Identifying and modelling the determinants of woody plant invasion of lowland heath



    Corresponding author
    1. Centre for Ecology and Hydrology, Dorset, Winfrith Technology Centre, Winfrith Newburgh, Dorchester, Dorset DT2 8ZD, UK, and
    2. School of Biological Sciences, University of Liverpool, Liverpool, Merseyside, L69 3BX, UK
    Search for more papers by this author
  • P. D. PUTWAIN,

    1. School of Biological Sciences, University of Liverpool, Liverpool, Merseyside, L69 3BX, UK
    Search for more papers by this author
  • N. R. WEBB

    1. Centre for Ecology and Hydrology, Dorset, Winfrith Technology Centre, Winfrith Newburgh, Dorchester, Dorset DT2 8ZD, UK, and
    Search for more papers by this author

P. Manning, NERC Centre for Population Biology, Imperial College London, Silwood Park Campus, Ascot, Berkshire SL5 0NE, UK (tel. +44 207 5 942482; fax +44 1344 873173; e-mail


  • 1The invasion of Betula spp. represents the moment of transition between European lowland heath and scrub ecosystems. We hypothesized that Betula invasion was controlled by a multivariate threshold, comprising factors that may be subdivided into seed and safe-site limitations, and that phosphorus availability was a key determinant of the Betula safe-site.
  • 2We performed a multifactorial field experiment in which seed rain, P-availability and disturbance were manipulated. All treatments had significant effects on Betula seedling densities, with seed availability proving the single greatest limitation. P addition emerged as having highly significant effects 4 and 12 months after germination. Disturbance had an initially large positive effect that dwindled across time.
  • 3Detailed descriptions of seedling density data were achieved by converting two factorial variables (seed rain and P availability) to a continuous form, including numerous covariates, and producing ‘minimum adequate’ generalized linear models (MAMs). These provided accurate descriptions of the data (explained deviance typically > 70%) and delimited site conditions where transition was likely.
  • 4Seed availability remained the most important factor, but the increased number of parameters in models describing densities of more mature seedlings suggested a temporal accumulation of recruitment limitations. Statistical modelling allowed for the subdivision of disturbance effects into those associated with a reduction in vegetation densities and the remaining ‘direct’ effects, e.g. soil disturbance, which shifted over time from being positive to negative. The results also support the hypothesis that soil phosphorus sorption capacity (PSC) determines heath–scrub transition as many of the identified determinants are controlled, either directly or indirectly, by PSC.
  • 5The study demonstrates that a combination of experimental and statistical modelling approaches can provide a detailed description of the factors controlling early stage invasion and may therefore have considerable utility in ecosystem management. Although site specific, the models illuminate the mechanisms by which many larger scale processes, e.g. burning and grazing, regulate the persistence of lowland heath ecosystems.


Transitions between alternate ecosystem states are of fundamental importance in both community and ecosystems ecology (Scheffer et al. 2001). Such transitions are often abrupt and may depend upon the invasion of a single species or functional group, which modifies the environment to initiate change to the alternate state (Srivistava & Jeffries 1996; Chapin et al. 1997; Petraitis & Latham 1999; Edmunds & Carpenter 2001; Wardle 2002). Invasion of any species is likely to be determined by multiple factors relating to dispersal, colonization and reproductive output. In plant ecology these factors are, at smaller spatio-temporal scales, usually reduced into factors controlling the abundance of propagules (dispersal or seed limitation) and those controlling colonization success (safe-site or recruitment limitation). Acquiring an understanding of the factors controlling the invasion of species that trigger transitions is key to ecosystem management as it allows for the development of strategies in which the species causing the transition may be encouraged or precluded. Ecosystems may therefore be held in the desired state. In this paper, we describe results of an investigation of the factors determining the invasion of species (Betula spp.) that induce ecosystem state change and present empirical models describing colonization success.

The last century saw a rapid decline in the area of European lowland heath ecosystems, with the major cause shifting in recent years from direct conversion to agriculture and forestry to transitions towards grassland in continental Europe and woodland and scrub in the UK (Aerts & Heil 1993; Rose et al. 1999). A number of woody plant species may invade lowland heath but none is more prevalent, or the cause of such widespread change, as Betula spp. (Mitchell et al. 1997). Colonization of heath by Betula (both B. pubescens (Ehrh.) and B. pendula (Roth)) represents the moment of transition between dwarf shrub and scrub ecosystems in which there are numerous, significant and resistant changes to both community composition and ecosystem properties (Atkinson 1992; Mitchell et al. 1998, 1999).

The customary view of heath persistence is that it depends upon either traditional management, e.g. grazing and burning (Webb 1998), or the presence of large free-ranging herbivores (e.g. Bokdam & Gleichman 2000). Although management appears largely responsible for between-heath differences in encroachment in the UK, there is considerable regional and patch scale variation in the likelihood of scrub invasion in unmanaged heath, a pattern that correlates with soil phosphorus sorption capacity (PSC) (Chapman et al. 1989a). Positive relationships between PSC and plant-available P have resulted in the hypothesis that Betula invasion is sensitive to small differences in P availability (Manning 2002). Such a mechanism would be consistent with recent observations on continental ombrotrophic bogs, which suggest P limitation of Betula invasion (Tomassen et al. 2003). In reality, however, Betula invasion is likely to be controlled by numerous variables in a site- and scale-dependent manner. Although the process of Betula recruitment has received considerable attention (Atkinson 1992), with numerous factors being proposed as important to seedling germination, growth and survival, the relative contribution of these factors has not been simultaneously assessed, precluding accurate description of the recruitment threshold.

We hypothesized that the likelihood of transition between lowland heath and Betula scrub states is determined by the degree of both seed and safe-site limitation and that a key axis in the determination of the Betula safe-site is P availability. These hypotheses were tested in a multifactorial field experiment in which phosphorus availability, seed rain and disturbance were manipulated. Two separate analyses were conducted, the first being an experimental test of treatment effects. In the second analysis treatment factors were converted into a continuous form and combined with covariate data to produce statistical models describing seedling densities. This approach aimed to incorporate a rigorous assessment of the manipulated factors and a detailed description of the ‘invasion threshold’, the overall aim being the production of a general, semi-mechanistic and multivariate model of invasion and therefore heathland ecosystem persistence.


study site

The experimental site was located on the Arne peninsula, Dorset (UK National Grid Reference SY 976 894). The site has been managed as a nature reserve since 1966, before which it was grazed by livestock and cleared of trees for timber use. Unlike many UK lowland heaths Arne has not experienced widespread Betula invasion, even when left unmanaged (Pickess et al. 1992).

The vegetation of the site is a relatively dry Erica tetralix-Sphagnum compactum (M16) community of the British National Vegetation Classification (Rodwell 1992); it is dominated by Erica tetralix (L.), Molinia caerulea ((L.) Moench) and Calluna vulgaris ((L.) Hull). The dwarf shrub species appear to have undergone the full cycle described by Watt (1947) as degenerate, pioneer and building phase vegetation are found together, resulting in a wide variety of vegetation densities and some small gaps ≥ 25 cm2. Four experimental blocks ran parallel to a linear strip of Betula at distances of 13–24 m. Aerial photographs show that these trees were absent in 1924 when the area now covered by Betula was an agricultural field. The site was not actively managed in the 8 years before the experiment's inception, but the Arne Nature Reserve (498 ha) supports a large population (c. 300) of Sika Deer (Cervus nippon), the total range size of which is unknown. Mean soil PSC values of the site were found to be 231 mg P kg−1 and 323 mg P kg−1 at the 0–50 mm and 50–150 mm depths, respectively. These values are low for UK lowland heath, where PSC values typically exceed 1000 mg P kg−1 (Chapman et al. 1989a). In June 1999, before the onset of experimentation, the experimental plots were cleared of all Betula. In this period the site also experienced a heather beetle (Lochmaea suturalis (Thoms.)) outbreak that seriously damaged all and killed > 50% of the existing C. vulgaris bushes on the site.

experimental design

The experimental manipulations: seed addition (three levels), phosphorus (P) addition (four levels) and disturbance (two levels), were arranged in an orthogonal factorial, randomized block design with each of the four blocks consisting of a row of 2 × 1 m plots with their short axis parallel to the direction of the row, so that all plots within a block were approximately equidistant to the nearest source of seed rain. As there were two replicates per block there were a total of 192 plots with eight replicates per treatment combination. There was a 50-cm guard row between plots and a 1-m guard row between blocks so the experiment covered a flat area of 78 × 11 m. The four P addition treatment levels were: control, one, two and three additions, and were comprised of individual additions of triple super phosphate at levels of 17.6 kg−1 P ha−1. These were applied manually in June 1999, July 1999 and January 2000. Each addition was approximately equivalent to that released in the burning of a 30-year-old heath of this area (Chapman 1967).

Seed was applied at three densities: control (background only), background plus 50 viable seeds m−2 and background plus 250 viable seeds m−2. These application rates coincide with typical seedbank densities of Betula spp. in heathland ecosystems (Thompson et al. 1997). Betula pubescens seed was collected from three heath sites across southern England in September 1999: Horsell Common, Surrey, National Grid reference TQ014598 (38 trees), Beaulieu River, New Forest, SU380060 (54 trees) and Crichton's Heath, Dorset, SY974896 (65 trees). By using the results of tetrazolium viability tests (Bullock 1996) and seed counts, a seed mix was prepared that contained equal proportions of viable seeds from all three sites and had a mean seed viability of 9%. A seed mix of multisite origin ensured that the emphasis of seed addition was on the effect of seed numbers per se as opposed to the genotypic suitability of the invader. Seed addition occurred in early November 1999. This coincided with the cessation of the natural seed rain period.

The aim of the disturbance treatment was to simulate many general components of various heathland disturbances, e.g. plant death and soil mixing, and was applied manually with a mattock. In June 1999 disturbed plots were struck and turned for 150 seconds m−2. In this time much of the vegetation was destroyed and the top 100 mm of soil was disrupted. The effect of the P addition and disturbance treatments on long-term plant available P, as estimated by ammonium oxalate extractable P (Pox), was studied in a sub-experiment located in specially allocated plots randomly inserted amongst those of the main experiment. In this experiment, which is described in detail elsewhere (Manning 2002), the P addition and disturbance treatments were applied factorially to smaller 50 × 50 cm plots at the same times and with the same levels. Soil samples of the 0–50 mm depth were taken in November 2000 and analysed for Pox according to the method of Pote et al. (1996) Disturbance had no significant effects on Pox. P addition had a significant effect on Pox (two-way anova, F3,56 = 4.95, P ≤ 0.01). Comparison of means found the difference to occur between the control (Pox = 103 mg P kg−1) and high treatment level plots (two additions Pox = 191 mg P kg−1 or three additions Pox = 189 mg P kg−1).

In addition to the three manipulated factors several covariates were measured. Soil cores were taken in December 1999 on a regular sampling grid of 72 positions, which maximized coverage of the site. Soil of the 0–50 mm depth was analysed for soil organic matter content (SOM) (% dry mass), moisture content (% total) and Pox (mg P kg−1). SOM and moisture content were also recorded for the 50–150 mm depth.

Background levels of Betula seed rain were monitored using seed traps positioned at regular intervals between the experimental plots. The traps used were cards coated in weatherproof glue, and positioned horizontally 50 cm above the ground. Trapping occurred between July and November 1999 and coincided with natural seed shedding. Trapped seed was tested for viability by visual assessment because the glue precluded the use of tetrazolium tests. Seed with a full oily content and spongy white tissue was classified as viable. The reliability of this method was confirmed by conducting both visual and tetrazolium tests on the same 40 seeds. Results matched perfectly (100% agreement) so natural and experimentally added seed rain values were compared with confidence. As the proportion of viable seeds in the seed rain was low, estimated background seed rain was calculated by multiplying the total seed rain of each trap by the mean percentage viability of each sampling period. There were no apparent spatial trends in seed viability.

Vegetation recording was conducted, in a random order, using a point quadrat method shortly after the disturbance treatment had been applied, July and August 1999. Each plot was recorded with a 4 × 9 grid of 60-mm fibre canes representing the central area of 160 × 60 cm. Presence/absence data were collected for all plant species, including mosses and lichens, at each 1-cm interval of each cane; standing dead material (necromass) was also recorded, but not to species level. The basal substrate penetrated by each cane was recorded as being either a moss or lichen species, organic (litter or humus) or mineral material.

The abundance and stem diameter at base (to the nearest 0.5 mm) of Betula seedlings, including cotyledon stage plants, were recorded in June, August and September 2000, and again in May 2001. Stem diameter was used to derive basal area. The species of Betula was not recorded because Betula spp. seedlings cannot be differentiated when young (Atkinson 1992). Conducting the final survey in May 2001 allowed the 2001 seedling cohort to be identified and subsequently removed from the analysis. An additional investigation into the effect of the measured variables on the likelihood of attack by mammalian herbivores was conducted by recording whether each seedling displayed mammalian herbivore damage. This was classified as damage to the plant stem; small losses of leaf tissue were attributed to invertebrate herbivores.


Evaluation of treatment effects was conducted using seedling count data of the four surveys in a fixed form analysis of deviance general linear model with the four factors, block, seed addition, P addition and disturbance, in an orthogonal design. Analysis was conducted in S + 6 for Windows (Anonymous 2001). A log-link and Poisson errors were specified, block was treated as a fixed factor and therefore block–treatment interactions were not calculated.

statistical modelling

A more detailed description of the determinants of Betula colonization was achieved by fitting general linear regression models to seedling data, where the descriptors were continuous measurements of the environment derived from the experimental site. All descriptors were hypothesized as important determinants of Betula recruitment. Individual plot measurements were not available for all variables. Therefore, kriging interpolation (Robertson 1987; Rossi et al. 1992) of data collected from the sampling grids covering the site was conducted, using the spatial statistics module of S + 6 for Windows. This approach reduced sampling effort and provided theoretically accurate covariate estimates for background levels of seed rain, soil moisture Pox and SOM. As anisotropy (directional dependence of spatial covariance structure) was not detected in any of the variables omnidirectional semivariograms were calculated for all distance lags up to 40 m, with the constraint of each lag estimate being based on ≥ 30 distance pairs. The best model for each empirical variogram was selected by choosing the fitted model with the lowest r2 (fitted by weighted non-linear least squares) from three variogram models: spherical, exponential and Gaussian (McBratney & Webster 1986). A grid of point estimates containing values for the central position of each square metre of the plots was then calculated by ordinary kriging with the selected model, which was, for all variables, the spherical model. Each plot's covariate value was obtained by taking the mean of the two central points. For two variables, Pox and SOM 50–150 mm, the distances in the empirical variograms were restricted to 12 m and 20 m, respectively, as preliminary analysis indicated genuine spatial structure within this range that was clouded by the re-occurrence of similar patches at larger scales.

Statistical models were fitted to seedling density data with the GLM procedure of S + 6 for Windows, and using the guidelines for model simplification described by Crawley (1993, 2002) to produce a minimum adequate model (MAM). This ‘best’ model has the smallest minimal residual deviance (analogous to r2) possible, with the constraint of all parameters being statistically significant. The MAMs were arrived at through the fitting of a maximal model (by maximum likelihood) containing all possible parameters (all descriptor variables and their interactions), but excluding higher order interaction terms, followed by the stepwise deletion of parameters that do not account for a significant proportion of the deviance. The process began with interaction terms. For each step of the simplification process each parameter was removed and then reinserted into the model after deviance change was noted. Once this had been done for all parameters the parameter accounting for the least deviance was removed, and the process was repeated until only significant terms remained. This approach avoids many of the ‘order-sensitive’ problems of automatic stepwise procedures (Crawley 2002). Significance was estimated using deletion tests in which a parameter is removed and the corresponding change in deviance is assessed with likelihood ratio tests. Models fitted to seedling density data had a log link and Poisson error, as appropriate for count data. The likelihood ratio deletion tests were performed using F statistics to account for over-dispersion (i.e. the maximal models residual deviance/residual d.f. was > 1). Once the MAM was fitted, the relative contribution of each explanatory variable was obtained through its deletion from the model, and measured as the change in explained deviance over the total explained deviance. We have expressed this value as deviance change on deletion (%DCD). DCD values for main effects include the effect of removing any interaction terms that contained the main effect.

Two factorial variables, P addition and seed addition, were converted into a continuous form to reduce parameterization and increase the generality of the fitted models. For a total seed rain estimate, background seed rain was added to the values of those that were added experimentally. A single P availability value was calculated by taking the mean Pox availability of the various P and disturbance treatment combinations, as estimated from the sub-experiment described above, and adding the difference from mean background Pox levels, as estimated from the kriged grid, for each individual plot. The third of the factorial treatments, disturbance, was subdivided into several variables including proportional substrate cover, above-ground biomass densities and necromass density. Because substrate and vegetation data were collected shortly after applying the disturbance treatment, disturbance effects could be considered independent of vegetation and substrate effects. Excessive correlation between plant biomass and disturbance was avoided as de-vegetated beetle attacked, disturbance resistant and naturally low density plots resulted in a variety of vegetation densities in both control and disturbed plots.

The requirement for species-specific vegetation terms was assessed using deletion tests to compare the fit of models containing either individual species variables or the sum total of all live ‘hits’ (biomass density). The resulting parameter values cannot be considered as representing competition alone because observed effects of neighbouring plant biomass on Betula seedling densities will represent a net balance between facilitation and inhibition (Callaway & Walker 1997), incorporating indirect effects such as soil environment.

A statistical model describing the proportion of seedlings attacked by mammalian herbivores in September 2000 was formulated using the same model simplification procedures. The model was fitted by weighted regression using a logit link function, binomial error distribution, the proportion of seedlings attacked as the response and individual sample sizes (total number of seedlings in the plot) as weights. The data were not over-dispersed and so χ2 statistics were used in the likelihood ratio tests of the model simplification process.


Analysis of deviance found that all three treatments, disturbance, P addition and seed addition, had highly significant effects on Betula colonization success (Table 1, Fig. 1a,b). Seedling density was greatest in disturbed, high P, high seed rain plots in all surveys. There was some degree of density-dependent compensation, but this was insufficient to have an exactly compensatory effect (Fig. 1a,b). Despite the overall trend of initial invasion success (1 month after germination) equating with longer-term success (after 1 year) there were considerable temporal changes in the effect strength of the treatments (Table 1, Fig. 1a,b). Disturbance effects peaked in August 2000 (F1,165 = 31.03, P = << 0.01, explained deviance (ED) = 7.5%) when seedlings were approximately 3 months old. As the experiment progressed the benevolence of disturbed conditions declined (September 2000, F1,165 = 27.30, P = << 0.01, ED = 6.5%; May 2001, F1,165 = 10.99, P ≤ 0.01, ED = 3.2%). The effects of P addition at the germination stage (June 2000 survey) were positive but not significant (F3,165 = 2.17, P ≥ 0.05, ED = 1.6%; Fig. 1a). Several months later (August 2000) the seedlings showed a far greater response (F3,165 = 11.64, P = << 0.01, ED = 8.5%). The magnitude of this effect declined later in the experiment (September 2000, F3,165 = 9.00, P = << 0.01, ED = 6.4%) but P remained an important determinant of seedling density for the entire year over which the surveys were conducted (May 2001, F3,165 = 8.89, P = << 0.01, ED = 7.7%; Fig. 1b). Seed addition was the largest treatment effect throughout the study period, suggesting considerable seed limitation. Seed addition effects were particularly pronounced at the germination stages (June 2000, F2,165 = 98.20, P = << 0.01, ED = 47.5%) but had declined substantially 1 year after germination (May 2001, F2,165 = 55.92, P = << 0.01, ED = 32.1%; Fig. 1b).

Table 1.  Analysis of deviance of experimental treatment effects on Betula seedling densities with a fixed form GLM (Poisson error/log link). Interaction terms were non-significant and have not been presented here. **P < 0.01, ***P << 0.01, NS = P ≥ 0.05, null d.f. = 191, residual d.f. = 165
Factord.f.June 2000August 2000September 2000May 2001
Block3 7.09*** 8.86*** 9.10*** 6.69***
P addition3 2.17NS11.64*** 9.00*** 8.89***
Seed addition298.20***70.82***82.84***55.92***
Figure 1.

The effect of seed addition, disturbance and P addition on Betula seedling densities for (a) 1 month (June 2000) and (b) 1-year-old seedlings (May 2001). s1 = control, s2 = 50 seeds m−2, s3 = 250 seeds m−2, c = control, d = disturbance. Error bars represent the standard error of the mean.

statistical models

Examination of variation in the variables used in the modelling process revealed several general patterns, including a tendency for kriged variables to display little variation (an integral property of the kriging process) and for manipulated variables to display the greatest variability (Table 2). Variation in some variables, e.g. soil water content and organic substrate cover, was low (Table 2), and is unlikely to represent the full variety of naturally occurring heathland conditions.

Table 2.  Descriptive statistics: mean, standard deviation (SD) and coefficient of variation (CV), for the continuous variables used in the seedling density statistical models
Mean vegetation height (cm) 11.1  2.60.24
Molinia caerulea density (hits m−2) 15.1 20.31.35
Erica tetralix density (hits m−2) 20.2 15.50.77
Calluna vulgaris density (hits m−2) 10.5 14.71.40
Necromass density (hits m−2) 78.8 33.50.43
Biomass density (hits m−2) 46.3 27.40.59
Total moss and lichen density (hits m−2)  4.7  7.91.66
P availability (Pox) (mg P kg−1)156.0 41.30.27
Water content 0–50 mm (% total soil mass) 68.2  5.10.07
Water content 50–150 mm (% total soil mass) 33.7  3.10.09
SOM 0–50 mm (% dry mass) 52.7 10.80.20
SOM 50–150 mm (% dry mass) 13.3  3.00.22
Mineral substrate (% cover)  8.1 12.11.49
Organic substrate (% cover) 86.3 12.40.14
Total seed rain (seeds m−2)125.1109.40.88

The model fitted to the June 2000 survey data (Table 3), which had Poisson error and a log link function, explained 70.9% of the deviance (ED) in seedling densities shortly after germination. The variable of greatest influence was total seed rain (DCD = 46.1%, P = << 0.01), suggesting that seed limitation was of greater importance than safe-site limitation at this stage. The densities of Molinia caerulea (DCD = 14.3%, P = << 0.01) and Erica tetralix (DCD = 6.3%, P = << 0.01) were found to have significant negative relationships, suggesting inhibition of germination and/or very early seedling establishment where these species were present. C. vulgaris densities did not significantly affect seedling numbers. Disturbance failed to elicit any significant effects other than through the reduction of E. tetralix and M. caerulea densities (data available on request). Most edaphic variables, e.g. SOM and substrate type, were also non-significant. Soil water content in the 0–50 mm depth (DCD = 0.98%, P ≤ 0.05) and P availability (DCD = 1.3%, P ≤ 0.05) did, however, demonstrate positive relationships with seedling numbers that accounted for small, unique portions of the deviance. The significant P availability effect appears contrary to the results of the factorial analysis but merely reflects the lower parameterization of the P variate in this model, d.f. = 1 as opposed to 3. Statistical modelling also detected a small interaction between seed and safe-site limitation, as represented by the interaction between total seed rain and E. tetralix density (DCD = 0.8%, P ≤ 0.05). This probably reflects the proportionally lower coincidence of seed and safe-site when both are limiting.

Table 3.  Minimum adequate model (MAM) (log link, Poisson error) describing Betula seedling densities (m−2) approximately 1 month after germination (June 2000). Null model deviance = 1565 on 191 d.f., MAM deviance = 456 on 185 d.f. Explained deviance = 70.9%. Deleted terms: disturbance, biomass density (hits m−2), all substrate measurements (% cover), total moss and lichen density (hits m−2), necromass density (hits m−2), mean vegetation height (cm), water content 50–150 mm (% total mass), soil organic matter (SOM) 0–50 mm (% dry mass), SOM 50–150 mm (% dry mass), Calluna vulgaris density (hits m−2). DCD = deviance change when deleted from MAM
 Parameter valueStandard error of parameter valueDCDP (F-test)
Water content 0–50 mm (% total mass)  0.02480.0063  0.98%  < 0.05
P availability (Pox) (mg P kg−1)  0.00290.00065  1.31%  < 0.05
Molinia caerulea density (hits m−2)−0.0320.004414.33%<< 0.01
Erica tetralix density (hits m−2)−0.0030.0026  6.25%<< 0.01
Total seed rain (seeds m−2)  0.00790.000446.1%<< 0.01
E. tetralix× Total seed rain−0.000070.00002  0.84%  < 0.05

The MAM fitted to the September 2000 survey data (log link, Poisson error, 73.9% ED) retained most of the variables contained in the June 2000 model, but also included additional variables, reflecting determinants of seedling survival (Table 4). Temporal variation of the determinants of seedling densities was revealed by both the variables retained and the changes in their parameter values and the proportion of ED unique to them. Although there was a general increase in the number and importance of variables associated with safe-site limitation, total seed rain remained the factor with the single greatest contribution to ED (DCD = 40.0%, P = << 0.01).

Table 4.  Minimum adequate model (MAM) (log link, Poisson error) describing Betula seedling densities (m−2) approximately 4 months after germination (September 2000). Null model deviance = 1212 on 191 d.f., MAM deviance = 316 on 181 d.f. Explained deviance = 73.9%. Deleted terms: biomass density (hits m−2), all substrate measurements (% cover), total moss and lichen density (hits m−2), necromass density (hits m−2), water content 50–150 mm (% total mass), water content 0–50 mm (% total mass), SOM 50–150 mm (% dry mass). DCD = deviance change when deleted from MAM
 Parameter valueStandard error of parameter valueDCDP (F-test)
Calluna vulgaris density (hits m−2)−0.01750.0031 2.93%<< 0.01
Molinia caerulea density (hits m−2)−0.02120.005613.52%<< 0.01
Erica tetralix density (hits m−2)−0.05240.004611.78%<< 0.01
P availability (Pox) (mg P kg−1)  0.00730.0009 5.81%<< 0.01
Disturbance (binomial variable)  0.01210.112 2.38%<< 0.01
Total seed rain (seeds m−2)  0.00730.000540.00%<< 0.01
Mean vegetation height (cm)  0.12340.021 2.74%<< 0.01
SOM (% dry mass) (0–50 mm)  0.01080.0039 0.63%  < 0.05
Disturbance × Total seed rain  0.00280.0008 0.92%  < 0.05
M. caerulea× Total seed rain−0.00010.00002 0.69%  < 0.05

The net effect of the surrounding vegetation remained strongly negative, with the densities of all three dominant plant species, C. vulgaris (DCD = 2.9%, P = << 0.01), M. caerulea (DCD = 13.5%, P = << 0.01) and E. tetralix (DCD = 11.78%, P = << 0.01), having highly significant effects on Betula seedling density. Mean vegetation height had a significant positive relationship (DCD = 2.7%, P = << 0.01) that was independent of vegetation density effects, thus revealing that plots with shorter vegetation were less invasable than those with taller vegetation of the same density. Disturbance accounted for a small portion of the variance in seedling density (DCD = 2.4%, P = << 0.01) and this was independent of its effect on vegetation density. P availability had increased in importance since June 2000; high P plots contained higher seedling densities (DCD = 5.8%, P = << 0.01). Seedling densities also responded positively to the SOM content of the topsoil (0–50 mm) (DCD = 0.6%, P ≤ 0.05). Two interactions were detected, between safe-site and seed limitation: disturbance × total seed rain (DCD = 0.9%, P ≤ 0.05) and M. caerulea density × total seed rain (DCD = 0.7%, P ≤ 0.05), again a greater proportional effect of adding seed where safe-sites are non-limiting.

The MAM fitted to the May 2001 survey data (approximately 1 year after germination) had a log link, Poisson error and 14 significant parameters. It accounted for 69.2% of the total deviance (Table 5). Although broadly similar to the September 2000 model, partially compensatory mortality in higher density plots reduced the perceived impact (shallower slope, and lower intercept within the models) of many variables. Despite this, total seed rain remained the most important descriptor (DCD = 31.7%, P = << 0.01). Interactions with disturbance (DCD = 0.9%, P ≤ 0.05) and M. caerulea density (DCD = 0.8%, P ≤ 0.05, Fig. 2a) were also retained. The disturbance × total seed rain interaction had, however, switched to being negative (i.e. survivorship was proportionately lower in disturbed plots), a pattern that may be explained by the seedling density and size dependent actions of herbivores (see below). Main effects of disturbance also shifted to being net negative when isolated from vegetation reduction effects (DCD = 4.5%, P = << 0.01), a finding that is consistent with the temporal decrease in the significance of the disturbance treatment in analysis of deviance.

Table 5.  Minimum adequate model (MAM) (log link, Poisson error) describing Betula seedling densities (m−2) approximately 1 year after germination (May 2001). Null model deviance = 824 on 191 d.f., MAM deviance = 254 on 177 d.f. Explained deviance = 69.2%. Deleted terms: biomass density (hits m−2), all substrate measurements (% cover), total moss and lichen density (hits m−2), necromass density (hits m−2), water content 50–150 mm (% total mass), water content 0–50 mm (% total mass). DCD = deviance change when deleted from MAM
 Parameter valueStandard error of parameter valueDCDP (F-test)
Calluna vulgaris density (hits m−2)−0.01850.004 5.33%<< 0.01
Molinia caerulea density (hits m−2)−0.01610.00611.29%<< 0.01
Erica tetralix density (hits m−2)−0.18260.03612.64%<< 0.01
P availability (Pox) (mg P kg−1)  0.00450.0018 6.82%<< 0.01
Disturbance (binomial variable)−0.19810.132 4.53%<< 0.01
Total seed rain (seeds m−2)  0.00800.000631.74%<< 0.01
Mean vegetation height (cm)  0.07050.035 3.62%<< 0.01
SOM (% dry mass) (0–50 mm)  0.01910.0054 1.53%   < 0.01
SOM (% dry mass) (50–150 mm)−0.4760.0191 0.76%   < 0.05
E. tetralix× P availability  0.00040.0001 0.86%   < 0.05
C. vulgaris×M. caerulea−0.00210.0008 0.98%   < 0.05
Disturbance × Total seed rain−0.00130.0005 0.86%   < 0.05
E. tetralix× Mean vegetation height  0.00520.0016 1.09%   < 0.05
M. caerulea× Total seed rain  0.00070.00003 0.76%   < 0.05
Figure 2.

Fitted response surface of Betula seedling density in May 2001 (1-year-old seedlings) as affected by (a) Molinia caerulea and total seed rain density, (b) mean vegetation height and Erica tetralix density, and (c) Erica tetralix density and P availabitity (Pox), displaying interactions between these variables. In each case all other variables in the fitted model (Table 5) were held constant at their mean.

Vegetation densities retained strong effects upon seedling densities and were found to interact (DCD = 1.0%, P ≤ 0.05) in the case of M. caerulea density (main effect DCD = 11.3%, P = << 0.01) and C. vulgaris (main effect DCD = 5.3%, P = << 0.01), revealing that seedling densities were particularly high where the abundance of both species was low. E. tetralix density had the greatest negative effect on Betula densities of all the identified factors (DCD = 12.6%, P = << 0.01). Mean vegetation height also had a negative effect (DCD = 3.6%, P = << 0.01), and interacted with E. tetralix density (DCD = 1.1%, P ≤ 0.05) to produce extremely low predicted seedling densities where E. tetralix was dense and short (see Fig. 2b). Temporal changes in model structure appear to suggest an increase in the importance of edaphic factors in determining seedling densities. There were positive effects of SOM and P availability in the 0–50 mm depth (SOM, DCD = 1.5%, P ≤ 0.01; P availability, DCD = 6.8%, P = << 0.01) and a small negative effect of SOM at the 50–150 mm depth ((DCD = 0.8, P ≤ 0.05). An E. tetralix density × P availability interaction (DCD = 0.9%, P ≤ 0.05) seems to reflect a reduction of the negative effects of dense E. tetralix where P availability was high (Fig. 2c).


Observation suggests that a major, but unquantified, factor determining seedling numbers was predation by mammalian herbivores. By May 2001 40% of seedlings displayed mammalian herbivore damage and many seedlings may have been destroyed altogether; 64.6% of the deviance in the proportion of seedlings attacked (in September 2000) was described by a statistical model with binomial error and a logit link (Table 6). Herbivory was greatest where seedlings were large, aggregated and open to attack. All significant variables, with the exception of C. vulgaris and E. tetralix densities, correlated with open conditions and so deviance overlap was considerable. Seedling basal area was the key determinant (DCD = 33.3%, P = << 0.01; Fig. 3). Additional effects included positive density-dependent predation (seedling density, DCD = 1.8%, P ≤ 0.05) (Fig. 3) and a shelter or avoidance effect of the surrounding species that appears to be independent of growth suppression effects. This was stronger for E. tetralix density (DCD = 2.6%, P ≤ 0.01) than for C. vulgaris density (DCD = 1.5%, P ≤ 0.05). A greater proportion of seedlings were attacked where the vegetation was taller (mean vegetation height, DCD = 1.1%, P ≤ 0.05), possibly due to greater openness of the vegetation (apparency) or attraction of herbivores to these areas. Disturbance, M. caerulea density and P availability had no significant direct effects.

Table 6.  Minimum adequate model (MAM) (binomial error, logit link) describing the determinants of mammalian herbivore attack, as represented by the proportion of 4-month-old Betula seedlings displaying herbivore damage (September 2000). Null model deviance = 362 on 151 d.f., MAM deviance = 128 on 146 d.f. Explained deviance = 64.6%. Deleted terms: Molinia caerulea density (hits m−2), P availability (Pox) (mg P kg−1), necromass density (hits m−2), and disturbance. DCD = deviance change when deleted from MAM
 Parameter valueStandard error of parameter valueDCDP2-test)
Mean seedling basal area (mm2)  1.6340.1633.3%<< 0.01
Mean vegetation height (cm)  0.0690.034 1.1%   < 0.05
Seedling density (m−2)  0.01850.007 1.8%   < 0.05
Erica tetralix density (hits m−2)−0.02620.0087 2.6%   < 0.01
Calluna vulgaris density (hits m−2)−0.01220.0053 1.5%   < 0.05
Figure 3.

Fitted response surface for the proportion of Betula seedlings in September 2000 displaying evidence of mammalian herbivore attack as described by seedling basal area and seedling density. All other variables in the fitted model (Table 6) were held constant at their mean.


The experimental results agree with most plant population studies (Eriksson & Ehrlén 1992; Turnbull et al. 2000; Mazia et al. 2001) in finding both seed and safe-site limitation to affect the likelihood of invasion. They also demonstrate that phosphorus availability can play a major role in the determination of the Betula safe-site. Statistical models provided descriptions of the seedling data that lacked experimental rigour but were general, close fitting, less parameterized and more detailed. Modelling also highlighted the increasing specificity of conditions for establishment as the seedlings passed through a series of ‘environmental filters’. Though it may be argued that this could be an artefact of increasing variance in the response variable there are clear biological explanations for the observed patterns, e.g. most soil factors remain unimportant before the exhaustion of seed reserves.

The multidimensionality of the models highlights the requirement to provide multivariate descriptions of the ‘conditions’ (Scheffer et al. 2001) at which ecological state shifts occur. The fitted models can be thought of as representing descriptions of the invasion threshold for Betula, which is conceptually similar to the basin of attraction around the F2 conditions of state shift, described by Scheffer et al. (2001). Numerical combinations of predictor variables resulting in seedling densities of zero represent stable, uninvasable heath conditions. Conditions resulting in predicted seedling densities > 0 represent the conditions within which transition may potentially occur.

Despite the close fit of the models, they are dependent upon the assumption that seedling densities are indicative of the likelihood of transition to scrub. Seedling presence does not always equate with adult survivors (Turnbull et al. 2000) and so further demographic monitoring will be required to validate the model. The other major drawback in using our statistical models as descriptors of the threshold is their site specificity. Although attempts were made to include a variety of heathland conditions, the variance and effect of each identified determinant, and unmeasured site conditions, may differ between sites and regions.

determinants of invasion

Seed availability was the single most limiting identified factor, even a full year after germination. Seed limitation may emerge as a result of low adult density and reproductive output, dispersal failure, low reproductive output (Clark et al. 1998), germination failure and seed predation. The main cause at this site was most likely to be wind direction; the prevailing wind blew seeds in the opposite direction to the experimental plots. Short distance dispersal limitation and strongly directional colonization in Betula has been previously observed by Arradóttir et al. (1997), who only found safe-site saturation at distances < 4 m from parent trees. Such results, when coupled with those presented here, question the dogma of hugely effective dispersal in Betula spp., and suggest that considerable seed limitation is present over many heath patches, which are often much further from seed sources than our experimental plots.

A number of factors, e.g. vegetation density and P availability, were identified that could be classified as axes of the Betula safe-site or regeneration niche (sensu Grubb 1977). These explained a large proportion of the deviance in Betula seedling densities, especially for more mature seedlings. Most studies looking at safe-site limitation have defined suitable microsites as areas where disturbance has occurred (Eriksson & Ehrlén 1992). In this study, disturbance increased seedling densities, but statistical modelling revealed that, in the long term, positive effects were the product of competition reduction and that other disturbance effects were negative. These negative effects were probably associated with both the greater size and apparency of seedlings attracting herbivores, and abiotic stresses, e.g. frost heaving. Viewing safe-sites as merely any disturbed conditions is therefore an oversimplification as the regeneration niche of any plant species is likely to be complex and multidimensional.

The largest safe-site effects were the net negative effects of the site's three principal competitor species. Although parameter values for each species depend upon the biomass represented by each ‘hit’, the models suggest that Betula is most negatively affected by E. tetralix. This response may represent both direct competition and a low nutrient supply rate. Van Vuuren et al. (1992) found that N mineralization rates were 4.4 g N m−2 y−1 in E. tetralix soil compared with 7.8 g N m−2 y−1 in Molinia soil, and in dry soil Calluna communities, 6.2 g N m−2 y−1. Total P and P mineralization may also be low in the soil of E. tetralix (Berendse et al. 1989; Van Vuuren et al. 1992). Temporal increase in the relative importance of E. tetralix may reflect seasonality in the biomass allocation of M. caerulea. Seedlings surviving the winter in dense M. caerulea are likely to have benefited from the shelter of its litter, whilst avoiding the incessant competition suffered by seedlings growing with E. tetralix. Although the net effect of M. caerulea was less than that of E. tetralix, its abundance was able to explain considerable portions of the variance in Betula seedling densities, particularly soon after germination. M. caerulea effects are likely to contain influences of both competition and litter effects; germination was extremely poor where M. caerulea litter was present. A similar inhibition of seed germination by litter has been observed in a number of studies (e.g. Mazia et al. 2001; Wilby & Brown 2001). Failure to detect this effect with necromass density is probably a result of no species specificity in the measure, i.e. only M. caerulea litter inhibited germination. Strong competition effects of dense M. caerulea are supported by the results of Bokdam & Gleichman's (2000) longer-term study, which found that Betula seedlings established, but did not survive, in ungrazed M. caerulea tussocks. The third dominant, C. vulgaris, had a lesser effect on Betula seedling densities, possibly as a result of heather beetle damage making C. vulgaris less competitive. Taller vegetation was also more invasible. This may result from the lower likelihood of short vegetation containing light gaps than tall vegetation of the same density or higher nutrient uptake per unit biomass in the short, compact, building phase. These results, when synthesized, suggest that the invasability of heath communities by Betula is dependent on not only vegetation biomass and growth phase, as found in upland heaths by Gong & Gimingham (1984), but also its species composition.

The detection of a small but significant positive relationship with SOM is likely to represent, in the early stages, a positive association with soil water content, and for later seedlings, N availability. Although N availability was not directly measured, Berendse (1990) found that SOM content and M. caerulea abundance explained 84% of the variance in N mineralization rate in Dutch heath soils. As both these variables were measured, N effects are likely to be included within these terms. Recent research, conducted in ombrotrophic bog microcosms, concluded that growth of B. pubescens was P limited above mineralization rates of 0.25 g N m−2 y−1 (Tomassen et al. 2003). The role of N in determining Betula invasion may therefore be minor in wet heath ecosystems (see N-mineralization rates above).

A key result of this experiment is the finding that Betula seedling densities are, in natural field conditions, responsive to differences in available P, a variable that can be partly explained by PSC (Manning 2002). This adds support to the hypothesis of Chapman et al. (1989a), that the likelihood of transition in lowland heath ecosystems depends upon soil PSC. Benefit from P fertilization is likely to result from low inorganic P availability in heathlands and the ecophysiology of both resident and invading species. Although both dwarf shrub and Betula species utilize inorganic and organic P sources their proportional use of these pools and response to increased supply rate differ considerably. Ericoid species primarily utilize organic nutrient sources via mycorrhizas (Read 1983), and may not exhibit the morphological plasticity that is typical of Betula spp. (Harrison & Helliwell 1979). In fertilized soils, and possibly on high PSC, high SOM heaths, inorganic P levels may exceed those over which ericoid species may respond. This hypothesis is supported by the finding of Helsper et al. (1983), who found that P addition of 57 kg P ha−1 (equivalent to the maximum total in this study) had little effect on C. vulgaris-dominated communities. A similar relationship occurs in Dutch heaths in which M. caerulea is more responsive to elevated concentrations of inorganic nitrogen (Aerts & Berendse 1988; Aerts 1989; Aerts & Chapin 2000) than its dwarf shrub competitors, resulting in species replacement and transition to grassland.

This apparently simple model is likely to be confounded by the presence and activity of ectomycorrhizal fungi on Betula roots (Atkinson 1992). Recent research has found that Betula seedlings may, via mycorrhiza, acquire significant portions of their P from organic sources (Perez-Moreno & Read 2001a,b). Although this indicates that ectomycorrhizal Betula seedlings bypass inorganic P limitation, Jonsson et al. (2001) and Newton & Pigott (1991) found an inverse relationship between soil fertility and ectomycorrhizal colonization. Other studies have found a poor correlation between ectomycorrhizal colonization and both P uptake (Harrison & Helliwell 1979) and seedling growth (Newton 1991). A mechanistically consistent synthesis of these results would be that Betula seedlings access organic P sources via their mycorrhizal partners where inorganic P availability is low, and that a reduction of this dependency in high P conditions frees carbon resources that may be utilized in overcoming alternative stresses, thus boosting survival.


Herbivore effects on plant recruitment could not be quantified directly but the fitted model suggested that herbivores had distinct preferences, which explain the density dependence in seedling survival. Plots with high seed rain, low vegetation density and high P availability produced large seedlings at high densities and were subsequently selected by herbivores. The other, contrasting, effects of E. tetralix and C. vulgaris are likely to be related to the browsing behaviour of the deer at the site. E. tetralix, for instance, which displayed a negative relationship with herbivory, is thought to be unpalatable (Bannister 1966) and may provide physical protection. The results demonstrate some correspondence with those of Rao et al. (2003), who found that browsing by mountain hares on B. pubescens was more likely where plant stem diameter was large. However, they also found converse trends, e.g. plants grown in tall vegetation were less susceptible to attack.

The conditions favouring herbivore attack in this study suggest that the Betula regeneration niche is narrow when herbivores are present. This niche appears, at the Arne site, to be qualitatively best represented by fertilized plots severely attacked by heather beetles. In these conditions seedlings seem to avoid the pressures of herbivory and competition but benefit from protection by dead standing woody material, and nutrients released by L. suturalis (Heil & Diemont 1983; Brunsting & Heil 1985). The determinants of Betula invasion, in the absence of herbivores, may be adequately described by an early stage seedling density model (e.g. September 2000, Table 5). In such situations competition effects are likely to remain important until the plant overtops the dwarf shrub vegetation. Where herbivores are present at high densities, temporary release may be necessary if invasion is to occur. Because seedlings surviving the early stages of colonization, e.g. 1 year, may be able to persist in a ‘bonsai’ state within dwarf shrub vegetation for as long as 31 years (Kinnaird 1974), and larger seedlings have a capacity to resprout, invasion may occur in different conditions to those in which initial colonization occurred.

Caution is required in the extrapolation of these results to larger spatial and temporal scales as the determinants that vary, and thus explain variation, at one scale may be uniform at another. Despite this drawback, the fitted models are capable of illuminating landscape scale trends as many of the major, large-scale processes of heathlands, e.g. grazing regimes, burning and beetle outbreaks, as well as soil and climate trends, have significant impacts on the identified determinants over large areas. Soil PSC, for instance, affects P availability directly but may also drive rates of vegetation and soil development (Chapman et al. 1989b; Manning 2002). Lowland heath in high PSC regions is likely to reach the low competition, high SOM, high P state more rapidly than in low PSC regions and will therefore require more intensive or regular management if it is to be held in a non-invasable condition. Burning of degenerate stage heath in high PSC regions is, for example, likely to result in hot fires that destroy the dwarf shrub rootstock and greatly increase nutrient availability (Manning 2002).

Landscape-scale seed limitation is likely to be largely dependent on the degree of safe-site and seed limitation in the previous generation. Seed and safe-site limitation may therefore interact significantly at the landscape scale, resulting in outcomes varying between an ‘epidemic’ of transition and stasis. Although iterative simulation modelling (e.g. Clark et al. 2001) would be required to see how these processes operate, it is likely that transition may be continuous or step-like depending on the system's susceptibility to change, e.g. PSC and disturbance frequency.

Despite drawbacks the methods used here appear capable of delimiting conditions in which invasion may occur and may therefore be of general utility in ecosystem management. What is now required is a confirmation of the models’ generality, a task that will require multisite and scale studies, covering a wider range of heath conditions. An understanding of natural variation in determinant factors at larger scales is also needed before the identity and role of the determinants are quantitatively confirmed.


We would like to thank the Royal Society for the Protection of Birds and English China Clays International for site access. The staff of CEH Dorset provided considerable support. Anonymous referees provided useful comments on the manuscript. PM was funded by a NERC/CASE CEH studentship.