Northern GIG Lake Macrophyte Boundary Comparisons

Northern GIG lake macrophyte boundary comparisons

N Willby, University of Stirling, UK

31 October 2011.

This document reports the comparison of bias at class boundaries for those methods being considered in NGIG. Two comparisons are considered, viz;

1. All low and moderate alkalinity lakes, both clear and humic (i.e. types 101,102, 201 and 202) across all 5 MS (FI, IE, NO, S and UK), and

2. Clear and humic high alkalinity lakes (i.e. types 301 and 302) across IE, NO and UK only.

The bias shown here has been checked using 13 different models which perform the country correction, calculate the common metric EQR, or model the relationship between MSEQR and ICMEQR in subtly different ways and all have been found to produce more or less the same outcome.

Intercalibration is based on an Option 2 approach. The IC common metric is based on species scores where each species is represented by its arithmetic mean annual mean TP value in those lakes where it occurs in a large cross-GIG data set (n=1600). The TP value for each species was then log transformed and the series was then rescaled from 1-10. The ICM value for a site represents the average score of the taxa recorded there. A site specific ICM EQR was calculated by using a General Linear Model to predict the Expected ICM in an international set of benchmark sites, with log alkalinity and altitude acting as covariates and ‘colour’ (clear or humic) and ‘country of origin’ acting as fixed effects. The use of country in this way achieves benchmark standardisation and ensures that the ICM EQR is independent of country. The ICM EQR was then calculated as:

ICM EQR = (Observed ICM – 10)/(Expected ICM – 10)

A linear regression between the MS EQR and the ICM EQR was then used to establish the position of the MS boundaries on the ICM EQR scale. Within this regression all sites with fewer than four ICM scoring taxa were excluded. High alkalinity lakes with alkalinity exceeding 4 meq/L were also excluded as these were virtually confined to a single MS and were outside the window of environmental conditions covered by the benchmark sites. Because the amount of data contributed by each country to each of the six IC types was either relatively small, or the gradient of ICM EQR or MS EQR values covered by each type separately was quite narrow, we used a global model to relate ICM EQR to MS EQR with the IC type acting as an additional fixed effect. This ensures that the within-MS gradient is maximised and that bias in relation to each individual type can be quantified, although it is only considered appropriate to use the average bias across types for the purposes of harmonisation due to the specific uncertainty associated with the type effect in these models. The ICM EQR v MS EQR relationship was highly significant (p = <0.0001), r20.39 in all cases. These models are illustrated in Figure 1.

Figure 1.Relationships between national EQR and ICMEQR.

Low and moderate alkalinity lakes

Prior to calculation of bias the UK method was modified by adjusting the method it uses to calculate the EQR. This has the effect of making the UK more precautionary, especially with respect to the GM boundary. At the September N-GIG meeting in Oslo it appeared that the UK was persistently too relaxed at the GM boundary. The HG and GM boundaries of the S method were also each lowered by 0.05 units as offered prior to the commencement of Phase 2 as S appeared to have an unduly precautionary classification relative to all other MS.

With these changes in place the picture of class bias is as shown in Table 1. This is a summary of the outputs averaged across the 4 lake types. The full analysis can be found in Option2Compare_NGIG.xlxs which is attached (sheets ‘101_102_201_202’ and ‘harm_101_102_201_202’). This is based on a spreadsheet devised by Geoff Phillips. Evidently most MS lie well within 0.25 classes of the global mean position on the ICM EQR scale with only S fractionally too precautionary.

Table 1. Average class bias per MS prior to boundary harmonisation

NO / UK / IE / S / FI
HG / -0.10 / 0.09 / -0.13 / 0.39 / -0.03
GM / -0.22 / -0.16 / -0.15 / 0.33 / 0.22

To achieve a situation where the bias for all countries is ±0.25 classes the following are required:

NO: No general action is required, although it is noted that the negative bias associated the GM boundary in type 201 (moderate alkalinity clear) is quite significant (-0.7 classes). Since NO have type specific boundaries a small increase in the position of the GM boundary for this type would seem appropriate. Currently NO have different reference values on the TIc scale for types 201 and 202 but the same values of TIc at the HG and GM boundaries for these types. It is expected that type 201 would have higher values than 202 reflecting the higher reference TIc of type 201.At this point NO volunteered a new set of generally more precautionary class boundaries which were applied to the global mean ICM EQR established using the original MS EQR boundaries.

UK: No action required

IE: No action required

S: Currently still slightly too stringent at both boundaries. S would need to lower their HG and GM boundaries by a further 0.02 and 0.01 units respectively for their bias to fall inside the band (allowing for rounding errors). This would amount to a total reduction in 0.07 and 0.06 units respectively from the values used in Phase 1. S have indicated that any further reduction in the position of their boundaries would undermine the boundary setting protocol which they adopted. Consequently they will remain slightly precautionary.

FI: No action required

If these changes are put into place the bias for each country averaged across the four types considered will appear as shown in Table 2.

Table 2. Class bias in lake types 101, 102, 201 an 202 after boundary adjustments by individual MS.

NO / UK / IE / S / FI
HG / -0.06 / 0.09 / -0.13 / 0.39 / -0.03
GM / -0.12 / -0.16 / -0.15 / 0.33 / 0.22

Through minor (optional) modifications to the position of class boundaries we therefore consider the classification of clear and humic, low and moderate alkalinity lakes in N-GIG to be harmonised.

High alkalinity lakes

Prior to calculation of bias the UK method was modified by adjusting the method it uses to calculate the EQR. This has the effect of making the UK more precautionary, especially with respect to the GM boundary. No other changes were made prior to comparison.

With these changes in place the picture of class bias is as shown in Table 3. This is a summary of the outputs averaged across the clear and humic high alkalinity lake types for each MS. The full analysis can be found in Option2Compare_NGIG.xlxs which is attached (sheets ‘301_302’ and ‘Harm_301_302’). In this comparison NO is strikingly precautionary, while IE and to a lesser extent UK are slightly too relaxed. Consideration of the individual lake types reveals that NO is especially precautionary with respect to type 302, whilst IE is more relaxed with respect to type 301.

Table 3. Average class bias per MS prior to boundary harmonisation

NO / UK / IE
HG / 0.69 / -0.11 / -0.34
GM / 0.69 / -0.35 / -0.33

To achieve a harmonised classification of HA lakes across NO, IE and UK significant class boundary changes would be needed by NO and more minor changes by IE and UK.

In the case of NO the changes could be distributed evenly across 301 and 302, in which case both boundaries would need to be lowered by 0.08 EQR units. As an alternative a larger change could be made to the boundaries of 302 and a smaller change to those for 301. This would entail lowering the HG and GM boundaries for type 301 by 0.06 and 0.05 EQR units respectively. A larger modification of 0.1 and 0.11 EQR units at HG and GM respectively would then be needed in type 302.

For IE it would be necessary to raise the HG boundary to 0.94 from its present value of 0.9 and to raise the GM boundary by a more modest 0.01 units. For UK the GM boundary would need to be raised by 0.02 units.

If the above changes were to be implemented the class bias would then appear as follows (Table 4).Note if the bias still fractionally exceeds ±0.25 this is because a larger change in EQR would reduce the bias to a value that is even further below 0.25 than it is above 0.25. Such changes would seem to be unnecessarily excessive.

Table 4. Average class bias per MS after boundary harmonisation in HA lakes

NO / UK / IE
HG / 0.26 / -0.11 / -0.25
GM / 0.26 / -0.25 / -0.25

Although it is possible to find a solution where the three MS achieve comparability in their classifications of HA lakes the scale of adjustment needed by NO raise questions over either the comparability of these lakes, even after benchmark standardisation, or over the boundary setting protocol employed by NO. All methods have a very strong relationship with the ICM EQR although this is somewhat weaker in the case of IE (Figure x). At this point NO volunteered a small reduction in several class boundaries which would render it fractionally less precautionary (bias reduced to 0.60 and 0.63 at HG and GM boundaries respectively) from 0.69 at but not by a sufficient amount to move it within the harmonisation band.

In view of the findings based on three MSa bilateral comparison between UK and IE (against the common metric EQR) was then undertaken. In both cases the bias is well inside the harmonisation band and neither MS would need to make any adjustment to their boundaries (Table 5; top). In a bilateral comparison between UK and NO (Table 5, middle), UK would be too relaxed at GM and NO too precautionary at both boundaries (Table 5, middle). In a direct comparison of NO and IE one country will obviously appear far too precautionary and the other too relaxed (Table 5, bottom).

Table 5. Bilateral comparisons of class bias at HG and GM boundaries in type 301 and 302 lakes.

NO / UK / IE
HG / 0.09 / -0.08
GM / -0.06 / 0.08
NO / UK / IE
HG / 0.47 / -0.22
GM / 0.56 / -0.47
NO / UK / IE
HG / 0.57 / -0.43
GM / 0.49 / -0.58

Given the close agreement between UK and IE it could possibly be argued from these collective pairwise comparisons that the classification of HA lakes by NO is unduly precautionary. For additional information a direct comparison is attached of IE versus UK EQRs when the UK method is applied to IE data in the N-GIG dataset (Figure 3). An Option 3 approach could not be carried out since a reciprocal application of the IE method to UK lake data was not possible. This comparison confirms that the classification of UK is slightly more precautionary with respect to the HG boundary but that general agreement is very good.

Figure 3.Direct pairwise comparison of IE EQR versus UK EQR for NGIG lakes in IE. IE boundaries solid green and blue, UK dashed.

It is possible that NO appears very precautionary due to the lack of a comparable Scandinavian MS in this exercise. Since the agreement between the methods of UK and IE is so close we consider it preferable to consider these methods harmonised. Moreover, the UK boundaries for HA lakes have already been fully tested and intercalibrated in CB-GIG where they were found to be relatively precautionary. NO can then chose to adopt a precautionary classification, subject to separate testing against the classifications of another Scandinavian MS at a later date if and when suitable data becomes available.