Skip to contents

Some calibrations depend on a second variable, such as the concentration of an ion. This vignette shows how to include it as a second predictor:

D47 = alpha + beta × Temperature + gamma × Ion

It builds on vignette("bayesian", package = "bayclumpr").

Preparing the data

The calibration dataset needs the usual columns plus Ion. If Ion is measured with error, we also need its uncertainty (1 SD). We simulate 100 observations where Ion raises D47 by 0.005 per unit. We then add measurement error to Ion with an SD of 0.1.

set.seed(3)
ds <- cal.dataset(error = "S1", nobs = 100)

trueIon <- rnorm(nrow(ds))
ds$D47 <- ds$D47 + 0.005 * trueIon
ds$IonError <- 0.1
ds$Ion <- trueIon + rnorm(nrow(ds), 0, ds$IonError)

Fitting the calibration

cal.ion.bayesian() accounts for errors in Temperature, D47, and Ion. IonError can be a single value for all samples or one value per sample.

fit <- cal.ion.bayesian(calibrationData = ds, IonError = ds$IonError)
fit
#> Bayesian clumped isotope calibration
#> Priors: Flat 
#> 
#> BLM1_fit
#>  parameter estimate        SE   median    lower    upper
#>      alpha 0.272221 0.0101978 0.272289 0.252790 0.292457
#>       beta 0.036662 0.0008437 0.036654 0.035025 0.038262
#>      gamma 0.005016 0.0012738 0.005045 0.002435 0.007419
#>      sigma 0.013616 0.0010132 0.013565 0.011776 0.015795
#> 
#> Convergence
#>     model max_Rhat min_ESS divergences
#>  BLM1_fit   0.9997    2474           0

gamma is the slope for Ion. Its posterior should be close to the simulated value of 0.005.

When Ion is known without error, we leave out IonError:

fitKnown <- cal.ion.bayesian(calibrationData = ds)
fitKnown
#> Bayesian clumped isotope calibration
#> Priors: Flat 
#> 
#> BLM1_fit
#>  parameter estimate        SE   median    lower    upper
#>      alpha 0.272166 0.0097646 0.272049 0.253372 0.291367
#>       beta 0.036670 0.0008032 0.036674 0.035068 0.038200
#>      gamma 0.005044 0.0012516 0.005059 0.002507 0.007482
#>      sigma 0.013589 0.0010361 0.013542 0.011749 0.015711
#> 
#> Convergence
#>     model max_Rhat min_ESS divergences
#>  BLM1_fit   0.9997    2634           0

Reconstructing temperatures

The reconstruction dataset needs Ion for each sample, and IonError when we account for its uncertainty:

recData <- data.frame(
  Sample = paste("Sample", 1:3),
  D47 = c(0.6, 0.7, 0.8),
  D47error = 0.01,
  Ion = c(-1, 0, 1),
  IonError = 0.1
)

With useIonError = TRUE, the error in Ion is included in the reconstructed temperatures:

rec.ion.bayesian(calModel = fit, recData = recData, useIonError = TRUE)
#>     Sample D47 D47error  meanTemp SD       SE   median      lower      upper
#> 1 Sample 1 0.6     0.01 59.156851 NA 8.567143 58.98611  43.048420 76.4799251
#> 2 Sample 2 0.7     0.01 19.920086 NA 5.793169 19.65157   9.238422 32.0191140
#> 3 Sample 3 0.8     0.01 -8.176717 NA 4.342052 -8.29918 -16.420039  0.6563953

With useIonError = FALSE, Ion is treated as known:

rec.ion.bayesian(calModel = fit, recData = recData, useIonError = FALSE)
#>     Sample D47 D47error  meanTemp SD       SE    median     lower     upper
#> 1 Sample 1 0.6     0.01 59.118286 NA 8.613112 58.910189  43.18238 77.029470
#> 2 Sample 2 0.7     0.01 19.722917 NA 5.792888 19.650943   8.92384 31.400810
#> 3 Sample 3 0.8     0.01 -8.261464 NA 4.338728 -8.299877 -16.58234  0.606063

The columns are the same as in rec.bayesian(). meanTemp is the posterior mean temperature in °C, and SE is its standard error. SD is NA here because the dataset has neither N nor D47sd. Add one of them to get the uncertainty of a single replicate. median, lower, and upper describe the 95% posterior interval, based on SE. The uncertainty includes the calibration, the scatter around the calibration line, and the error in D47.