Edge disturbance shapes liana diversity and abundance but not liana‐tree interaction network patterns in moist semi‐deciduous forests, Ghana

Abstract Edge disturbance can drive liana community changes and alter liana‐tree interaction networks, with ramifications for forest functioning. Understanding edge effects on liana community structure and liana‐tree interactions is therefore essential for forest management and conservation. We evaluated the response patterns of liana community structure and liana‐tree interaction structure to forest edge in two moist semi‐deciduous forests in Ghana (Asenanyo and Suhuma Forest Reserves: AFR and SFR, respectively). Liana community structure and liana‐tree interactions were assessed in 24 50 × 50 m randomly located plots in three forest sites (edge, interior and deep‐interior) established at 0–50 m, 200 m and 400 m from edge. Edge effects positively and negatively influenced liana diversity in forest edges of AFR and SFR, respectively. There was a positive influence of edge disturbance on liana abundance in both forests. We observed anti‐nested structure in all the liana‐tree networks in AFR, while no nestedness was observed in the networks in SFR. The networks in both forests were less connected, and thus more modular and specialised than their null models. Many liana and tree species were specialised, with specialisation tending to be symmetrical. The plant species played different roles in relation to modularity. Most of the species acted as peripherals (specialists), with only a few species having structural importance to the networks. The latter species group consisted of connectors (generalists) and hubs (highly connected generalists). Some of the species showed consistency in their roles across the sites, while the roles of other species changed. Generally, liana species co‐occurred randomly on tree species in all the forest sites, except edge site in AFR where lianas showed positive co‐occurrence. Our findings deepen our understanding of the response of liana communities and liana‐tree interactions to forest edge disturbance, which are useful for managing forest edge.


| INTRODUC TI ON
Lianas are woody climbing plants that are rooted in the soil, but use trees or shrubs to climb to the canopy. Lianas are common in tropical forests, but they are generally more abundant and diverse in disturbed areas such as forest edges and canopy gaps (Zhu & Cao, 2010). Nonetheless, some authors found that lianas were less abundant in some disturbed forests (Addo-Fordjour et al., 2012).
Irrespective of the trend, liana community structure often changes with disturbance in forest ecosystems (Bongers et al., 2020). Human disturbance of forests results in fragmentation (Harper et al., 2005), causing changes in forest structure and microclimatic conditions at edges (Magnago et al., 2015). Such edge-induced changes tend to be favourable to disturbance-adapted, light-demanding species such as lianas (see Hawthorne, 1996;Laurance et al., 2001), but generally disadvantageous to trees (Laurance et al., 2006). The disturbanceadapted nature of lianas makes them an ideal group of plants which can be used to test edge effects. Previous studies reported that edge effects enhanced liana diversity and abundance in some forests (Addo-Fordjour et al., 2021;Campbell et al., 2018;Laurance et al., 2001;Ofosu-Bamfo et al., 2019), but others did not detect changes in liana diversity in response to edge (Mohandass et al., 2014;Ofosu-Bamfo et al., 2019). Several properties of forest edge such as edge size, edge type, and surrounding matrix type can mediate edge effects on plant community structure (Martino, 2015), and be responsible for the varied responses of community structure to edges in different forests. For example, liana and tree communities may show different responses to forest edge (Addo-Fordjour & Owusu-Boadi, 2016). Such edge-induced changes can alter liana-tree interactions, as reported for plant-animal networks (Fagan et al., 1999;Porensky, 2011). Nonetheless, there is scarcity of information on the response of liana-tree interaction network patterns to forest edge. Studying edge effects on liana community structure and liana-tree interactions can reveal interesting and unique findings that can contribute towards the development of edge theory.
Species interact to form complex networks of biological communities (Hagen et al., 2012). The species interactions that occur within networks tend to shape ecological communities and drive evolution (Fontaine et al., 2011;Jacobsen et al., 2018). Ecological network approach has been used to study species interactions in more detail, revealing much more information on community structure (Watts et al., 2016). The use of ecological networks thus improves our knowledge on community ecology and the evolutionary processes shaping biological communities (Losapio et al., 2019).
Thus, understanding of network patterns would make it possible to predict the ecological and evolutionary consequences of networks.
For instance, networks exhibiting modular structure are expected to show a higher stability and robustness, as such a structure would limit diffusion of perturbations through the networks (Thébault & Fontaine, 2010). Médoc et al. (2017) reported that nestedness increases the stability of networks. Moreover, Thébault and Fontaine (2010) revealed that nestedness increases the stability of mutualistic networks, but destabilises antagonistic networks. With regards to evolution, nestedness increases the variation of individual fitness, resulting in a core of species that drive the evolution of the whole community (Bascompte et al., 2003;Cantor et al., 2017;Gómez et al., 2011). Similarly, modularity may also enhance evolution by allowing certain modules to evolve independently of other organisms (see Hansen, 2003).
In spite of the usefulness of ecological network approach as outlined above, it is scarcely used in liana studies, resulting in limited knowledge on liana-tree interaction networks, and lack of consensus regarding the interaction patterns. Previous studies used different network metrics to characterise liana-tree interactions. For example, nestedness, a network pattern in which the interactions of less connected species form proper subsets of the interactions of more connected species (Bascompte et al., 2003;Landi et al., 2018;Ponisio et al., 2019), has been used to characterise the structure of liana-tree networks. Different patterns of nestedness are reported in literature for liana-tree networks including nested (Sfair et al., 2010) and non-nested (Addo-Fordjour & Afram, 2021;Addo-Fordjour et al., 2016, 2021Blick & Burns, 2009;Magrach et al., 2016;Ofosu-Bamfo et al., 2019) structures. Among the studies that did not find nested structure in liana-tree networks, some reported anti-nested structure which depicts non-random assembly (Addo-Fordjour & Afram, 2021;Addo-Fordjour et al., 2021;Blick & Burns, 2009;Magrach et al., 2016), while others observed non-significant nestedness that shows random assembly (Addo-Fordjour et al., 2016;Ofosu-Bamfo et al., 2019). Ecological networks can also be compartmentalised into modules whose members interact more among themselves (Carstensen et al., 2016). This phenomenon referred to as modularity, is predicted to stabilise ecological networks (Massol et al., 2017;Thébault & Fontaine, 2010). Species within modular networks perform distinct topological roles, with implications for forest management and conservation (Olesen et al., 2007). Sfair et al. (2010) did not find modular structure in their networks, but Addo-Fordjour and Afram (2021) recorded significant modular structure in liana-tree networks.
Specialisation at the network and species levels can cause non-nested and modular organisation of species (Addo-Fordjour & Afram, 2021;Castledine et al., 2020;Médoc et al., 2017). Thus, in liana-tree networks in which coevolution leads to specialisation (Sfair et al., 2015), the networks may tend to be non-nested and/ or modular. Another important metric used to characterise network structure is species co-occurrence, which describes the frequency of pairs of liana species to co-occur on the same phorophyte species

T A X O N O M Y C L A S S I F I C A T I O N
Community ecology (Zulqarnain et al., 2016). Species co-occurrence patterns are useful in inferring the ecological and evolutionary history of liana species, as closely related species tend to have similar niches that increase their chances of co-occurrence (Zulqarnain et al., 2016). Like the above-mentioned network metrics, mixed patterns of liana species co-occurrence have been reported in literature, which include positive co-occurrence (Addo-Fordjour et al., 2016;Zulqarnain et al., 2016), negative co-occurrence (Blick & Burns, 2009, 2011, and random co-occurrence (Addo-Fordjour et al., 2016). With the mixed findings on the structure of liana-tree interactions in literature, there is the need for more studies to be conducted to determine the most consistent patterns. Knowledge of co-occurrence patterns is important for increasing our understanding of species interactions and predicting community stability and maintenance, and ecosystem functioning, all of which are useful in forest conservation (Vizentin-Bugoni et al., 2016).
This study determined the response patterns of liana community assemblages and structure of liana-tree interaction networks to edge in two moist semi-deciduous forests in Ghana. The forest edges we studied were surrounded by large matrices of crop farmlands, thus making the edges much exposed. The nature and size of land matrix bordering forest edges play a key role in determining the intensity of edge effects on plant community structure (Aragón et al., 2015). To this end, edges bordered by wide land matrices are expected to exert stronger effects on plant communities than edges surrounded by narrow area of land (Addo-Fordjour & Owusu-Boadi, 2016). In reality, the existence of a marked contrast in the physiognomy and structure between a forest edge and its surrounding land matrix causes variation in the microclimatic conditions of that forest edge and the interior site (Aragόn et al., 2015). Based on the above, we expected edge effects on liana assemblages and lianatree interaction patterns in the two moist semi-deciduous forests.
Edge disturbance permits greater penetration of sunlight into forest edges, and also increases forest edge dryness (Thier & Wesenberg, 2016), both of which can favour liana proliferation. On the basis of the above, we tested the following hypotheses: 1. Liana diversity and abundance would be higher in edge site than non-edge sites.
2. We expected that as edge disturbance enhances liana abundance at the forest edge, network connectance will increase, resulting in less specialised, nested and non-modular network structures in edge site, while the networks in the non-edge sites will be less connected, more specialised, non-nested and modular.
3. Edge effects will cause shifts in topological roles of liana and tree species due to changes in the distribution and abundance of the species.
4. As sunlight and dry conditions are elevated at edge sites relative to the non-edge sites, competition of lianas for the resources in edge site may be lower. Moreover, as edge effects tend to cause tree mortality at forest edges (Murcia, 1995), the number of available host species may reduce, increasing liana infestation per host. Thus, we expected that liana species in edge sites would show positive co-occurrence on host trees, while the species in non-edge sites will randomly co-occur on their hosts.
The findings of our study would be useful in the management of forest edges and conservation of edge species. Our study seeks to add valuable information to literature, thus helping to obtain general patterns of liana assemblages and structure of liana-tree interactions in relation to edge effects. These findings can contribute to the development of a theory on edge effects in view of the fact that there is dearth of information on the role of edge disturbance in shaping the patterns of liana-tree network structure in forests.

| Study areas
We conducted the study in two moist semi-deciduous tropical for- SFR, respectively, suggest that climatic differences between edge and interior sites of AFR could be more pronounced than those in SFR. Based on the above differences in the structure of the two forests, we expected the patterns of edge effects on liana communities in the forests to differ.

| Asenanyo forest reserve
Asenanyo forest reserve is a production forest that was established in the year 1940 and covers an area of 22,800 ha in the Ashanti Region of Ghana (Wiafe, 2014). It is of the moist semi-deciduous forest ecosystem, with the dominant tree species being Celtis mildbraedii, Triplochiton scleroxylon and Entandrophragma spp. (Forest Services Division, 2010a;Wiafe, 2014). The forest has a bimodal rainy season from April to October (maximum rainfall: May-June; minimum rainfall: September-October) and a dry season from November to March. Annual rainfall range is 1250-500 mm (Hall & Swaine, 1981). Temperature in the reserve ranges from an average of 30.5°C to 21°C, with a mean annual relative humidity of about 84%. AFR has about 20 admitted farms scattered throughout the reserve, the size of each averaging approximately 5 ha (Forest Services Division, 2010a). The reserve also has one admitted community occupying an area of about 955.70 ha (Forest Services Division, 2010a).

| Suhuma forest reserve
SFR is also a production forest of about 36,030 ha located in the Sefwi Wiawso Forest District (Hawthorne & Abu-Juan, 1995).
There are 24 admitted farms in the reserve each averaging 11.5 ha (total 276 ha) and one admitted community covering an area of 389 ha (Forest Services Division, 2010b). The reserve is exposed to active logging. Its canopy is discontinuous due to excessive logging activity but still has emergent trees that may reach heights of about 40 m. The forest lies within the moist semi-deciduous forest zone in Ghana, and thus its vegetation is dominated by tree species such as C. mildbraeddii, Baphia nitida, Nesogordonia papaverifera, Microdesmis puberula, Khaya ivoriensis, Daniella ogea and Dacryodes klaineana (Hall & Swaine, 1981). The forest reserve experiences two distinct seasons: the dry season and the rainy season. The rainy season is from April to October, whereas December to March marks the dry season. Average annual rainfall is between 1300 and 1600 mm. Mean annual temperature ranges between 26 and 29°C, and relative humidity is usually above 90% in the rainy season and falls to 60% during the dry season (Forest Services Division, 2010b).

| Sampling design and data collection
A total of eight 50 × 50 m plots were randomly established in each of three forest sites, namely, edge, interior and deep-interior.
Each forest site had two randomly demarcated and independent sampling areas, each of which contained four plots. The edge site was defined as 0-50 m from the forest edge, while interior and deep-interior sites were 200 m and 400 m from the forest edge, respectively. Variable penetration distances of edge have been reported in previous studies. These studies revealed that edges can extend up to 100 m from the forest edge, while other studies also detected edge effects up to 300 m (Flaspohler et al., 2001;Gascon et al., 2000;Laurance et al., 2018;Liu & Taylor, 2002). Thus, we set our two interior sites 100 m beyond each of the aforementioned edge penetration distance limit, resulting in 200 m and 400 m distances from the forest edge.
We surveyed and identified all lianas with diameter (at 1.30 m from the rooting base) ≥1 cm as well as trees (diameter at breast height ≥10 cm) that carried lianas in the plots. The minimum interplot distance in the sampling areas was 150 m. Plant species were identified by a plant taxonomist, and through the use of herbarium specimens and identification guides (Hawthorne, 1990;Hawthorne & Jongkind, 2006).

| Community structure
We used species richness, Shannon diversity index and species evenness to characterise liana diversity in the forest sites.
A rarefaction-extrapolation technique was used to standardise species richness based on a constant number of individuals using iNEXT package in R. We computed Shannon diversity index and species evenness with PAST statistical package version 2.17c (Hammer et al., 2001) and tested the significance of the differences in the indices among the forest sites using permutation tests in the PAST software. Computation of Shannon diversity index (H′) and species evenness index (E) was based on the following equations: where, pi = proportion of the ith species, and Inpi = natural log of pi, S = species richness Community abundance of lianas was compared among the forest sites by running nested ANOVA, where sampling area was nested within forest site. We employed aov function in the stats package in R to perform the nested ANOVA.
Using the equation of Harper et al. (2005Harper et al. ( , 2015, we calculated

| Network structure of liana-tree interactions
Liana-tree network structure was quantified using the following network metrics: (1) connectance and specialisation asymmetry, (2) degree of specialisation (H2', d'), (3) nestedness, (4) modularity, (5) module connectivity and interactions (c and z values), (6) species cooccurrence. We used quantitative liana-tree species matrices except in the species co-occurrence test where binary matrices were employed. Each of matrices was made up of liana species assigned to rows and tree species assigned to columns. We also represented the various networks in graphs using plotweb function in the bipartite package in R.
1. Network connectance and specialisation asymmetry Weighted connectance was calculated to express network connectance in the study. It is defined as the linkage density divided by number of species in the network (van Altena et al., 2016;Dormann, 2021). The values of weighted connectance range from 0 (no connectance) to 1 (perfectly connected). Weighted connectance was run with the networklevel function in the bipartite package.
Similarly, the networklevel function was used to calculate specialisation asymmetry of the networks.

Degree of specialisation
The degree of specialisation was determined for the various networks and the individual species in the networks as follows: Using the H2' index, we quantified network specialisation of the various forest sites. The index measures the extent to which observed interactions deviate from the interactions that would be expected given the marginal totals of the interactions per species (Blüthgen et al., 2006). Generally, higher values of the H2' index indicate that the species in the network are more selective, resulting in higher specialisation of the network. The index ranges from 0 (no specialisation) to 1 (complete specialisation). The H2' index was run with H2fun function in the bipartite package.
The degree of species specialisation was determined by calculating d' index, using dfun function in the bipartite package. This index is defined as the deviation from a conformity expected by the overall utilisation of potential partners (Blüthgen et al., 2007).

Nestedness
Nestedness occurs when the more specialist species interact only with subsets of the species interacting with the more generalist species (Bascompte et al., 2003;Ponisio et al., 2019). This means that generalists interact with one another, and specialists tend to interact with generalists, but specialist-specialist interactions are often absent (Bascompte et al., 2003). We calculated weighted nestedness metric, WNODF with the network-level function in bipartite package in R (Dormann, 2021), in accordance with the nestedness equation of Almeida-Neto and Ulrich (2010). The WNODF metric ranges from 0 (fully non-nested) to 100 (fully nested). There are two forms of non-nested pattern described in literature: (1) when nestedness value is consistent with the null model expectation, and (2) when nestedness value is significantly less than that of the null model. The aforementioned patterns of nestedness refer to two different community assemblies (random and non-random assembly, respectively) and therefore must be distinguished. We therefore used anti-nestedness to refer to the situation where observed nestedness values were significantly lower than those expected by chance, whereas we referred to networks that presented observed nestedness values which were consistent with null model expectation as not nested.

Modularity
We measured modularity index (Q) with the DIRTLPAwb+ algorithm using computeModules function within the bipartite package (Beckett, 2016). Modularity measures the tendency of a network to form modules of interacting species, which interact more with one another than with species of other modules (Carstensen et al., 2016;Dormann, 2021). The Q index ranges from 0 for networks with clustering not different from random to 1 for networks with perfect modules. The Q index calculation followed the equations in Newman (2006).

Test of statistical significance of the metrics
The above mentioned network metrics (i.e. connectance, degree of specialisation, nestedness, modularity) were tested for their statistical significance by generating 1,000 null models and comparing them with the observed metric values using the Patefield algorithm (Patefield, 1981) in the bipartite package.

Module connectivity and interactions
The topological roles of liana and tree species with respect to network modularity were assessed based on the number of links of the species. We achieved this by calculating the weighted standardised among-module connectivity (c) and within-module interactions (z), using species strength of interaction (Watts et al., 2016).
To obtain the corresponding appropriate c and z thresholds for the species, we generated 100 null models of the original networks using DIRTLPAwb + algorithm, and 95% quantiles as thresholds of with their topological roles. The relationships were expressed in scatter plots using ggscatter function in ggpubr package of R. The correlation analysis was run on log-transformed data.

Species co-occurrence
Liana species co-occurrence patterns were determined with the cooc_null_model function from EcoSimR package (Gotelli et al., 2015). We used the C-score metric, which is the average number of checkerboards for two species (Stone & Roberts, 1990), to measure species co-occurrence. The metric was calculated according to the equation described by Almeida-Neto and Ulrich (2010). To assess the patterns of co-occurrence, 10,000 null models were generated by the quasiswap algorithm and compared with the observed cscore values. The c-score measures the tendency of species to not co-occur (Stone & Roberts, 1990). Thus, the greater the c-score in relation to the null model, the greater the tendency of the species to not co-occur (i.e. segregation) and the smaller the c-score value in relation to the null model, the higher the tendency of species to co-occur (i.e. aggregation).

| Liana community structure
There were more liana species in edge site (40 species) than interior site (35 species), which in turn had more species than deep-interior site (30 species) in AFR (Table 1) The contribution of the five most abundant liana species to the total liana abundance in the forest sites of AFR were as follows: edge -54%, interior -55.1% and deep-interior -59.8% (Table 1; Appendix S1). In the case of SFR, the five most abundant liana species contributed 53.9, 53.1 and 37.5% of the total liana stems in edge, interior and deep-interior sites, respectively. Liana abundance dif- MEI in AFR ranged from −1 to 0.92 (Table 1)  Connectance of the three networks was significantly lower than that of the null models ( Table 2). The specialisation asymmetric values of the networks in AFR were close to zero, indicating weak asymmetry. The specialisation asymmetry value of interior site network in AFR was consistent with that of the null model; those of the other networks were significantly higher than randomised expectations. The networks in SFR did not only show weak asymmetry, but they also did not differ significantly from that expected by chance. proportions of lianas than tree species showed higher specialisation than the null models.

| Network metrics
In AFR, the observed nestedness metric values were significantly lower than the means of the null model in the three forest sites (Table 2). Likewise, the liana-tree networks were less connected than the null models of the three networks. However, the three networks were more modular and specialised compared to the null networks.
The significant modularity of the networks resulted in the formation of a number of modules in edge site (14 modules), which was more than the number of modules in deep-interior (11 modules), which in turn, was more than that in interior site (7 modules

| Species topological roles in the networks
In

| Species co-occurrence of lianas
The matrix in edge site of AFR showed positive co-occurrence pattern, as the observed c-score of the matrix in edge site was significantly lower than the mean of the null model (   2015), which did not show modularity. In SFR, all the three networks were not nested but modular. Though the two nestedness patterns shown by the networks in the AFR and SFR refer to nonnested structure, that of the former depicts non-random assembly of species, whereas the latter indicates random assembly of species.
We argue that a clear distinction should be made between the two types of non-nestedness in network studies so that the distribution pattern of each of them would be fully understood. The absence of nestedness in AFR and SFR may be due to differences in liana species ability to colonise host trees and/or the use of defence strategies of hosts to avoid lianas (Addo-Fordjour et al., 2016;Genini et al., 2012).
As a recap, a nested structure is formed when there are interactions involving generalists and generalists, and specialists and generalists, but no interaction of specialists and specialists (Landi et al., 2018). Staniczenko et al. (2013) showed that for a nested quantitative network, interactions of generalist-generalist species are strongest, followed by those of generalist-specialist species, with no specialistspecialist interactions (or when present with much weaker interactions). Thus, for a nested structure to occur in a quantitative network like ours, there should be a good number of specialist and generalist species undergoing interactions. However, in our networks, we observed only a few generalists of lianas and trees that interacted, but with many specialist species interacting among themselves. This situation increased the likelihood of specialist-specialist interactions at the expense of generalist-generalist and generalist-specialist interactions, resulting in the absence of nested structure in the various networks. A similar trend was observed in mycorrhizal networks (Jacquemyn et al., 2015). The specialist-specialist interactions in our networks may account for the non-asymmetry and weak asymmetry of the networks. This finding shows that our networks tended to be more symmetric in their interactions, a trend which causes absence of nestedness and significant modularity in ecological networks (Guimarães et al., 2007). Overall, the findings on liana-tree network structure reported in the current and previous studies show that there is no universal pattern in the structure of liana-tree interactions. The patterns obtained may be dependent on the network complexity, and species traits and abundance, which are known to influence the organisation of liana-tree interactions (Sfair et al., 2010(Sfair et al., , 2018. The existence of high modular structure in the various networks may increase their stability and robustness by limiting diffusion of perturbations through the networks (Thébault & Fontaine, 2010). This may explain why the patterns of network structure in edge site were consistent with those in interior and deep-interior sites, irrespective of disturbance at edge site. The presence of modular structure in our networks may help conserve the networks of species interaction, which in turn, may lead to the conservation and maintenance of ecosystem functioning. The modular structure can enhance the stability of the liana communities in the various sites, and increase their robustness to perturbations (Olesen et al., 2007).
When lianas are connected to many trees within a community, they tend to make the trees susceptible to fall, because during natural or artificial disturbance, lianas connected to falling trees may pull down other trees connected to them. However, module formation in networks may limit the pulling effects of lianas on trees to only affected modules, thereby conserving species in the other modules.
Liana-tree interaction tends to be antagonistic, as lianas act as structural parasites of trees and compete intensely with trees for resources (Sfair et al., 2015(Sfair et al., , 2018. Species of antagonistic networks often evolve high specialisation in order to survive the antagonism of the interactions (Maliet et al., 2020). Our results revealed strong species and network specialisation in the forest sites, which implies the existence of strong liana-host specificity across the various networks in the two forest. Network specialisation and host specificity have been reported to cause non-nestedness and modularity in networks (Cordeiro et al., 2020;Dallas & Cornelius, 2015;Maliet et al., 2020;Wardhaugh et al., 2015). Given this information, the non-nested and modular structure observed in our networks may be driven by specialisation of the networks and host specificity of the liana species. The specialisation in the liana-tree networks may F I G U R E 7 Relationships between liana species abundance and their number of interactions in edge, interior and deep-interior sites of (a) Asenanyo Forest Reserve, and (b) Suhuma Forest Reserve be related to co-evolution in lineages of lianas and trees in the networks (Sfair et al., 2015). The possibility of co-evolution of lianas and trees in our networks is supported by Ponisio and M'Gonigle (2017), who showed that ecological communities that co-evolve become more anti-nested and modular over time. Montoya et al. (2015) found out that functional group diversity increases with modularity in complex networks, and that functional groups form modules in communities. In this regard, the presence of high number of modules per network in the forest sites may reflect the existence of different liana functional groups that interact with tree communities in the forests. Such networks with high level of modularity may possess increased resistance to disturbance (Olesen et al., 2007;Saunders & Rader, 2019). Differences in colonisation rates in fish parasites were found as a cause of anti-nested structure in such networks (Poulin & Guégan, 2000). In each of the networks, different liana species showed varying degree of specialisation, while others exhibited generalisation. This phenomenon suggests that the rates of colonisation differ markedly among the species, with highly specialised species having lower rate of colonisation, while species with low specialisation, or generalisation exhibit higher colonisation rate. In this regard, like the parasite-fish networks (Poulin & Guégan, 2000), the anti-

| Species co-occurrence of lianas
Generally, lianas were assembled randomly on their hosts in most of the forest sites, suggesting that chance events rather than edge disturbance, determined liana distribution on trees. Thus, we argue that the liana communities might have been assembled on trees by stochastic processes including host characteristics. Our finding is consistent with that reported in a semi-deciduous forest in Brazil (Zulqarnain et al., 2016). Contrary to the above, liana species in edge site of AFR showed positive species co-occurrence on their hosts.
Since this network was organised into modules, the positive co-

| Implications for forest management and conservation
This study presents findings on the response of liana communities and the structure of liana-tree interaction networks to edge disturbance. Our findings highlight that the severity of edge effects on liana species diversity may be influenced by land-use history prior to edge disturbance. This shows that edge effects on liana species diversity may not be universal but site-specific, depending on historical events of the forest. As this information is an exception to the general understanding that forest edges enhance liana diversity, it expands our knowledge about liana species diversity response to forest edge, which potentially, could contribute towards the development of a more inclusive theory on forest edge. Our study also shows that fragmentation of already disturbed sites may hamper liana species diversity due to edge effects. The presence of antinested structure in the networks of our forest sites could have arisen from strong selection for liana-host specificity (Dickie et al., 2017;Sfair et al., 2010), given that most of the liana and tree species were specialists. The specificity among the liana and tree species may reflect in the formation of modules in the forest sites. With this development, future disturbance in the sites may be localised to specific modules, thus resulting in the conservation of other species in the networks (Thébault & Fontaine, 2010

| CON CLUS ION
The findings of the study revealed considerable edge effects on liana diversity and abundance in the two moist semi-deciduous forests. The response of liana diversity to edge effects was positive in AFR, while a negative response was recorded in SFR.
Despite the enhanced abundance in edge site of the two forests, the patterns of liana-tree network structure of edge site were similar to those in interior and deep-interior sites. The networks in AFR showed anti-nested structure, while the networks in SFR revealed a nestedness pattern which was consistent with the null models. All the networks in the two forests were less connected, but modular and specialised. Lianas were mostly randomly distributed on host trees in all the forest sites except edge site in SFR.
Topologically, the majority of liana and tree species were peripherals (i.e. specialist), but a few species tended to be generalists, acting as connectors, module hubs and network hubs. The role of most of the species did not change from one site to another, even though the topological role of a few species changed from one site to another. Overall, our study shows that liana community structure was more susceptible to forest edge than liana-tree network structure. The findings of our study corroborate previous studies, and also present unique findings related to liana-tree network structure. Our findings which enhance our understanding of liana-tree interactions, have conservation implications relating to stability and robustness of the networks. Finally, the findings of the present study can potentially contribute to the development of a comprehensive theory on edge effects.

A K N OWLED G EM ENTS
The authors are grateful to the Forestry Commission of Ghana for providing access to the forest reserves.

CO N FLI C T O F I NTE R E S T
The authors declare no conflict of interest. Writing -review & editing (supporting).

DATA AVA I L A B I L I T Y S TAT E M E N T
The data associated with this study is available at dryad (https://doi. org/10.5061/dryad.crjdfn35t).