Trend Identification (Variability)

Many factors can produce nonstationarity in annual maximum series (AMS) data, including changes in climate, land use/cover, and water management. This vignette demonstrates how to use the ffaframework to check for evidence of nonstationarity in the variability of a time series.

List of Tests

Mean Trend Tests

Function	Purpose
eda_mk_test	Tests for a monotonic trend (Mann-Kendall test).
eda_bbmk_test	Tests for a monotonic trend under serial correlation (block-bootstrap MK test).

Stationarity Tests

Function	Purpose
eda_spearman_test	Tests for serial correlation (Spearman test).
eda_kpss_test	Tests for a stochastic trend (KPSS test).
eda_pp_test	Tests for a deterministic trend (Phillips-Perron test).

Variability Trend Tests

Function	Purpose
MW-MK Test	Tests for a trend in the variability (moving-window MK test)
eda_white_test	Tests for time-dependence in the variability (White test).

Trend Estimation (Mean and Variability)

Function	Purpose
eda_sens_trend	Estimates slope and intercept of a linear trend (Sen’s slope estimator).
eda_runs_test	Evaluates residuals’ structure under linear model assumptions (Wald-Wolfowitz runs test).

Case Study

This vignette will explore data from the Mission Creek near East Kelowna (08NM116) hydrological monitoring station. The remoteness of this station means that trends annual maxima are caused by changes in climate as opposed to changes in land use or cover. Data for this station is provided as Application_5.csv in the ffaframework package.

library(ffaframework)

csv_path <- system.file("extdata", "Application_5.csv", package = "ffaframework")
df <- read.csv(csv_path, comment.char = "#")
df <- subset(df, !is.na(max)) # Remove missing values

head(df)
#>   year  max
#> 1 1949 49.3
#> 2 1950 52.1
#> 3 1951 49.3
#> 4 1952 50.7
#> 5 1953 62.3
#> 6 1954 36.2

plot_ams_data(df$max, df$year, title = "Mission Creek near East Kelowna (08NM116)")

Assessing Trends in the Variance

Moving-Window Mann-Kendall (MWMK) Test

The MWMK test is used to detect trends in the variability of a time series. First, a moving-window algorithm is used to estimate the variability of the AMS data. Then, the Mann-Kendall test is applied to the series of standard deviations to check for a trend.

The ams_mw_variability estimates the moving-window standard deviations and the eda_mk_test function performs the Mann-Kendall test.

mw <- ams_mw_variability(df$max, df$year)
mwmk_test <- eda_mk_test(mw$std)
print(mwmk_test$p_value)
#> [1] 0.01600616

Conclusion: At a p-value of 0.016, we reject the null hypothesis. There is evidence of a linear trend in the variability in the data.

White Test

The White test checks for heteroskedasticity, or general time-dependence in the variability. The null hypothesis is homoskedasticity, or constant variability in the data.

white_test <- eda_white_test(df$max, df$year)
print(white_test$p_value)
#> [1] 0.00125362

Conclusion: At a p-value of 0.0012, we reject the null hypothesis. There is statistical evidence of heteroskedasticity.

Sen’s Trend Estimator

While the previous tests provide evidence for a monotonic trend in the variability, they do not estimate the slope or intercept of this trend. We can estimate the monotonic trend using Sen’s trend estimator, which uses a nonparametric approach that is robust to outliers.

eda_sens_trend() takes two arguments: the data (either an annual maximum series or vector of standard deviations) and the corresponding vector of years. Then, plot_sens_trend() can be used to generate a plot of the results. It requires the AMS data and corresponding vector of years. It also takes an optional trend in the mean (mean_trend), or a trend in the variability (variability_trend).

variability_trend <- eda_sens_trend(mw$std, mw$year)

plot_sens_trend(df$max, df$year, variability_trend = variability_trend)

Note: The covariate is computed using the formula \((\text{years} - 1900) / 100\).

Runs Test

Sen’s trend estimator assumed that the nonstationarity is linear. The Wald-Wolfowitz runs test assess the feasibility of the linearity assumption by checking the residuals for randomness. If the residuals are random (the null hypothesis), there is evidence that the underlying trend is linear. The eda_runs_test() function takes the output of eda_sens_trend() as an argument.

runs_test <- eda_runs_test(variability_trend)

print(runs_test$p_value)
#> [1] 0.2658385

plot_runs_test(runs_test, title = "Runs Test (AMS Variability)")

Conclusion: At a p-value of \(0.266\), we fail to reject the null hypothesis. There is evidence that the residuals are random and a linear trend is suitable for the data.

Conclusion

The MWMK and White tests find evidence of nonstationarity in the variability of the AMS. The runs test confirms that a linear model is suitable for the nonstationarity. Flood frequency analysis of this dataset requires a time-dependent probability model.

Recommendation: Use NS-FFA.