We will introduce the basic functionality of assignR using data bundled with the package. We’ll review how to access data for known-origin biological samples and environmental models, use these to fit and apply functions estimating the probability of sample origin across a study region, and summarize these results to answer research and conservation questions. We’ll also demonstrate a quality analysis tool useful in study design, method comparison, and uncertainty analysis.

Let’s do the same for a growing season precipitation H isoscape for North America. Notice this is a spatial raster (SpatRaster) with two layers, the mean prediction and a standard error of the prediction. The layers are from waterisotopes.org, and their resolution has been reduced to speed up processing in these examples. Full-resolution isoscapes of several different types can be downloaded using the getIsoscapes function (refer to the help page for details).

The package includes a database of H and O isotope data for known origin samples (knownOrig.rda), which consists of three features (sites, samples, and sources). Let’s load it and have a look. First we’ll get the names of the data fields available in the tables.

The sites feature is a spatial object that records the geographic location of all sites from which samples are available.

Load H isotope data for North American Loggerhead Shrike from the package database.

By default, the subOrigData function transforms all data to a common reference scale (defined by the standard materials and assigned, calibrated values for those; by default VSMOW-SLAP) using data from co-analysis of different laboratory standards (see Magozzi et al., 2021). The calibrations used are documented in the function’s return object.

Transformation is important when blending data from different labs or papers because different reference scales have been used to calibrate published data and these calibrations are not always comparable. In this case all the data come from one paper:

If we didn’t want to transform the data, and instead wished to use the reference scale from the original publication, we can specify that in our call to subOrigData. Keep in mind that any subsequent analyses using these data will be based on this calibration scale: for example, if you wish to assign samples of unknown origin, the values for those samples should be reported on the same scale.

For a real application you would want to explore the database to find measurements that are appropriate to your study system (same or similar taxon, geographic region, measurement approach, etc.) or collect and import known-origin data that are specific to your system.

Single-isoscape Analysis

We need to start by assessing how the environmental (precipitation) isoscape values correlate with the sample values. calRaster fits a linear model relating the precipitation isoscape values to sample values, and applies it to produce a calibrated, sample-type specific isoscape.

d2h_Ll = calRaster(known = Ll_d, isoscape = d2h_lrNA, mask = naMap)

## 
## 
## ---------------------------------------
##         ------------------------------------------
## rescale function uses linear regression model, 
##         the summary of this model is:
## -------------------------------------------
##         --------------------------------------
## 
## Call:
## lm(formula = tissue.iso ~ isoscape.iso[, 1], weights = tissue.iso.wt)
## 
## Residuals:
##      Min       1Q   Median       3Q      Max 
## -149.640  -16.474    0.743   19.946  107.625 
## 
## Coefficients:
##                   Estimate Std. Error t value Pr(>|t|)    
## (Intercept)        2.20058    1.81952   1.209    0.227    
## isoscape.iso[, 1]  1.12628    0.04019  28.026   <2e-16 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Residual standard error: 32.41 on 522 degrees of freedom
## Multiple R-squared:  0.6007, Adjusted R-squared:    0.6 
## F-statistic: 785.4 on 1 and 522 DF,  p-value: < 2.2e-16

## NULL

Let’s create some hypothetical samples to use in demonstrating how we can evaluate the probability that the samples originated from different parts of the isoscape. The isotope values are drawn from a random distribution with a standard deviation of 8 per mil, which is a pretty reasonable variance for conspecific residents at a single location. We’ll also add made-up values for the analytical uncertainty for each sample and a column recording the calibration scale used for our measurements. If you had real measured data for your study samples you would load them here, instead.

id = letters[1:5]
set.seed(123)
d2H = rnorm(5, -110, 8)
d2H.sd = runif(5, 1.5, 2.5)
d2H_cal = rep("UT_H_1", 5)
Ll_un = data.frame(id, d2H, d2H.sd, d2H_cal)
print(Ll_un)

##   id        d2H   d2H.sd d2H_cal
## 1  a -114.48381 2.456833  UT_H_1
## 2  b -111.84142 1.953334  UT_H_1
## 3  c  -97.53033 2.177571  UT_H_1
## 4  d -109.43593 2.072633  UT_H_1
## 5  e -108.96570 1.602925  UT_H_1

As discussed above, one issue that must be considered with any organic H or O isotope data is the reference scale used by the laboratory producing the data. The reference scale for your unknown samples should be the same as that for the known origin data used in calRaster. Remember that the scale for our known origin data d is OldEC.1_H_1. Let’s assume that our fake data were normalized to the UT_H_1 scale. The refTrans function allows us to convert between the two.

Ll_un = refTrans(Ll_un, ref_scale = "OldEC.1_H_1")
print(Ll_un)

## $data
##   id       d2H   d2H.sd     d2H_cal
## 1  a -127.8800 3.374984 OldEC.1_H_1
## 2  b -125.0795 2.970657 OldEC.1_H_1
## 3  c -109.6746 2.808334 OldEC.1_H_1
## 4  d -122.5874 3.016841 OldEC.1_H_1
## 5  e -121.7123 2.594943 OldEC.1_H_1
## 
## $chains
## $chains[[1]]
## [1] "UT_H_1"      "UT_H_2"      "US_H_5"      "US_H_6"      "VSMOW_H"    
## [6] "EC_H_9"      "EC_H_7"      "OldEC.1_H_1"
## 
## 
## attr(,"class")
## [1] "refTrans"

Notice that both the d2H values and the uncertainties have been updated to reflect the scale transformation.

Now we will produce posterior probability density maps for the unknown samples. For reference on the Bayesian inversion method see Wunder, 2010

Ll_prob = pdRaster(d2h_Ll, Ll_un)

## NULL

Cell values in these maps are small because each cell’s value represents the probability that this one cell, out of all of them on the map, is the actual origin of the sample. Together, all cell values on the map sum to ‘1’, reflecting the assumption that the sample originated somewhere in the study area. Let’s check this for sample ‘a’.

global(Ll_prob[[1]], 'sum', na.rm = TRUE)

##   sum
## a   1

Check out the help page for pdRaster for additional options, including the use of informative prior probabilities.

Multi-isoscape Analysis

We can also use multiple isoscapes to (potentially) add power to our analyses. We will start by calibrating a H isoscape for the monarch butterfly, Danaus plexippus.

Dp_d = subOrigData(taxon = "Danaus plexippus")

## 150 samples are found from 31 sites

## 150 samples from 31 sites in the transformed dataset

d2h_Dp = calRaster(Dp_d, d2h_lrNA)

## 
## 
## ---------------------------------------
##         ------------------------------------------
## rescale function uses linear regression model, 
##         the summary of this model is:
## -------------------------------------------
##         --------------------------------------
## 
## Call:
## lm(formula = tissue.iso ~ isoscape.iso[, 1], weights = tissue.iso.wt)
## 
## Weighted Residuals:
##     Min      1Q  Median      3Q     Max 
## -11.030  -3.301   0.118   3.086   9.086 
## 
## Coefficients:
##                    Estimate Std. Error t value Pr(>|t|)    
## (Intercept)       -51.98592    1.95937  -26.53   <2e-16 ***
## isoscape.iso[, 1]   0.74467    0.04055   18.37   <2e-16 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Residual standard error: 4.438 on 148 degrees of freedom
## Multiple R-squared:  0.695,  Adjusted R-squared:  0.693 
## F-statistic: 337.3 on 1 and 148 DF,  p-value: < 2.2e-16

## NULL

Our second isoscape represents ⁸⁷Sr/⁸⁶Sr values across our study region, the state of Michigan. It was published by Bataille and Bowen, 2012, obtained from waterisotopes.org, cropped and aggregated to coarser resolution, and a rough estimate of uncertainty added.

In this case, we do not have any known-origin tissue samples to work with. However, our isoscape was developed to approximate the bioavailable Sr pool, and Sr isotopes are not strongly fractionated in food webs. Thus, our analysis will assume that the isoscape provides a good representation of the expected Sr values for our study species without calibration.

Let’s look at the Sr isoscape and compare it with our butterfly H isoscape.

plot(sr_MI$weathered.mean)

crs(sr_MI, describe = TRUE)

##      name authority code area         extent
## 1 unknown      <NA> <NA> <NA> NA, NA, NA, NA

crs(d2h_Dp$isoscape.rescale, describe = TRUE)

##     name authority code  area             extent
## 1 WGS 84      EPSG 4326 World -180, 180, 90, -90

Notice that the we have two different spatial data objects, one for Sr and one for d2H, and that they have different extents and projections. In order to conduct a multi-isotope analysis, we’ll first combine these into a single object using the isoStack function. In addition to combining the objects, this function resolves differences in their projection, resolution, and extent. It’s always a good idea to check that the properties of the isoStack components are consistent with your expectations.

Dp_multi = isoStack(d2h_Dp, sr_MI)
lapply(Dp_multi, crs, describe = TRUE)

## [[1]]
##     name authority code  area             extent
## 1 WGS 84      EPSG 4326 World -180, 180, 90, -90
## 
## [[2]]
##     name authority code  area             extent
## 1 WGS 84      EPSG 4326 World -180, 180, 90, -90

Now we’ll generate a couple of hypothetical unknown samples to use in our analysis. It is important that our isotopic markers appear here in the same order as in the isoStack object we created above.

Dp_unk = data.frame("ID" = c("A", "B"), "d2H" = c(-86, -96), "Sr" = c(0.7089, 0.7375))

We are ready to make our probability maps. First let’s see how our posterior probabilities would look if we only used the hydrogen isotope data.

Dp_pd_Honly = pdRaster(Dp_multi[[1]], Dp_unk[,-3])

## NULL

We see pretty clear distinctions between the two samples, driven by a strong SW-NE gradient in the tissue isoscape H values across the state.

What if we add the Sr information to the analysis? The syntax for running pdRaster is the same, but now we provide our isoStack object in place of the single isoscape. The function will use the spatial covariance of the isoscape values to approximate the error covariance for the two (or more) markers and return posterior probabilities based on the multivariate normal probability density function evaluated at each grid cell.

Dp_pd_multi = pdRaster(Dp_multi, Dp_unk)

## NULL

Note that the addition of Sr data greatly strengthens the geographic constraints on our hypothetical unknown samples: the difference between the highest and lowest posterior probabilities is much larger than with H only, and the pattern of high probabilities reflects the regionalization characteristic of the Sr isoscape. This is especially true for sample B, which has a fairly distinctive, high ⁸⁷Sr/⁸⁶Sr value.

Post-hoc Analysis

Odds Ratio

Many of the functions in assignR are designed to help you analyze and draw inference from the posterior probability surfaces we’ve created above. For the following examples we’ll return to our single-isoscape, Loggerhead shrike analysis, but the tools work identically for multi-isoscape results.

The oddsRatio tool compares the posterior probabilities for two different locations or regions. This might be useful in answering real-world questions…for example “is this sample more likely from France or Spain?”, or “how likely is this hypothesized location relative to other possibilities?”.

Let’s compare probabilities for two spatial areas - the states of Utah and New Mexico. First we’ll extract the state boundaries from package data and plot them.

s1 = states[states$STATE_ABBR == "UT",]
s2 = states[states$STATE_ABBR == "NM",]
plot(naMap)
plot(s1, col = c("red"), add = TRUE)
plot(s2, col = c("blue"), add = TRUE)

Now we can get the odds ratio for the two regions. The result reports the odds ratio for the regions (first relative to second) for each of the 5 unknown samples plus the ratio of the areas of the regions. If the isotope values (& prior) were completely uninformative the odds ratios would equal the ratio of areas.

s12 = rbind(s1, s2)
oddsRatio(Ll_prob, s12)

## $`P1/P2 odds ratio`
##          a        b        c        d        e
## 1 133.2879 102.0869 23.61812 80.53221 74.09955
## 
## $`Ratio of numbers of cells in two polygons`
## [1] 1

Here you can see that even though Utah is quite a bit smaller the isotopic evidence suggests it’s much more likely to be the origin of each sample. This result is consistent with what you might infer from a first-order comparison of the state map with the posterior probability maps, above.

Comparisons can also be made using points. Let’s create two points (one in each of the Plover regions) and compare their odds. This result also shows the odds ratio for each point relative to the most- and least-likely grid cells on the posterior probability map.

pp1 = c(-112,40)
pp2 = c(-105,33)
pp12 = vect(rbind(pp1,pp2))
crs(pp12) = crs(naMap)
oddsRatio(Ll_prob, pp12)

## $`P1/P2 odds ratio`
##          a        b        c        d        e
## 1 133.2879 102.0869 23.61812 80.53221 74.09955
## 
## $`Odds relative to the max/min pixel`
##    ratioToMax.a ratioToMax.b ratioToMax.c ratioToMax.d ratioToMax.e
## P1  0.143349129  0.190848034    0.6216964  0.243880063   0.26513164
## P2  0.001077453  0.001874017    0.0265338  0.003037688   0.00358998
##    ratioToMin.a ratioToMin.b ratioToMin.c ratioToMin.d ratioToMin.e
## P1    746070024    359728180    6506160.4    187955684    149643448
## P2      5607676      3532323     277680.8      2341113      2026227

The odds of the first point being the location of origin are pretty high for each sample, and much higher than for the second point.

Distance and Direction

A common goal in movement research is to characterize the distance or direction of movement for individuals. The wDist tool and it’s helper methods are designed to leverage the information in the posterior probability surfaces for this purpose.

The analyses conducted in assignR cannot determine a single unique location of origin for a given sample, but the do give the probability that each location on the map is the location of origin. If we know the collection location for a sample, we can calculate the distance and direction between each possible location of origin and the collection site, and weighting these by their posterior probability generate a distribution (and statistics for that distribution) describing the distance and direction of travel.

Let’s do a weighted distance analysis for our first two unknown origin loggerhead shrike samples. Since these are pretend samples, we’ll pretend that the two point locations we defined above for the oddsRatio analysis are the locations at which these samples were collected. Here are those locations plotted with the corresponding posterior probability maps.

# View the data
plot(Ll_prob[[1]], main = names(Ll_prob)[1])
points(pp12[1], cex = 2)

plot(Ll_prob[[2]], main = names(Ll_prob)[2])
points(pp12[2], cex = 2)

Now let’s run the analysis and use the functions c and plot to view the summary statistics and distributions returned by wDist.

wd = wDist(Ll_prob[[1:2]], pp12)
c(wd)[c(1,2,4,6,8,10,12,14,16)] #only showing select columns for formatting!

##   Sample_ID wMeanDist w10Dist w50Dist w90Dist wMeanBear  w10Bear   w50Bear
## 1         a   2876759 1290783 3150518 4138590 -169.8939 115.5482 -159.2225
## 2         b   3518594 2008390 3618131 4953292  178.3471 113.3519  177.4966
##     w90Bear
## 1 -105.5870
## 2 -118.4943

plot(wd)

## NULL

Comparing these statistics and plots with the data shows how the wDist metrics nicely summarize the direction and distance of movement. Both individuals almost certainly moved south from their location of origin to the collection location. Individual a’s migration may have been a little bit shorter than b’s, and in a more southwesterly direction, patterns that are dominated more by the difference in collection locations than the probability surfaces for location of origin. Also notice the multi-modal distance distribution for individual a…these can be common in wDist summaries so it’s a good ideal to look at the distributions themselves before choosing and interpreting summary statistics.

Assignment

Researchers often want to classify their study area in to regions that are and are not likely to be the origin of the sample (effectively ‘assigning’ the sample to a part of the area). This requires choosing a subjective threshold to define how much of the study domain is represented in the assignment region. qtlRaster offers two choices.

Let’s extract 10% of the study area, giving maps that show the 10% of grid cells with the highest posterior probability for each sample.

qtlRaster(Ll_prob, threshold = 0.1)

## class       : SpatRaster 
## dimensions  : 13, 28, 5  (nrow, ncol, nlyr)
## resolution  : 3.999998, 3.999998  (x, y)
## extent      : -164.6667, -52.66672, 20.99996, 72.99993  (xmin, xmax, ymin, ymax)
## coord. ref. : lon/lat WGS 84 (EPSG:4326) 
## source(s)   : memory
## names       :     a,     b,     c,     d,     e 
## min values  : FALSE, FALSE, FALSE, FALSE, FALSE 
## max values  :  TRUE,  TRUE,  TRUE,  TRUE,  TRUE

Now we’ll instead extract 80% of the posterior probability density, giving maps that show the smallest region within which there is an 80% chance each sample originated.

qtlRaster(Ll_prob, threshold = 0.8, thresholdType = "prob")

## class       : SpatRaster 
## dimensions  : 13, 28, 5  (nrow, ncol, nlyr)
## resolution  : 3.999998, 3.999998  (x, y)
## extent      : -164.6667, -52.66672, 20.99996, 72.99993  (xmin, xmax, ymin, ymax)
## coord. ref. : lon/lat WGS 84 (EPSG:4326) 
## source(s)   : memory
## names       :     a,     b,     c,     d,     e 
## min values  : FALSE, FALSE, FALSE, FALSE, FALSE 
## max values  :  TRUE,  TRUE,  TRUE,  TRUE,  TRUE

Comparing the two results, the probability-based assignment regions are broader. This suggests that we’ll need to assign to more than 10% of the study area if we want to correctly assign 80% or more of our samples. We’ll revisit this below and see how we can chose thresholds that are as specific as possible while achieving a desired level of assignment ‘quality’.

Summarization

Most studies involve multiple unknown samples, and often it is desirable to summarize the results from these individuals. jointP and unionP offer two options for summarizing posterior probabilities from multiple samples.

jointP calculates the probability that all samples came from each grid cell in the analysis area. Note that this summarization will only be useful if all samples are truly derived from a single population of common geographic origin.

jointP(Ll_prob)

## class       : SpatRaster 
## dimensions  : 13, 28, 1  (nrow, ncol, nlyr)
## resolution  : 3.999998, 3.999998  (x, y)
## extent      : -164.6667, -52.66672, 20.99996, 72.99993  (xmin, xmax, ymin, ymax)
## coord. ref. : lon/lat WGS 84 (EPSG:4326) 
## source(s)   : memory
## name        : Joint_Probability 
## min value   :      8.142348e-46 
## max value   :      2.647340e-02

unionP calculates the probability that any sample came from each grid cell in the analysis area. In this case we’ll save the output to a variable for later use.

Ll_up = unionP(Ll_prob)

The results from unionP highlight a broader region, as you might expect.

Any of the other post-hoc analysis tools can be applied to the summarized results. Here we’ll use qtlRaster to identify the 10% of the study area that is most likely to be the origin of one or more samples.

qtlRaster(Ll_up, threshold = 0.1)

## class       : SpatRaster 
## dimensions  : 13, 28, 1  (nrow, ncol, nlyr)
## resolution  : 3.999998, 3.999998  (x, y)
## extent      : -164.6667, -52.66672, 20.99996, 72.99993  (xmin, xmax, ymin, ymax)
## coord. ref. : lon/lat WGS 84 (EPSG:4326) 
## source(s)   : memory
## name        :  lyr1 
## min value   : FALSE 
## max value   :  TRUE

Quality Analysis

How good are the geographic assignments? What area or probability threshold should be used? Is it better to use isoscape A or B for my analysis? The QA function is designed to help answer these questions.

QA uses known-origin data to test the quality of isotope-based assignments and returns a set of metrics from this test. The default method conducts a split-sample test, iteratively splitting the dataset and using part to calibrate the isoscape(s) and the rest to evaluate assignment quality. The option recal = FALSE allows QA to be run without the calRaster calibration step. This provides a less complete assessment of methodological error but allows evaluation of assignments to tissue isoscapes made outside of the QA function, for example those calibrated using a different known-origin dataset or made through spatial modeling of tissue data, directly.

We will run quality assessment on the Loggerhead shrike known-origin dataset and precipitation isoscape. These analyses take some time to run, depending on the number of stations and iterations used.

qa1 = QA(Ll_d, d2h_lrNA, valiStation = 8, valiTime = 4, by = 5, mask = naMap, name = "normal")

##

We can plot the result using plot.

plot(qa1)

The first three panels show three metrics, granularity (higher is better), bias (closer to 1:1 is better), and sensitivity (higher is better). The second plot shows the posterior probabilities at the known locations of origin relative to random (=1, higher is better). More information is provided in Ma et al., 2020.

A researcher might refer to the sensitivity plot, for example, to assess what qtlRaster area threshold would be required to obtain 90% correct assignments in their study system. Here it’s somewhere between 0.25 and 0.3.

How would using a different isoscape or different known origin dataset affect the analysis? Multiple QA objects can be compared to make these types of assessments.

Let’s modify our isoscape to add some random noise.

dv = values(d2h_lrNA[[1]])
dv = dv + rnorm(length(dv), 0, 15)
d2h_fuzzy = setValues(d2h_lrNA[[1]], dv)
plot(d2h_fuzzy)

We’ll combine the fuzzy isoscape with the uncertainty layer from the original isoscape, then rerun QA using the new version. Obviously this is not something you’d do in real work, but as an example it allows us to ask the question “how would the quality of my assignments change if my isoscape predictions were of reduced quality?”.

d2h_fuzzy = c(d2h_fuzzy, d2h_lrNA[[2]])
qa2 = QA(Ll_d, d2h_fuzzy, valiStation = 8, valiTime = 4, by = 5, mask = naMap, name = "fuzzy")

##

Now we can plot to compare.

plot(qa1, qa2)

## NULL

Assignments made using the fuzzy isoscape are generally poorer than those made without fuzzing. Hopefully that’s not a surprise, but you might encounter cases where decisions about how to design your project or conduct your data analysis do have previously unknown or unexpected consequences. These types of comparisons can help reveal them!

Questions or comments? gabe.bowen@utah.edu

assignR Examples

Gabe Bowen, Chao Ma

September 01, 2023