Using Elemental Analyses and Multivariate Statistics to Identify the Off-site Dispersion from Informal e- waste Processing

Electronic waste (e-waste) is informally processed and recycled in Agbogbloshie in Accra (Ghana), which may be the largest such site in West Africa. This industry can lead to significant environmental contamination. In this study, surface dust samples were collected at a range of sites within Accra to establish the offsite consequences of such activities. Fiftyone samples were collected and analysed for 69 elements by ICP-mass spectrometry after nitric acid digestion. The data indicated a significant enrichment in metals associated with 2 solder and copper wire at the site itself and a downwind dispersion of this source material to a distance of approximately 2.0 km. Chlorine and bromine were also elevated at this site as residues from polyvinyl chloride combustion and flame retardants respectively. The elemental composition indicated that only low technology electrical equipment was being treated this way. Multivariate statistical analyses by principal components analysis and polytopic vector analysis identified three sources contributing to the system; i) burn site residue dispersing within 2 km from the source site, ii) marine matter on the beaches alone and iii) the baseline soil conditions of the city of Accra. Risk ratios and hazard quotients developed from the measured concentrations indicated that copper was providing the greatest risk to inhabitants in most cases although nickel, vanadium, chromium and zinc also contributed.


Introduction
Waste electronic and electrical equipment (WEEE or e-waste) is generated at a significant rate across the globe 1 . The export of such waste from developed to less developed nations is restricted under the Basel Convention 2 . Despite this, there is a significant transfer of such waste to locations like Agbogbloshie in Accra, Ghana where the metals are recovered by low technology methods for future use 3 . E-waste may contain a diversity of elements that, once recovered, have a significant resale value. These include gold, silver, copper, some of the 3 platinum group elements (PGM) and the rare earth elements (REEs) 4 . The methods of recovery employed in Ghana are relatively simple (the burning off of plastics to release the metals) and may lead to significant atmospheric contamination. Health studies 5 also indicate that the burners may be exposed to dioxins, mercury and heavy metals.
Previous studies of the metal contaminants present in the soil at the e-waste site have indicated significant enrichment of elements associated with electronic components and a number of these can also be quantified in hair from workers and residents near the reprocessing site 6 . The stable isotopes of lead indicated that there may be several sources of this element, including food and soil, which also contribute to these hair samples.
Another study 7 analysed metals from the Agbogbloshie site and calculated the risk factors to workers. In this study, the risk factors were shown to exceed the US EPA cancer risk value 8 ; recommendations were subsequently made for governmental intervention to improve the conditions. In a study using a portable X-ray fluorescence device, elevated concentrations were measured across the site itself 9 and hazard quotients indicated that the concentrations exceeded the 1.0 threshold on the recycling site. Of the nine metals analysed, Pb, Cu, As and Sb posed the greatest risk.
Along with metals, a number of organic contaminants can be found at the same location.
For example, studies have shown enhanced polybrominated diphenyl ethers (PBDEs) 10 ; these compounds are usually present in electronic equipment as a flame retardant. In addition to bromine, chlorine may also be present in a number of forms; this is due to the thermal decomposition of PVC (polyvinyl chloride) plastic in both the e-waste and covering 4 the wire that is burned to release the copper 11 . Other organic compounds may be in the ewaste itself or formed during the thermal decomposition processes.
The pathway between the source of contaminants at Agbogbloshie and receptors in wider city of Accra is through the atmospheric transport of gases, vapours and particles. Airborne particulate matter (PM) is both a vector for contaminant transport and has a significant impact on human health compared to other common air pollutants such as ground level ozone and carbon monoxide and reductions have benefits to health 12 . Particulate matter may comprise a heterogeneous mixture of solid and liquid particles suspended in air that varies continuously in size and chemical composition in space and time. The chemical composition of PM is diverse and includes nitrates, sulphates, elemental and organic carbon, organic compounds (e.g. polycyclic aromatic hydrocarbons), biological compounds (e.g. pollen and bacteria) and metals (e.g. iron, copper, nickel, zinc and vanadium). Various elements are found in different size ranges of atmospheric particles. Elements such as Al, Si, Ca, Fe, Na, Mg and K are mainly associated with the coarse particulate matter (>3µm) in ambient air. Ba, Be, Cr, Sn, Ni, Sr, U, and V are found in an intermediate range of about 1-5 µm whereas As, Sb, Se, Cd, Pb and Zn are present dominantly in the fine fraction (0.3-0.8 µm) 13,14 . This is relevant both due to the negative health effects of PM and the potential for long-range transport. The principal aim of this investigation was to determine the extent of offsite metal contamination derived from the Agbogbloshie e-waste burning site, by taking dust samples as a proxy for atmospheric fallout. This will use the multi-element chemical signature associated with the source(s) to quantify the dispersion into the city of Accra.

Experimental
Many of the elements of interest, such as those with human health implications, are principally in the solid phase and will settle out of the atmosphere relatively quickly at a rate dependent on the particle density and meteorological conditions. Atmospheric sampling to intercept the airborne particles was not feasible over such a large area and within a sensible timescale. Samples at the burn site itself were taken from the surface of the soils. These "soils" contained residues from the burning of e-waste as well as fragments from the source materials. Samples were collected along three major transects downwind of the burning site at Agbogbloshie (Figure 1). An additional transect along the strandline of the beach was also sampled to investigate the movement of contaminants through river discharge and longshore drift towards the east. All the sampling locations can be seen in Figure 1 and   A number of approaches may be used to determine the contribution from any one source within the samples. In a mixing approach, known chemical signatures are mixed together in different proportions until the best fit to the observed data is obtained. This relies on having samples of all sources appropriate to the system to mix together. An alternative approach is unmixing where the environmental data are used to determine what the original sources must have been to give rise to the observed conditions 15 . In this study, polytopic vector analysis (PVA), an unmixing approach, was used to determine (a) the most appropriate 8 number of discrete sources within the dataset, (b) the chemical composition of these sources and (c) the contribution that each source makes to every sample.
Since the primary wind direction is from the south and west 16 , three transects were chosen that ran to the east, the north east and the north of the site. The Agbogbloshie processing site is located towards the western side of Accra and is, therefore, upwind of the majority of the residential districts. A further short transect was sampled towards the west to determine the upwind condition and a final transect was sampled along the shoreline, primarily to the east of the outlet from the Korle Lagoon and Odaw River. In total, 51 sites were sampled including the burn site and the adjacent residential area of Old Fadama ( Figure 2). The location of each site was recorded with a GPS and photographs were taken at each location to provide context when considering the results.
Dust was sampled from hard surfaces along the three city transects using a fine brush taking care to avoid resuspension into the air. Samples were collected from hard surfaces such as roadsides and regions contaminated with local soils were avoided. The dusts were double sealed into plastic bags and stored at room temperature until analysed. In addition to the dust samples in the city, surface samples were collected from the upper surfaces on the Agbogbloshie waste site. These latter materials were not soils in the true sense of the word and were comprised of mixed matter from many sources.
Digestion of the samples was performed with a closed-vessel microwave technique system (UltraCLAVE, Milestone, Italy). The choice of nitric acid for digestion was a compromise as this produces the greatest number of elements in solution that can be quanitified in an ICP-HR-MS analysis and it better reflects the potential (bio)-availability of these elements within the human lung. Since we are also primarily interested in the chemical signatures and not the total concentration, a consistent approach such as this was used. Approximately 0.1 g of each sample was weighed accurately on a Mettler PG503 balance and 5 ml nitric acid (HNO3, s.p.) and 3 ml deionized water was added. The samples were digested according to a 65 min. temperature programme, with stepwise heating to 250°C and a holding time of 30 minutes at 250°C. After cooling, the digests were quantitatively transferred to polypropylene tubes and diluted to a total volume of 50 ml with deionized water.
A total of 69 elements were quantified in the 51 samples with 11 of these samples being extracted three separate times. The 11 pseudo-replicate analyses were to establish the degree of homogeneity of the samples and the repeatability of the analysis. In total, 73 separate analyses were completed: • For determination of Hg, aliquots of 25 ml were diluted with deionized water to 50 ml after addition of hydrochloric acid (HCl, s.p.) to a total concentration of 0.5% (v/v), and analysis was performed on an Agilent 7700x series ICP-MS octopole reaction system.
Indium was added to the sample line as an internal standard at a constant rate of approximately 1 µg.ml -1 .
• For determination of the remaining 68 elements (from Li to U), aliquots of 1.0 ml, 0.1 ml and 0.01 ml, respectively, were diluted to 10 ml using deionized water or 1% HNO3 for a total acid matrix of 1%. Rhenium (at 1 ng.ml -1 ) was added as an internal standard.
• All elements, including Hg, in all samples, were quantified by external calibration using multi element mixtures from Teknolab AS and Spectrascan all made from high purity NIST traceable primary element solutions of 99.99% or better. All calibration solutions (except for Hg) were prepared in nitric acid solutions free of chloride to avoid common Additionally, to reduce risk of matrix interferences and signal suppression due to matrix effects the analytes were determined in diluted samples.
• Procedural blanks were produced using the same reagents and microwave extraction but without any sample in the Teflon extraction tubes. These samples were used to determine the limits of quantification for the dust sample extracts (LoQ = 10 * Standard Deviation of the procedural blank).
• Quality control (QC) samples were used to check the recovery and external calibration procedures. Certified reference materials (BCR 176R, fly ash and MODAS-2, bottom sediment (Institute of Nuclear Chemistry and Technology, IChTJ, Poland)) were digested alongside the samples to check the efficiency of the digestion. Results from analysis of CRMs are presented in supplementary information. Blank samples were run after highly concentrated samples to check for appropriate washout and carry-over between samples. All samples, standards, blanks, QC-samples and CRM had rhenium (or indium for Hg) added as an internal standard at a concentration of 1 ng.ml -1 in a HNO3 acidified The dispersion of the elements from the burn site was investigated through a number of multivariate statistical methods such as principal components analysis (PCA), the projection to latent structures through partial least squares (PLS) and polytopic vector analysis (PVA) 15 .
In each case, the data were converted to proportions as this removes the concentration effect and enables the underlying chemical signature to be determined. PCA and PLS were performed with SIMCA-P from Umetrics and PVA was conducted using software from Dr Robert Ehrlich.

Instrumental Variability
As part of the calibration of the ICP-MS, quality control standard solutions of known concentration were routinely analysed as part of the quantification procedure. These results also provide a measure of the variability associated with the instrument. For the major metals in solution, the coefficient of variability was 6% for replicate analyses.

Concentrations
The concentrations of individual elements measured across all samples ranged over many orders of magnitude, as anticipated from the crustal abundances of each 17 . The measured concentrations are presented in the Supplementary Information and as a box and whisker plot and these superficially mirror the expected crustal abundances, also presented on the same figure 18 . The enrichment or depletion relative to these crustal abundances were calculated from the mean concentration for each element across all samples. Several elements were enriched (Figure 3.) and these are the ones that can be linked to a potential presence in electrical wastes. Table 1 lists the correspondence of these enriched elements to various potential sources such as wiring, solder and catalytic converters. The mean, median, range and standard deviation for each element can be seen in Table 2.

Element
Potential  As an example representative element for the solder in the recycled e-wastes, the spatial distribution of tin is presented in Figure 4. High concentration values were found at the burn site and lower concentrations were measured upwind of the site and towards the ends of the downwind transects. All of the elements listed in Table 1

Source Apportionment
Polytopic vector analysis (PVA) was carried out to determine the least amount of potential sources, also called end members, explaining the measured elemental concentrations. One of the outputs from the PVA model is the amount of variance explained for each element for each run solution. The model was run with 2 to 10 end members (EMs) and the individual fits to each element are present in Figure 9. This figure shows the following features: 23 1. Two elements (Ca and Fe) are well explained (>80% of the variance) with just two EMs indicating they probably only have two sources in the system.

A substantial number of elements have a significant improvement in their explained
variance when the number of EMs is increased from two to three. This includes most of the REEs, Sr and V, which increase to ~80% of the variance explained with three EMs. This indicates these elements may have three sources in the system.
3. There are a number of other large improvements in fit, most notably, at from five and six EMs. Smaller improvements in fit are obtained with even more EMs.
4. Many of the heavier elements including Pt, Au, Th and U are poorly explained even after 10 EMs have been fitted.
5. The mean fit rises from 21% at 2 EMs, 37% at 3 EMs, 42% at 4 EMs, 51% at 5 EMs and 58% at 6 EMs. These mean fits are relatively small compared to other systems and may be due to poorly explained elements reducing the mean. One of the issues with all unmixing models is determining the correct number of EMs that cover all of the actual sources, whilst not attempting to fit the noise in the data 19 . Before proceeding to determine the composition and contribution for each EM, the number of elements was reduced to remove those elements that might be considered as randomly distributed within the system and not associated with any particular source. These contribute to the noise in the system and reduce the overall predictability of the model.
Four different approaches were used to determine which of the elements should be removed.
1. The dataset was examined and the elements with more than 50% of the samples exhibiting concentrations below the limit of detection were removed. In this case, only four elements fitted this category (Ru, Ta, Os and Au) and there was little improvement in the overall model as a result.
2. The coefficient of variation for each element was considered (see Table 2) and a decision made on whether high values indicated either significant enrichment at the burn site above a "normal" baseline condition in the other samples or low and a random variability in the entire dataset. Having considered these, it was determined that excluding elements through this approach would not be robust and objective.
3. The PCA loadings were considered ( Figure 8) and the ten elements near the origin with loadings between -0.05 and +0.05 on PC1 and -0.1 and +0.1 on PC2 were removed. The PVA model was rerun and there was an improvement in the fits and the clarity in terms of the number of end members but not sufficiently to warrant pursuing this approach further.
4. The initial fit for each element in the PVA end member plot (Figure 9) was considered and all elements that failed to have 50% of their variance explained 26 within ten EMs were deemed to be due to "noise" or due to a random distribution in the samples. In this case, 14 elements were removed leaving 55 elements in the analysis. Many of these elements were the same as those that had been considered under the other schemes. Notably, most of the heavy elements (Pt, Au, Th, U and Ir) were removed and not used in the further PVA models. This multi-criteria approach improves the confidence that these elements are not contributing significantly to the overall chemical composition of the sources.
The PVA unmixing model approach was rerun with the reduced dataset and the results showed that the variance for all elements were explained to 50% or greater after 10 EMs.
The scree plot of the mean explained variance indicated two possible end member solutions; one with three sources and a second with six. Both models were run and the data compared. In the three end member solution, the end members could be identified, on the basis of their geographical distribution and chemical composition, as the baseline soils of Accra, the marine environment and the burn site. The six end member solution also had the baseline soils and marine environments but divided the burn site source into four reinforcing the heterogeneous nature of the soils on this site (Table 3). In this case, the first three end members all had representative samples within the dataset that contained greater than 90% of that one pure source (i.e. they had a chemical composition close to that of one of the sources). For sources 5 and 6, the typifying samples only had a 36% contribution from those sources indicating that it was most probably a mixture of several different sources. It was, therefore, determined that the three end member solution would 27 be the most appropriate for this system and the results of this model are correspondingly discussed below.
The chemical composition of the three end members can be seen in Figure 10.   (Figure 2). The distribution of the three end members at the sites within Accra can be seen in Figure   11. End member 1, which is typical of the baseline soils, is the major source to all samples in the three transects downwind of the burn site as well as the one upwind to the west. EM2 is only present to any significant degree along the shoreline and it make little to no contribution to the samples within the city. This suggests there is little sea to land transfer 30 of contaminants at this location. EM3 has its origin at the burn site and there are measureable contributions in the dust samples to the east (downwind) of the source. An enlargement of the central region around the burn site, showing the EM3 distribution, can be seen in Figure 12. Although the values at the burn site may exceed 80%, the contribution this source makes to the samples from the city is small with samples containing between 20

Location
and 40% to the east. The maximal range of influence from this source for these elements is between 1.5 and 2.0 km. No enrichment was seen in the upwind samples.

31
The baseline soils The marine component Burn site residues

Pseudo-replicate (Intra-sample) Variability
Eleven samples were chosen for three separate sub-sample analyses (with separate digestions and analysis) to assess the homogeneity of the samples. At least one sample from each transect was chosen for pseudo-replicate variability testing, along with three taken from within the Agbogbloshie processing site. The results indicated that intra-sample variability existed, with a mean coefficient of variation (CoV) across all metals and all pseudo-replicates of 30% (median of 21%). Elements present at the lowest concentrations tended to have the greatest coefficient of variability although this was not a statistically significant trend. The larger CoV values were driven partly by values close to the limit of detection. The CoVs exhibited a log-normal distribution ( Figure 13). The relatively high values indicate the heterogeneous nature of many of the samples as the elements appear to be concentrated in discrete particles rather than being surface bound.

Discussion
Overall, the elemental concentrations measured in this work are in line with other studies that have focused on the burn site 6 . The elemental composition of the samples from the burn site was reflective of the materials used in electronic items (Table 1). Many of the metals in these samples are common in solder. Solder for electronic circuit boards typically used to contain a mixture of 63% tin and 37% lead but, due to the toxicity of the lead, it has been replaced with a number of alternative metals including copper, silver, bismuth, indium, zinc and antimony 20 . Although tantalum was excluded from the multivariate statistical analyses due to its low abundance and occurrence in samples, it was present in two samples at the burn site possibly due to its use in capacitors. The chlorine is likely to have been derived from the PVC and bromine from flame retardants. Some of the antimony may also be associated with flame retardants 21 .
REEs that are typically present in newer high technology equipment such as mobile telephones, high density magnets and display equipment 22 were essentially absent in the burn site residues, indicating that only low technology equipment was processed at the Agbogbloshie burn site. This conclusion is supported by observations made when sampling that the bulk of the waste was relatively low technology items with many white goods (e.g. fridges and freezers) and vehicles. The burners were principally interested in recovering the copper and it appeared as if the wiring looms from old vehicles were being stripped out and the PVC coatings burned off. The newer, high technology, equipment was potentially being removed from the waste stream before it reached Agbogbloshie and was recycled elsewhere. Items within the waste stream that could potentially be reused or repurposed were also removed and were available for sale on the streets of Accra.
The burning aspect of metal recovery was conducted on a designated area of the Agbogbloshie site (site 27 on the maps, see Figure 2  Accra is a busy city with much commerce taking place on the streets. As a consequence, the city streets are relatively dusty with a range of additional potential sources of elements. The concentrations of elements in these city dusts are not dissimilar to the crustal abundances overall, but the REEs are relatively depleted and several elements that can be directly related to the e-waste are enriched. These same elements are also diagnostic of the burn site residues in the PCA and their occurrence can be linked to the nature of the e-waste. Other studies have also identified Cu, Cd and Pb as indicative of these informal e-waste processing sites (e.g. Israel 25 and China 26 ) and more developed European processors 27 .
Copper is present at an elevated concentration in the e-waste materials and is generally the target of the recycling community. The remaining elements are generally not recovered.
Human health may be at risk from a wide and diverse range of elements and compounds.
The exposure routes will also vary but, in general, will encompass inhalation of gases and fine particles as well as the ingestion of food items. The risk to human health from exposure to the elements measured in this study is difficult to quantify without measurement of a number of exposure factors and blood markers, not performed within this study. Thus, as a proxy for risk, a ratio was developed using the maximum permissible concentrations (MPC) in soils 28 . The ratios were calculated at each site and for each metal with an MPC value. For those samples with three (pseudo)-replicates, the mean value was calculated. At several sites, more than one element had a risk ratio greater than 1; since there was no clear approach as to how to combine these risks, the element exhibiting the highest risk was presented.
Results for the risks are shown in Figure 15; in this figure, the colour of the spot is indicative of the element that gave the highest risk and the size of the spot is relative (log10 scale) to the value of the ratio. The greatest risk is from copper, with greatest risk ratios at the burn sites at Agbogbloshie. It is clear that there are some effects of these dusts in the city leading to an increased human health risk, which is to be expected given the dispersion patterns observed in the PVA. What is less explainable is the presence of a few elevated values towards the end of the transects. These are due to nickel at site 05, zinc at sites 03, 34 and 36 and, more unusually chromium at site 50. The source of these elements at these locations is not clear.  In addition to the ratios developed above, the hazard quotient approach of the US Environmental Protection Agency 8 was also used. This is the same approach used by Itai and co-workers 9 . The concentrations were applied to the following equation using the coefficients from Itai et al.: Where C is the concentration of the element (mg.kg -1 ); IRS is the soil ingestion rate (100 mg.day -1 ); EF is the exposure frequency (number of days per year exposed to the dusts = 365 days per year); ED is the exposure duration and 30 years has been used here to be consistent with 9 ; RfD0 is the chronic oral reference dose (mg.kg -1 .day -1 ); BW is the body weight of the exposed person and 62 kg has been used; and AT is the average time of exposure (ED x 365 days per year).
The RfD0 values used were from Itai et al. 9 .
Results showed that the greatest hazard quotient was for copper (17.7) at the burn site. All values in excess of 1.0 were at the Agbogbloshie site, and not in the city, except for copper at site 44. Other elements that exceed 1.0 include Pb, Sb and Zn. These data are consistent with those of Itai et al. 9 although the values calculated here are greater than their reported quotients. This may be due to our sampling of the surface dusts that may have a greater surface area compared to bulk soils.
Overall, there are significantly elevated metal concentrations and associated risks with the soils and dusts on the Agbogbloshie e-waste processing site. This is not unexpected given the nature of the activities that are performed on site. Out of the elements analysed here, a significant number exceed the nominal MPC but copper, the target of e-waste recovery at Agbogbloshie, leads to the greatest risk. It must be stressed that this does not include the considerable potential for organic compounds present with e-waste or formed during the combustion of the plastics on site 29 . It appears as if the dispersion of the elements measured here from the burn site do not penetrate far into the city and the major zone of exposure is within 2.0 km of the site in an easterly direction.
The chemical composition of the marine samples from the shoreline are different from those at the e-waste site and the city, which indicates that materials from the waste site do not enter the river and are washed out into the sea. This correlates with observations made during sampling, that there was a considerable amount of plastic litter on the shoreline but little to no matter from the e-waste site. The majority of the transect samples also indicate a baseline composition unaffected by the waste site.

Conclusions
The type of e-waste being processed was of a low technology level and included discarded white goods and end-of-life vehicles. This was reflected in the chemical composition of the elements found in surface samples taken from the burn site, with the major components related to both tin-lead solder (Sn, Pb) and lead-free solder (Sn, Cu, Ag, Bi, In, Zn and Sb).
Other elements found from samples taken on-site included chlorine potentially derived from the burned plastic residue and bromine from flame retardants.
The mechanism of processing e-waste at Agbogbloshie involves low temperature combustion of waste PU foam as a fuel, with little thermal lift in the generated smoke. This means that the dispersion of these elements is rather limited, and the subsequent contributions can only be quantified up to 1.5 km from the site in a downwind (easterly) direction. The risk to human health from these elements is elevated at the burn site itself and in samples within a relatively short distance of the combustion. For copper, risk ratios up to 2745 and hazard quotients up to 17.7 were calculated. Other elements presenting elevated risk are zinc, nickel, vanadium, chromium, antimony and lead.
Analysis of the entire dataset by a range of multivariate statistical methods indicated that ~15 elements had no specific association with any one source in the region and were poorly explained. Of the four approaches used to identify which elements could be removed to improve source discrimination, the most appropriate was excluding the elements that had <50% of the variance explained after 10 end members had been fitted in PVA. The number of potential sources within the city is most likely to be three; the baseline soils of Accra, the marine environment and the burn site. For the burn site, however, there are a number of potential sources as the area is not homogeneous and elements may be resuspended and dispersed from each.