A change point is an abrupt shift or a temporal pattern switch (e.g., beginning of a trend) in the time series. In FFA, change points indicate inhomogeneous periods, meaning a single model may not represent the entire record adequately. Instead, piecewise analysis should be applied to each homogeneous sub-period. The FFA framework provides two statistical tests for change point detection:

This vignette will demonstrate how these statistical tests can be used together to robustly identify change points in the data.

Case Study

This vignette will explore the Kootenai River at Porthill (08NH021) station, located on the border of British Columbia and Idaho. The Porthill station is also downstream of the Libby Dam, which finished construction in 1972. Data for this station is provided as Application_2.csv in the ffaframework package.

library(ffaframework)
#> Loading required package: ggplot2
#> Loading required package: patchwork

csv_path <- system.file("extdata", "Application_2.csv", package = "ffaframework")
df <- read.csv(csv_path, comment.char = "#")
df <- subset(df, !is.na(max)) # Remove missing values

head(df)
#>   year  max
#> 1 1928 2350
#> 2 1929 1680
#> 3 1930 1730
#> 4 1931 1470
#> 5 1932 2190
#> 6 1933 2640

plot_ams_data(df$max, df$year, title = "Kootenai River at Porthill (08NH021)")

The Pettitt Test

This rank-based test detects a single abrupt change in the median of a time series. The null hypothesis assumes no change point.

Use the eda_pettitt_test function to perform the test. It requires two arguments:

pettitt_test <- eda_pettitt_test(df$max, df$year)

print(pettitt_test$p_value)
#> [1] 0

print(pettitt_test$change_year)
#> [1] 1972

plot_pettitt_test(pettitt_test)

Conclusion: A p-value of <0.001 provides strong evidence of a change point in the year 1972.

The MKS Test

The Mann-Kendall-Sneyers (MKS) test identifies trend changes in the data.

Use eda_mks_test with the same arguments as above.

mks_test <- eda_mks_test(df$max, df$year)

print(mks_test$p_value)
#> [1] 0.01495225

print(mks_test$change_df$year)
#> [1] 1960 1985

plot_mks_test(mks_test)

Conclusion: At a p-value of 0.015, there is evidence of trend changes in 1960 and 1985.

Note: Since the MKS test can identify multiple change points, the reported p-value is determined using the most significant change point.

Interpreting and Selecting Change Points

In this example, the Pettitt and MKS tests both suggest structural changes in the time series.

Consider the following guidelines when choosing where to split the data:

  1. Incorporate domain knowledge and case-specific understanding. For example, if we knew that a water regulation structure was built in 1972 (the Libby dam, for this dataset), this would support the Pettitt test result.
  2. Avoid overpartitioning. The Pettitt and MKS tests operate independently and may detect multiple change points. To reduce sample size issues and uncertainty, retain only significant change points with physical justification.
  3. Prioritize based on p-value. Lower p-values indicate stronger evidence and should be given more weight.