Comment on Goedhart and Huijsmans (2017)

Full text

LETTER TO THE EDITOR
Comment on Goedhart and Huijsmans (2017)
Goedhart & Huijsmans (2017), in their paper ‘Accounting for
uncertainties in ammonia emission from manure applied to
grassland’, which critiques Hanekamp et al. (2017), have
performed a valuable service in emphasizing the need to
estimate uncertainty in ammonia (NH
3
) emissions calculated
from the Ryden & McNeill (R&M) model (1984). Studies
published prior to Hanekamp et al. (2017) have presented
results from the R&M model as if they were certain and
precise. Both Goedhart & Huijsmans and Hanekamp et al.
have shown that there is much greater uncertainty than
was heretofore appreciated and communicated. Goedhart &
Huijsmans and Hanekamp et al. only differ in the manner of
estimating this uncertainty. Both approaches have strengths
and shortcomings. We agree with the conclusion of Goedhart
& Huijsmans that further work on modifying the R&M model
will lead to better probabilistic predictions of ammonia
emissions.
Goedhart & Huijsmans agree that ‘there are sources of
uncertainty in the relationships of wind speed and NH
3
concentration with height’; their equations (2) and (3) and
Hanekamp et al.’s (2a) and (2b). The parameters (D, E) and
(A,B) are unknown and estimated from the data. The central
idea of Hanekamp et al. was to incorporate uncertainty in
those parameter estimates into the predictive uncertainty of
ammonia emissions. By their method of estimating predictive
uncertainty in the observations, Goedhart & Huijsmans
would agree with this necessity.
All probability models are conditional on the information
assumed. This was true in the R&M model, and it is true in the
estimates of predictive uncertainty in the Hanekamp et al.
approach and in the approach chosen by Goedhart &
Huijsmans. All three different approaches give rise to different
estimates of uncertainty. To reiterate, in the original R&M
model, this uncertainty is set at zero. Goedhart & Huijsmans
claim that the Hanekamp et al. approach is ‘fundamentally
flawed’. We disagree: considering both our approach and the
approach of Goedhart and Huijsmans, we observe that these
are just different. Our approach is more general than Goedhart
& Huijsmans, but that is because of the considerations of the
R&M model itself. Indeed, as we show below, and because of
the ad hoc nature of the R&M model, the approach that
Goedhart & Huijsmans advocate often fails. Goedhart &
Huijsmans recognize this, but dismiss this flaw by saying that
they will address it later: “Due to this misfit, application of the
parametric bootstrap sometimes resulted in a small proportion
of unrealistically large simulated emissions. While this did not
affect the 95% bootstrap confidence intervals, it indicates that
the linear regression model in equation (3) does not fit well. ...
We are currently working on an alternative to the R&M model
that does not have this flaw”.
The R&M model predicts 0 or negative values of NH
3
and
thereby fails as can be seen in Figure 1 of Hanekamp et al.
(2017; depicted above). This is because of the linear nature
of the regression. Hanekamp et al. note this limitation, while
also pointing out that the actual measures of wind speed and
NH
3
concentration fall outside the predictive bounds of the
regression more often than standard confidence intervals
predict (95% confidence intervals were used). Therefore, a
simple, and as acknowledged liberal, calculation of
confidence bounds was used in an attempt to capture some
of the limitations of the R&M model.
Instead, Goedhart & Huijsmans simulate parameters (D,
E) and (A,B) driven by the results of the regression on actual
data. Once a set of parameters are found in the simulation,
they are inserted into equations (2) and (3), and estimates of
wind and ammonia emissions are derived. The problem with
this is, as Goedhart & Huijsmans recognize, that many
simulations of (D,E) and (A,B) give near 0 values of D and
E, leading to enormous (near infinite), physically impossible
values for the observations. The equations (2) and (3), when
inverted, just fail. (Hanekamp et al. recognized this
possibility in their analysis of their Figure 1).
In any simulation approach, the mean of the simulation
is used to give a point estimate. Estimating values of
ammonia emissions in simulations with (near) infinite values
gives mean estimates that are also enormous. It is strange
that Goedhart & Huijsmans did not discover this. We
surmise they used the standard point estimates of (D,E)
and (A,B) directly from the regression and not from their
recommended overly sensitive simulation approach. Overall,
the simulation approach of Goedhart & Huijsmans will not
work for the R&M model.
Goedhart & Huijsmans also undertake an analysis of
pairwise experiments, by first stating correctly that the “basic
wind speed and concentration data needed for such an
analysis are unfortunately no longer available”. They
proceed then to in effect ‘recreate’ the data and perform a
new simulation. All of Goedhart & Huijsmans’s calculations
and analysis using their recreated data appear correct; we
have no dispute with their methods. Although Goedhart &
Huijsmans state, based on their simulations, that “narrow
band application and shallow injection have, on average,
much smaller ammonia emissions than surface broadcast
spreading, even when accounting for large uncertainties”,
they overlook the fact that the atmospheric concentrations,
which we discussed extensively, do not exemplify such
©2017 British Society of Soil Science 603
Soil Use and Management, December 2017, 33, 603–604 doi: 10.1111/sum.12393
SoilUse
and Management

reductions in any way. But since anybody can devise data
that subsequently are not tested against known atmospheric
concentrations, and since any set of data put into a
simulation only gives back what was put in (as shown in
Briggs; 2016; pp. 103–108), this analysis by Goedhart &
Huijsmans cannot give new insight into varying ammonia
emissions related to different manuring techniques.
In conclusion, we agree with Goedhart & Huijsmans that
the R&M ad hoc equations should be abandoned. A new
model, one that produces physically possible results, and one
on which an agreed statistical method for generating
predictive uncertainty bounds can be found, should be
devised. Additionally, a new set of emission experiments,
where new and actual data is captured to replace the lost
experimental data, should be undertaken.
W. M. BRIGGS
1
,J.C.HANEKAMP
2,3
&M.CROK
4
1
Independent Researcher
2
University College Roosevelt, Middelburg, the Netherlands
3
Environmental Health Sciences, University of
Massachusetts, Amherst, MA, USA
4
Independent Researcher
E-mail: j.hanekamp@ucr.nl; hjaap@xs4all.nl
References
Briggs, W.M. 2016. Uncertainty –The soul of modeling, probability &
statistics. Springer, Switzerland.
Goedhart, P.W. & Huijsmans, J.F.M. 2017. Accounting for uncertainties
in ammonia emission from manure applied to grassland. Soil Use and
Management,33,595–602.
Hanekamp, J.C., Briggs, W.M. & Crok, M. 2017. A volatile
discourse –reviewing aspects of ammonia emissions, models and
atmospheric concentrations in The Netherlands. Soil Use and
Management,33, 276–287.
Ryden, J.C. & McNeill, J.E. 1984. Application of the
micrometeorological mass balance method to the determination of
ammonia loss from a grazed sward. Journal of the Science of Food
and Agriculture,35, 1297–1310.
©2017 British Society of Soil Science, Soil Use and Management,33, 603–604
604 Letter to the Editor