value interpolated to that coverage for every seed. ```{r hill-coverage, eval = requireNamespace("ggplot2", quietly = TRUE)} hc <- spaccHillCoverage(species, coords, q = c(0, 1, 2), target_coverage = 0.9, n_seeds = 20, progress = FALSE) print(hc) ``` The `standardized` list gives the seed-averaged effective number of species at 90 percent coverage for each order. Richness ($q = 0$) is highest and most seed-variable, while $q = 2$ is lowest and most stable, the expected ordering ${}^{0}D \ge {}^{1}D \ge {}^{2}D$. ```{r hill-coverage-std} sapply(hc$standardized, mean) ``` ```{r hill-coverage-plot, eval = requireNamespace("ggplot2", quietly = TRUE)} plot(hc, xaxis = "coverage") + transparent ``` Reading diversity against coverage rather than against sites puts all three curves on a completeness axis. The vertical gap between the $q = 0$ and $q = 2$ curves at a fixed coverage measures the unevenness of the community: a wide gap means a few dominant species and a long rare tail. Standardizing all three orders to a single coverage is what makes the comparison across $q$ fair. If each order were read at the end of its own accumulation, the orders would sit at slightly different completeness levels and the contrast between them would mix a real diversity difference with a residual effort difference. Fixing coverage first removes that confound, leaving the spacing between ${}^{0}D$, ${}^{1}D$, and ${}^{2}D$ as a clean signal of evenness. The same machinery scales to comparing several communities: standardize every community to the same coverage, then read off the diversity profile for each, and the profiles are directly comparable because they share both the completeness and the set of orders. ## Spatial subsampling Spatial autocorrelation inflates apparent accumulation when nearby sites hold similar species. Two tools reduce it. `subsample()` returns site indices to keep under a grid, random, or thinning rule; the grid method overlays cells of a chosen size and retains one random site per cell, which spreads the retained sites out. `spatialRarefaction()` instead runs distance-weighted permutations and returns the spatially constrained expectation directly. ```{r subsample} set.seed(7) keep <- subsample(coords, method = "grid", cell_size = 0.2, seed = 7) length(keep) ``` A grid at cell size 0.2 over the unit square thins 60 sites down to roughly the number of occupied cells. Comparing the accumulation curve on the thinned set against the full set shows how much of the original rise came from spatially redundant neighbours. ```{r thinned-vs-full, eval = requireNamespace("ggplot2", quietly = TRUE)} full <- spacc(species, coords, n_seeds = 30, progress = FALSE) thin <- spacc(species[keep, ], coords[keep, ], n_seeds = 30, progress = FALSE) df_full <- transform(as.data.frame(full), set = "Full") df_thin <- transform(as.data.frame(thin), set = "Thinned (grid)") ``` ```{r thinned-plot, eval = requireNamespace("ggplot2", quietly = TRUE)} ggplot2::ggplot(rbind(df_full, df_thin), ggplot2::aes(sites, mean, color = set)) + ggplot2::geom_line(linewidth = 1) + ggplot2::labs(x = "Sites", y = "Species", color = NULL) + transparent ``` The thinned curve reaches a similar asymptote with fewer sites, since the discarded sites were spatial near-duplicates that added little new information. `spatialRarefaction()` gives the same correction analytically, returning a data frame with `sites`, `mean`, `sd`, and quantile bounds from distance-weighted draws. ```{r spatial-rare} sr <- spatialRarefaction(species, coords, n_perm = 50) tail(sr[, c("sites", "mean", "lower", "upper")], 2) ``` ## Comparison on the same data The three standardizations can disagree about which sites are more diverse, because they hold different quantities constant. The clearest demonstration uses the effort tiers from the simulation. The low-effort and high-effort tiers share the same 20-species community but differ fivefold in individuals per site. Raw richness, richness at matched individuals, and richness at matched coverage tell three different stories. ```{r tier-split} lo <- species[lambdas == 1, ] hi <- species[lambdas == 5, ] raw <- c(low = sum(colSums(lo) > 0), high = sum(colSums(hi) > 0)) raw ``` Matched effort uses individual-based rarefaction at a common subsample size set by the smaller pool. Each tier is rarefied to the same number of individuals, so the comparison no longer rewards the deeper-sampled tier. ```{r matched-effort} n_match <- min(sum(lo), sum(hi)) S_lo <- rarefy(lo, n_individuals = n_match)$expected S_hi <- rarefy(hi, n_individuals = n_match)$expected c(low = round(S_lo, 1), high = round(S_hi, 1), n = n_match) ``` Matched coverage uses coverage-based standardization at a common completeness. The two tiers are interpolated to the same coverage so the comparison rewards equal proportional representation rather than equal individuals. ```{r matched-coverage} co_lo <- spaccCoverage(lo, coords[lambdas == 1, ], n_seeds = 20, progress = FALSE) co_hi <- spaccCoverage(hi, coords[lambdas == 5, ], n_seeds = 20, progress = FALSE) S_lo_c <- mean(interpolateCoverage(co_lo, target = 0.9)[, 1]) S_hi_c <- mean(interpolateCoverage(co_hi, target = 0.9)[, 1]) c(low = round(S_lo_c, 1), high = round(S_hi_c, 1)) ``` Collecting the three views into one table makes the shift in conclusion explicit. Raw counts favour the high-effort tier by the largest margin; matched effort and matched coverage shrink or reverse that margin, because both tiers draw from the same community and the difference was an artefact of sampling depth. ```{r comparison-table} data.frame( standardization = c("Raw richness", "Matched individuals", "Matched coverage (0.90)"), low = c(raw["low"], round(S_lo, 1), round(S_lo_c, 1)), high = c(raw["high"], round(S_hi, 1), round(S_hi_c, 1)), row.names = NULL ) ``` The lesson is that a richness comparison is only as meaningful as the quantity held constant. Reporting raw richness across uneven samples measures effort; reporting it at matched effort or matched coverage measures the community. When the standardized margins are small, the apparent difference in the raw data was sampling, not biology. The three rows answer three different questions about the same two tiers. The raw row asks how many species were recorded, which is a statement about the survey, not the community. The matched-individuals row asks how many species each tier would show if both had been sampled to the same depth, which removes the fivefold difference in effort and brings the two tiers close together. The matched-coverage row asks how many species each tier holds at equal completeness, which can favour the low-effort tier if its smaller sample happened to be more even. Since both tiers were generated from the same 20-species pool, the correct answer is that they are equally diverse, and the standardized rows recover that far better than the raw row. In field data the truth is unknown, so the discipline is to report the standardized comparison and to state which quantity was held constant, rather than letting the reader assume raw counts are comparable. A short note on which standardization to trust when they disagree. Matched individuals and matched coverage will not always agree, and the difference is informative rather than a contradiction. When they diverge, the tiers differ in evenness as well as in effort: equalizing individuals leaves the more even community at higher coverage, while equalizing coverage gives it fewer individuals. Coverage is usually the more defensible target for comparing communities that differ in evenness, because equal completeness is closer to the ecological question of how much of each community has been seen. Matched individuals remains the right target when the comparison is explicitly about a fixed sampling budget rather than about completeness. ## Practical guidance Standardization choices come down to what varies across the samples being compared and whether the data carry the abundance information coverage needs. **Rules of thumb.** - Target coverage of 0.90 to 0.95 for most comparisons. Below 0.85 the rare tail is too poorly sampled for the coverage estimate to be stable; at 0.99 the curves are near their asymptote and the comparison adds little over a richness estimator. - Keep at least 30 to 50 individuals per site before trusting an individual-based rarefaction; below that the curve is dominated by the first few draws and the bootstrap interval is wide. - Limit extrapolation to roughly 2 to 3 times the observed sample size, or to coverage no more than a few points past the observed maximum. Beyond that the Chao asymptotics extrapolate the rare tail well past the data. - Use at least 20 to 30 seeds in `spaccCoverage()` and `spaccHillCoverage()` so the across-seed standard deviation is a usable uncertainty estimate rather than noise from too few starting points. - Down-weight the rare tail with $q = 1$ or $q = 2$ when singletons are suspected to be identification errors; richness ($q = 0$) is the order most sensitive to that noise. **Minimum sample size.** Individual-based rarefaction needs enough individuals that the curve has begun to flatten; 30 to 50 per site is a working floor, and comparisons should rarefy down to the smallest pool, not up. Coverage-based methods need enough singletons and doubletons to estimate $f_1$ and $f_2$; a sample with fewer than about 10 species or no doubletons gives an unstable coverage estimate. Sample-based curves (`mao_tau`, `coleman`) need at least 10 to 15 sites before the expectation and its variance are meaningful. **Decision table.** | Situation | Recommended method | Why | |-----------|--------------------|-----| | Sites differ in number of individuals, abundances available | Coverage-based (`spaccCoverage`, `interpolateCoverage`) | Compares at equal completeness | | Effort varies but only counts of individuals matter | Individual-based (`rarefy`) | Hurlbert expectation at matched $n$ | | Effort varies through plot count, presence/absence only | Sample-based (`mao_tau`, `coleman`) | Standardizes on number of sites | | Need several diversity orders at once | `spaccHillCoverage` | Tracks $q = 0, 1, 2$ and coverage together | | Nearby sites are spatially redundant | `subsample(method = "grid")`, `spatialRarefaction` | Reduces autocorrelation inflation | | Want richness past observed completeness | `extrapolateCoverage` | Chao asymptotic projection | **When not to use coverage-based methods.** Coverage is defined through $f_1$, $f_2$, and $N$, so it requires genuine abundance counts. Presence/absence data have no singletons or doubletons in the abundance sense, and the Chao-Jost estimator cannot be formed; sample-based rarefaction on sites is the correct tool there. Coverage is also unreliable when the sample is tiny or has no doubletons, since the denominator $(N-1) f_1 + 2 f_2$ then rests on a single frequency class. And when sampling effort is already equal across sites, standardization is unnecessary; raw richness is directly comparable and adding a rarefaction step only discards information. ## References - Chao, A. & Jost, L. (2012). Coverage-based rarefaction and extrapolation: standardizing samples by completeness rather than size. *Ecology*, 93, 2533-2547. - Chao, A., Gotelli, N.J., Hsieh, T.C., Sander, E.L., Ma, K.H., Colwell, R.K. & Ellison, A.M. (2014). Rarefaction and extrapolation with Hill numbers. *Ecological Monographs*, 84, 45-67. - Coleman, B.D., Mares, M.A., Willig, M.R. & Hsieh, Y.H. (1982). Randomness, area, and species richness. *Ecology*, 63, 1121-1133. - Colwell, R.K., Mao, C.X. & Chang, J. (2004). Interpolating, extrapolating, and comparing incidence-based species accumulation curves. *Ecology*, 85, 2717-2727. - Hurlbert, S.H. (1971). The nonconcept of species diversity: a critique and alternative parameters. *Ecology*, 52, 577-586.