The effects of vegetative type, edges, fire history, rainfall and management in fire- 1 maintained habitat 2

. The combined effects of fire history, climate, and landscape (e.g., on habitat specialists need greater focus in fire ecology studies, which usually only emphasize 15 characteristics of the most recent fire. Florida scrub-jays are an imperiled, territorial species that 16 prefer medium (1.2-1.7 m) shrub heights, which are dynamic because of frequent fires. We 17 measured short, medium, and tall habitat quality states annually within 10 ha grid cells (that 18 represented potential territories) because fires and vegetative recovery cause annual variation in 19 habitat quality. We used multistate models and model selection to test competing hypotheses 20 about how transition probabilities vary between states as functions of environmental covariates. 21 Covariates included vegetative type, edges (e.g., roads, forests), precipitation, openings (gaps 22 between shrubs), mechanical cutting, and fire characteristics. Fire characteristics not only 23 included an annual presence/absence of fire covariate, but also fire history covariates: time since 24 the previous fire, the longest fire-free interval, and the number of repeated fires. Statistical 25 models with support included many covariates for each transition probability, often including 26 fire history, interactions and nonlinear relationships. Tall territories resulted from 28 years of fire 27 suppression and habitat fragmentation that reduced the spread of fires across landscapes. Despite 28 35 years of habitat restoration and prescribed fires, half the territories remained tall suggesting a 29 regime shift to a less desirable habitat condition. Edges reduced the effectiveness of fires in 30 setting degraded scrub and flatwoods into earlier successional states making mechanical cutting 31 an important tool to compliment frequent prescribed fires. 32

We identified a suite of covariates predicted to influence the transition probabilities between 165 habitat states. The static (non-time varying) covariate "oak" identified potential territories 166 intersecting well-drained soils, and "flatwoods" identified potential territories that included 167 moderately drained soils with smaller patches of oaks than are found on well-drained soils. The 168 static covariate "edge "recorded whether a man-made or forest edge intersected a potential territory. The dynamic (time-varying) covariate "open" was a categorical variable that 170 distinguished whether scrub in potential territories was open (>50% or area had open sandy gaps 171 between shrubs) or closed (few gaps). The dynamic covariate "cutting" distinguished territories 172 where at least ¼ of the territory was subject that year to mechanical cutting of trees and shrubs. 173 The fire history of territories differed across the KSC/MINWR landscape; in order to 174 encompass these differences we incorporated several dynamic covariates related to the fire 175 histories that we predicted had an influence on transition probabilities. The dynamic covariate 176 "fire" distinguished territories where at least ¼ of the territory burned that year based on remote 177 sensing and fire records (Shao and Duncan 2007). We chose 3 fire history variables that reflected 178 different hypotheses about how fire history would affect habitat transitions that we predicted to 179 be uncorrelated with fire and each other. The first fire history covariate time-since-fire, "TSF", 180 represented the number of years without fire prior to the current interval and therefore 181 independent of "fire" during the current interval. We selected TSF because growth is more rapid 182 soon after fires and then slows (Schmalzer 2003 196 We used multistate models to analyze annual transition probabilities between states (Fig. 1)  Each multistate model consisted of a likelihood combining 3 multinomials, 1 for each of the 203 3 states. We estimated transition probabilities for state changes (e.g., short-to-medium); 204 transition probability estimates of states remaining the same between years (e.g., short-to-short) 205 were estimated by subtracting the transition probability estimates of state changes from 1.0. The 206 transition of short-to-tall was constrained to zero because it didn't occur, which enhanced with data on these management efforts.

225
No simple models that included only vegetation type (oak/flatwoods) or habitat management 226 actions (fire/no fire, cutting/no cutting) had support, whereas 6 more complicated models had > 227 99% of the empirical support (Table 1). There was much similarity amongst the top models, but 228 differences among transition probabilities regarding which covariates were important. Models 229 with support included all covariates that were previously important (oak, edge, fire, and cutting) 230 and new covariates involving fire history (time-since-fire, longest fire free interval, number of 231 fires in previous 10 years, openings) and annual rainfall (standardized precipitation index).

232
Table 2 compares our a priori predictions with results for all covariates important in the best 233 supported model. We presented only the best model because other top models were similar and 234 β's absent in the top model involved interactions that had standard errors many times greater 235 than mean estimates. The best model had mean β's with CI's that overlapped zero, but excluding 236 these effects during post hoc analyses produced models without AICc support. The main effect for β's describing the oak versus flatwoods often had CI that overlapped zero, but we included 238 the oak versus flatwoods effect because all supported models included an oak versus flatwood 239 effect, and many covariates had different effects depending on whether the site was oak or 240 flatwoods ( Table 2). All supported models had an effect for whether fire occurred in the current 241 interval, but the CI's overlapped zero for transitions from earlier to later successional states. transitions between short-to-medium, medium-to-short and for medium-to-tall, which were 247 transition probabilities that previously had low sample sizes and therefore limited our abilities to 248 test many covariate effects. In this study, transitions from short and medium to other states were 249 more common because the states had greater relative abundances, and we estimated transition covariates, and many new covariate effects (fire history, precipitation, openings) for many 290 transitions (e.g., short-to-medium, medium-to-short, medium-to-tall).

291
Restoration and management actions (e.g., fire, cutting) had important effects, but 292 environmental factors often had greater effects on transition probabilities (e.g., edge effects on 293 tall-to-medium: Table 2). Edges (roads and forests), primarily resulting from anthropogenic 294 factors, were among the most influential factor across transition probabilities; these strong

306
Fire history effects were practical to study in our system because fires occurred every few 307 years instead of decades, or longer. The presence/absence of fire during an annual interval 308 usually had a larger effect than fire history variables, except that time since fire (TSF) had a great 309 effect for the short-to-medium transition in oak. The presence/absence of fire might have had a 310 lesser impact than TSF because short scrub generally lacks enough fuels to burn extensively.

311
The TSF nonlinear effect could be explained by growth being most rapid a few years after fire 312 (Schmalzer and Hinkle 1992). 313 We predicted stronger effects from other fire history covariates, especially for the length 314 of the fire free interval (LFI) because scrub is difficult to restore once it is unburned for >20 315 years. We expected that LFI would be important because increasing underground biomass and tall-to medium transitions contradicting our a priori hypothesis, but the effect was poorly 321 estimated (CI overlapped zero). One explanation might be that it takes at least 3-5 years for 322 enough fuels to accumulate to carry fire and greater fuel levels might be needed for fires to burn 323 severely enough to spread into tall scrub patches. Another explanation is that fires did not burn

357
Restoration programs often focus on conducting enough management to get past a threshold 358 that alters system behavior, removing feedback loops that lead to a degraded state, and enhancing 359 feedback loops that produce a desired stable state (Suding 2011). Burning during the best season 360 to stimulate grasses and promote fire spread might be advantageous, but fire managers in our 361 study sites had extreme limitations regarding meteorological and operational constraints that 362 made burning only during the dry-to-wet transitional season (e.g., May) difficult. Cutting was an 363 expensive tool, and it may be unreasonable to remove most forest edges and edges associated 364 with human landscape features so that prioritization of management efforts becomes necessary. 365 We have observed many habitat and population management restoration successes in particular Funding for prescribed fires often focuses on maximizing fuels reduction allowing fuels to 372 accumulate until fires can burn them extensively, causing potential Florida scrub-jay territories 373 to have a large sink (short, tall) to source (medium) habitat ratio, as occurred herein. We believe 374 "optimal habitat management" might be better at reducing catastrophic fire risk than fuels 375 management in priority areas (Breininger et al. 2014a). In an optimal habitat management 376 strategy, prescribed fires would be initiated sooner than a fuels reduction strategy by attempting 377 prescribed fires before all fuels are likely to ignite thus creating transitory openings and 378 heterogeneity among shrub patches at the territory scale. In habitat occupied by Florida scrub-379 jays near carrying capacity, optimal habitat management would attempt mosaic fires to provide 380 some unburned patches to serve as nest sites, provide acorns, and areas to escape predators. Such   Superscripts refer to particular transition probabilities between states S = short, M =medium, T = tall. All top models included the effects of edge and oak × fire for all ѱ. All top models included a quadratic relationship for time since fire (TSF) in oak for ѱ SM ; the covariate TSF was not supported in flatwoods when predicting ѱ SM . All top models included a linear TSF effect for ѱ MT  Transition probability superscripts were SM for short to medium, MS for medium to short, MT for medium to tall, TS for tall to short, TM for tall to medium. Superscripts that included oak had a β specific to oak, superscripts that included flatwoods had a β specific to flatwoods and superscripts without oak or flatwoods had a single β for both oak and flatwoods transitions. Abbreviations TSF = Time-since-fire, LFI = longest fire interval, SPI = standardized precipitation index. Nonlinear relationships were represented by a squared term (quadratic). Table 3. Annual transition probability estimates (95% CI).
year short-medium medium-short medium-tall tall-short tall-medium