Getting started with toeclim

What is time of emergence?

Climate records combine a slowly evolving forced signal (for example greenhouse-gas driven warming) with internal variability (weather, ENSO, decadal oscillations). The time of emergence (ToE) is the moment the signal becomes large relative to that variability, so that a change can be told apart from natural fluctuations (Hawkins and Sutton, 2012).

toeclim implements a transparent, configurable version of this idea and attaches uncertainty to it:

1. estimate a smooth signal relative to a baseline mean;
2. estimate noise from a baseline period;
3. declare emergence when the signal-to-noise ratio exceeds k for at least persist consecutive years (optionally through to the end of the record);
4. bootstrap the residuals to see how much of the answer is just noise.

A single series

d <- toe_simulate(n_years = 150, slope = 0.03, onset = 1970,
                  sigma = 0.3, phi = 0.5, seed = 10)
head(d, 3)
#>   year forced   member_01
#> 1 1900      0 -0.03289004
#> 2 1901      0  0.09251191
#> 3 1902      0 -0.22392926

The forced response is flat until 1970 and then warms at 0.03 units per year, so the true signal is known.

toe_trend(d$member_01, d$year)
#> Sen's slope / Mann-Kendall trend test
#>   n = 150 
#>   Slope: 0.01612 per unit time,  95 % CI [ 0.01409, 0.01814 ]
#>   Mann-Kendall: S = 8584 , Z = 14.1 , p = <2e-16 
#>   Lag-1 autocorrelation (detrended): 0.803 -> trend-free prewhitening applied

fit <- toe_emerge(d$member_01, d$year, baseline = c(1900, 1950),
                  n_boot = 200, seed = 1)
fit
#> Time of emergence
#>   Signal: quadratic | threshold k = 2 | persistence = 5 (window)
#>   Baseline: 1900-1950 | noise (baseline) sigma = 0.281
#>   ToE: 1993 
#>   Bootstrap (ar1, n = 200): 90% interval [1983.95, 2001.00], 100% of replicates emerge

plot(fit)

The shaded orange band is the bootstrap interval for the emergence year. The grey band shows the baseline mean plus or minus k noise standard deviations.

How likely is emergence by a given date?

toe_prob(fit, times = c(1990, 2000, 2010, 2020))
#>   time prob
#> 1 1990 0.39
#> 2 2000 0.94
#> 3 2010 1.00
#> 4 2020 1.00

Sensitivity to subjective choices

The threshold k and the persistence requirement are choices, not physics. toe_sensitivity() shows how much they matter.

toe_sensitivity(d$member_01, d$year, baseline = c(1900, 1950),
                k = c(1, 2, 3), persist = c(1, 5, 10))
#>   k persist  toe
#> 1 1       1 1980
#> 2 2       1 1993
#> 3 3       1 2004
#> 4 1       5 1980
#> 5 2       5 1993
#> 6 3       5 2004
#> 7 1      10 1980
#> 8 2      10 1993
#> 9 3      10 2004

Ensembles

For large ensembles, apply the method to every member and inspect the spread. The bundled toe_example data set has ten synthetic members.

members <- toe_example[, grep("^member", names(toe_example))]
ens <- toe_ensemble(members, toe_example$year, baseline = c(1900, 1960))
ens
#> Ensemble time of emergence (10 members)
#>   Emerged: 10 / 10
#>   Median ToE: 1998.5  | range: 1988 - 2015  | 5-95%: 1989 - 2015 
#>   Pooled (ensemble-mean signal): 2001

plot(ens)

The pooled estimate uses the ensemble-mean signal, which has little noise of its own, against the typical single-realisation noise. It is usually earlier than most members: the ensemble mean reveals the forced response before any single realisation can.

Preparing real data

Monthly data can be converted to seasonal-cycle-free anomalies and then to annual means:

dates <- seq(as.Date("1950-01-01"), as.Date("2019-12-01"), by = "month")
mon <- as.integer(format(dates, "%m"))
yr  <- as.integer(format(dates, "%Y"))
x <- 12 + 8 * sin(2 * pi * mon / 12) + 0.02 * pmax(0, yr - 1980) +
  rnorm(length(dates), sd = 0.8)

anom   <- toe_anomaly(x, dates, baseline = c(1950, 1979), by = "month")
annual <- toe_annual(anom$anomaly, dates)
head(annual, 3)
#>   year     value  n
#> 1 1950 0.1252714 12
#> 2 1951 0.1423281 12
#> 3 1952 0.2366975 12

toe_emerge(annual$value, annual$year, baseline = c(1950, 1979))
#> Time of emergence
#>   Signal: quadratic | threshold k = 2 | persistence = 5 (window)
#>   Baseline: 1950-1979 | noise (baseline) sigma = 0.244
#>   ToE: 2007

Notes and limitations

• The signal is a statistical smoother, not a physical attribution. A "loess" signal will follow low-frequency internal variability if the span is small.
• Noise from a short baseline is uncertain; toe_noise() reports the effective sample size.
• Persistence is counted in observations, so use regularly spaced data.
• Bootstrap intervals describe sampling uncertainty given the chosen definition, not structural uncertainty about the definition itself.

References

Hawkins, E. and Sutton, R. (2012). Time of emergence of climate signals. Geophysical Research Letters, 39, L01702. doi:10.1029/2011GL050087

Yue, S., Pilon, P., Phinney, B. and Cavadias, G. (2002). The influence of autocorrelation on the ability to detect trend in hydrological series. Hydrological Processes, 16, 1807-1829. doi:10.1002/hyp.1095