Detection biases yield misleading patterns of species persistence and colonization in fragmented landscapes

. Species occurrence patterns, and related processes of persistence, colonization and turnover, are increasingly being used to infer habitat suitability, predict species distributions, and measure biodiversity potential. The majority of these studies do not account for observational error in their analyses despite growing evidence suggesting that the sampling process can significantly influence species detection and subsequently, estimates of occurrence. We examined the potential biases of species occurrence patterns that can result from differences in detectability across species and habitat types using hierarchical multispecies occupancy models applied to a tropical bird community in an agricultural fragmented landscape. Our results suggest that detection varies widely among species and habitat types. Not incorporating detectability severely biased occupancy dynamics for many species by overestimating turnover rates, producing misleading patterns of persistence and colonization of agricultural habitats, and misclassifying species into ecological categories (i.e., forest specialists and generalists). This is of serious concern, given that most research on the ability of agricultural lands to maintain current levels of biodiversity by and large does not correct for differences in detectability. We strongly urge researchers to apply an inferential framework which explicitly account for differences in detectability to fully characterize species-habitat relationships, correctly guide biodiversity conservation in human-modified landscapes, and generate more accurate predictions of species responses to future changes in environmental conditions.


INTRODUCTION
Characterizing spatial patterns of species distributions is a fundamental objective in theoretical and applied ecology (Andrewartha and Birch 1974, McIntire andFajardo 2009) as well as biodiversity conservation (Wiens and Graham 2005).Given that different population and community level processes can result in similar species distribution patterns, it is necessary to track changes in patterns over time to discern species habitat use and preferences.One approach is to explicitly model species occupancy in a given location as a function of persistence (i.e., the probability of a species occurring in a location given that it had occurred there in the previous time step) and colonization (i.e., the probability of a species occurring in a location given that it had not occurred there in the previous time step).This approach can greatly increase understanding of the dynamic processes that influence species distributions, improving knowledge on how species' distribution patterns are affected by environmental factors (Hanski 1998).The resulting information is central not only to population and community dynamics, but also in evaluating status and trends of populations of conservation concern.
The relevance and applicability of inferences based on dynamic occupancy models rests largely on the assumption that the occupancy state (i.e., species presence or absence) can be inferred accurately from detection-nondetection survey data.Substantial evidence indicates that observed patterns of occurrence can be influenced by differences in species ''detectability'' (i.e., the probability of observing a species or individual when present) during the sampling process (Tyre et al. 2003).The growing body of literature on this topic demonstrates that heterogeneity in detectability among species and habitat types can mask patterns of species occupancy for insects (Pellet 2008), amphibians (Mazerolle et al. 2005), reptiles (Ke ´ry 2002) and birds (Boulinier et al. 1998) and can lead to mischaracterizations of the true state of occupancy (MacKenzie et al. 2003, MacKenzie et al. 2006, Royle and Dorazio 2008).
Although the deleterious effects of ignoring detectability in occupancy models have been shown for a wide variety of species, the effects of not incorporating detectability on estimates of persistence and colonization have received far less attention (but see Nichols et al. 1998).Using simulations, Moilanen (2002) demonstrated how false zeros (i.e., failure to observe a species when it is in fact present) resulted in a misclassification of the occupancy status of habitat patches, leading to an overestimation of species' migratory distances, dispersal ability, and projected persistence of the metapopulation.More recently, Risk et al. (2011) showed that correcting for imperfect detection in metapopulations of two Rail bird species resulted in higher occupancy rates and lower species turnover rates.However, the circumstances under which estimates of occupancy dynamics are affected by differences in detectability among species and habitat types remain largely uncharacterized, and the potential exists for serious misinterpretations when using detection/nondetection data for management or conservation purposes (MacKenzie 2006).For example, agricultural lands with higher detection probabilities could be misidentified as ''high quality'' or of conservation value based on high rates of persistence and/or colonization, when in fact the opposite may be true if detection is simply lower in higher quality habitats (e.g., dense forest).Determining the true conservation potential of human-modified or secondary habitats (e.g., agricultural matrix, habitats at higher elevations) in response to primary habitat loss or climate change necessitates a clear understanding of colonization rates.Yet, patterns of colonization may be masked when apparent contrasts in rates are merely reflective of differences in the observation process, and not representative of the true biological process.Low or variable detection rates across habitats and species can lead to an underestimation or misclassification of species richness and distribution patterns, resulting in inaccurate assessments of biodiversity (Boulinier et al. 1998).Yet, current research characterizing species distributions and habitat relationships using detection/nondetection data, by and large, assumes that detection during sampling is either perfect (e.g., Gaublomme et al. 2008) or at least consistent across habitat types (i.e., Caplat and Fonderflick 2009) and species (i.e., Cerezo et al. 2010).
In this paper, we assess when, and to what extent, failure to account for imperfect detection during the sampling process can lead to erroneous conclusions about species' occupancy dynamics, as inferred from estimates of persistence and colonization using detection/nondetection data.We contrast inferences made for a tropical bird community in a fragmented landscape in Costa Rica (Ruiz-Gutierrez et al. 2010) under two separate data processing protocols and associated models.In both scenarios, we used a hierarchical multispecies occupancy model to estimate species-specific probabilities of persistence and colonization, stratified by habitat (forest vs. non-forest) (Dorazio andRoyle 2005, Dorazio et al. 2006).In the first model, we accounted for potential differences in detection by estimating species-specific detection probabilities separately in forest and surrounding agricultural, non-forest habitats.In the second model, we did not include a detection component, thereby assuming that detection was perfect across all species in both forest and nonforest habitat types.We demonstrate how detec-tion biases can lead to erroneous inferences on the processes shaping species distributions, including species habitat associations, for a large and diverse community.We discuss the implications of our results in the context of current ecological research, and highlight the importance of considering sampling biases for effective biodiversity conservation planning.

Avian community data and modeling framework
The data come from a tropical bird community near Las Cruces Biological Station in southwestern Costa Rica, which were surveyed annually from May-September, 2004-2008.Data collection occurred in 21 centrally located point-count stations within seven remnant forest fragments (ranging 1.4 to 25 ha) and a larger forest reserve (227 ha) and 10 stations that were situated in the surrounding non-forest habitat comprised of small-scale parcels of pasture, coffee, and secondary growth (Borgella and Gavin 2005).Each year, the point count surveys were conducted for three consecutive days, rotating the visitation order and starting time across years (see Ruiz-Gutierrez et al. 2010 for further sampling details).
A total of 212 species were observed across five years, but for our comparative analyses, we incorporated all species that were observed at least 30 times (n ¼ 73) with a minimum of five observations in both forest and non-forest habitat to achieve parameter identifiability (Ruiz-Gutierrez et al. 2010).Due to these constraints imposed by the data, our approach differs from that of other multispecies models (i.e., Dorazio et al. 2006) in that we do not formally extend our inferences to all species in the community or include those species that were not observed.Each species was categorized a priori by degree of forest dependency using the classification system developed by Stiles (1985) based on decades of observations and research on the birds of Costa Rica.Species in the ''HIGH FD'' category (n ¼ 13) are considered highly forest dependent, specialized, and largely restricted to forest habitat types.Species in the ''MED FD'' category (n ¼ 39) are thought to use both forest and non-forest habitat types, and thus depend moderately on forest habitat.Species in the ''LOW FD'' category (n ¼ 21) are associated with non-forest, urban and agricultural landscapes (Stiles 1985).
We compared the results from the model in Ruiz-Gutierrez et al. (2010), a hierarchical multispecies model, which estimated species and group-level detection, persistence and colonization, to results from a model in which we assumed that detection was perfect.Ruiz-Gutierrez et al. ( 2010) specified the species-specific occurrence model by assuming that occupancy, z i,j,t (a value of 1 if present and 0 otherwise), for species i at site j in year t was a Bernoulli random variable such that, z i,j,t ; Bern (w i,j,t ) where w i,j,t is the probability of occurrence for species i in site j, during year t.They modeled w i,j,t using the logitlink function: The parameters u1 i and u2 i are the probabilities, on the logit scale, that species i was present in year t given that it was there in the previous year (t À 1) for forest and non-forest habitats (hab j ¼ 1 if site j was in a forest fragment and zero otherwise) (Royle andKe ´ry 2007, Martin et al. 2009).Because it was not possible to determine the occupancy status of species prior to the first year of sampling and sampling in the first survey year occurred at a more limited number of sites than in subsequent years, it was assumed that z i,j,0 ¼ 1 for all species (Ruiz-Gutierrez et al. 2010), which is a slightly different parameterization than the dynamic occupancy models presented in MacKenzie et al. (2003).Thus u1 i and u2 i represent the combined species-specific probabilities of occupancy in the first year of sampling and the probability of occupancy given that the species was present the previous year for all subsequent years.We refer to u1 i and u2 i as the probabilities of persistence in forest and nonforest habitats, respectively.The parameters w1 i and w2 i are the probabilities, again on the logit scale, that species i was present given that it was not there in the previous year, in forest and nonforest habitat, respectively.We refer to w1 i and w2 i as the probabilities of colonization.We used the species-specific estimates of persistence and colonization to calculate each species' turnover rate according to Royle and Ke ´ry (2007).Turn-over is a metric used to describe the rate of change of species occupancy (Nichols et al. 1998) and is defined in this context as the probability that a randomly selected occupied site at time t was not occupied at t À 1 (i.e., colonized at time t).Although we were not specifically interested in the effects of elevation on occupancy, the model accounted for species distributional limits related to elevation with linear (a1 i ) and squared (a2 i ) effects of elevation (Ke ´ry et al. 2009).
If a species was detected at a given site j in any of the three survey days (replicates are denoted as k), then it is assumed that the species was in fact present at site j that year (this is the usual closure assumption).Otherwise, the species could have been present and undetected, or the species was truly absent (MacKenzie et al. 2002).Ruiz-Gutierrez et al. (2010) utilized the repeated sampling protocol (k ¼ 3) to differentiate nondetection from true absences by modeling the observation process to estimate species detection probabilities during each sampling replicate (MacKenzie et al. 2006).Thus detection, x i,j,t,k , for species i at site j in year t during sampling occasion k was a 1 if a species was detected and a zero otherwise.Detection was also assumed to be a Bernoulli random variable, x i,j,t,k ; Bern( p j Á z i,j,t ), where p j is the detection probability of species i at site j (assumed constant over years and sampling periods) given that species i was in fact present at site j in year t.When a species is not present z i,j,t ¼ 0, the probability of observing a species is zero.Detection was similarly modeled using a logit-link function and accounted for species habitat effects: where v1 i and v2 i are the species specific detection probabilities in forest and non-forest habitats (on the logit scale), respectively.
In a separate analysis, we modified the model in Ruiz-Gutierrez et al. ( 2010) by removing the detection component and thus assuming perfect detectability for all species in both forest and nonforest ( p j ¼ 1).We combined data across all sampling periods in a year at each survey location to generate a vector of assumed occurrence of each species such that z i,j,t ¼ 1 if x i,j,t,k ¼ 1 for k ¼ 1, 2, or 3 (i.e., Cerezo et al. 2010).Using this set of combined data, we used the occurrence model only to estimate species-specific persis-tence and colonization probabilities (including species-specific elevation effects) and similarly derived turnover rates.
A benefit of the multispecies modeling approach, and indeed of multi-level models generally (Gelman and Hill 2007), is that estimates of occupancy dynamics can be made more precise by linking species-specific parameters within groups through a hierarchical component (Zipkin et al. 2009).This can lead to an improved composite analysis of species groups (in this case LOW, MED, and HIGH FD species) as well as more precise estimates for individual species (Sauer and Link 2002, Ke ´ry and Royle 2008, Russell et al. 2009).Both the model that incorporated detection and the one that assumed perfect detection included an additional, grouplevel component that linked together speciesspecific parameters, within FD categories, with a common distribution (Dorazio et al. 2006, Royle andDorazio 2008).Thus species-specific estimates of detection (when it was included), persistence, and colonization were assumed to be related exclusively to other species within their forest dependency (FD) category, stratified by habitat type (Ruiz-Gutierrez et al. 2010).More specifically, each of the species-specific estimates (v1 i , v2 i , u1 i , u2 i , w1 i , w2 i ) were assumed to be drawn from a common normal distribution (separate distributions for each parameter, stratified by FD category) and the mean (across species within an FD category) and standard deviation (among species) were estimated across each FD group.

Comparing results between models
The models were analyzed using a Bayesian approach with the programs R (R Development Core Team 2009) and WinBUGS (Spiegelhalter et al. 1999) with identical specifications and noninformative priors.We ran three chains of the model for 10,000 iterations each after a burn-in of length 5,000 and thinned the model by 5. We assessed convergence of the model using r-hat (Gelman and Hill 2007), which compares the variations within chains to the variation among the chains.Analysis using a Bayesian framework is ideal for our models because Markov Chain Monte Carlo (MCMC), as it is implemented in WinBUGS, creates a (posterior) distribution of estimates for each parameter.We used these posterior distribu-tions to make comparisons of parameter values at the species and group levels between models.For both models, we generated posterior distributions for each species-and group-level parameter by summarizing 3000 random samples drawn from iterations of the MCMC algorithm.WinBUGS model code can be found in the Supplement; specifications of the model that includes detection can be found in Ruiz-Gutierrez et al. 2010.Sample data, WinBUGS code files, and R implementation and interpretation files can be found in the following website: A Hierarchical Approach to Multi-species Inference hhttp://137.227.242.23/pubanalysis/communitymodelingi.
Our interests lie in comparing how parameter values varied (1) between models for identical parameters (e.g., quantifying differences in estimates of persistence for LOW FD species in forest habitat between the model that corrected for detection and the one that did not) and ( 2) within models between habitat types (e.g., quantifying differences in persistence for LOW FD in forest vs. non-forest habitat in the model that did not include detection).We focused on group-level persistence, colonization, detection and (derived) turnover parameters and quantified these differences using two approaches.To accomplish our first objective, we compared estimates of persistence, colonization and turnover for each species group (FD category), in forest and non-forest habitats, by calculating the Kolmogorov-Smirnov statistic to determine the percent overlap of the parameters' posterior distributions, as generated in the models with and without detection (Burnham and Anderson 1998).Specifically, we determined the proportion of MCMC iterations that occurred within the overlap of the 95% interior range of both posterior distributions (in models with and without detection) for a given parameter.To accomplish our second objective and determine the directional differences in group-level parameter estimates between forest and non-forest within models, we computed the proportion of the MCMC iterations in which one parameter was greater than the other (Ruiz-Gutierrez et al. 2010).A value that is close to 0.5 suggests there is no difference in parameter estimates.Extreme deviations from 0.5 imply comparatively less overlap in posterior distributions (e.g., a value of 0.95 is interpreted as a 95% probability that one parameter value is greater than the other).

Turnover rates
Species turnover rates were overestimated across all three categories of forest dependency and both habitat types when detection was assumed to be perfect (Fig. 2D-F).The posterior distributions of mean group-level turnover rates under both models did not overlap for any FD category in forest and non-forest habitat types, except for a slight overlap for the HIGH FD group in non-forest (Table 1; Fig. 2D).Mean group-level turnover rates were relatively low (,0.4) and both models estimated higher turnover rates in forest relative to non-forest for the LOW FD group (Fig. 2F).

Persistence and colonization
Species persistence was underestimated when detectability was not included in the model across all three categories of forest dependency and both habitat types (Fig. 3A-C).The 95% credible intervals of the posterior distributions of mean group-level persistence in the models with and without detection did not overlap for any category of forest dependency in both forest and non-forest habitat types, except for a slight overlap in non-forest habitat for the HIGH FD group (Table 1, Fig. 3A).Despite a significant underestimation of persistence when detection was assumed to be perfect, both models estimated higher persistence in forest relative to nonforest for the HIGH and MED FD categories (Table 2).However, the model that did not account for detection estimated higher persistence in non-forest compared to forest habitats for the LOW FD category (Table 2).In the model that accounted for imperfect detection, persistence was not different in forest and non-forest habitats for the LOW FD category (Table 2, Fig. 3C).
There was comparatively more overlap between models in estimates of group-level colonization in both forest and non-forest habitats (Table 1).In forest habitats, colonization was generally overestimated in the model that did not include detection for the HIGH FD category (Fig. 3D), underestimated for the LOW FD category (Fig. 3F), and on par for the MED FD category (Fig. 3E).In non-forest habitat, colonization was more similar for HIGH and MED FD categories (Fig. 3D, E) but overestimated for the LOW FD category when detection was assumed to be perfect (Fig. 3F).Similar to our results for persistence, both models estimated higher probabilities of colonization of forest relative to nonforest habitats for the HIGH and MED FD categories (Table 2).For the LOW FD category, the model that assumed perfect detection estimated that colonization was greater in non-forest relative to forest (Table 2, Fig. 3F).Yet, results from the model with imperfect detection showed the LOW FD category had higher probabilities of colonization in forest relative to non-forest (Table 2, Fig. 3F).Thus, not correcting for biases in detection lead to a false result that LOW FD species have higher colonization probabilities in available non-forest habitats, when in fact all species categories had higher colonization probabilities in forest compared to non-forest habitats (Fig. 3D-F).
Species-specific estimates of colonization appear to be correlated to persistence in both forest (Fig. 4A) and non-forest habitats (Fig. 4B) when estimates are not corrected for detection biases.Yet, when differences in detection are accounted for, there is little evidence that a direct relationship exists between species-specific estimates of persistence and colonization, and there is higher variability in colonization compared to persis- tence, especially in forest habitat.Failing to account for detection also segregates estimates of persistence and colonization in non-forest by category of forest dependency: species in the HIGH category were clustered around low values of both persistence and colonization, followed by species in the MED, and subsequently the LOW category (Fig. 4B).In the model that accounts for species-and habitat-specific detection, estimates of persistence and colonization were much more variable in non-forest and were not segregated by forest dependency category (Fig. 4B).Species-specific estimates of occupancy dynamics can be found in the supplemental information for Ruiz- Gutierrez et al. (2010).

DISCUSSION
Our results indicate that detection varies widely by species and habitat type, to the extent that not incorporating heterogeneity in detection probabilities can lead to a biased estimation of occupancy dynamics for many species by causing (1) an overestimation of species turnover, (2) misleading patterns of persistence and colonization, (3) false correlations between species-specific persistence and colonization, and (4) the misclassification of species into ecological categories (i.e., forest specialists and generalists).These patterns of biases may not have been evident had we applied a static occupancy model (i.e., estimated a single occupancy probability constant across all years), focused on only a limited number of species with sufficient data for  v www.esajournals.orgsingle species models (e.g., a set of indicator of species), or assumed that all species within a group would respond in kind.The extent of bias was even more severe when data were analyzed using just one annual point count (not shown), as per the practice described by Bibby et al. (1992) and recommended in Ralph et al. (1995).This is of important concern when considering that using just one annual or seasonal point count is the most widely applied protocol for monitoring bird communities (Forcey et al. 2006).Our results lend support to the utility of dynamic, multispecies occupancy models to characterize spatial patterns of habitat use across species within a community (Russell et al. 2009, Ruiz-Gutierrez et al. 2010, Zipkin et al. 2010).
The detection biases that we uncovered further suggest that the burden of proof in research based on observational (or count) data should be on demonstrating that detection is perfect or equal across species and habitats for inferences based on detection/nondetection data (MacKenzie and Kendall 2002).The inclusion of detectability may be especially relevant for species-rich communities in structurally complex ecosystems (e.g., the threatened and heavily monitored tropics), where species detection rates are likely to contrast sharply with one another, and may be further stratified by habitat type.Failing to account for detectability is therefore likely to be a major pitfall to quantifying biodiversity across fragmented and human-modified landscapes in diverse regions, which is not commonly taken under consideration (Gotelli andColwell 2001, Fahrig et al. 2010).

Ecological and conservation implications of detection biases
For most species, we found significant differences in estimates of persistence, colonization and turnover driven by differences in detectability across habitat types.At the group level, turnover was systematically overestimated while persistence was underestimated in both forest and matrix habitat types.Failing to account for heterogeneity in species detection rates resulted in faulty ecological knowledge regarding the functional relationship between species and their environment, and could consequently misguide management and conservation initiatives.Working under the widely used assumption that spatial distribution patterns are the end result of species habitat selection (to procure gains in lifetime reproductive success), biased estimates of persistence and colonization can lead to errors in identifying biological scenarios under which this assumption is violated (e.g., source-sink systems, ecological traps) (Runge et al. 2006).Fig. 4. Species-specific mean posterior estimates of persistence and colonization for forest (A) and non-forest (B) habitat, as estimated using the hierarchical multi-species model that accounts for differences in detectability (black) and the model that assumed perfect detection (grey), by category of forest dependency (HIGH: circle, MED: triangle and LOW: cross FD).The light grey trend lines represent the relationship between persistence and colonization for the model that does not include detection.There was no relationship between persistence and colonization when detection was included in the model.v www.esajournals.orgFor example, incorrect estimates of relatively higher persistence and/or colonization in agricultural land-uses could misclassify these lands as ecological traps if reproductive success for LOW FD species was found to be lower in these relative to forest habitats (Schlaepfer et al. 2002).
The biased estimates of persistence and colonization in non-forest habitat for LOW FD species might incorrectly attribute a high biodiversity conservation potential to these matrix habitats, when in reality forest habitats appear to be driving the persistence of most bird species in this tropical community.Much research has demonstrated that conservation of biodiversity in protected areas and remnant habitats necessitates incentives aimed at maintaining or improving the quality and suitability of the surrounding matrix habitats (Prugh et al. 2008, Franklin and Lindenmayer 2009, Fahrig et al. 2010).Yet, our results suggest that in order to maximize conservation returns, conservation policy should be based on unbiased estimates of persistence and colonization of primary habitats (e.g., forested areas) as well as surrounding habitat types for multiple species.Such an approach can quantify the direct relationship between anthropogenic land-uses and species colonization, and identify specific characteristics of these secondary habitats which can be targeted and manipulated to ultimately facilitate movement of species and increase habitat provisions (i.e., agri-environment schemes).
Caution should further be exercised when utilizing estimates of species persistence as indicators of current and predicted habitat use.Although persistence and colonization may be correlated in some landscapes, detection biases resulted in an incorrect correlation between species-specific estimates of persistence and colonization for a large proportion of species in our community of tropical birds.By correcting presence-absence information for differences in detectability among species and habitat types, we found that the suite of habitat variables for which a given species shows a high degree of persistence might not be good predictors of the habitat types or environmental conditions under which the same species would colonize, if available.Correlation between persistence and colonization is an implicit assumption of habitat suitability models (HSM), a set of increasingly popular tools used to predict the likelihood of species occurrences across a range of current and projected environmental variables (Hirzel and Le Lay 2008).Similar to occupancy models, HSM, based on detection only or detection/nondetection data (HSM-D), presume that spatial distributions of species are representative of their true ecological requirements by implicitly assuming that the observed realized niche has a high degree of overlap with the fundamental niche of the species (Guisan and Thuiller 2005).This suggests that the environmental conditions or habitat types where a species has a high probability of being detected defines a set of conditions used to extrapolate where the species is likely to occur under current or future conditions.When HSM-D models are used to predict changes in species distributions as a result of future habitat loss or climate change, the additional underlying assumption of the statistical model is that the conditions where species are represent the conditions that they are most likely to colonize after a disturbance event.In other words, HSM-D assume that static detection is a direct measure of the likelihood of future occurrence (i.e., colonization), and this could be the reason behind their high vulnerability to false absences in suitable habitats, a potential result of unequal detectability (Gu and Swihart 2004, Guisan and Thuiller 2005, Thuiller and Munkemuller 2010).We found a high degree of variability in colonization that was uncorrelated to speciesspecific estimates of persistence when detectability was included in the estimation process.Since HSM-D models do not explicitly correct for errors in the observational process, our results may elucidate an additional and important source of bias in the predictive accuracy of HSM-D (Guisan and Thuiller 2005), an issue which has not been previously addressed.Our overall low estimates of colonization further suggest that if species need to gradually colonize other habitats (e.g., as a result of predicted climate change and/or loss of habitat), they may do so at relatively low rates.Although data for HSM are often limited, inclusion of dynamic processes, to the extent possible, could increase the accuracy of projections of future species distributions under climate change scenarios by allowing for explicit predictions regarding the functional relationship of environmental condiv www.esajournals.orgtions and species persistence and colonization.The correction of differences in detectability in these dynamics models can also greatly enhance the overall accuracy of predictions of HSM-D, resulting in much improved management practices for conservation planning of reserve networks and biological corridors for multiple species, and increasing the probability of mitigating current rates of biodiversity loss (Nicholson et al. 2006).

Misclassification of species into ecological categories
The apparent segregation of species-specific persistence and colonization estimates in nonforest habitat by category of forest dependency when detection is assumed to be perfect (Fig. 3B) suggests that the classification system developed by Stiles (1985) was influenced by how detectable these species were in non-forest habitats.For example, species with higher detection probabilities in the agricultural matrix, like the tropical kingbird (Tyrannus melancholicus; detection forest: 0.09 (60.06 SD), detection non-forest: 0.41 (60.07 SD)), which perches visibly, calls frequently, and exhibits aerial displays to capture insects in open areas, is traditionally assumed to not ''need'' or ''prefer'' forest.Yet our results show that this may simply be because kingbirds are not as easily detected in the forest canopy where they also forage for insects.Other species that are rarely detected outside of forested habitats, such as the blue-crowned manakin (Pipra coronata, detection forest: 0.48 (60.07 SD), detection non-forest: 0.23 (60.13 SD)) are often assumed to be more affected by changes in the amount and configuration of available forest.Our results suggest that the availability of forest habitat is important to the survival for many, if not all, species including those that are typically associated with matrix habitats.The contrast between predefined habitat requirements for species (especially in the LOW FD category) and our results highlighting the need of forest for community persistence, is problematic considering that a significant proportion of applied research focuses only on species classified as habitat specialists (i.e., vulnerable) in birds (e.g., Julliard et al. 2006), mammals (e.g., Puettker et al. 2008) and insects (e.g., Rand and Tscharntke 2007).Traditionally, classification schemes are often derived solely from observational ap-proaches that do not account for detection biases (Azeria et al. 2007, Pardini et al. 2009).Research based on such classification schemes may misrepresent true patterns of habitat use for a large number of species in a community, and misidentify or underestimate which species will be most impacted by environmental changes.

Conclusions
Our results highlight the ecological and conservation consequences of relying on inaccurate or misinformed estimates of community dynamics and related studies.We utilized an analytical approach to correct for such biases, yielding a more accurate representation of the functional relationships of species, communities and their environments.When detection rates vary widely for species in different habitat types, occupancy dynamics are likely to be biased for multiple species, leading to overestimation of species turnover rates, and false and misleading patterns of persistence and colonization.We believe that similar biases might exist in other taxa in which detectability has been proven to be habitat specific (e.g., frogs, mammals).The accuracy and utility of observational data for making informed conservation and management decisions will be maximized when approaches: (1) account for heterogeneity in species detection, (2) focus on the dynamic processes of species turnover, persistence and colonization rather than static species occurrence patterns, and (3) include as many species as possible such that inferences are not based on arbitrary species groupings and categorizations.Such approaches will yield stronger predictions of how species may respond to future changes in environmental and climatic conditions, and ultimately inform high-quality management and conservation of biodiversity.

Fig. 1 .
Fig. 1.Species-specific mean posterior estimates of detection in forest and non-forest habitats, as estimated using the hierarchical multispecies model, for 73 species in southwestern Costa Rica.The bars represent the standard deviations (SD) of each estimate.

Fig. 2 .
Fig. 2. Density plots for the posterior distributions of the community-level means for detection (A-C) and turnover (D-F), as estimated using the hierarchical multispecies model correcting for detection biases (dark lines) and assuming perfect detection (grey lines).Solid lines represent distributions for forest and dotted lines are nonforest.

Fig. 3 .
Fig. 3. Density plots for the posterior distributions of the community-level means for persistence (A-C) and colonization (D-F), as estimated using the hierarchical multispecies model correcting for detection biases (dark lines) and assuming perfect detection (grey lines).Solid lines represent distributions for forest and dotted lines are non-forest.

Table 1 .
Proportion of posterior distribution overlap of the 95% interior range of group-level parameter estimates in models with and without detection.Parameter value overlap of persistence, colonization and turnover are shown for both forest and nonforest habitat types for each category of forest dependency (HIGH, MED and LOW FD).

Table 2 .
Probability that group-level parameter estimates of persistence and colonization are higher in forest compared to non-forest habitat, for models with and without detection, by category of forest dependency (HIGH, MED and LOW FD).The bold values indicate when the model without detection misidentified the directional relationship.