Questions about calibration and best practices for isoriX

[Christophe Duplais]

Dear Alexandre,

 

We are working on assigning site of origin for corn earworm moths (migratory agricultural pest) using δ²H in the wing to understand overwintering in northern U.S. and would greatly appreciate your advice on calibration strategy within isoriX which is an amazing tool!

 

Our dataset includes 78 field-collected adults from five U.S. states over two years (2023, 2024). For calibration, we reared larvae in eight states (54 adults total, with several replicates per site per year 2023 and 2024) and measured δ²H in wing at the adult stage to match field collection for the overwintering study.

 

We are encountering substantial heterogeneity in calibration replicates (see ppt). At some sites the within-site variance is small, whereas at others it is large. When we remove apparent outliers (named local, with > 1 or 2 SD from site-level means), the regression slope and intercept shift markedly. For example, in 2023 with all calibration data points the slope was  = 0.52 ± 0.14, but after removing outliers the slope was  = 0.62 ± 0.13. Another example, in 2024, local calibration yields slope = 0.32 ± 0.12, whereas in 2023 the slope was  = 0.52 ± 0.14. We also generated isoscapes using combined vs. year-specific calibrations and by switching calibration years for assignments (see summary slides Calibration combination), and the geographic assignments vary, sometimes substantially, among versions.

 

We would value your perspective on several points:

 

· How should we interpret large within-site variance in reared individuals? Do you view this primarily as biological or analytical noise, or rather as evidence that differences in variance among sites should be modeled explicitly (e.g., using a mixed or hierarchical framework) instead of removing those samples as outliers?

 

· What is best practice regarding apparent outliers in calibration datasets of this size? Our exploratory filtering (1–2 SD) did not improve assignment resolution or confidence and can change slopes of calibration curves.

 

· Which is preferable: (i) fitting year-specific calibration models and propagating year as a grouping factor, or (ii) pooling years and modelling year as a random effect within a single calibration? Initially, we had hoped to use one calibration curve across years, but the year-specific slopes and intercepts differ enough to meaningfully alter the resulting isoscapes.

 

· More generally, what level of variability in the slopes and intercepts would you consider acceptable for robust assignment under isoriX, given our sample sizes? (We will double the number samples, and we can also add more calibration samples as our fridge is full of moths!) 

 

We would greatly appreciate your feedback as we explore this beautiful dataset, to ensure that we are doing things properly.

Thank so much for your help!

Kind regards,

Christophe

Christophe Duplais (he/him/his)

Associate Professor

Barton Laboratory

15 Castle Creek Dr.

Geneva, NY 14456

 

Department of Entomology website

Cornell AgriTech website

Cornell University                                               

 

‘Diversity is a fact. Equity is a choice. Inclusion is an action. Belonging is an outcome.’

Arthur Chan

[Alexandre Courtiol]

Dear Christophe,

Thanks a lot for your interest in IsoriX.

My first question is how did you perform the calibration? Did you use IsoriX for that and which calibration method did you select?

That matters because certain methods in IsoriX do account for the uncertainty associated with δ²H in the environment.

If you did use the default method and did the calibration in IsoriX, then I would certainly not filter out any data since "outliers" may not be truly outliers but calibration points where there is a lot of uncertainty in the underlying isoscape.

Moreover, the computation of variance components used in the assignment do consider that the distributions used for calibration are not truncated.

Ideally, there would be enough observations to buffer possible measurement errors but those errors tend to be small anyhow.

About the difference between years: do they correspond to year-specific isoscapes or do you use the same isoscape?

Depending on that, it could be that you are forcing one isoscape although there was a difference in the true isoscapes across years.

In summary, I may be wrong (there are a lot of factors at play), but I think that the "issues" stem from the isoscape.

So could you please let me know how you built the isoscape? Did you apply a weighing by precipitation or not, did you combine several years or not?

Best,

Alex

--
You received this message because you are subscribed to the Google Groups "IsoriX" group.
To unsubscribe from this group and stop receiving emails from it, send an email to isorix+un...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/isorix/9b054af1-cc79-4c3e-80a9-531393b779een%40googlegroups.com.

--

Alexandre Courtiol, www.datazoogang.de

[buk...@udel.edu]

Hello,

My name is Bukola Molake. I am the master's student working on the isoscape portion of this project. I am certainly okay with moving this discussion to the mailing list if Dr. Duplais and my PI, Dr. Crossley, are as well. 

Yes, I used IsoriX to calibrate and the default "wild" method. We did not take measurements of the isotopic composition of the environment, only the raised moths. 

 We typically had 3 known individuals per location for our calibration samples. Would this be more limiting, or would the spread (locations where moths were raised) have a greater effect on buffering errors?

Precalibrations, the isoscape is the same and not year or season-specific. I created multiple isoscapes, testing to see if the resolution or result would change with calibration samples. Initially, each year of unknown samples used the corresponding year of known samples to calibrate (original). We used an aggregate of all known samples to calibrate (combination), and we also tested using different years of known samples to calibrate (switch). I was unsure if the variation in results meant that our data was compromised. It seems that the variation and outliers are not an issue, but rather a lack of samples? These examples are in the attached presentation of isoscapes. Yes, we used precipitation averages.

Thanks,

Bukola

[buk...@udel.edu]

I realized I did not attach the presentation examples.

Bukola