2503
In this section, statistical analysis of a heterogeneous compilation of existing data on
MBT-2504
derived RDF/SRF, derived under diverging boundary measurement conditions, is presented.
2505
Statistical analysis is challenging for many reasons, including paucity of data and different
2506
statistics reported in the literature data (e.g. mean or median to estimate location). TABLE 18
2507
summarises background information on data sources. Available data is mainly found in
2508
reports5, 32, 190, 194
, along with a few peer-reviewed publications164, 215, 234
. These data series
2509
are mostly derived from quality assurance systems internally implemented by MBT plant
2510
operators to satisfy end-user contractual requirements and/or to demonstrate compliance with
2511
national standards. The RDF/SRF has been produced from varying input materials (different
2512
countries/regions and residual or mixed waste collection schemes), treated in MBT plants
2513
with different design and operational configurations (but predominantly biodrying), prepared
2514
from different fractions of the input waste (e.g. partially including or excluding the biomass
2515
fraction), and prepared to different national standards and end-user requirements (mainly
2516
cement kilns, but also dedicated FBC and power plants).
2517
2518
<<Table 18>>
2519
2520
Different objectives and methodologies applied alongside the entire measurement process,
2521
including sampling plans, sub-sampling and sample preparation, analytical determination
2522
techniques, and dissimilar statistical analysis before final reporting of data in the literature
2523
restrict the potential for meaningful comparison of data by further statistical analysis.
2524
Additionally, not all the diversifying methodological details of measurement are clearly
2525
stated for each case. It has been proposed that the number of analyses performed for each
2526
data series can cautiously be used to guarantee a minimum reliability of the data190. The
2527
majority of the data sets used in our analysis are based on >10 samples/analyses (lower limit
2528
proposed by CEN).
2529
The effect of sampling plans on specific data series is anticipated to be critical, and much
2530
higher than the uncertainties introduced by differences at the analytical stage. The CEN
2531
guidance for SRF sampling plans237DD CEN/TS 15442:2006 incorporated elements of the
2532
latest developments in Gy’s theory of sampling238-241, developed over the last fifty years242.
2533
Enhanced reliability can be expected for data series that followed the sampling theory, such
2534
as the TAUW investigation194; compliance with national quality assurance systems, such as
2535
the Italian Eco-deco data5; and constituting average measurements over long time periods
2536
(Nehlsen plant)5; those that have been assessed had independently, such as the
Herhof-2537
Stabilate®SRF, investigated by Niederdränk et al.215.
2538
Publicly available data is statistically analysed to produce the most representative and
2539
thorough overview of the achieved quality by MBT-derived RDF/SRF to date. Available
2540
statistics of input data on RDF/SRF sample properties is largely limited to: (1) measure of
2541
(central) location of the sample population (mean and/or median); (2) measure of the upper
2542
limit (80thpercentile and/or max) values; and (3) limited reported values for their spread
2543
(typically in the form of standard deviation).
2544
The selection and validity of statistical tools for describing properties of waste is highly
2545
debatable. Certain aspects of the ongoing debate are addressed by van Tubergen et al.190and
2546
Pehlken et al.78. Simplifying, it can be generally assumed that population distributions (or
2547
probability density function curves) for many of the properties of waste-derived products
2548
exhibit skewness and lengthy distribution tails (long-tailedness)78. Typically positive
2549
skewness is evident.
2550
During the pre-normative research for CEN SRF standards TAUW investigated the type
2551
of statistical distribution that is suitable to describe the properties of SRF. The study included
2552
outputs from biodrying plants and from mechanical sorting of the high CV fraction of
2553
residual waste, analysing 35 samples for each case. Results of the Kolmogorov-Smirnov test
2554
for Cl, Hg, Cu, and Cr have shown that for both plant cases the type of distribution that best
2555
fits these properties was log-normal rather than normal, with the exception of Hg for
2556
biodrying which was normally distributed194. However, the CEN technical committee draft
2557
has opted for Cl to be classified by the mean rather than the median, implying a normal
2558
distribution would suffice; indeed there is a certain degree of overlap between the two forms
2559
of distribution. Conversely, Hg has to be classified according to both median and 80th
2560
percentile values, implying that a skewed long-tailed distribution is anticipated.
2561
Other data reported on SRF from Eco-deco and Nehlsen biodrying processes5indirectly
2562
verify that many properties are not fully normally distributed. Data series for which values
2563
for both mean and median are available show varying degrees of difference between these
2564
two statistics, establishing that their distributions are skewed, hence not normal. For instance,
2565
Pb and Cr show clear positive skewness in all three available data sets; whilst Hg shows
2566
mixed behaviour, including one case of negative skewness.
2567
The central trend and the spread of the values for each property are estimated by
2568
calculating and graphically presenting in box-plots the minimum, lower quartile, median,
2569
upper quartile and maximum value for both median and 80thpercentile of input data.
2570
Calculations and graphics were made with Statistica 8®243, following typical box-plot
2571
convetions (see FIGURE 20). Preferably statistical analysis should be performed only on
2572
median and 80thpercentile input SRF data statistics, but in practice means were also used to
2573
indicate the location. Median and 80thpercentile are measures of location and upper values
2574
respectively, independent of the form of distribution and robust to outliers. They are
2575
therefore suitable to describe properties of RDF/SRF that show log-normal distribution or
2576
skewness and longtailedness in general. In the case of normally distributed property, median
2577
is identical to the arithmetic mean. High percentiles, like the 80thpercentile, are particularly
2578
useful for estimating the upper values of concentrations of trace metals of concern (e.g. Cd),
2579
as demanded in practice. The alternative would be to employ data points that are statistics
2580
based on normally distributed populations (arithmetic mean, standard deviation etc);
2581
however, this assumption does not hold for many of the waste measurands, especially for
2582
analytes expected to be present at trace concentrations.
2583
However, median values were not available for all the cases; as a compromise mean
2584
values were also used to indicate the location of input data series. For properties that exhibit
2585
positive skewness, the overall average central trend of the median values is biased towards
2586
higher values when means are used instead of medians. However, the differences between
2587
medians and means, despite evidence for non-symmetrically distributed properties, are not
2588
significant in all cases, as shown by the cases for which comparative data are available. The
2589
number of means used instead of medians for each property are shown in TABLE 19. For Pb
2590
only mean values were used, as there were very limited median values. Given that Pb exhibits
2591
positive skewness, the average location of Pb is then overestimated.
2592
The significance of the results is also restricted by the fact that data series of different
2593
reliability are given equal, non-weighed treatment (e.g. values resulting from long and short
2594
time series results); and by the low number of available data-series on RDF/SRF properties,
2595
ranging from 14 to 4 (properties with fewer data points available are not analysed). As
2596
received (ar) values were converted to dry basis (d) for the properties necessary to do so, by
2597
using the median(/mean) moisture content values, where available. As received net calorific
2598
values were transformed into dry basis by applying the proposed CEN formula244for the
2599
calculation of analyses to different bases245. This has rendered certain NCV (d) to be much
2600
higher than values reported in the literature (e.g. Eco-deco values5), possibly calculated by a
2601
different conversion formula. Nevertheless, given the limitations, calculation of medians and
2602
use of box-plots can provide the most appropriate, accurate and clearly presented account for
2603
the RDF/SRF properties.
2604