Trophic adjustments in regional assessments
Introduction
A key assumption behind the regional assessments of contaminants in biota is that the assessment criteria are comparable across trophic levels, in the sense that the ratios of the fitted concentration to the assessment criteria are similar for e.g. fish and shellfish. This report examines this assumption for mercury concentrations relative to the QSsp, PCB concentrations relative to the EAC and PBDE concentrations relative to the FEQG. The report illustrates the consequences for regional assessments of status when the assumption is not met and investigates alternative regional models that account for trophic differences in concentration.
Mercury
Summary plots
The following plot shows the ratio of the mean mercury concentration in the final monitoring year to the QSsp for each fish and shellfish timeseries. The points are coloured by tissue:
- muscle = magenta
- liver = red
- soft body = blue
- tail muscle (one timeseries) = cyan
so shades of red indicate fish and shades of blue indicate shellfish. There is clearly a difference between fish and shellfish, with shellfish samples being more likely to have concentrations below the QSsp.
An alternative way of presenting this is to plot the concentration ratios against the trophic level of each species (obtained from FishBase, although Crangon crangon and Ruditapes philippinarum were accidentally omitted). Again, the difference between fish and shellfish is marked, but there is no obvious trophic effect across fish species.
The current regional assessment assumes that the mean concentration ratio (in any given region) is the same across species. However, the residual plot below refutes this assumption. In the plot the species are ordered by trophic level, with ties ordered alphabetically. Again, there is no evidence of a trophic effect across fish species.
Incorporating species effects
The lack of trophic adjustment matters because, at present, the status in each region is estimated conditional on the species monitored in that region. A region is therefore more likely to have acceptable environmental status if monitoring revolves around shellfish.
What can be done about this? A quick solution is to add a species random effect to the regional model. This gives a model of the form
- fixed = region
- random = station + species + station.series
The model has no explicit adjustment for trophic effects, but the regional effects are now standardised on a species that is typical of monitoring activity across the whole OSPAR area. The regional effects are now consistent, even though we don’t really know what species (or trophic level) they are adjusted to.
Here is the impact on the regional assessment. The first tab shows the current approach. The second tab shows the effects when there is a term accounting for the species effect. Regions such as the Greenland-Scotland ridge and the Irish and Scottish West Coast, where there is a lot of shellfish monitoring, now have a status that is ‘redder’ than before. Note that a side-effect of adding in a species random effect is that the standard errors on the regional effects are also larger (correctly so, although we might be able to do better if we consider conditional rather than marginal predictions).
Unadjusted
Adjusted to average species
Trophic models
Now let’s see whether we can exploit the trophic level information. We also investigate whether there are any differences between muscle and liver in fish. First, we fit a model with both a trophic_level and a tissue effect:
- fixed = region + tissue + trophic_level
- random = station + species + station.series
where tissue differentiates liver from muscle samples. Here is the model summary
Multivariate Meta-Analysis Model (k = 449; method: REML)
Variance Components:
estim sqrt nlvls fixed factor
sigma^2.1 0.1831 0.4278 380 no station
sigma^2.2 0.2764 0.5258 13 no species
sigma^2.3 0.0279 0.1671 449 no station_series
Test for Residual Heterogeneity:
QE(df = 434) = 9817.7893, p-val < .0001
Test of Moderators (coefficients 1:15):
QM(df = 15) = 173.6660, p-val < .0001
Model Results:
estimate se zval pval ci.lb ci.ub
regionIberian Sea -1.0090 0.6482 -1.5567 0.1195 -2.2794 0.2614
regionNorthern Bay of Biscay -0.9838 0.6413 -1.5340 0.1250 -2.2408 0.2732
regionCeltic Sea -1.4576 0.6420 -2.2705 0.0232 -2.7158 -0.1993 *
regionIrish Sea -1.1169 0.6425 -1.7385 0.0821 -2.3761 0.1423 .
regionIrish and Scottish West Coast -1.5326 0.6427 -2.3846 0.0171 -2.7924 -0.2729 *
regionChannel -1.1061 0.6431 -1.7199 0.0855 -2.3666 0.1544 .
regionSouthern North Sea -0.9273 0.6422 -1.4440 0.1487 -2.1858 0.3313
regionSkagerrak and Kattegat -1.2260 0.6473 -1.8941 0.0582 -2.4947 0.0427 .
regionNorthern North Sea -0.7229 0.6435 -1.1233 0.2613 -1.9841 0.5384
regionNorwegian Trench -1.0169 0.6530 -1.5571 0.1194 -2.2968 0.2631
regionNorwegian Sea -1.4909 0.6769 -2.2025 0.0276 -2.8176 -0.1642 *
regionGreenland-Scotland ridge -2.1046 0.6582 -3.1972 0.0014 -3.3947 -0.8144 **
regionBarents Sea -1.9155 0.6640 -2.8849 0.0039 -3.2170 -0.6141 **
tissueLI 0.0540 0.0474 1.1409 0.2539 -0.0388 0.1469
trophic_level 0.6668 0.2105 3.1678 0.0015 0.2542 1.0793 **
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1The parameter estimates show that there is
- no significant difference between fish liver and fish muscle concentration ratios (tissueLI: p = 0.2539)
- a significant trophic level effect (trophic_level: p = 0.0015)
Let’s refit the model without the (non-significant) tissue effect to get a more parsimonious model:
- fixed = region + trophic_level
- random = station + species + station.series
Multivariate Meta-Analysis Model (k = 449; method: REML)
Variance Components:
estim sqrt nlvls fixed factor
sigma^2.1 0.1829 0.4277 380 no station
sigma^2.2 0.2826 0.5316 13 no species
sigma^2.3 0.0276 0.1663 449 no station_series
Test for Residual Heterogeneity:
QE(df = 435) = 9938.7229, p-val < .0001
Test of Moderators (coefficients 1:14):
QM(df = 14) = 171.8536, p-val < .0001
Model Results:
estimate se zval pval ci.lb ci.ub
regionIberian Sea -1.0128 0.6542 -1.5483 0.1216 -2.2950 0.2693
regionNorthern Bay of Biscay -0.9872 0.6474 -1.5248 0.1273 -2.2562 0.2817
regionCeltic Sea -1.4605 0.6480 -2.2538 0.0242 -2.7307 -0.1904 *
regionIrish Sea -1.1191 0.6485 -1.7257 0.0844 -2.3902 0.1519 .
regionIrish and Scottish West Coast -1.5362 0.6488 -2.3677 0.0179 -2.8078 -0.2646 *
regionChannel -1.1089 0.6492 -1.7082 0.0876 -2.3813 0.1634 .
regionSouthern North Sea -0.9357 0.6482 -1.4436 0.1488 -2.2061 0.3347
regionSkagerrak and Kattegat -1.2305 0.6533 -1.8834 0.0596 -2.5109 0.0500 .
regionNorthern North Sea -0.7239 0.6496 -1.1144 0.2651 -1.9970 0.5493
regionNorwegian Trench -1.0213 0.6590 -1.5497 0.1212 -2.3129 0.2703
regionNorwegian Sea -1.4950 0.6826 -2.1901 0.0285 -2.8329 -0.1571 *
regionGreenland-Scotland ridge -2.1084 0.6641 -3.1746 0.0015 -3.4101 -0.8067 **
regionBarents Sea -1.9196 0.6698 -2.8659 0.0042 -3.2325 -0.6068 **
trophic_level 0.6697 0.2124 3.1525 0.0016 0.2533 1.0861 **
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1The trophic level effect is estimated to be 2.0 with 95% confidence limits 1.3 and 3.0. Thus, concentrations are estimated to roughly double with every unit increase in trophic level.
However, the exploratory plots suggested that there wasn’t much of a difference in concentration ratios between fish species. We might therefore consider a model which has a factor (called class, say) which differentiates fish and shellfish (instead of a linear trophic level effect); i.e.Â
- fixed = region + class
- random = station + species + station.series
Multivariate Meta-Analysis Model (k = 449; method: REML)
Variance Components:
estim sqrt nlvls fixed factor
sigma^2.1 0.1832 0.4280 380 no station
sigma^2.2 0.1647 0.4059 13 no species
sigma^2.3 0.0278 0.1667 449 no station_series
Test for Residual Heterogeneity:
QE(df = 435) = 7707.2092, p-val < .0001
Test of Moderators (coefficients 1:14):
QM(df = 14) = 200.2043, p-val < .0001
Model Results:
estimate se zval pval ci.lb ci.ub
regionIberian Sea 1.4335 0.1866 7.6809 <.0001 1.0677 1.7993 ***
regionNorthern Bay of Biscay 1.4581 0.1885 7.7345 <.0001 1.0886 1.8276 ***
regionCeltic Sea 0.9821 0.1854 5.2973 <.0001 0.6187 1.3455 ***
regionIrish Sea 1.3175 0.1742 7.5610 <.0001 0.9760 1.6590 ***
regionIrish and Scottish West Coast 0.9045 0.1862 4.8587 <.0001 0.5396 1.2693 ***
regionChannel 1.3309 0.1838 7.2422 <.0001 0.9707 1.6911 ***
regionSouthern North Sea 1.4977 0.1743 8.5916 <.0001 1.1560 1.8394 ***
regionSkagerrak and Kattegat 1.2038 0.1754 6.8636 <.0001 0.8600 1.5475 ***
regionNorthern North Sea 1.7125 0.1774 9.6541 <.0001 1.3648 2.0601 ***
regionNorwegian Trench 1.4219 0.2009 7.0784 <.0001 1.0282 1.8156 ***
regionNorwegian Sea 0.9516 0.2590 3.6739 0.0002 0.4440 1.4593 ***
regionGreenland-Scotland ridge 0.3352 0.2178 1.5392 0.1238 -0.0916 0.7620
regionBarents Sea 0.5271 0.2227 2.3668 0.0179 0.0906 0.9636 *
classShellfish -1.1710 0.2592 -4.5182 <.0001 -1.6789 -0.6630 ***
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1The class effect suggests that the trophic magnification moving from shellfish to fish is 3.2 with 95% confidence limits 1.9 and 5.4.
For a more formal comparison, we can look at the AIC of each model. We will throw in the model with no trophic level or class effect as well.
- M1: region
- M2: region + class
- M3: region + trophic_level
AIC
M2 621.8752
M3 626.9407
M1 633.5628Model M2 is clearly the preferred model. This has a separate level for fish and shellfish, but does not differentiate between fish of different trophic levels.
Trophic adjustments
Here is the impact on the regional estimates of status. The plots are:
- Unadjusted: no trophic adjustment and no species random effect (what we have used in the past)
- Species adjusted: species random effect (included for completeness) based on model M1
- Shellfish adjusted: adjusted to shellfish (making everything a bit greener) based on model M2
- Fish adjusted: adjusted to fish (making everything a bit redder) based on model M2
- Trophic adjusted: adjusted to trophic level 3 based on model M3
Unadjusted
Species adjusted
Shellfish adjusted
Fish adjusted
Trophic adjusted
PCBS
PCBs are assessed against EACs which are expressed on a lipid weight basis. For simplicity, we investigate a regional assessment of CB153, rather than the full regional assessment (of a suite of PCB concentrations) which takes a long time to run. CB153 typically dominates the PCB concentration profile and is less beset by issues with less-than values.
Summary plots
Stripplot
Scatter plot
Residual plot
Note that the residuals seem pretty well distributed about zero apart from the highest trophic fish. This suggests that any trophic effect might be nonlinear.
Trophic models
First, here’s a quick look to see if there is any evidence of a tissue effect - but there isn’t (tissueMU: p = 0.2323)
Multivariate Meta-Analysis Model (k = 347; method: REML)
Variance Components:
estim sqrt nlvls fixed factor
sigma^2.1 0.5912 0.7689 333 no station
sigma^2.2 0.4291 0.6550 12 no species
sigma^2.3 0.1080 0.3287 347 no station_series
Test for Residual Heterogeneity:
QE(df = 332) = 10001.2934, p-val < .0001
Test of Moderators (coefficients 1:15):
QM(df = 15) = 472.3240, p-val < .0001
Model Results:
estimate se zval pval ci.lb ci.ub
regionIberian Sea -4.0256 0.9363 -4.2994 <.0001 -5.8608 -2.1905 ***
regionNorthern Bay of Biscay -3.7089 0.9207 -4.0284 <.0001 -5.5135 -1.9044 ***
regionCeltic Sea -4.7572 0.9174 -5.1854 <.0001 -6.5553 -2.9591 ***
regionIrish Sea -4.1356 0.9250 -4.4707 <.0001 -5.9487 -2.3226 ***
regionIrish and Scottish West Coast -5.9735 0.9189 -6.5005 <.0001 -7.7746 -4.1725 ***
regionChannel -3.3766 0.9271 -3.6420 0.0003 -5.1938 -1.5595 ***
regionSouthern North Sea -3.9335 0.9334 -4.2141 <.0001 -5.7630 -2.1041 ***
regionSkagerrak and Kattegat -4.6392 0.9409 -4.9307 <.0001 -6.4833 -2.7951 ***
regionNorthern North Sea -4.4378 0.9289 -4.7775 <.0001 -6.2584 -2.6172 ***
regionNorwegian Trench -5.1554 0.9477 -5.4399 <.0001 -7.0128 -3.2979 ***
regionNorwegian Sea -4.6367 1.0212 -4.5403 <.0001 -6.6383 -2.6351 ***
regionGreenland-Scotland ridge -6.6220 0.9593 -6.9028 <.0001 -8.5023 -4.7418 ***
regionBarents Sea -6.0173 0.9838 -6.1162 <.0001 -7.9455 -4.0890 ***
tissueMU -0.4414 0.3696 -1.1945 0.2323 -1.1658 0.2829
trophic_level 0.5363 0.3050 1.7581 0.0787 -0.0616 1.1341 .
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1Now let’s fit a series of models and compare their model fits by AIC. The first three models are the same as before. The extra model M4 fits a smoother in trophic_level on 2 degrees of freedom.
- M1: region
- M2: region + class
- M3: region + trophic_level
- M4: region + s(trophic_level, 2)
AIC
M3 923.8704
M1 924.1877
M2 925.1585
M4 925.5556M3 is the best model in terms of AIC, indicating a linear trophic level effect. However, there is not a lot beetween the four models, suggesting that evidence for a trophic level effect is weak at best.
Here are the parameter estimates from the best model:
Multivariate Meta-Analysis Model (k = 347; method: REML)
Variance Components:
estim sqrt nlvls fixed factor
sigma^2.1 0.6271 0.7919 333 no station
sigma^2.2 0.4834 0.6953 12 no species
sigma^2.3 0.0752 0.2743 347 no station_series
Test for Residual Heterogeneity:
QE(df = 333) = 10116.3440, p-val < .0001
Test of Moderators (coefficients 1:14):
QM(df = 14) = 454.6862, p-val < .0001
Model Results:
estimate se zval pval ci.lb ci.ub
regionIberian Sea -3.8383 0.9601 -3.9978 <.0001 -5.7200 -1.9565 ***
regionNorthern Bay of Biscay -3.5042 0.9461 -3.7037 0.0002 -5.3585 -1.6498 ***
regionCeltic Sea -4.5681 0.9427 -4.8459 <.0001 -6.4157 -2.7205 ***
regionIrish Sea -3.9490 0.9489 -4.1615 <.0001 -5.8089 -2.0891 ***
regionIrish and Scottish West Coast -5.7895 0.9446 -6.1293 <.0001 -7.6408 -3.9382 ***
regionChannel -3.1934 0.9516 -3.3558 0.0008 -5.0586 -1.3283 ***
regionSouthern North Sea -3.7394 0.9566 -3.9090 <.0001 -5.6143 -1.8644 ***
regionSkagerrak and Kattegat -4.4825 0.9672 -4.6346 <.0001 -6.3781 -2.5869 ***
regionNorthern North Sea -4.2489 0.9526 -4.4605 <.0001 -6.1159 -2.3819 ***
regionNorwegian Trench -4.9779 0.9714 -5.1246 <.0001 -6.8818 -3.0740 ***
regionNorwegian Sea -4.4715 1.0441 -4.2827 <.0001 -6.5178 -2.4251 ***
regionGreenland-Scotland ridge -6.4493 0.9832 -6.5592 <.0001 -8.3764 -4.5222 ***
regionBarents Sea -5.8518 1.0074 -5.8091 <.0001 -7.8261 -3.8774 ***
trophic_level 0.4496 0.3083 1.4583 0.1448 -0.1547 1.0538
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1There is only weak evidence for a trophic level effect (p = 0.1448). The estimate of the effect is 1.6 with 95% confidence limits 0.8 and 2.9.
Trophic adjustments
Here is the impact on the regional estimates of status. The plots are:
- Unadjusted: no trophic adjustment and no species random effect (what we have used in the past)
- Species adjusted: species random effect (included for completeness) based on model M1
- Trophic adjusted: adjusted to trophic level 3 based on model M3
Unadjusted
Species adjusted
Trophic adjusted
PBDEs
PBDEs are assessed against the FEQG expressed on a lipid weight basis. We will focus on BDE47 as the compound which dominates the profile and is least likely to be affected by less-than issues.
Summary plots
Stripplot
Scatter plot
Residual plot
Trophic models
First, here’s a quick look to see if there is any evidence of a tissue effect. There is (tissueMU: p = 0.0003), but this could be because nearly all the muscle samples are in herring and eelpout and the tissue effect is confounded with species. Looking back at the mercury and CB153 data, the concentrations in herring have always been lower than in other species in the same region (Skaggerak and Kattegat). We’ll ignore this for now and return to it in a later report.
Multivariate Meta-Analysis Model (k = 241; method: REML)
Variance Components:
estim sqrt nlvls fixed factor
sigma^2.1 0.5155 0.7180 231 no station
sigma^2.2 0.0566 0.2379 11 no species
sigma^2.3 0.0633 0.2516 241 no station_series
Test for Residual Heterogeneity:
QE(df = 227) = 4429.6567, p-val < .0001
Test of Moderators (coefficients 1:14):
QM(df = 14) = 2692.3869, p-val < .0001
Model Results:
estimate se zval pval ci.lb ci.ub
regionIberian Sea -6.5322 0.5135 -12.7204 <.0001 -7.5386 -5.5257 ***
regionNorthern Bay of Biscay -8.1162 0.4910 -16.5286 <.0001 -9.0786 -7.1537 ***
regionCeltic Sea -7.6913 0.4958 -15.5142 <.0001 -8.6630 -6.7196 ***
regionIrish Sea -6.2970 0.5266 -11.9573 <.0001 -7.3292 -5.2649 ***
regionIrish and Scottish West Coast -7.0497 0.5043 -13.9791 <.0001 -8.0381 -6.0613 ***
regionChannel -7.9308 0.5161 -15.3674 <.0001 -8.9423 -6.9193 ***
regionSouthern North Sea -7.3085 0.5620 -13.0036 <.0001 -8.4100 -6.2069 ***
regionSkagerrak and Kattegat -7.3820 0.5831 -12.6610 <.0001 -8.5248 -6.2393 ***
regionNorthern North Sea -6.6301 0.6034 -10.9888 <.0001 -7.8127 -5.4476 ***
regionNorwegian Trench -6.7570 0.6772 -9.9779 <.0001 -8.0843 -5.4297 ***
regionGreenland-Scotland ridge -8.1363 0.6509 -12.5000 <.0001 -9.4121 -6.8606 ***
regionBarents Sea -8.0743 0.6772 -11.9232 <.0001 -9.4016 -6.7471 ***
tissueMU -1.3019 0.3614 -3.6026 0.0003 -2.0101 -0.5936 ***
trophic_level 0.7191 0.1831 3.9265 <.0001 0.3601 1.0780 ***
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1Now let’s fit a series of models and compare their model fits by AIC.
- M1: region
- M2: region + class
- M3: region + trophic_level
- M4: region + s(trophic_level, 2)
AIC
M3 612.3422
M4 613.2215
M2 615.4454
M1 616.2566Again M3 is the best model in terms of AIC. Here are the parameter estimates:
Multivariate Meta-Analysis Model (k = 241; method: REML)
Variance Components:
estim sqrt nlvls fixed factor
sigma^2.1 0.4200 0.6481 231 no station
sigma^2.2 0.1574 0.3968 11 no species
sigma^2.3 0.1550 0.3937 241 no station_series
Test for Residual Heterogeneity:
QE(df = 228) = 4435.3603, p-val < .0001
Test of Moderators (coefficients 1:13):
QM(df = 13) = 1410.2767, p-val < .0001
Model Results:
estimate se zval pval ci.lb ci.ub
regionIberian Sea -6.2318 0.6781 -9.1907 <.0001 -7.5608 -4.9028 ***
regionNorthern Bay of Biscay -7.8024 0.6595 -11.8311 <.0001 -9.0950 -6.5098 ***
regionCeltic Sea -7.3481 0.6581 -11.1651 <.0001 -8.6381 -6.0582 ***
regionIrish Sea -5.9106 0.6833 -8.6506 <.0001 -7.2497 -4.5714 ***
regionIrish and Scottish West Coast -6.7436 0.6660 -10.1256 <.0001 -8.0489 -5.4382 ***
regionChannel -7.6064 0.6781 -11.2166 <.0001 -8.9355 -6.2773 ***
regionSouthern North Sea -6.8342 0.7048 -9.6963 <.0001 -8.2156 -5.4528 ***
regionSkagerrak and Kattegat -7.3527 0.7489 -9.8173 <.0001 -8.8206 -5.8847 ***
regionNorthern North Sea -6.1540 0.7375 -8.3441 <.0001 -7.5995 -4.7085 ***
regionNorwegian Trench -6.5141 0.8126 -8.0164 <.0001 -8.1067 -4.9214 ***
regionGreenland-Scotland ridge -7.8776 0.7891 -9.9828 <.0001 -9.4243 -6.3310 ***
regionBarents Sea -7.8340 0.8129 -9.6367 <.0001 -9.4273 -6.2407 ***
trophic_level 0.5460 0.2298 2.3763 0.0175 0.0957 0.9963 *
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1There is a significant trophic level effect (p = 0.0175). The estimate of the effect is 1.7 with 95% confidence limits 1.1 and 2.7.
Trophic adjustments
Here is the impact on the regional estimates of status. The plots are:
- Unadjusted: no trophic adjustment and no species random effect (what we have used in the past)
- Species adjusted: species random effect (included for completeness) based on model M1
- Trophic adjusted: adjusted to trophic level 3 based on model M3
Unadjusted
Species adjusted
Trophic adjusted
