Modelling Beyond a Straight Line
This project uses synthetic data, built to resemble monthly water-quality monitoring records.
Trend analysis in environmental monitoring often starts with a straight line or a Mann-Kendall test. Many records curve, plateau, break, or shift level. This vignette works through methods for those records, building from the regression framework up to semi-parametric models, with each method paired with a case study analyte that needs it. The focus is on describing what happened in the observed record.
Why it matters for a decision
Model choice changes what gets reported and what action follows. A plateau read as ongoing growth can lead to a treatment upgrade nobody needed. A step change read as a gradual trend can send an investigation toward the wrong cause. A flexible model fit to noise can find a trend that isn’t there, and a model that is too simple can miss a real change.
What it covers
Framework
- The regression framework: estimand, link, systematic component, and how ordinary least squares fits inside it
- Transformations: how to tell whether a series needs a log-response transform, a feature-engineered regressor, or a different link function
Methods, each with a case study
- Piecewise regression for a slope break (Total Nitrogen)
- Changepoint detection for a level shift at an unknown date (Specific Conductance)
- k-NN regression and LOESS to see the shape of a rise-then-plateau before choosing a model (Total Phosphorus)
- GAMs for step-like, non-monotonic change (Total Copper), including simultaneous intervals that show where the record is rising, falling, or flat, compared with a single-slope summary
- GAMMs for a seasonal cycle plus trend with autocorrelated residuals (Dissolved Oxygen)
Choosing a model
- Linear vs. nonlinear comparison on two control analytes, one with a straight-line trend (Total Chloride) and one with no trend (Total Zinc), to check whether added flexibility is justified
- Mechanism vs. shape: how the question, what is known about how the data arose, and whether the answer has to reach past the record decide which model belongs in a report. Worked through on Total Phosphorus, with a contaminant plume example where a process-based model and a forecast are justified
Design
The site has one method per chapter, each built on a case study. The home page includes a table that maps the question you’re trying to answer (exploratory, confirmatory, forensic, and so on) to the chapter to start from. The site uses a locked R package environment (renv) and an automated build, so every model, diagnostic, and figure can be regenerated from the code.