How much of the world is woody?


  • Richard G. FitzJohn,

    1. Biodiversity Research Centre and Department of Zoology, University of British Columbia, Vancouver, BC, Canada
    2. Department of Biological Sciences, Macquarie University, Sydney, NSW, Australia
    Search for more papers by this author
    • These authors contributed equally.

  • Matthew W. Pennell,

    Corresponding author
    1. Department of Biological Sciences and Institute for Bioinformatics and Evolutionary Studies, University of Idaho, Moscow, ID, USA
    2. National Evolutionary Synthesis Center, Durham, NC, 27705, USA
    Search for more papers by this author
    • These authors contributed equally.

  • Amy E. Zanne,

    1. Department of Biological Sciences, George Washington University, Washington, DC, 20052, USA
    2. Center for Conservation and Sustainable, Development Missouri Botanical Garden, Louis, MO, USA
    Search for more papers by this author
  • Peter F. Stevens,

    1. Department of Biology, University of Missouri, St. Louis, MO, USA
    2. Missouri Botanical Garden, St Louis, MO, USA
    Search for more papers by this author
  • David C. Tank,

    1. Department of Biological Sciences and Institute for Bioinformatics and Evolutionary Studies, University of Idaho, Moscow, ID, USA
    Search for more papers by this author
  • William K. Cornwell

    1. Department of Systems Ecology, VU University, Amsterdam, The Netherlands
    2. Evolution & Ecology Research Centre, School of Biological Earth and Environmental Sciences, University of New South Wales, Sydney, NSW, Australia
    Search for more papers by this author


  1. The question posed by the title of this study is a basic one, and it is surprising that the answer is not known. Recently, assembled trait data sets provide an opportunity to address this, but scaling these data sets to the global scale is challenging because of sampling bias. Although we currently know the growth form of tens of thousands of species, these data are not a random sample of global diversity; some clades are exhaustively characterized, while others we know little to nothing about.
  2. Starting with a data base of woodiness for 39 313 species of vascular plants (12% of taxonomically resolved species, 59% of which were woody), we estimated the status of the remaining taxonomically resolved species by randomization. To compare the results of our method to conventional wisdom, we informally surveyed a broad community of biologists. No consensus answer to the question existed, with estimates ranging from 1% to 90% (mean: 31.7%).
  3. After accounting for sampling bias, we estimated the proportion of woodiness among the world's vascular plants to be between 45% and 48%. This was much lower than a simple mean of our data set and much higher than the conventional wisdom.
  4. Synthesis. Alongside an understanding of global taxonomic diversity (i.e. number of species globally), building a functional understanding of global diversity is an important emerging research direction. This approach represents a novel way to account for sampling bias in functional trait data sets and to answer basic questions about functional diversity at a global scale.


The distinction between a woody and non-woody growth form is probably the most profound contrast among terrestrial plants and ecosystems: for instance, a forest is dominated by woody taxa, while a grassland is dominated by herbs. The recognition of the fundamental importance of this divide dates back at least to Enquiry into Plants by Theophrastus of Eresus (371–287 BC), a student of Plato and Aristotle, who began his investigation into plant form and function by classifying the hundreds of plants in his garden into woody and herbaceous categories (Theophrastus 1916).

The last two thousand years of research into wood since Theophrastus classified his garden have uncovered its origin in the early Devonian (∼400 Mya; Gerrienne et al. 2011); that prevalence of woodiness varies with climate (Moles et al. 2009); that wood has been lost many times in diverse groups, both extant and extinct (Judd, Sanders & Donoghue 1994), often as an adaptation to freezing temperatures (Zanne et al. 2014); that it has also been gained many times, particularly on island systems (Carlquist 1974; Givnish 1998); and that many different forms of pseudo-woody growth habit have appeared across groups that have lost true woodiness or diverged before true woodiness evolved (Cornwell et al. 2009). We know about its mechanical properties and developmental pathways, its patterns of decomposition and their effects on ecosystem function (Cornwell et al. 2009) and that woody and herbaceous species have markedly different rates of molecular evolution (Smith & Donoghue 2008). However, we have no idea about what proportion of species in the world are actually woody.

Recently, assembled functional trait data sets provide an opportunity to address this question. However, such data sets are, almost without exception, biased samples of global diversity. Researchers collect data for specific questions on a local scale, and assembling these local data sets creates a useful resource (Kattge et al. 2011). But as with GenBank's assembly of genetic data (Smith et al. 2011), the simple compilation of data is not an unbiased sample, and these initial sampling biases will, in turn, bias downstream analyses. Understanding and accounting for the biases in these data sets are an important and necessary next step.

We sought to develop an approach that accounts for this bias. In doing so, we were able to re-ask Theophrastus' 2000-year-old question at a global scale: how many of the world's plant species are woody? We also sought to understand how well scientists were able to overcome this bias and make a reasonable estimate. To do this, we took the unconventional approach of coupling our analysis with an informal survey in which we asked our question to the broader community of botanists and other biologists.

Materials and methods

Data set

We used a recently assembled data base with growth form data for 49 061 vascular plant species (i.e. lycopods, ferns, gymnosperms and angiosperms), which is the largest such data base assembled to date (Zanne, et al. 2013, 2014, available on the Dryad data repository; doi: 10.5061/dryad.63q27/2). This data base uses a functional definition of woodiness: woody species have a prominent above-ground stem that persists through time and changing environmental conditions and herbaceous species lack such a stem – a definition originally suggested by Gray (1887). Zanne et al. (2014) chose this simple definition because it best characterized the functional aspect of growth form that they investigated, allowing them to compare species that maintain an above-ground stem through freezing conditions to ephemeral species that avoid freezing conditions. More precise definitions that rely on lignin content and/or secondary vascular tissue from a bifacial cambium are problematic because there are many exceptions depending on tissue type, times of development or environmental conditions (Groover 2005; Spicer & Groover 2010; Rowe & Paul-Victor 2012). Because our analyses and survey were based on this data base, we present this functional definition of woodiness here for clarity (see Zanne et al. (2014) for a discussion of the various definitions of woodiness, their merits and pitfalls). Note that in addition to species producing secondary xylem, this definition classifies, among other groups, palms, tree ferns and bamboo as woody.

As with all large data assemblies, the underlying data sets were collected for a variety of research goals. For example, a number of the data sets come from forestry inventories, which, of course, are biased towards recording woody species. Other sources of sampling bias, including geographically restricted sampling in many sub-data sets, may be less obvious but nonetheless may have major implications for the inferences drawn from aggregate data bases.

Because the effort to organize plant taxonomy, especially synonymy, is ongoing, there was uncertainty regarding the status of many plant names. To bring species binomials to a common taxonomy among data sets, names were matched against accepted names in The Plant List (2014). Any binomials not found in this list were matched against the International Plant Name Index ( and Tropicos ( Potential synonymy in binomials arising from the three lists was investigated using the Plant List tools (The Plant List 2014). As a result of this cleaning, the number of species in the final data set was reduced from 49 061 to 39 313.

Theophrastus recognized both the fundamental importance of the distinction between woody and herbaceous plants and that this distinction is in some cases difficult to make. There are two ways that species were recorded as ‘variable’ in form (Beaulieu, O'Meara & Donoghue 2013). First, different records of a single species may conflict in growth form (having both records of woodiness and herbaceousness); this affected 307 of the 39 313 species in the data base. Secondly, 546 species (1.4%) were coded as variable. Following Beaulieu, O'Meara & Donoghue (2013), we coded species in these groups as ‘woody’ or ‘herbaceous’ when a majority of records were either ‘woody’ or ‘herbaceous’, respectively, and for these species, records of ‘variable’ do not contribute to the analysis. Our final data base for the main analysis contained 38 810 records with both information on woodiness and documented taxonomy – 15 957 herbs and 22 853 woody species. This included records from all flowering plant orders currently accepted by APG III (The Angiosperm Phylogeny Group 2009) and the fern taxonomy of Stevens (2001), covering 15 232 genera and 465 families. The 503 species excluded at this step had identical numbers of records of being woody and herbaceous. We also ran analyses where we coded growth forms by treating species with any record of woody or variable as ‘woody’ (and similarly for herbaceous), using all 39 313 species. Neither of these cases is likely to be biologically realistic but allowed us to evaluate the maximal possible effect of mis-coding variable species.

Estimating the percentage of species that are woody

To estimate the percentage of species that are woody, we cannot simply use the fraction of species within our trait data base that are woody (22 853 of 38 810 = 59%) as these records represent a biased sample of vascular plants. For example, most Orchidaceae are probably herbaceous; we have only one record of woodiness among the 1537 species for which we have data. However, the fraction of Orchidaceae species with known data (1537 of 27 801 = 6%) is much lower than the overall rate of knowledge for all vascular plants (38 810 of 316 143 = 12%), which will upwardly bias the global estimate of woodiness. Conversely, systematic under-sampling of tropical species would bias the global woodiness estimate downwards, as tropical floras are thought to harbour a greater proportion of woody species than temperate ones (Moles et al. 2009).

We developed a simple method to account for this sampling bias when estimating the percentage of woody species. In our approach, we treat each genus separately, and in all cases, we know that there are nw woody and nh herbaceous species and a total of N species in the genus. For example, the genus Microcoelia (Orchidaceae) has 30 species in total, and we know that 12 are herbaceous and none are known to be woody (N = 30, nw = 0, nh = 12). We do not know the state of the remaining 18 species, so the true number of woody species, Nw, must lie between 0 and 18. In general, we cannot assume that these species are all herbaceous, even though both biological and mathematical intuition suggest that most of them will be.

We used two different approaches for imputing the values of these unknown species. First, we assumed that the known species were sampled without replacement from a pool of species with Nw woody and Nh herbaceous species (Nw + Nh = N), following a hypergeometric distribution. The probability that x of the species of unknown state are woody (x = 0,1,…,N − nw − nh) is proportional to

display math(1)

Under this sampling model, the more species for which we do not have data, the greater the uncertainty in our estimates for the proportion of species which are woody. For Microcoelia this model gives a 42% probability that all species are herbaceous, and a 90% chance that at most 3 species are woody. This approach probably overestimates the number of woody species in this case, and in other cases, where all known species are woody [e.g. Actinidia (Ericaceae)], it will probably underestimate the number of species that are woody. We see this as corresponding to a weak prior on the shape of the distribution of the fraction of woody species within a genus and will refer to this as the ‘weak prior’ approach because it weakly constrains the state of missing species.

However, the distribution of woodiness among genera and families is strongly bimodal; most genera are either all woody or all herbaceous (Fig. 1, Fig. S1 in Supporting Information, and Sinnott & Bailey 1915). Among the 791 genera with at least 10 records, 411 are entirely woody, 271 are entirely herbaceous, and only 58 have between 10% and 90% woody species. Qualitatively similar patterns hold at both the level of family and order, although the distribution becomes progressively less bimodal as one moves up the taxonomic hierarchy (Figs S1 and S2). As a result, knowing the state of a handful of species within a genus can give a reasonable guess at the state of remaining species.

Figure 1.

Distribution of the percentage of woodiness among genera. The distribution of the percentage of species that are woody within a genus is strongly bimodal among genera (panel a – showing genera with at least 10 species only). The two different sampling approaches generate distributions that differ in their bimodality (panel b). If we sample species with replacement from some pool, with a weak prior on the fraction of woodiness within the pool, then we generate a broad distribution with many polymorphic genera (blue line). Sampling with replacement, assuming that species are drawn from a pool of species that has a fraction of woody species equal to the observed fraction of woodiness, generates a strongly bimodal distribution (red line).

To model the other extreme of sampling, we used an approach where we computed the observed fraction of woody species

display math

and sampled the state of the unobserved species using a binomial distribution, which represents the case of sampling with replacement. In this case, the probability that x of the species is woody is:

display math(2)

In cases where all known species are woody (or herbaceous as in Microcoelia), this will assign all unknown species to be woody (or herbaceous). For such genera, increasing the number of unobserved species will not increase the uncertainty in the estimate, in contrast to the weak prior sampling approach. We therefore see the binomial sampling approach as corresponding to a very strong prior on the bimodal distribution of woodiness among genera, and we will refer to this as the ‘strong prior’ approach because it more strongly constrains the state of missing species within genera with no known polymorphism. While neither of these approaches is ‘correct’, they probably span the extremes of possible outcomes. In polymorphic genera, the two approaches will give similar results, especially where the number of unknown species is relatively large.

For genera where there was no information on woodiness for any species, we sampled a fraction of species that might be woody from the empirical distribution of woodiness fractions among genera within the same order. We did this after imputing the missing species values within those other genera. So, if a genus is found in an order with genera that had woodiness fractions of {0,0,0.1,1}, we would have approximately a 50% chance of sampling a 0% woodiness fraction for a genus, with probabilities from 0.1 to 1 being fairly evenly spread. Given this woodiness fraction, we then sampled the number of species that are woody from a binomial distribution with this fraction and the number of species in the genus as its parameters.

In addition to the number of species known to be woody and herbaceous, we also require an estimate of the number of species per genus. For this, we used the number of accepted names within each genus in the Plant List (The Plant List 2014). The taxonomic resources were compiled by Zanne et al. (2014) are on available on Dryad (Zanne et al. 2013).

For each genus, we sampled the states of unobserved species, from either the hypergeometric or binomial distribution, parameterized from the observed data for that genus. For each sample, we can then combine these estimates to compute the number (or fraction) of species that are woody at higher taxonomic levels (family, order or vascular plants). We repeated this sampling 1000 times to generate distributions of the number (or fraction) of species that are woody. The code and data to replicate this analysis are available on GitHub at All analyses were conducted in R v3.0.2 (R Core Development Team 2013).


In estimating the number of species within Angiosperm families, Joppa, Roberts & Pimm (2010) found that expert opinion generally agreed closely with estimates from a statistical model. We were interested in whether a consensus answer existed – even if not formalized in the literature – and if so, whether it was consistent with our estimates. We created an English-language survey (which we also translated into Portuguese) asking for an estimate of the percentage of species that are woody according to the above definition. We also asked respondents to indicate their level of familiarity with plants, level of formal training and the country in which they received their training. We sent out the survey to several internet mailing lists and social media websites (see Appendix for details on the survey).


Across all vascular plants, we estimated the fraction of woody species to be between 45% and 48%. Specifically, using our strong prior sampling approach (binomial distribution), we estimated 45.6% of species are woody (95% confidence interval of 45.3–45.9%), and with the weak prior (hypergeometric distribution) approach, we estimated 47.6% (95% CI of 46.9–48.2%) (Fig. S3). The different approaches generated different distributions of the per-genus percentage of woodiness (Fig. 1), with a less strongly bimodal distribution using the weak prior approach. (See Figs S1 and S2 for the distributions at the level of families and orders, respectively) However, the two different approaches (strong vs. weak priors) led to similar phylogenetic distributions of estimated woodiness (Fig. 2 vs. Fig. S4), differing only in the details. We have compiled a table of the estimated number of woody species under both sampling approaches for all genera, families and orders included in our analysis. This is included in the Supplementary Material and is available on the Dryad data repository (FitzJohn et al. 2014).

Figure 2.

Distribution of the percentage of woodiness among orders of vascular plants. Each tip represents an order, with the width of the sector proportional to the square root of the number of recognized species in that order (data from accepted names in The Plant List 2014). The bars around the perimeter indicate the percentage of woody (black) and herbaceous (white) species, estimated using the ‘strong prior’ (binomial) approach. Using the ‘weak prior’ (hypergeometric) approach generally leads to an estimated percentage that is closer to 50% (see Fig. 1 and Fig. S4). Phylogeny from Zanne et al. (2014) (available on Dryad; 10.5061/dryad.63q27/3). Orders not placed by APG III (The Angiosperm Phylogeny Group 2009) are not displayed. We note that there is some discrepancy between the Zanne et al. tree and previous well-supported phylogenetic hypotheses (e.g. Soltis et al. 2011), most notably, in the position of the Magnoliids; however, the higher level relationships do not influence any of the analyses.

As stated above, neither of these sampling approaches is ‘correct’. However, as the observed distribution of woodiness fraction among genera is itself strongly bimodal, we believe that the true result lies closer to 45% than to 47%. A more sophisticated hierarchical modelling approach could lead to a more precise answer, but we feel that our values probably span the range of estimates that such an approach would generate. And in any case, we felt that addressing a simple question warranted a simple approach.

Different codings of variable species (see above) significantly moved our estimates, despite affecting a small minority of species. Coding all variable species as woody, our estimates increased by 1.6–47.1% with the strong prior approach and by 1–48.6% with the weak prior approach (Fig. S5). Similarly, with coding all variable species as herbaceous, the fraction of woody species decreased by 1.9–43.7% under a strong prior and by 1.3–46.3% under a weak prior (Fig. S5).

There was strikingly little consensus among researchers as to the percentage of species that are woody. We received 292 responses from 29 countries, with estimates that ranged from 1% to 90% with a mean of 31.7% (Fig. S6). The lowest estimate from our analyses (45% woody) is greater than 81% of our survey estimates. We found little effect of respondents' level of training on their estimate (Fig. 3). There was a significant effect of the respondent's familiarity with plants on the estimates, primarily driven by respondents with little botanical familiarity (the ‘What's a Plant?’ category in the survey), whose estimates tended to be lower (less woody) than the estimates of those with more familiarity. However, excluding respondents with little familiarity with plants had virtually no effect on the mean estimate of respondents (32.4% excluding this category as compared to 31.7% with them included). Restricting survey responses to only respondents at least ‘familiar’ with plants, and with at least an undergraduate degree in botany or a related field (143 responses), only increased the mean survey estimate to 32.9%.

Figure 3.

Distribution of responses to the survey question ‘What percentage of the world's vascular plant species are woody?’. Responses are divided by familiarity with plants (panel a) and formal training in botany or a related discipline (panel b). The mean and 95% confidence intervals for our estimates of the proportion of woody species from the empirical data are depicted by the horizontal shaded rectangles; the blue upper rectangle corresponds to the ‘weak prior’ approach, and the red lower rectangle corresponds to the ‘strong prior’ approach (see Appendix for details).

Before carrying out the survey, we had hypothesized that researchers from tropical regions may perceive the world as woodier than researchers from more temperate regions due to the latitudinal gradient in woodiness (Moles et al. 2009). Indeed, there was an effect of being in a tropical country, with the estimates from tropical countries being slightly higher than those from temperate countries (P = 0.02), but this effect was very small (r2 = 0.02, Fig. S6).


Our estimates of woodiness differed from both the survey and the simple mean of the global data base: neither simple statistics nor biologists' intuition was accurate in this case. The difference from community knowledge is in striking contrast to Joppa, Roberts & Pimm (2010), who found that expert opinion on the number of species within different Angiosperm groups agreed closely with results based on analyses of data and their bias.

The respondents to our survey perceived there to be substantially fewer woody species in the world than there probably are. This herb-centric view of the world may arise from the importance of our (mostly herbaceous) cultivated crops, or the fact that people – including most researchers – likely spend more time in the garden than in the forest, and especially not in tropical forests where diversity is high and disproportionately woody.

Our estimates of the percentage of species that are woody (45/48%) differ from the raw estimate based on species in our data base (59%). This difference is caused by the interaction between biased sampling and clustered trait data at a variety of taxonomic scales. The distribution of woodiness is bimodal among genera, and the distribution of sizes of those genera differs with woodiness. Genera that are primarily herbaceous (less than 10% woody species for genera with at least 10 records) were on average larger than primarily woody genera (more than 90% woody species), with a mean of 214 species compared to 151 (See Fig. S7). This means that even a random sampling above the level of species will lead to a biased estimate.

The effect of sampling bias within our data base on the estimate is amplified by the distribution of woodiness at higher taxonomic levels, with families or even orders often being predominantly either woody or herbaceous (Fig. 2 and Sinnott & Bailey 1915). There are two major clades that are primarily herbaceous – the monocots (Monocotyledons) and ferns (Monilophyta). However, there are many primarily herbaceous clades nested within woody clades, and vice versa, which makes the combination of taxonomic and functional information crucial for answering this type of question.

We also found that the way in which we handled variable species significantly altered the estimates. That changing the state of such a relatively small number of species has the potential to alter inferences made at a global scale is rather surprising. Two points regarding this are worth noting here. First, we reiterate that our alternate coding schemes (all variable species coded as herbaceous and all variable species coding as woody) are rather extreme and unlikely to be biologically realistic. Secondly, while these alternate coding schemes certainly affected the estimates, the magnitude of their effect is much less than that of the overall sampling bias in the original data base.

Higher order classifications are at least as much a product of human pattern matching as biological processes. Genera correspond to the morphological discontinuities among species that humans deem important (Scotland & Sanderson 2004), which likely includes woodiness (e.g. Hutchinson 1973). The relative rarity of genera with significant numbers of both woody and herbaceous species (Fig. 1) reinforces the importance of this trait. A significant, but unaccounted for, source of error is the likely non-random woodiness of undiscovered species. We would predict that there are likely more herbs to be discovered than woody plants; larger genera tend to be more herbaceous (Fig. S7), and we think it is more likely that new species are yet to be described in these large groups. In principle, rarefaction analysis could estimate the number of species remaining to be discovered in different groups, but this is not possible for many plant clades (Costello, Wilson & Houlding 2011); for many clades, the ‘collecting curve’ shows little sign of saturation, which is required for such an analysis.

Sampling biases are pervasive in ecological data sets and need to be addressed when using them for analyses. Global data bases of functional traits (e.g. TRY; Kattge et al. 2011) are central to biodiversity research, but through no fault of the data base collator, they are inevitably biased in terms of taxonomic breadth, and this may have serious consequences for the reliability of inferences drawn from them. For example, for woodiness, the economic importance of forestry species likely leads to their over-sampling in this data set. This sampling bias also affects many commonly used methods in ecological and evolutionary research (e.g. Ackerly 2000; Nakagawa & Freckleton 2008; Pennell & Harmon 2013; Pakeman 2014) in addition to its well-understood effects on conventional statistics. In our case, taking the data at face-value, we would have greatly overestimated the global percentage of woody species. Inferring the global frequency of any trait would face the same problem. For example, the ecologically important traits of nitrogen-fixing, mycorrhizal symbioses and pollinator syndrome are strongly taxonomically structured, and we would expect raw estimates to be biased in the same way that woodiness was. Our approach was developed for binary traits but similar approaches could be developed for multi-state categorical or continuous traits.

In addition to improving an estimate of the mean, the methods in this paper can also be used to generate a probability of each unobserved species being woody. Thus, it can be used as a type of taxonomically informed data imputation. Recently, two related approaches have been developed to do just this, both focusing on continuous traits (Guénard, Legendre & Peres-Neto 2013; Swenson 2014). While their details differ, both approaches are model based in that they impute trait values for missing species based on the fitted parameters of phylogenetic models estimated from the species already in the data base. This is conceptually different from our approach; we do not assume any model for the evolution of woodiness, such as the ‘Mk’ model (Pagel 1994), which is commonly used to model discrete characters evolving on a phylogeny. Both types of approaches – using taxonomic categories (this study) vs. modelling trait evolution along a phylogeny – have advantages and disadvantages. One disadvantage of a modelling-based approach is that if the sampling is biased with respect to the character states, the parameter estimates themselves will be biased, leading to an incorrect estimation of the states for the remaining species. While our approach avoids this issue, we ignore potentially useful information on the phylogenetic relationships within genera and branch lengths separating lineages.

Concluding remarks

As a result of centuries of effort, we now have an increasingly complete understanding of taxonomic diversity. More recent developments in assembling global trait data bases offer the promise of gaining similar insights into the functional diversity of the earth's biota. While the question we ask in this paper – what proportion of the world's flora is woody? – is simple, answering it required dealing with the pervasive biases that will be present in most large data sets. Researchers should be aware that because of these biases and the phylogenetically structured distribution of traits, the law of large numbers will not apply and that estimates from trait data bases will not converge on the true value. Our approach is just one of many potential ways to address these biases; we hope that our analysis encourages others to think critically and creatively about the problem. Just as Theophrastus' garden was a non-random sample of the Greek flora, our trait data bases contain diverse biases; accounting for them will be important in making inferences about broad–scale ecological and evolutionary patterns and processes.


We thank the members of the Tempo and Mode of Plant Trait Evolution working group for contributing to project development, members of the broader community who took the time to fill out and comment on our survey and Rafael Maia for translating our survey and helping us to distribute it. In particular, we thank Jon Eastman for developing the taxonomic resources we used for this study. We thank Dales Indian Cuisine in Durham, NC, for providing the buffet lunch over which this project was brought to life. This work was supported by the National Evolutionary Synthesis Center (NESCent), NSF #EF- 0905606, Macquarie University Genes to Geoscience Research Centre through the working group. RGF was supported by a Vanier Commonwealth Graduate Scholarship from the Natural Sciences and Engineering Research Council of Canada (NSERC). MWP was supported by a NESCent graduate fellowship and a University of Idaho Bioinformatics and Computational Biology graduate fellowship. WKC was supported by Netherlands Organisation for Scientific Research (NWO) through its Open Competition Program of the section Earth and Life Sciences (ALW) Grant No. 820.01.016.

Data accessibility

Previously published resources

Data produced in this study

  • Results from analyses: included as a supplemental file and available on Dryad (FitzJohn et al. 2014).
  • Survey results: included as a supplemental file and available on Dryad (FitzJohn et al. 2014).
  • R scripts: available on the project GitHub repository (