Is PRC ordinary PCA plotted over time?

No. It is a constrained, reference-oriented analysis related to redundancy analysis, with the temporal structure explicitly represented.

Must every variable move in the curve's direction?

No. Positive and negative weights indicate different relationships to the selected axis.

Does the first curve show every effect?

No. Additional axes or individual responses may contain important information not captured by the first axis.

AI Glossary / Statistics and probability / Advanced

Principal response curve

Definition

A principal response curve, or PRC, summarizes treatment-related changes in several response variables over time relative to a reference treatment or control. The method is a special form of redundancy analysis developed for repeated observations of ecological communities. It separates the common temporal pattern from treatment-associated departures and displays a leading multivariate response pattern as a curve. Variable or species weights help interpret how individual responses relate to that pattern. It is not the same as a principal curve fitted through a cloud of points.

What it means

Experiments can measure many outcomes repeatedly, making separate time plots difficult to interpret. PRC offers a compact view of the dominant treatment-related trajectory after accounting for temporal changes shared with the control. A curve departing from zero indicates a pattern relative to that reference in the fitted representation. The sign and magnitude must be interpreted together with variable weights, scaling and transformations. A single curve may summarize only part of the response, so its apparent simplicity should not obscure additional patterns or uncertainty.

How it works

Arrange a multivariate response matrix with treatment, time and experimental-unit information. Choose transformations and scaling appropriate to the response variables, then fit the specified constrained ordination model with time effects accounted for. Extract a treatment-by-time curve for a selected axis and corresponding response weights. Examine the proportion of relevant variation summarized and whether additional axes matter. Statistical tests and resampling must respect experimental randomization and repeated observations, rather than freely permuting dependent rows. Document the reference treatment and software conventions because score scaling and sign choices affect the plotted coordinates.

Three practical examples

Ecology: summarize how a biological community changes after a treatment compared with concurrent controls.
Environmental monitoring: display multivariate departures at different sampling times while retaining the relationship to individual measured variables.
Experimental research: explore coordinated changes in several outcomes under a repeated-observation design, provided the PRC model and inference scheme are appropriate to the study rather than selected only for an attractive plot.

Think of it this way

Think of a shared background melody with a treatment adding a different pattern over time. The curve summarizes that additional pattern, while the variable weights indicate which measurements move with it or against it. One dominant pattern can simplify listening, but it does not reproduce every detail of the full performance.

Worked example

For an original interpretation example, suppose a fitted single-axis PRC approximation on transformed responses is treatment-minus-control response = c_t times b_j. At one time let c_t = 0.5 and three variable weights be 2, -1 and 0. The fitted transformed differences are then 1, -0.5 and 0. These stipulated coefficients illustrate interpretation; they are not the output of a fitted real dataset. If the transformation is log(y+1), a difference of 1 corresponds to multiplying y+1 by exp(1), not multiplying raw y by exp(1). For a control value 9, the corresponding fitted treatment value is (9+1)exp(1)-1, approximately 26.1828. For a control value 19 and difference -0.5, it is 20exp(-0.5)-1, approximately 11.1306. Reversing both the curve sign and every weight sign leaves all products unchanged. These calculations illustrate why curve signs and back-transformations require careful explanation.

Illustrative case study

In a fictional case, Meadow Research measures several community responses in replicated control and treatment systems over time. Analysts use a PRC display to summarize a prominent departure and inspect the variable weights to identify which responses contribute. They also check additional axes and keep replicate information in the permutation design. The report labels the transformation and reference clearly, avoiding a claim that every response follows the displayed curve equally or that a visual separation alone establishes statistical significance.

Common misconceptions and mistakes

A PRC ordinate is not automatically a percentage change in every response. A near-zero weight means the variable is weakly represented by that axis, not necessarily unaffected by treatment in all respects. Sign reversal of a matched curve and weight vector does not reverse the substantive result. A single axis does not capture all treatment-related variation. Repeated measurements are not independent simply because they appear as separate rows. For log-transformed data, the exact transformation determines the back-transformation, and uncertainty and residual variation complicate interpreting a transformed mean as an arithmetic mean on the original scale.

Business use

PRC can help research teams communicate a complex longitudinal multivariate experiment, especially when a clear control trajectory is available. The business or research decision should remain connected to meaningful original outcomes, not only the ordination display. Report experimental replication, transformations, the amount of variation summarized and the inference procedure. Investigate important variables individually when the joint pattern suggests a practical concern, using an appropriate analysis plan. The method is a summary of a specified model and design, so it should not replace checks of measurement quality, treatment implementation or generalizability.

Check your understanding

A curve score changes sign and every corresponding variable weight also changes sign. Do the fitted products change?

Answer: No. Their products stay the same, so the represented pattern is unchanged.

Related terms

Redundancy analysis
constrained ordination
repeated measures
multivariate response
treatment-by-time interaction
species score

Learn next

Learn how constrained ordination differs from unconstrained PCA, then study score scaling, transformations and permutation schemes that respect repeated experimental observations.

Sources

Van den Brink and ter Braak, Principal response curves: Analysis of time-dependent multivariate responses of biological community to stress: https://onlinelibrary.wiley.com/doi/abs/10.1002/etc.5620180207 Vegan developers, Principal Response Curves documentation: https://vegandevs.github.io/vegan/reference/prc.html

Previous: Principal curve Next: Principal stratification