We conducted one search for each of the key responses, (1) Fire AND season AND plant recruitment, resulting in 550 papers, of which 370 were unique. The first screening determined if each paper was relevant to the topic, through the reading of the title and abstract; 667 were removed at this stage. The remaining 329 were thoroughly examined, of which 277 were excluded based on criteria outlined below, leaving 52 articles included in this meta-analysis19,43,50,52,55,59,60,61,62,63,64,65,66,67,68,69,70,71,72,73,74,75,76,77,78,79,80,81,82,83,84,85,86,87,88,89,90,91,92,93,94,95,96,97,98,99,100,101,102,103,104,105, which produced a data set of 133 pairwise comparisons.
2); (3) report on plant survival, reproduction or recruitment (articles that measured diversity, richness or evenness were discarded); and (4) contain data on means (or pairwise F-statistics reported within ANOVA tables), estimation of variation (including standard deviation or standard error > 0) and a clearly defined and stated sample size for each experimental group. From each of the papers, we also extracted secondary characteristics that we aimed to examine as correlates of plant response to fire season. Due to limited and inconsistent reporting of fire and inter-fire metrics (Fig.
To assess how impactful these interactions were on plant responses, we recorded all mentions of these fire metrics from each of the sampled articles. For each pairwise comparison, Hedges’ g was calculated, along with an estimated standard error and 95% confidence bounds around the mean effect using the “esc” package109 within the R statistical programme110.
We implemented three random effects models over each of the three topics of interest (i.e., recruitment, reproduction and survival). Instead, we assumed that effect size will differ between climate types, vegetation communities and plant functional groups. This was implemented only when Hartung-Knapp confidence intervals were narrower than CIs from classic random effects models that used DerSimonian-Laird estimations111. To further understand differences within and between fire seasons for each topic, we assigned each pairwise comparison to a subgroup based on the climate type and vegetation type of the study ecosystem and the plant functional group to which each species belongs that was reported within the comparisons.
Vegetation type can be broadly described as forest, woodland, savanna, shrubland or grassland, and reflects variation in local geological and climate interactions that define the extent of distinct ecosystems. Many plant species that inhabit fire-prone ecosystems have developed traits that support recruitment, reproduction and survival under historical fire regimes15,29,31.
The distribution of these fire-adapted traits within and between ecosystems is driven, at least in part, by the long-term characteristics of the fire regime29,120,121. In this analysis, we focused on three functional traits best represented in the literature relating to recruitment, reproduction and survival. The most common functional group analysed in our data set was resprouting plants, occurring in 29 articles. From the 53 reviewed articles, the resulting data set contained122 pairwise comparisons, with effect sizes (Hedges’ g) ranging from −37 to 80.
Using a random effects model resulted in a global mean, excluding outliers, reported in Hedges’ g of 0.3505 (95% CI: 0.0005; 0.7005, P-value = 0.04). These data comprise research from six continents, including a large representation from North America, Australia, Mediterranean Europe and South Africa, and across 10 climate regions, including 42 comparisons from humid subtropical (Cfa) climate regions. There were 29 pairwise comparisons between fires at different times within the historical fire seasons with a large proportion of these comparisons (N = 12, ~40%) coming from climates with predictable seasonal rainfall, e.g., tropical savannas (Aw). For each of the three topics (i.e., survival, reproduction and recruitment), we assessed each set of responses for publication bias and p-hacking. Publication bias was assessed by visually identifying asymmetry in funnel plots and applying Egger’s regressions125. The P-curve can provide evidence of P-hacking if there is an overrepresented and skewed proportion of P-values close to 0.05126.
Combined, the P-curve analysis and the Egger’s regression show no evidence of publication bias or P-hacking.