Recently I read the publication ‘Weak spatiotemporal response of prey to predation risk in a freely interacting system’ (Cusack, J.J. et al., 2019) and the following text is my response to the study.
However, I am very new to statistics and, therefore, I cannot guarantee that I have understood the analysis aspects most correctly.
This publication highlighted a statistical analysis problem which I had not considered before, namely, the necessity to adopt a null model.
If I understand correctly, the null model is needed because we are studying landscapes and communities in which predation already exists and has existed for some time.
Null model allows to determine if something has changed at all by comparing the current status to a simulated situation in which no change has occurred.
The trouble with studies attempting to assess the impact of predation on the prey behaviour lies with the fact that the prey behaviour has already been altered by predation and we cannot fully evaluate the degree of alteration of prey behaviour by predator presence because we cannot compare it to some ‘pristine’ or ‘pre-predator’ state which would allow us to determine how elk moved on these landscapes without predators around.
Any analysis of current prey behaviour is already an analysis of prey behaviour under predation pressure.
In order to draw potential comparisons, null model implies the use of randomly generated prey potential movements and use of landscape.
These random generations are, as far as I understand, based on observed prey movements that, in order to be able to detect significant changes when compared to wolf-impacted behaviour, have been randomized.
It is interesting to follow publications where such comparisons are possible to be made by analyzing data that has been gathered before and after the recolonization of a species – in cases when baseline data has been collected prior to the change the effects of which are being evaluated (e.g., the cougar and wolf interaction study where cougar space and time use patterns were known prior to wolf recolonization and could be, consequently, compared to the patterns after the recolonization – Atwood, T.C, et al., 2007).
As stated before, in order to resolve this problem (lack of knowledge of how the prey behaved without wolves on the landscape because data could only be obtained with wolves already on the landscape), null model is used with random movement generation from the existing movement patterns.
However, I do not wholly understand why these random movements are considered adequate substitutes for actual prey movement without predation pressure, especially, when degree of avoidance is rather crucial.
Namely, elk do not move randomly when wolves do not live on the landscape and the randomization of observed elk movement after wolf recolonization might not reflect elk use of space properly without wolves affecting this use.
Elk move with respect to food (including preferred food which is seasonally variable), with respect to their social grouping patterns and hierarchies, with respect to shelter, to water, to other predators (especially during calving) and with respect to weather and snow accumulation.
As mentioned in the publication by Cusack, J.J. et al., 2019, snow (winter conditions) might make the elk response to predation less pronounced because it changes (limits) where elk can access food, where they can travel more easily, where trails can be established and used by many elk (if the snow is very deep).
This leads to a reduced choice – elk simply have to use the available habitats regardless of predation risk because there are no or there are few alternatives.
Thus, predator effects on prey movement would be less pronounced in such periods when prey cannot compensate for loss of some foraging grounds or migration routes.
Such movements might be more predictable and less altered by predation effects than movements during vegetation growth season.
But even movements in spring – summer – autumn are hardly random and because wolf presence can influence, for example, social grouping and access, I do not believe that randomization of observed elk movement with wolves on landscape are reflective of elk movement without wolves on landscape when elk would have different grouping, access and other factors impacting their behaviour but they would not be restricted by predation, either.
Thus I have been wondering how the randomization of the null model works in studies concerning predator effects on prey movement, e.g., landscape of fear studies.
This is of utmost importance because such estimates found the basis of, for example, trophic cascade research that allows us to appreciate the role of large predators in our ecosystems.
Inappropriately designed null models could mask the effects of predators on prey.
If the effects of predators on prey movement after predator recolonization are analyzed through comparison to random movements by prey with predators already on landscape, it might be impossible to determine the true adjustments by the prey to the predator presence.
Elk movements are non-random but if they have to be randomized in order to provide a baseline, this randomization should not be derived from a system which has already been impacted.
Thus, my proposition is that the null model should not be randomized at all but rather based on assessment of average behaviour patterns by prey toward certain landscape or resource attributes that have been adapted for the specific study site.
I have been wondering if null model could be worked out in some systems where wolves have not yet colonized and where elk movement ‘mean values’ could be estimated with reference to some common landmarks and resources (water, shelter, optimal / suboptimal forage etc.).
Of course, elk behaviour changes from place to place but if some averages or percentages were obtained related to landmarks and resources that all elk (/red deer) are using everywhere, the null model then could be based on a more natural elk spatial behaviour without predation restraint because elk behaviour is not random and random behaviour cannot attribute weight to the actual degree of avoidance and adjustment.
This points at two problems:
- Elk movement is never random and random movement cannot characterize the true prey response;
- Elk adjustments to predation pressure cannot be truly evaluated if we compare them not to free choices by elk made without predation restraints but to already restrained elk movements.
Otherwise – do we not fail to take into account the elk preferences (selectivity) for certain sites or resources which govern their movement and also their avoidance or inability to use certain parts of habitats?
For example, elk might exhibit more pronounced aversion from wolves if their choices were more limited than random but, at the same time, free from predation risk.
If the randomization, for example, assumes that the whole area of elk-use is available to all elk (regardless of age, status etc.) and that most parts of it can be used with equal ease at all times, but in reality snow hinders movement in most parts and only three routes are available throughout most of the winter and two of those routes are affected (even if slightly) by predator avoidance strategies – this would indicate at stronger avoidance reaction than if we assumed that there were many more random routes available.
Also, for example, if some areas cannot be used in certain times due to wind exposure or because they accumulate more snow (northern aspect) but they can be used at other times (in afternoons, when wind blows from other direction, during periods when temperature has considerably risen etc.) or they can be used by females and not by males (as is common in red deer species in Europe regarding lower-quality, shorter swards), then this also reduces the temporary availability of space to the elk, and the reactions to predator temporal patterns might appear more pronounced when these restrictions have been accounted for.
For example, if some site can only be used in afternoons or mornings etc. due to some climate/weather circumstance, then alteration of this use by predator avoidance (also in the mornings when predator activity tends to peak) would be more significant than with respect to sites that can be used at all times.
If elk are unable to use some of such areas with wolves on the landscape while they would likely be using them without wolves on the landscape – randomization from observed elk movements with wolves on the landscape would conceal the actual choices that elk have made after the reappearance of the wolves; and randomization as such would veil the true selectivity that has been altered.
I think that the null model should be more complex and that the use of space (also in time) should be correlated to a more specified behaviour by varied groups of individuals during different times with respect to various landmarks / resources that is exhibited typically by elk without predators on the landscape.
Notes by Ieva 