Temporal Trend Analysis
Long-running water quality monitoring programs accumulate years of results across multiple sampling areas. This vignette demonstrates a reproducible pipeline for assessing whether an analyte is trending over time, and whether that trend differs by monitoring area.
All program names and site identifiers have been replaced with generic labels; synthetic data is used so the vignette runs standalone.
What it does
- Multi-year time-series overview figures across analytes and monitoring areas
- Two-way ANOVA (
result ~ year * area) with iterative studentized-residual outlier removal - Assumption checks: Breusch-Pagan heteroscedasticity test, Shapiro-Wilk normality test
- Interaction significance testing that decides between the additive model and
emmeanspost-hoc pairwise comparisons - Slope extraction and back-transformation to a %/year rate of change, since a significant year effect only says a trend exists, not how big it is
- Mann-Kendall trend test and Sen’s slope estimator, run per monitoring area, as an assumption-light cross-check on both the direction and magnitude of the ANOVA trend
- Significance annotation on post-hoc comparison figures, driven by the actual model p-values
Design
The pipeline is organized into modular R functions (trend_model.R, mann_kendall.R, plot_trend.R, utils.R) and documented via an R Markdown vignette, including a flowchart figure of the model’s decision path (outlier removal → assumptions → interaction test → additive/interaction branch). Four case-study analytes exercise different branches of the pipeline: a shared real trend, a flat null case, a trend confined to one area, and a flat case with injected high-leverage outliers.