Using spatial pattern analysis to distinguish causes of mortality: an example from kelp in north-eastern New Zealand


  • Present address: Department of Zoology, University of Melbourne, PO Box 138, Queenscliff, Victoria 3225, Australia (e-mail

*Correspondence: Russell G. Cole, NIWA, PO Box 893, Nelson, New Zealand (fax + 64 3-548 1716; e-mail r.cole@NIWA.CRI.NZ).


1 Spatial analysis techniques were used to differentiate between climate-induced and pathogen-induced mass mortalities of the kelp Ecklonia radiata in north-eastern New Zealand. We predicted that climate-induced effects would generate broad-scale patterns, whereas pathogen-induced mortality would be traceable among neighbouring thalli.

2 Spatial autocorrelation analysis was performed on the proportion of E. radiata affected by dieback in quadrats during an initial mortality event in 1991. The absence of any consistent spatial scale of affected thalli between 10 and 100 m suggested that small-scale spread of an agent might be occurring.

3 Individual thalli were therefore mapped at two sites during a subsequent mortality event in 1992/93, and the degree of damage recorded. Spatial analyses found little evidence of aggregation of either intact or affected thalli at scales of 1–150 cm.

4 The relative spatial patterns of healthy and affected plants in mapped quadrats during the 1992/93 mortality provided little evidence of spatial association or repulsion between these broad damage categories.

5 The large-scale mortality of 1992/93 was consistent with a physiological response to broad-scale light deprivation, although other agents, perhaps both a virus and amphipod grazing, might also have been involved. Potentially complex interactions among the candidate agents render interpretation of the spatial patterns difficult.


Large plants play a major role in structuring communities. They contribute directly by photosynthesizing materials, and indirectly by providing habitat complexity. The contribution of trees, the primary ‘habitat-formers’ (sensuJones & Andrew 1993) in terrestrial ecosystems, to generating habitat heterogeneity is relatively well-known (Runkle 1985). In marine systems, macro-algae perform a similar role by extending habitat from the substratum into the water column (Schiel & Foster 1986; Jones & Andrew 1993). Macro-algae alter both the physical and biological characteristics of their surrounding environment (Reed & Foster 1984; Eckman et al. 1989) and have important flow-on effects for marine food chains (Bray & Ebeling 1975; Duggins et al. 1989; Horn 1989). Macro-algal stands are dynamic in space and time (Dayton et al. 1984, 1992), and these changes may have profound effects on the community as a whole.

Macro-algae may be subject to disturbance at a range of scales. For example, regional changes in climatic conditions (such as El Nino effects and upwelling; Dayton et al. 1984, 1992) may stress plants over a very large area (Tegner & Dayton 1987). Storms may remove sizeable tracts of seaweed from the substratum, although effects may be greatest at a particular depth range (Seymour et al. 1989). In contrast, grazing (primarily by sea urchins) generates more localized patchiness (reviewed by Pringle 1986; Scheibling 1996). Quantifying the spatial scales of abundance in macro-algal stands is therefore important for two reasons. First, the scale of patchiness may indicate which disturbance agents are exerting the strongest effects on the community. Secondly, in order to make predictions about the importance of disturbance on the community it is necessary to ascertain the grain and extent of the disturbance regime.

On exposed shores in north-eastern New Zealand, Ecklonia radiata (C.Ag.) J.Ag. is the dominant subtidal laminarian macro-alga, co-occurring with fucoids at depths of 3–5 m, and forming monospecific stands at depths greater than about 10 m (Choat & Schiel 1982; Schiel 1988, 1990). It is a stipitate kelp (reaching stipe lengths of 1 m or more) that affords food (Schiel 1982; Choat & Clements 1992) and shelter (Jones 1984; Choat & Ayling 1987; Taylor & Cole 1994) to many organisms. As with most temperate macro-algae, the main sources of mortality are storms (Novaczek 1984) and grazers (Schiel 1982). However, two instances of mass mortality of E. radiata observed recently at Goat Island near Leigh (36°16′S, 178°48′E) (Cole & Babcock 1996) were not associated with either storms or sea urchin grazing, and Easton et al. (1995) found virus-like particles on affected thalli. We employed spatial analyses to determine the pattern of mortality and distinguish between two possible causal agents: (i) light limitation associated with plankton blooms, and (ii) a pathogen transferred through the population.

Our hypotheses were that if a large-scale factor such as light limitation was responsible for the mortality, it would act synchronously across the entire depth zone, and we would be able to detect this in a large-scale analysis. In contrast, if a disease vector was responsible for the mortality we would be able to detect its transmission at smaller spatial scales. We therefore combined a broad-scale analysis carried out in 1991, in which we recorded morbidity in a grid in E. radiata forest with a fine-scale study in 1993. We analysed the broad-scale data using spatial autocorrelation (Ripley 1981; Legendre 1993), which enabled scales of pattern ranging from 10 m to 100 m to be quantified. The fine-scale study comprised univariate and bivariate point-pattern analysis (Diggle 1983; Ripley 1987) and provided resolutions of patterns at scales of 0.01–1.5 m.


Grid study – 1991

In May 1991, thalli of Ecklonia in Alphabet Bay (within Goat Island Bay, near Leigh, New Zealand) suffered ‘dieback’. During dieback, the primary lamina erodes from the distal end, with attendant loss of the secondary laminae, which contribute most of the photosynthetic tissue and biomass of the kelp (Fig. 1) (see Cole & Babcock 1996 for further description and photographs). During the May 1991 dieback event, we delineated a rectangular grid within the affected area in Alphabet Bay (Cole & Babcock 1996). The grid (100 m parallel to shore and 50 m perpendicular to shore) was sampled with 1-m2 quadrats (n = 121) at 10-m intervals alongshore and 5-m intervals offshore, over a 3-day sampling period (11–13 May 1991). We counted the numbers of intact and affected thalli in each quadrat, and calculated the proportion of affected plants in each quadrat in which thalli occurred. Estimates of population size and structure of intact and affected thalli were obtained from random quadrat samples in the area.

Figure 1.

Decay stages recorded for Ecklonia radiata in 1992/93.

We analysed these data by calculating the spatial autocorrelation of the proportion of affected thalli. We used Moran's I, the most widely used coefficient of spatial autocorrelation, which measures the spatially weighted correlation between values of a variable (in this case percentage morbidity occurrence) in quadrats at different distances apart (Sokal & Oden 1978a,b; Upton & Fingleton 1985; Legendre 1993). The Spatial Autocorrelation Analysis Package (SAAP, see table 3 in Legendre 1993) was used to calculate the correlogram.

Mapped thalli – 1993

Analysis of the 1991 patterns from the grid at Alphabet Bay showed evidence of patchiness at the lower end of the scales tested, so when a further mortality event occurred during 1992/93, we studied finer-scale processes by mapping individual thalli. We chose two sites within Goat Island Bay (Knot Rock and Martins Rock, c. 500 m west of Alphabet Bay) with flat sandstone ledges to minimize the effects of topographic complexity on the accuracy of distance measurements. In May/June 1993 thalli were mapped in 3 × 3-m quadrats using the method of Underwood (1977). Quadrats (n = 3 per site) were established by nailing plastic tags into the substratum in areas 10–11 m deep where E. radiata was present. The quadrat size was constrained by the upper limit of reliable measurement with graduated tapes. The position of each thallus was measured to 1 cm accuracy and its condition categorized into one of the five decay classes (Fig. 1).

Point-pattern analysis was used to decompose the observed distribution of all thalli within each quadrat into separate fractions of pattern attributable to the distribution of the damaged thalli, the intact thalli and their interaction. A significant departure from randomness of the portion due to interaction would provide evidence of small-scale spatial structuring of the mortality agent. Software written in QBasic (Duncan 1991) was used to calculate K(t) statistics. Univariate patterns (i.e. those derived from total thalli, intact only, or damaged only) were calculated using the second-order univariate property from the function (Diggle 1983):


where λ is the intensity (analogous to the mean) of the number of ‘events’, in this case macro-algal thalli, and E represents the expected value. This function represents the number of thalli we would expect within a certain distance of a plant, given the number of thalli within the quadrat. Following Haase (1995) we re-expressed K(t) as:


because we were interested in the deviation from complete spatial randomness.

If the observed value was greater than that expected from a random distribution, then the thalli were clumped at that scale. Conversely, if the observed value was less than the predicted value, thalli can be said to be overdispersed at that scale. Confidence envelopes for both analyses were calculated by 99 randomizations for each separation.

Bivariate pattern (i.e. due to the interaction between intact and damaged thalli) was described by the second-order function (Diggle 1983):


Again, this was re-expressed in the form:


If the observed value was greater than that expected from a random distribution, then the two types of event were positively associated. Conversely, if the observed distribution was less than the random expectation then the two event types were negatively associated.

The analysis of mapped data was complicated by the presence of boundaries; calculations made using individuals near the edge could bias the value of the test statistic. The use of buffer zones around each quadrat would have been impractical; the diving involved was at the limits of what we could accomplish physically. Consequently, we used the edge correction of Getis & Franklin (1987). Haase (1995) concluded that underestimates of K(t) might arise in some circumstances, but that the discrepancies introduced by the Getis & Franklin method increased at larger t-values. As we limited our analysis to half the quadrat length, this was unlikely to be important.

We classified morbidity into two classes: intact and damaged thalli (i.e. all damage categories were pooled). We have shown that once an Ecklonia thallus enters the reduced stage, it inevitably dies after moving through primary only, half stipe and holdfast stages (R. C. Babcock and R. G. Cole, unpublished data). Thus morbidity is a binary event; the seaweed either exhibits symptoms or it does not. Reduced individuals will inevitably move to a later morbidity stage, whereas intact individuals may or may not. Sample sizes for intact individuals were low in some cases, and this is reflected in the width of the confidence intervals.

The 1991 survey had suggested that larger (longer-stiped) individuals were more affected. Stipe length (for all thalli with complete stipes, i.e. primary only, reduced primary and intact) was therefore measured again in 1993, in three of the six quadrats (Knot 3, Martin 2, Martin 3). Spatial autocorrelation was used to analyse the stipe lengths of those plants to investigate whether plants with long stipes were clumped. Macro-algal growth may be positively correlated with density (Schiel & Choat 1980), and as a consequence we expected that longer stipes would be clumped a priori. Spatial autocorrelation analysis of stipe length tested whether clumping of affected thalli could be simply due to clumping of individuals of similar sizes.


Grid study – 1991

The pattern of morbidity at Alphabet Bay appeared patchy at scales of 10 m or more, although there was a tendency for the entire western (deeper) edge of the grid to have high proportions of affected plants (Fig. 2). However, there were also patches of high damage (> 80% of thalli) near the shallower eastern border of the study area, and a broad band of damage (> 60% of thalli) ran onshore–offshore through the centre of the grid. Patches of low morbidity occurred on the southern and northern borders. The apparent patchiness was not statistically significant at any scale, although there was a pattern of alternating positive and negative autocorrelation over the three smallest distance classes, indicating weak patchiness at scales of 10–30 m (Fig. 3). Most stipes in the area were less than 80 cm long, with lower numbers in the 10–20 cm size class (Fig. 4). Individuals with longer stipes were most affected; more than half the plants with stipes greater than 60 cm lacked laminae (Fig. 4).

Figure 2.

Map of Ecklonia radiata morbidity (percentage of thalli affected per 1-m2 quadrat) at Alphabet Bay, New Zealand, May 1991. The percentage of thalli affected is plotted in 20% contour intervals, with increasing percentages affected identified by darker shading. Depth along eastern (inshore) border c. 9 m, western (offshore) border c. 12 m.

Figure 3.

Correlogram of Moran's I, for the percentage of Ecklonia radiata thalli affected in the Alphabet Bay grid, New Zealand, May 1991. The horizontal dotted line indicates the expected value. No Moran's I-values were statistically significant.

Figure 4.

Numbers of Ecklonia radiata thalli in 10-cm stipe length classes that were affected (reduced primary lamina or lacking all laminae) or unaffected by dieback, Alphabet Bay, New Zealand, May 1991.

Mapped thalli – 1993

A similar pattern of decreased proportions of intact thalli and increased proportions of greater damage among longer stipes was evident following the 1992/93 outbreak (Fig. 5). Maps of the quadrats suggested clumping of particular damage categories (Fig. 6) and were therefore analysed with point-pattern analyses. No consistent intersite differences were evident in thallus distributions. The pattern of all thalli combined showed a random distribution in one quadrat at each site (Knot 3, Martin 3) while the remainder were clumped at varying scales (Fig. 7a). Thalli in Knot 1 were clumped at small scales (0–45 cm) and random at larger scales, whereas those in Martin 1 and Martin 2 were clumped at intermediate and large scales (> 40 cm) but randomly distributed at the finest scales (Fig. 7a). Thalli in Knot 2 varied between clumped and randomly distributed over the entire range of separations analysed (Fig. 7a).

Figure 5.

Percentage of three morbidity categories within thalli of each size class. Sizes were pooled across three mapped quadrats (Knot 3, Martin 2, Martin 3) measured in May/June 1993.

Figure 6.

Maps of morbidity stages of Ecklonia radiata in quadrats at Knot Rock and Martins Rock, Goat Island, New Zealand, measured in May/June 1993. I = intact; R = reduced; P = primary only; H = half stipe; F = holdfast. See Fig. 1 for details of morbidity stages.

Figure 7.

Plots of L(t) vs. spatial separation for each quadrat in which Ecklonia radiata was mapped. Dotted lines = confidence intervals based on 99 simulations. (a) Total thalli pooled; (b) intact thalli; (c) damaged thalli; (d) L12(t) vs. spatial separation for interaction between intact and damaged thalli.

Analysis of individual morbidity classes did not indicate any consistent pattern of aggregation. Intact thalli at some sites were randomly (Martin 2, Martin 3) distributed at all scales from 10 cm to 150 cm, but there was no consistency in the scale of clumping at the other sites: intact thalli at Knot 1 were marginally aggregated at 20 cm, at Knot 2 at all scales above 40 cm (Fig. 7b), whereas intact thalli at Martin 1 were aggregated at both small scales (< 10 cm) and larger scales (> 100 cm) (Fig. 7b). Due to the low numbers of intact thalli in Martin 2, the confidence intervals were very broad (Fig. 7b). Thalli with symptoms of dieback were generally clumped at most scales, although those at Knot 3 and Martin 3 were randomly distributed (Fig. 7c). Affected thalli at Knot 1 were only clumped at scales of less than 60 cm (Fig. 7c), in contrast to the patterns at Knot 2 (all scales), Martin 1 (> 40 cm) and Martin 2 (> 20 cm) (Fig. 7c).

Despite some indication that intact and affected plants were aggregated, no bivariate associations were apparent at any scale in any quadrat except in Knot 1, where negative associations between intact and affected thalli occurred between c. 20 cm and 60 cm (Fig. 7d). Overall this suggested that the aggregations evident in the univariate analyses occurred independently of damage category, and hence were likely to reflect the spatial pattern of the thalli before dieback rather than representing the transmission history of some putative pathogen. However, the confidence limits for most of the bivariate analyses were broad, due to the small sample sizes for intact thalli in most quadrats.

Because larger plants were affected more by the mortality, our failure to detect either negative or positive associations between affected and unaffected plants might have been due to the spatial distribution of different-sized thalli: if similar-sized plants were clumped then the patterns of mortality could be determined by one or both of size and proximity. To examine whether absence of pattern in the bivariate analyses could be due to spatial patterns in sizes, we calculated Moran's I (using inverse distance-squared weighting) for stipe lengths in the three quadrats where E. radiata stipes were measured. No consistent patterns were found in the pattern of I with respect to stipe length: in two quadrats (Martin 3 and Knot 3) a statistically non-significant positive autocorrelation indicated a possible clumping of similarly sized stipes, but stipes of similar sizes were negatively autocorrelated at Martin 2.


The scales of mortality patterns may suggest the scales at which the processes responsible for mortality operate (Mollison 1977; Chesson & Murdoch 1986). In this study, we identified two models that could explain both the source and dynamics of dieback in kelp forests. These models were distinguishable by the different scales at which morbidity patterns would be predicted to appear. First, a large-scale mortality agent such as light deprivation might be expected to generate broad-scale (> 10-m) mortality patterns (corresponding, for example, to changes in the depth at the sampling site) but would generate no fine-scale pattern. On the other hand, small-scale processes such as plant–plant pathogen transmission (Easton et al. 1995) might be expected to generate a small-scale (0.01–10-m) pattern.

Roughly 50% of the kelp thalli were affected in the 1991 outbreak, and kelps with long stipes were affected more by dieback. Large-scale (10–100-m) analyses carried out during that outbreak indicated no broad-scale patterns of morbidity. Although small-scale (0.01–1.5-m) analyses following the 1992/93 outbreak found limited evidence of aggregation of both intact and affected plants within quadrats, the distribution of each was independent of the distribution of the other damage category. In other words, the aggregation reflected the existing spatial pattern of the thalli rather than the spread of a pathogen from a single affected individual. No clear spatial patterns of morbidity were detected.

The 1991 dieback did not appear to be caused by light deprivation, and this was further supported by the fact that the outbreak did not exhibit the medium- to large-scale pattern of variation with depth that was subsequently observed during the 1992/93 event (Cole & Babcock 1996). No dramatic climatic aberrations were apparent during the outbreak (Francis & Evans 1993; cf. Rhodes et al. 1993), which was localized. Of six sites near Leigh visited during May 1991, only Alphabet Bay exhibited dieback (R. G. Cole, personal observation). We therefore suggest that localized environmental stress (freshwater runoff and the reduction of nutrients in warm coastal water during the preceding summer) operating at the scale of 100–1000 m acted to diminish kelp health at Alphabet Bay at this time.

The lack of broad-scale pattern in 1991 suggested that smaller scale processes might have contributed to dieback, and thus we mapped individual plants following the 1992/93 event. However, the spatial patterns at this scale revealed little evidence of spread of any factor between individuals. Although plants of all sizes suffered mortality, larger individuals were proportionally more affected, and it is possible that size-selectivity may have obscured any underlying small-scale patterns. If, for instance, direct physical contact is necessary for pathogen spread, this might occur more rapidly among larger plants because they have greater orbits. Although Easton et al. (1995) found evidence for involvement of a virus in kelp mortality, distinguishing primary and secondary agents is problematic and, without a sound knowledge of the primary agent and its temporal patterns of expression, we cannot be certain that the spatial pattern does not reflect pathogen spread after initial infection of longer-stiped individuals within the population of E. radiata.

It is possible that the two mortality outbreaks had different ultimate causes because the 1991 event was restricted to a single bay, whereas the 1992/93 event was more widespread. Synergistic interactions could also occur, with large-scale environmental stresses facilitating small-scale transmission of and susceptibility to disease. Large- and small-scale patterns should therefore be determined simultaneously to provide better discrimination. Finally, it should be borne in mind that not all plants may be susceptible to all mortality agents: larger plants may, for example, be more susceptible than small plants because they have lower polyphenolic concentrations (New South Wales data in table 3 of Steinberg 1989) that might influence grazer or pathogen activity (see below).

Numerous ecological factors may be linked to spatial proximity of plants (Real & McElhany 1996). Several studies (e.g. Schiel & Choat 1980) have found positive density-dependent growth of macro-algae, and if mortality is size-dependent this would produce the same spatial pattern of morbidity as a pathogen transmission model (i.e. ‘apparent contagion’; Campbell & Madden 1990). However, as we found no evidence of positive spatial autocorrelation of stipe lengths the observed pattern was not in this case simply an artefact of clumping of similar-sized plants.

While it is naive to expect that spatial analyses alone can be used to identify a disease-spreading process (Real & McElhany 1996), when used in association with biological information about the disease they are a cheap method of deciding which hypotheses are worthy of further investigation or experimental testing (Ord & Getis 1995; Lewis 1997). A major barrier to the effective use of spatial analysis to distinguish between mortality causes in macro-algae is the scarcity of information on pathogens and their vectors. Recent successes of such spatial analyses for terrestrial plants (Alexander & Antonovics 1988; Gilbert et al. 1994; Shykoff & Bucheli 1995) have depended on detailed information regarding the identity and biology of their pathogens. Recent investigations (T. Haggitt, Leigh Marine Laboratory, personal communication) suggest the association of an epifaunal amphipod with recurrent outbreaks. Although this association is consistent with kelp mortality being simply due to herbivory by amphipods, there may be more complex interactions: amphipods may track a pathogen, a pathogen may be spread by amphipods, or amphipod grazing might weaken plant defences or provide sites of infection. Holmes (1997) noted that diseases are more likely to persist in a population if that population is well-mixed, and Taylor (1998) suggested that night-time turnover of amphipods on clumped macrophytes was in the order of 55–69%, but because the ambit of individuals is unknown we cannot identify an appropriate scale at which to investigate that pattern. Finally, the temporal resolution afforded by transfer between damage classes is poorly known.

The environmental patchiness provided by dieback may, in turn, have important effects on the rocky reef ecosystem. Kelp removal may reduce numbers of epifaunal organisms (Taylor & Cole 1994) and alter community composition of fishes (Choat & Ayling 1987; Syms & Jones 1999). Detritus from kelp mortality may alter feeding patterns of the grazers that maintain biotic habitat structure (Duggins et al. 1989). In north-eastern New Zealand, kelp dieback may be a re-occurring, albeit irregular, event that has the potential to generate strong down-stream effects on subtidal ecosystems. Natural history information has been used in other systems to predict small-scale spatial patterns (Hewitt et al. 1996); information on rates of thallus decay, virus expression, amphipod birth, feeding, growth and migration will be required to allow further interpretation, and perhaps prediction of patterns of kelp mortality.


We thank P. Brown for assistance, and R.B. Taylor, R.C. Babcock, T.J. Anderson, the referees and the editors for their comments on the manuscript. This study was partly supported by a UGC scholarship (to R. G. Cole), and the University of Waikato's Department of Earth Sciences and National Institute of Water and Atmospheric Research Ltd, Nelson, provided facilities for manuscript preparation.

Received 12 November 1998revision accepted 11 May 1999