Introduction to Multivariate Statistics and Data Matrices
Introduction to Multivariate Statistics in Biology
Biology is inherently multivariate because differences between habitats, communities, and organisms are usually complex and involve multiple interacting factors.
Habitats or sites differ from one another across many axes, including annual rainfall, mean temperature, and the abundance of various specific species. Measuring these multiple variables simultaneously creates a multivariate dataset.
Organisms differ in numerous ways, such as appearance, behavior, physiology, and their life cycle. These are all separate variables that contribute to biological variation.
Multivariate analysis can also apply to a single type of variable measured multiple times.
Example: The influence of rainfall on plant growth. Instead of just looking at total annual rainfall, researchers might measure monthly rainfall, creating 12 variables (one for each month).
Example: Tracking changes in an organism’s behavior, appearance, or physiology as it develops from an embryo to sexual maturity and beyond.
Example: Seasonal changes that introduce multiple axes of variation relevant to biological questions.
Genetics provides a major source of multivariate data. In studies with thousands of genes measured across hundreds of individuals, each gene represents a variable.
The core definition of multivariate data is having multiple individual observations or objects (denoted as ) and multiple variables (denoted as ) measured for each of those objects.
The Multivariate Data Matrix and Notation
Data is typically organized into a data matrix, often referred to as .
An data matrix represents objects and variables.
Multivariate status is established whenever .
Example Data Set (Leaf Traits):
Reference: Originally published by Richard Allen in 1999; appears in the Quinn and Keogh textbook.
Variables (): Five leaf characteristics (e.g., specific leaf area, leaf nitrogen content, etc.).
Objects (): Eleven different species of plants.
The matrix allows us to define the position of points in -dimensional space. For this dataset, the dimension is .
Dependent vs. Independent Variables: These multiple variables may be dependent, independent, or a mixture of both. Their identity is defined by the specific biological question regardless of the matrix structure.
Visualizing Space: While we live in a three-dimensional world, visualizing data becomes difficult beyond two dimensions. Plotting points on an -axis and a -axis represents an dimensional space.
Subscript Notation:
We use subscripts to identify specific elements: refers to the -th object’s value for the -th variable.
is the first object’s value for the first variable.
Example from the leaf trait data: To find , count down to the third row/object (the species Erythronium americanum) and across to the second variable (leaf nitrogen content). The value is , meaning Erythronium americanum has an average leaf nitrogen content of .
Summarizing Univariate Data
When considering just one variable (), all objects can be described using two parameters: location and spread.
1. The Mean (Centroid):
Tells us where the data is distributed along the axis of that variable.
Formula:
For the leaf area data example, the mean is .
2. The Variance:
Tells us about the spread of the data around the mean.
Formula:
We square the deviations from the mean so that we only care about the distance from the mean, not whether the value is above or below it.
The numerator of the variance equation is the Sum of Squares (), which is also used in ANOVA to determine variance or spread in hypothesis testing.
Summarizing Multivariate Data: Covariance and Shape
When , a third parameter is required to describe the data: Covariance.
The Multivariate Centroid: In two-dimensional space, the centroid is the point representing the mean of variable 1 and the mean of variable 2 simultaneously.
Defining Covariance:
It quantifies the association between variables by taking the product of the deviations of each observation from their respective means.
Formula:
If on average objects do not deviate in a consistent pattern (one goes up while the other goes up, or one goes up while the other goes down), the covariance will be .
Unlike variance, we care about the sign (positive or negative) because it indicates the direction of the association.
Sums of Cross Products ():
This is the numerator of the covariance equation: .
The is also the numerator used for estimating the regression coefficient (beta).
Variation vs. Association:
Variation: Quantified by Sums of Squares or Variance. It occurs when multiple objects measured for the same variable have different values.
Association: Quantified by Sums of Cross Products or Covariances. It occurs when objects share patterns of deviation across two or more variables.
Covariance vs. Regression: Multivariate description focuses on shared variation without assuming cause and effect. There is no fixed independent () or dependent () axis, and covariance is not viewed as "rise over run."
The Variance-Covariance Matrix
Summary matrices (Variance-Covariance or Sums of Squares and Cross Products) organize data parameters into a square format.
Matrix Dimensions and Properties:
The matrix is square (); the number of rows and columns equals the number of variables.
The Diagonals: Contain the variances for each individual variable.
The Off-Diagonals: Contain the covariances between different variables.
Symmetry: The matrix is symmetrical because the covariance between variable 1 and 2 is identical to the covariance between variable 2 and 1 ().
Data Reduction: A matrix can condense all information from an data matrix into just three unique pieces of information (two variances and one covariance) regardless of how large is.
Physical Interpretation of Parameters:
Mean: Position in multivariate space.
Variance: Size (length of the axes of the data cloud).
Covariance: Shape (e.g., whether the data cloud forms a wide circle or a narrow ellipse).
Example Matrix Values (Leaf Traits):
Variance (Specific Leaf Area):
Variance (Leaf Nitrogen):
Covariance: (placed on both off-diagonal positions).
The Trace of a Matrix
The Trace is the sum of the variances on the diagonal of the variance-covariance matrix.
It represents the total size (total variation) of the entire -dimensional data cloud.
Comparing Traces (Example Case Study):
Study of 11 species measured in the USA vs. Canada for photosynthetic traits (Area-based net photosynthetic capacity () and leaf diffuse conductance ()).
USA Results:
Trace:
Canada Results:
variance was larger ( compared to USA).
variance was smaller ().
Total Trace:
Biological Conclusion: The photosynthetic capacity variables had greater overall variance among the plant species in the USA than in Canada, even though specific individual variances varied differently.
Correlation and Standardization
Problem with Covariance: The magnitude of covariance depends on the original scale of the variables, making it hard to compare different pairs of traits.
The Correlation ():
Provides a bounded metric from to .
It is independent of the original scale and the variances of the individual variables.
Formula for Correlation:
This is interpreted as the association (Sum of Cross Products) scaled by the amount of variance in the variables (product of Sums of Squares).
The Correlation Matrix:
Like the covariance matrix, it is square () and symmetrical.
The diagonal values are always because every variable is perfectly correlated with itself.
Comparing Associations (Leaf Examples):
Is a covariance of stronger than a covariance of ?
Correlation between Specific Leaf Area and Leaf Nitrogen: (very strong).
Correlation between photosynthetic traits: (weaker).
The closer the value is to , the stronger the association. A value of indicates no association.
The Coefficient of Determination in Multivariate Context
The Coefficient of Determination () is the square of the correlation.
It is a scale-independent metric used to understand how much variation is shared between variables.
Interpretation: It represents the percentage of variance in one variable explained by or shared with the variance in another variable.
Example calculation: For Specific Leaf Area and Leaf Nitrogen (), the is .
This means of the variation in leaf area is shared with the variation in nitrogen concentration.
Predictive Value: When variables share a huge amount of information through strong correlations, you can accurately estimate the value of one variable for a new object if you only know the value of the other accurately correlated variable.