library(tidyr)
library(dplyr)
library(ggplot2)
library(ggfortify)
## theme for ggplot
theme_set(theme_classic())
theme_update(text = element_text(size = 14))Data transformation for linear models
Day 7
Felix Nößler
Freie Universität Berlin @ Theoretical Ecology
January 23, 2024
Reproduce slides
1 When to transform data?
- The relationhip between the response and the predictor is non-linear/exponential \(\rightarrow\) log-transform the predictor variable
- You want an interpretable intercept \(\rightarrow\) center the predictor variables
- You want to compare the coefficients of different predictors \(\rightarrow\) standardize (center and divide by the standard deviation) the predictor variables
- You have quantities with several orders of magnitude \(\rightarrow\) log-transform the variables to be able to compare them
2 When to better not use transformations?
- often it’s better to use a generalized linear model (GLM) instead of a transformation of the response variable, especially if you want to test a hypothesis
see, for example: R. B. O’Hara and D. J. Kotze, “Do not log‐transform count data,” Methods in Ecology and Evolution, 2010. doi: 10.1111/j.2041-210x.2010.00021.x.
3 Transformation for visualisation
4 Workflow
flowchart LR A[fit model\nto raw\ndata] --> B[check\nassumptions] B -->|if not\nfulfilled| T[transform\ndata] T --> J J[fit model to\ntransformed\ndata] --> K[check\nassumptions] K --> P B -->|if fulfilled| P P[make\npredictions] --> S[plot raw\ndata and\npredictions] style T fill:#f80
Transformation for linear models