Literature search protocol
We conducted a systemic review of studies assessing the effects of fire on animal communities. In April 2018, we searched studies using the Scopus database employing the following combinations of keywords: (mammal* OR bird* OR reptil* OR arthropod*) AND (fire* OR (fire* AND vegetatio*) OR (burn*)) AND (richne* OR abundanc*). This search retrieved more than 14,000 documents, from which we selected only those scientific publications that were related to fire and fauna. Our initial screening and pre-selection process yielded 566 papers describing the effect of fire on fauna, or the faunal responses to past fire.
We scrutinized the 566 papers and the 200 studies evaluated by Pastro et al. (2014); and we selected only those meeting all the following criteria (Fig. 1): (a) Evaluate the effect of fire on the richness or the abundance of mammals, reptiles, amphibia, birds, and arthropods; (b) Have a before and after (i.e., comparison before and after the disturbance, Christie et al. 2019); a control-impact (i.e., comparison between control and fire-impacted sites after the disturbance, De Palma et al. 2018; Christie et al. 2019); or a before–after-control-impact (i.e., compare between control and fire-impacted sites before and after the disturbance Osenberg et al. 2006; Christie et al. 2019); (c) Possess a control treatment that had remained unburned for at least 25 years before the sampling date, since after that time, fire affected areas will present similar attributes to unburned areas (Henriques et al. 2006; Kelly et al. 2011); (d) Have information about the history of fire events (i.e., year and extension of the last fire); (e) Do not present an average value of the treatments with different fire histories (i.e., different years since the last fire event).
The selection process yielded 162 publications published between 1959 and 2018, 122 from the search in the Scopus database and 40 from Pastro et al. (2014) (Additional file 1: Appendix S1). These included 77 replicated studies (at least two control and two treatment plots or sites) and 85 unreplicated studies. We extracted the title, authors, authors affiliations, DOI, general conclusions, implications for conservation or management, and richness and abundance values from each study. In our database, we annotated both richness values and estimations, such as those performed using CHAO or rarefaction methods. CHAO focuses on comparing accumulation curves asymptote (Chao and Chiu 2016), whereas rarefaction methods standardize the size and cover of samples (Chao and Jost 2012). Both technics contribute to inferring species richness and comparing the species richness values of different communities obtained with different sampling techniques. As abundance descriptors, we annotated raw counts, density estimations, and flock sizes (described in some bird studies). Finally, studies of reptiles and amphibians were aggregated into one category ('herpetofauna'), as study cases usually evaluate responses in this way. Unless otherwise mentioned, the effect of fire on species richness and abundance was analyzed for 145 and 127 studies, respectively (Additional file 1: Appendix S1 and Additional file 2: Appendix S2).
Explanatory variables
From each study, we also extracted information of the following categorical predictors: fire type (managed or wildfire), biogeographic region (Afrotropical, Australian, Indomalayian, Nearctic, Neotropical, Palearctic, and Subantarctic), country, continent, and the year of the fire event under study. We followed Shlisky et al. (2007, 2009) to classify each study according to the ecological role of fire: (i) fire-dependent (i.e., savannas and conifer forests), those in which the biota evolved in and adapted to the presence of fire; thus, fire is necessary for the maintenance of biodiversity and ecological processes; (ii) fire-sensible, those in which biota has not evolved in and adapted to the presence of fire; also, where climatic conditions are not proper for fire ignition (i.e., high humidity zones, most of these ecosystems are located at the tropics); or (iii) fire-independent (i.e., deserts), those in which fire has a low probability of occurrence due to the lack of fuel sources (Shlisky et al. 2007, 2009).
In addition, six variables were recorded as descriptors of the fire regime for each study: (i) fire severity (high, medium, and low) according to the information provided by the authors of each study; (ii) the time elapsed between the fire and the sampling event (measured in years and coded in the database as 'years.from.fire'); (iii) the spatial extension of the fire (in hectares); (iv) the number of times the area had burned before the sampling event ('burned.times'); and, if more than one event occurred in the area; (v) the time interval in which fire events occurred (in years—'Interval.of.time'); and (vi) the time interval separating their occurrence (in years—'interval.between.fires').
Effect size
Following Pastro et al. (2014), we used the log response ratio between burnt and unburnt areas as effect size. This effect size was calculated for each study to show the magnitude of the effect of the fire. This is estimated as the log-transformed ratio between values at burnt and unburnt—ln(Xe/Xc), where Xe and Xc represent the species richness or abundance at burnt and unburnt treatments, respectively (Rosenberg et al. 2000; Salo et al. 2010). Positive effect sizes (positive CIs) indicate that fire increases species richness or abundance, while negative effect sizes (negative CIs) indicate a negative impact. An effect size ln(Xe/Xc) ~ 0 (with CI including 0) means that the fire has no effect. This metric was chosen over more traditional effect sizes, such as Hedges d or ln(R), because it does not require within-study variance (Salo et al. 2010). A large proportion of our data set consisted of unreplicated or pseudo-replicated studies in which within-study variance was not reported.
Data analyses
We first tested for context dependence on the responses due to fire type, fire ecology, biogeographical region, or community type. These four categorical variables are known to have some incidence on faunal responses; therefore, they can interact with the fire regime to produce unexpected responses (e.g., Pastro et al. 2014). The biogeographic region and the fire ecology, for instance, have a direct incidence on the faunal responses; species from fire-prone regions are generally more resilient to punctual fire events (Shlisky et al. 2009). The resilience of biological populations, on the other hand, is associated with traits, such as body size, diet, reproductive rate, and movement capacity (Sutherland and Dickman 1999; Santos et al. 2014), which exhibit a high variability among taxonomical groups (Litt and Steidl 2011). Finally, whether fires are set for management (prescribed fires) or wildfires (uncontrolled and spontaneous), the type of fire may interact with the fire regime, leading to synergistic responses. For example, prescribed fires can reduce fuel charges and favor vegetation types (Roberts et al. 2015), affecting fire regime components, such as history and severity.
After testing for context-dependency, we evaluated whether species richness or abundance respond to differences in the fire regime. As we had a skew distribution on some descriptors, we decided to recode them as categorical predictors, keeping a similar number of observations in each one. The time since the last fire event ('years.from.fire') was recoded in ten categories (< 0.5, 1, 2, 3, 4, 5, 6, 7, 7–15, > 15 years) that describe a gradient of recovery time. The greater the category, the less negative is expected to be the impact of fire on the fauna, as species would have more time to recolonize impacted zones. On the other hand, the number of times that the area had burned before the sampling event ('burned. times') was recoded in six categories (1, 2, 3, 4, 4–10, > 10 times) that describe a gradient of intensity. The greater the number of times, the bigger is expected to be the impact of fire on fauna. We could not evaluate the effect of the other components (area, the time interval in which fire events occurred, and the time interval separating their occurrence), for more than half of the studies did not include the information required to estimate them.
To evaluate how fauna responses to fire changed according to the factors explained above, we used linear mixed-models fitted by maximizing the restricted log-likelihood (REML). We fitted one model per predictor, including the categorical predictor as the fixed effect and the publication I.D. as the random effect. We set I.D. as the random term to allow the fixed effects to vary for each study (I.D.), since most studies compared burned treatments against a single control treatment, leading to 'pseudoreplicates' nested within each study location. We also exclude the intercept estimation from the models, since intercept models for categorical predictors with more than two levels use one level as the reference, with which all other levels are compared. Then, removing the intercept allows determining whether each level is significantly different from zero rather than the less interesting reference level.
We estimate 95% confidence intervals (CI) of the effect size ln(Xe/Xc) to assess the levels of each categorical predictor. A CI > 0 points to a positive effect (increase) on species richness or abundance. Meanwhile, a CI < 0 points to a negative effect (decrease) (Rosenberg et al. 2000). Confidence intervals were estimated using the "confint.merMod" function (bootstrapping percentile method with 1000 permutations). Comparisons among category levels were performed using a log-likelihood test using the function 'anova' of the 'lme4' packages.
All analyses were performed using R 3.6.1 (R Core Team 2020). All the data (Additional file 1: Appendix S1) and coding used in this study (Additional file 4: Appendix S4) are available. All linear mixed-effects models were fitted using the R package 'lme4' (Bates et al. 2015).
Publication bias or assessment of the risk of bias
As a standard quantile plot method was not possible for the unreplicated studies, we also used the funnel plot method by plotting the sample size of all the experiments against their effect size ln(Xe/Xc). No evidence of publication bias was found with either of these methods (Additional file 3: Appendix S3).