Paired sites, one subject to extensive control (treatment) and one with no control (non-treatment), were located at Coromandel (treatment: 2300 ha, 36°56′S, 175°37′E; non-treatment: 1200 ha, 36°58′S, 175°39′E) and Urewera (850 ha, 38°17′S, 177°11′E; 850 ha, 38°14′S, 177°11′E) in the North Island, and Haast in the South Island (1300 ha, 44°03′S, 168°47′E; 1300 ha, 44°06′S, 168°39′E). Each pair of sites had similar topography, altitude and forest composition (assessed using the Recce method; Hurst & Allen 2007). Extensive possum control was carried out at: Coromandel, 1995, 1999 and 2003 prior to the study and during the study in 2006; Haast, 1995–96 and 2000–01 prior to the study, but the next control operation was delayed until after the study period; and Urewera, 1993–94, partially in 1996–97, 1998–99 and 2002–03 prior to the study, and during the study in 2006–07.
Experimental protocols were based on techniques used by DOC to assess the abundance of possums (Trap-Catch Index (TCI); National Possum Control Agency 2010) and their impacts on tree canopies (Foliar Browse Index (FBI) and Foliage Cover Index (FCI); Payton, Pekelharing & Frampton 1999) and survival. Additional data were collected to characterize foraging patterns at a local scale because previous studies found significant site differences in possum impacts (Duncan et al. 2011). The experimental design included measurements of tree size using diameter at breast height (DBH), local availability of other food sources and possum browse at the patch scale (Fig. 1).
In each area, two tree species with foliage preferred by possums and one with foliage not preferred (Owen & Norton 1995; Sweetapple 2003; Sweetapple, Fraser & Knightbridge 2004) were selected as ‘indicators’ (Table 1) to assess the outcomes of extensive possum control. A criterion for selecting species was that they were relatively common and widespread throughout the area. Non-preferred species acted as experimental controls for non-selective disturbances such as storms, drought and earthquakes. Smaller numbers of a third highly preferred species were selected from sparse populations at Coromandel (Dysoxylum spectabile) and Haast (Fuchsia excorticata).
Table 1. Preferred and non-preferred tree species at each site, annual tree mortality at non-treatment and treatment sites [sample size n of trees that were tagged at the start of the study (Time1) and relocated during the remeasurement period (Time 2)]. Trees with unrecorded browse data (FCI and/or FBI), or that could not be relocated, at Time 2 were not used for analysis
|Coromandel|| Weinmannia silvicola ||WEISIL||Yes||0·18||225||0·28||213||0·525|
| Olearia rani ||OLERAN||Yes||2·17||221||0·94||238|| 0·014 |
| Dysoxylum spectabile ||DYSSPE||Yes||6·20||84||0·0|| 21|| 0·008 |
| Knightia excelsa ||KNIEXC||No||0·25||237||0·09||206||0·359|
|Haast|| Weinmannia racemosa ||WEIRAC||Yes||0·0||229||0·41||247||0·957|
| Schefflera digitata ||SCHDIG||Yes||15·53||200||12·16||197|| 0·040 |
| Fuchsia excorticata ||FUCEXC||Yes||21·20||102||19·73|| 3||0·5|
| Nothofagus menziesii ||NOTMEN||No||0·60||202||0·57||212||0·5|
|Urewera|| Weinmannia racemosa ||WEIRAC||Yes||6·87||218||0·11||237|| <0·001 |
| Beilschmiedia tawa ||BEITAW||Yes||0·33||229||0·40||249||0·621|
| Knightia excelsa ||KNIEXC||No||0·11||222||0·12||211||0·5|
Preliminary power analysis indicated that 200 trees of each species per site would be required to have an 80% chance of detecting a treatment effect in annual tree survival where the difference in annual mortality is ≥2% and the interval between measurements is ≥4 years (B. Reddiex unpublished data). Initial measurements (Time 1) and remeasurements (Time 2) were carried out in 2004 and 2009, respectively, at Coromandel and Haast, and in 2006 and 2010 at Urewera. Sampling transects, consisting of 5–15 plots at 50-m intervals, were located randomly in each site with randomly selected bearings. Transects and plot centres were marked and GPS coordinates recorded. Within 20 m of each plot centre, up to four trees of each indicator species at least 10 m apart where possible were permanently marked using numbered tags. Trees were included if they measured ≥10 cm DBH (1·4 m above the ground), and if the canopy was visible from the ground.
Browse damage was recorded using Foliar Browse Index (FBI) in five categories, 0–4, indicating that 0%, 1–25%, 26–50%, 51–75% or >75% of leaves, respectively, showed evidence of possum browse. Foliage cover was recorded using Foliage Cover Index (FCI), the fraction of sky occluded by leaves, observed from beneath the centre of the tree crown looking up. Occlusion was recorded in 10% categories, that is, 0·05 = 0–10%, 0·15 = 10–20%, etc. For each indicator tree, DBH, FCI and FBI were recorded at Time 1, and again at Time 2 for those trees still alive. At Time 2 the status (alive or dead) of each tree was recorded.
At each plot, the closest indicator tree to the plot centre was selected as the ‘focal tree’. The DBH, Alive/Dead status and FBI of all trees (with DBH ≥ 10 cm) within 5 m of the focal tree were measured (irrespective of species). These ‘neighbourhood trees’ were used to characterize a patch of forest that could be selected by foraging possums.
Possum Trap-Catch Index (TCI) was obtained at each site as near to Times 1 and 2 as possible following the standard protocol of randomly located lines within each area, each line consisting of 10 or 20 traps set for three nights (National Possum Control Agency 2010). TCI lines were located independently of the transects for measuring browse impacts. DOC measured TCI on treatment sites after each control operation; we used DOC trap lines that overlapped our study area to calculate mean TCI values (6–18 lines per site). In the non-treatment areas, we measured TCI using 7–10 lines per site, with some additional DOC measurements at Haast.
Although there was no possum control at the Haast treatment site between Times 1 and 2, we assumed it to be notionally ‘treated’ when testing for treatment effects. This is just within the likely period of reduced possum abundance, taking into account their expected recovery rate (Morgan & Hickling 2000).
Treatment effects on possum browse (FBI), canopy biomass (FCI) and tree mortality
Comparisons were carried out between treatment and non-treatment sites in each of the three areas, at Time 1 and Time 2 separately. Distributions of FBI values for each indicator species at each site and time were compared using Fisher's exact test for count data (Hypothesis 1a). To account for the low frequency of higher FBI values, we also compared the proportion of trees in the non-zero browse categories (‘any browse’) and the proportion of trees in FBI category 2 or higher (‘moderate-to-severe browse’). A one-sided binomial proportions test was used to determine whether the proportion of trees in each category was greater at the non-treatment site.
Distributions of FCI values for each indicator species at each site and time were compared using a one-sided t-test (Hypothesis 1b). The relationship between FBI and FCI was assessed using data pooled across treatment and area (where applicable) for each preferred species at Time 1.
Annual mortality of each tree species was compared between treatment and non-treatment sites using a one-sided t-test for the difference of two proportions (Hypothesis 1c). The observed annual mortality (m) for each tree species (by area and treatment) was calculated using N1(1−m)yrs=N2 where Nt is the number of trees alive in the sample at time t and yrs is the number of years between observations.
We examined whether tree mortality differed depending on FCI at Time 1 (Hypothesis 3). Due to the low number of recorded tree deaths, FCI values were grouped into three categories (low = <0·4; moderate = 0·41–0·7; high = 0·71–1·0) and pooled across treatment and non-treatment sites. Distributions of the FCI groups were compared, by species, between trees alive and dead at Time 2, using Fisher's exact test for count data.
Hierarchical models of possum browse and tree mortality
We used hierarchical models to generalize the relationships between browse and/or tree mortality and a range of covariates across tree species and sites. First, we modelled no browse, FBI = 0, vs. any browse, FBI > 0, on the possum-preferred ‘focal trees’ at Time 2 against the covariates: (1) Treatment/non-treatment, (2) PrefBio: sum of the biomass of preferred neighbourhood trees, excluding the focal tree (Hypothesis 2a), (3) NPrefBio: sum of the biomass of non-preferred neighbourhood trees surrounding the focal tree (Hypothesis 2b), (4) PrefEat: index of level of use of the forest patch by possums, calculated as the amount of preferred biomass consumed at each plot, that is, FBI × Biomass for each neighbourhood tree, summed across all neighbourhood trees in the plot (Hypothesis 2a), (5) DBH: index of size of focal tree (Hypothesis 2c) and (6) TCI: Trap-Catch Index at Time 2.
A complete set of covariate data were available from Time 2, and browse levels on the focal tree at Time 2 were assumed to depend on the current state, not the state 5 years ago. Each tree's biomass was calculated using an allometric relationship between DBH and biomass for New Zealand trees, where Biomass = 0·0406 × DBH1·53 (Richardson et al. 2009). Tree species additional to the indicator species were defined as being either preferred or non-preferred by possums based on published results (Owen & Norton 1995; Sweetapple 2003; Sweetapple, Fraser & Knightbridge 2004), possum browse from this study, and expert opinion (P. Sweetapple, pers. comm.).
For the seven main indicator species, both preferred and non-preferred, we investigated the relationship of annual mortality for the ‘focal trees’ against the covariates for their initial state and subsequent cumulative impacts: FCI at Time 1 (Hypothesis 3), DBH at Time 1 (Hypothesis 2c), treatment (Hypothesis 1) and average TCI from Times 1 and 2 (Hypothesis 1).
Browse and mortality were modelled using a Bayesian hierarchical approach, where for each covariate the coefficients for each tree species are distributed with a mean and variance from a higher-level ‘hyper-distribution’ (Royle & Dorazio 2008). This enabled us to model all species together, yet was flexible enough to allow the relationship between the covariates and FBI to differ between species. A range of models were specified separately for browse and mortality with various combinations of covariates mentioned above. The natural log of all continuous covariates (PrefBio, NPrefBio and PrefEat) was taken after adding a small value to accommodate zero values where applicable (1 for PrefBio and NPrefBio, 0·1 for PrefEat). To assist with convergence, covariates were standardized to have mean = 0 and SD = 1. Models were fitted using OpenBUGS 3·1 (Lunn et al. 2009), and compared using Deviance Information Criterion (DIC; Spiegelhalter et al. 2002).