Multidimensional Analyses for Testing Ecological, Ethnobiological, and Conservation Hypotheses

Abstract

This chapter outlines the commonest multidimensional analysis used in ecology, ethnobiology, and conservation. After mastering how to create a scientific research workflow based on a problem-based platform (Chap. 7, in this book), readers will learn how to visualize complex datasets (e.g., species or ethnospecies lists, several social-political or environmental variables, and so on), and test multivariate hypotheses. We provide a full reproducible example of every analysis discussed in this chapter as an Online Material. We further encourage students and researchers to move from a “description-based multidimensional analysis” (such as using unconstrained ordination, like PCA) to an explicit hypothesis-testing framework that can greatly improve learning and research programs.

1 Appendix for the Chapter Multidimensional Analysis for Testing Ecological, Ethnobiological, and Conservation Hypothesis

Thiago Gonçalves-Souza, Michel V. Garey, Fernando R. da Silva, Ulysses P. Albuquerque, and Diogo B. Provete.

21/05/2018

• 0.1 Loading the required packages

• 1 Unconstrained ordination

  • 1.0.1 Principal Component Analysis (PCA)

  • 1.0.2 Principal Coordinate Analysis

  • 1.0.3 Principal Components Regression

• 2 PERMANOVA

  • 2.0.1 EXAMPLE 1

  • 2.0.2 EXAMPLE 2

• 3 Community-Weighted Mean (CWM)

• 4 Constrained ordination

  • 4.0.1 RDA

• 5 Clustering

  • 5.0.1 Indicator value index

0.1 Loading the Required Packages

1 Unconstrained Ordination

Here, we will firstly use examplar datasets available in the R package datasets that already comes with any R installation.

Let’s look at the data. This dataset contains the masurement of sepals and petals of 50 flowers for each of the tree species of the genus Iris. You can see more by using /Library/Frameworks/R.framework/Versions/3.5/Resources/library/datasets/help/iris.

Next, we will run a PCA with a few packages that offer different outputs.

Finally, the plot:

The package factoextra has many nice plot capabilities. Here, we will plot the equilibrium circle to check the relative contribution of each variable to the formation of the axes:

The variables more or less contributed the same to the formation of the ordination axes.

And finally, the same analysis with the

The advantages of this package include: dealing with missing data and ability to easily plot the centroid of categorical variable in the ordination diagram. Let’s see how it runs:

As you can see, it automatically returns the biplot with the equilibrium circle and the individual factor map plotted with the centroid corresponding to the mean of each species in each dimension. With the equilibrium circle you can judge with each individual variable contributed more or less, depending on the length of the vector (arrow), to the formation of the axes in the reduced space. In this example, all four variables contributed more or less equally (see p. 447 of Legendre & Legendre 2012).

And we can easily see the correlation of each variable with the first two axes. This is useful when you want to use the axes in subsequent analyses, and ideally have to interpret how axes are summing up the information:

We will import a dataset published by Rodrigues et al. (2018) Plant Trait Dataset for Tree-Like Growth Forms Species of the Subtropical Atlantic Rain Forest in Brazil. Data 3(2), 16; https://doi.org/10.3390/data3020016. Available here . The dataset presents traits of leaf, branch, maximum potential height, seed mass, and dispersion syndrome of 117 tree species. For the sake of simplicity, we will use just a few traits and categorize half of the species as being used by local people and the other half not used.

Before jumping in and analyzing the data right away, let’s first take a closer look at it to better understand it. Let’s quickly check the relationship among variables with a correlation matrix:

Another ke aspect in any dataset is the amoung of missing values, usually coded as NAs.

As you can see, there’re a few data missing. So let’s choose variables with as few missing values as possible that are important to test our hypothesis.

Remember that FactoMineR::PCA can handle missing values by replacing them with the mean value of the column. But here, for the sake of simplicity we will just remove them:

Let’s visualize our new, reduced data set:

Now we will create a vector to categorize half of the species as preferred and the other half as nonpreferred, in order to test our initial hypothesis. Of course, this is just to demonstrate how the method works and doesn’t mean that those species are really used for any medicinal purpose.

Since, values were measured in different scales, we have to standardize them before running a PCA. Luckily, this is done by default internally in the function dudi.pca, arguments center = TRUE and scale=TRUE.

Now, the fun part, let’s see the results.

First, a basic graph showing the results, disconsidering the groups:

Let’s see how variables contributed to the formation of the axes:

Plant height (Hpot95_m) and Bark prop were not very useful for the formation of the axes. Here, the angle of the arrows informes about the correlation of that variable with a given axis. For example, SLA is closely related to axis 1, meaning that this variable contributed very much to the formation of the axis, but not much to the second axis. Interestingly, Hpot95_m is contributing equally to the formation of both axes. This means that data points (species) arranged to the right of the ordination plot (positive side) have a large SLA and are tall. Similarly, thos to the left are small trees and have small SLA, but have tick leaves, large seeds, and high BarkProp.

Now, what about the preferred vs. nonPreferred species?

We can see that our hypothesis (see main text) was shamefully rejected, since the morphology of the preferred was not different from the nonPreferred plants. Of course, again, this example was just to illustrate how the method works and are not based on any real dataset aimed to test this hypothesis.

A cleaner graph without the row labels allow us to better see that the groups totally overlap:

Here, we will use the plant species trait data available in the package cluster to illustrate how a PCoA works. This dataset contains 31 morphological and reproductive variables for 136 plant species along a urbanization gradient. Let’s begin by loading the data:

Notice that the there’re many traits coded as semiquantitative variables (ordered factor in R language), binary variables, and two continuous variables. Here, we will use the dist.ktab of the package ade4 to calculate the Gower dissimilarity coefficient. First, we have to inform the function the nature of the variables

Notice that according to the Broken stick criterion, the first 12 eigenvectors should be selected for further use, e.g., in a Principal Components Regression.

Let’s look at the biplot:

The relationship between environmental gradients and patterns of geographic variation in intraspecific body size of Boana faber

• Hypothesis: environmental gradients drive intraspecific body size of Boana faber

• Dependent variable: body size of Boana faber

• Independent variable: environmental variables (e.g., temperature and precipitation)

Let’s begin by importing the files

The first step is to implement the Principal Component Analysis (PCA), and then store the eigenvectors

We can see that the first axis explain 0.47% of the variation, while the second axis explain 0.28%, together they explain 0.75%.

Plotting the results in a biplot

We’ll select the first two axes of PCA to further use

Now, we’ll build several models of Principal Component Regressions (PCR) with different predictor variables for later use:

Using the Akaike Information Criteria corrected for small sample sizes (AICc) to select the best model

The model containing only the first axis of PCA is the most parcimonious with the lowest AICc value.

Now, let’s visualize the results:

The relationship between the number of fish species known by fishermen and their personal experiences

Hypothesis: personal experiences influence the number of fish species known by local fishermenDependent variable: number of fish species *Independent variable: personal experience (e.g., age, time as fisherman, number of fishing techniques used, if his father was also a fisherman)

Loading files:

Principal Component Analysis (PCA)

We can see that The first axis explain 0.57% of the variation, while the second axis explain 0.21%, together they explain 0.78%.

Visualizing the results:

Retaining the first two axes of PCA:

Building models of Principal Component Rrgressions (PCR) with different predictor variables

Using Akaike Information Criteria corrected for small samples (AICc) to select the best model

We can see that the model containing only the first PC is the most parcimonious one with the lowest AICc value.

Now, let’s visualize the results:

2 Permanova

It is an illustrative dataset that examines the hypothesis that land use alters the arthropod specie composition in the Cerrado biome.

• Hypothesis: land use alters the specie composition

• Dependent variable: arthropod species

• Independent variable: areas (pastures, sugarcanes, and Cerrado protected areas)

Importing the data with arthropod species composition:

To run a PERMANOVA, we first have to calculating a dissimilarity measure. Here, we’ll use the percentage different (aka. Bray-Curtis coefficient). We needd to exclude the first colum of data with treatments, that’s why the data[,-1].

PERMANOVA analysis uses the dissimilarity matrix as response variable and the column containing the treatments as predictor:

The results show that arthropod species composition are different among treatments with R² = 0.64; and P = 0.001. Now, we’ll perform a post-hoc test to find which treatments are different between them. For this, we will use the function pairwise_adonis from the package pairwiseAdonis, available in GitHub .

The results show that all areas are harboring different species composition. Now, let’s see the results using a Non-Metric Multidimensional Scaling (nMDS), which is the non-metric version of PCoA. This method will not be covered in the chapter.

Knowledge about plant species with medicinal uses by people living along an urban-rural gradient.

• Hypothesis: Urbanization decrease the knowledge of people about plants with medicinal uses

• Dependent variable: plant species

• Independent variable: people living in different areas (urban, periurban, and rural)

Loading the data with species composition of medicinal plants along urbanization gradient:

This time, we will use the Jaccard index of dissimilarity.

Running the PERMANOVA

The results show that knowledge about plant species is different across the gradient with R² = 0.47; and P = 0.001. Now, we’ll run a post-hoc test to find which treatments are different between them:

The results show that knowledge about plant species composition by people living in rural areas is different from people living in periurban and urban areas, but it is not different between people living in periurban and urban areas.

Performing Non-Metric Multidimensional Scaling (nMDS) to visualize the results:

The figure shows the difference in knowldege about plant species composition with medicinal uses.

3 Community-Weighted Mean (CWM)

Importing the datasets:

Preparing the data:

Running the CWM

Common errors while using the function functcomp: avoid using data.frame - so, use “as.matrix”. Species names are not the same (spell or order) in T and Y; so, use the following code to double check if they are correct. The result values should be all TRUE

Combining PERMANOVA with CWM to test hypothesis

Details about the permutation procedure can be found here

Here, we’ll run the permutations by strata to avoid pseudoreplication. First, let’s create the strata argument to run in the adonis function

Standardize environmental variables before the PERMANOVA if they are in different scales. You can do that directly with dplyr::mutate

Lastly, run the permanova with adonis function

Check the homogeneity of multivariate dispersions:

Visualize the results with a PCoA:

You can check which variables are correlated with the first two axes by using:

4 Constrained Ordination

We will use two exampled to illustrate how an RDA works.

Illustrative dataset that examines species-environment relationship.

• Hypothesis—The anurans species composition changes in response to environmental gradients.

• Dependent variable—anurans species

• Independent variable—environmental descriptors of habitat

First, let’s import all the data:

Then, let’s begin processing the data. We have to implement Hellinger transformation to ensure the matrix has Euclidean properties, because we’re dealing with species abundance with potentially many zeros.

The environmental variables were measured in different scales. So, we have to standardize them before consucting any analysis, to make them comparable:

Finally, let’s run an RDA

Calculating the P-value for the RDA model:

*P-value of the global model:

*P-value of rda axes:

*P-value of environmental variables:

Finally, let’s plot the results. This figure shows the variation in species composition constrained by environmental gradients

Illustrative dataset that tests whether cultivated medicinal plants change in relation to socioeconomic characteristics of households

• Hypothesis:Plant species with medicinal uses changes in response to socioeconomic characteristics

• Dependent variable: plant species

• Independent variable: socioeconomic characteristics of the household

Let’s begin by importing the data:

Data Processing:

Now, running the RDA

Calculating P value for the RDA model:

*P-value of global model

*P-value of rda axis

*P-value of environmental variables

5 Clustering

Illustrative dataset that evaluate which Ephemeroptera species are bioindicators of environmental quality.

• Hypothesis: Habitat contamination will affect different species resulting in the exclusion of some species in contaminated environments.

• Dependent variable: Ephemeroptera species

• Independent variable: preserved and anthropized streams

Let’s begin by importing the data:

We’ll divide the data into two parts, with a factor for preserved and modified areas:

Finally, let’s calculate the IndVal:

Illustrative dataset that examines the phytotherapy used by indigenous and non-indigenous populations

Hypothesis - Analyze if medicinal plants used to treat some illnesses are the same among indigenous populations and immigrants from the Southern, Brazil Dependent variable - phytotherapy species *Independent variable - populations (indigenous and European immigrants)

In this example we will used another package to calculate IndVal, which generates a different output

Creating the grouping variable to discriminate indigenous land from immigrant land:

Calculating IndVal index: