Using the right tool for the job: the difference between unsupervised and supervised analyses of multivariate...

Area of Science:

• Ecology
• Statistical analysis
• Data science

Background:

• Ecologists analyze data to identify variables associated with causes or consequences.
• Unsupervised methods (e.g., Principal Components Analysis, PCA) summarize data variation without considering a response variable.
• Supervised methods (e.g., Partial Least Squares, PLS) identify variable combinations that best explain causal relationships.

Purpose of the Study:

• To compare unsupervised (PCA) and supervised (PLS) analytical techniques in ecology.
• To illustrate the differences in outcomes between PCA and PLS using published and simulated datasets.
• To highlight situations where supervised methods are more appropriate than unsupervised methods for ecological data.

Main Methods:

• Analysis of a published ecological dataset using both PCA and PLS.
• Generation and analysis of simulated datasets to demonstrate differences between PCA and PLS.
• Comparison of the questions answered and insights gained from each method.

Main Results:

• PCA and PLS provide different answers, especially when key causal variables do not contribute significantly to overall data variation.
• For simulated data with many correlated, response-unrelated variables, PLS outperformed PCA in identifying relevant variables.
• Supervised methods like PLS are more effective than PCA in detecting associations with a response variable under specific data conditions.

Conclusions:

• PCA and PLS are not interchangeable and serve different analytical purposes in ecology.
• Supervised analyses (PLS) are crucial for accurately identifying ecologically relevant variables linked to specific outcomes.
• The overuse of PCA in ecology may stem from a lack of familiarity with more suitable supervised techniques like PLS.