A Statistical Analysis of the Stewardship Ontario Residential Waste Audit Program Bruce G. Wilson, University of New Brunswick, Fredericton, New Brunswick, Canada CONTACT Dr. Bruce G. Wilson, P.Eng. Department of Civil Engineering, University of New Brunswick P.O. Box 4400, Fredericton, NB, CANADA, E3B 5A3 Phone 1-506-453-5127 Fax : 1-506-453-5127 Email: wilsonbg@unb.ca INTRODUCTION Stewardship Ontario is the industry funding organization incorporated in Ontario, Canada to represent stewards of the printed papers and packaging that comprise Blue Box Wastes. Stewardship Ontario developed the Blue Box Program Plan (BBPP) in co-operation with Waste Diversion Ontario, a non-crown corporation created under the Waste Diversion Act on June 27, 2002. The BBPP was approved by the Ontario Minister of Environment on December 22, 2003 and commenced on February 1, 2004. The Act requires that stewards pay 50% of the net system costs of blue box recycling - about $62 million was distributed in 2007 to the 200 municipalities that provide/manage residential recycling programs in their communities. In addition to receiving and compiling the data reported to Stewardship Ontario by industry stewards, over the past three years, Stewardship Ontario has completed a series of residential waste characterization studies (audits) in single family and multi-residential buildings. Stewardship Ontario uses data from waste audits to: 1. assess waste generation for setting the fees paid by stewards; 2. determine the recovery performance of existing programs; and 3. assess opportunities and priorities for recovery improvement. The methodology for conducting the single family audits is described in Stewardship Ontario (2005) and a detailed methodology for all audit types is available on-line at http://www.stewardshipontario.ca/bluebox/eefund/projects/audits/waste_audit_own.htm. The results of approximately 1200 samples have been recorded in a database. Each sample represents 10 individual houses or seasonal residences or a single multi-family building, In 2008, Stewardship Ontario contracted with the University of New Brunswick to conduct a statistical analysis of the data and a review of the approach to determining generation rates. The project included a review of the reliability of the data and made recommendations on improvements to the methodology and sampling procedures. This paper summarizes the major findings and recommendations from the study. As part of the study, more than 80 relevant studies, manuals, and journal articles were considered (Wilson, 2008b). Based on that review, it appears that the Stewardship Ontario waste audit program is one of the largest (if not the largest) database of household waste generation information in the world. METHODOLOGY Data analysis was performed using Microsoft Excel 2002 and MINITAB 15 software. Summary statistics were calculated for 6 broad material categories and 11 specific material categories. In
total, 17 different material categories were analyzed as listed in Table 1. The data from Single Family areas and Multi-Family areas were analyzed separately. For each of the 17 material categories, descriptive statistics (mean, standard deviation, minimum, maximum, and skewness) were calculated. Data from each category were plotted on both normal and log-normal probability plots. Potential outliers in the data were identified by examining any data points that fell outside the expected confidence intervals on the probability plots. For each potential outlier, the original data was examined to try to determine if there was an error in the data. Finally, ANOVA analysis was used to determine if there are statistically significant differences in the data based variables such as the type of housing unit (Low Rise or High Rise), the Area Type (Large Urban or Small Urban and Rural), the Season (Spring, Summer, Fall or Winter), or the year (i.e. annual trends in the data). RESULTS Summary Statistics Table 1 presents summary statistics (mean, standard deviation, minimum, maximum, and skewness) for selected material categories (printed paper and packaging materials only) for Single Family households. (Some material categories such as organics, yard waste, textiles, and diapers were excluded from the analysis.) All of the data are positively skewed and some of the data are strongly skewed. Table 2 presents the same data for Multi-Family households. Table 1: Summary Statistics by Material Category Single Family Samples Standard Mean Number Category Deviation Minimum Maximum Skewness Kilograms per household per week 1 Printed Paper 2.5179 1.2176 0.1840 9.355 1.06 2 Paper Packaging 1.5908 0.8602 0.2555 11.028 3.42 3 Plastic 1.0727 0.3553 0.1610 3.305 0.90 4 Aluminum 0.1279 0.0702 0.0000 1.063 3.76 5 Steel 0.2328 0.1107 0.0100 1.138 1.46 6 Glass 0.7182 0.4280 0.0000 2.766 1.35 7 Newspaper 1.6050 0.9133 0.0525 8.439 1.59 8 Corrugated 0.6191 0.4752 0.0110 5.043 2.87 9 Boxboard / Cores 0.5021 0.1934 0.0490 2.181 1.24 10 PET Bottles & Jars 0.2015 0.0900 0.0130 0.759 1.03 11 HDPE Bottles & Jugs 0.1115 0.0546 0.0000 0.648 1.85 12 Aluminum Cans 0.0968 0.0550 0.0000 0.365 1.26 13 Steel Food & Beverage Cans 0.2047 0.0973 0.0000 0.806 0.88 14 Clear Glass Alcoholic 0.1631 0.1683 0.0000 1.204 2.18 15 Coloured Glass Alcoholic 0.2176 0.2375 0.0000 1.463 1.98 16 Clear Glass Food & Beverage 0.2968 0.1530 0.0000 1.540 1.39 17 Coloured Glass Food & Bev. 0.0407 0.0724 0.0000 0.723 3.71 Probability Plots Normal and log-normal probability plots were generated for each of the 17 material categories. These plots clearly showed that none of the data was well represented by a Normal probability distribution. This was expected, given the skewness of the data in Tables 1 and 2. However, lognormal probability plots suggested that, except for categories 14, 15, and 17, a log-normal distribution represents the data reasonably well, although some outliers may still exist. Figure 1
shows an example of a log-normal probability plot for corrugated packaging. For categories 14, 15, and 17 (three glass categories), a log-normal distribution does not fit the data. An example is shown in Figure 2. In fact, no probability distribution could be found to represent these data. This problem stems from the fact that no glass was found in a large proportion of the samples. For example, more than 30% of the samples found no Coloured Food & Beverage Glass. (Ontario introduced a depositreturn system for glass wine and spirit containers in Feb 2007. The impact of this change on alcoholic beverage glass figures for 2007 is unknown.) Table 2: Summary Statistics by Material Category Multi-Family Samples Number Category Mean Standard Minimum Maximum Skewness Deviation Kilograms per household per week 1 Printed Paper 1.7436 0.9668 0.0833 7.1200 1.76 2 Paper Packaging 1.2546 0.7208 0.1835 5.2967 1.88 3 Plastic 0.8382 0.5040 0.0904 6.2847 4.56 4 Aluminum 0.0865 0.0841 0.0045 1.4371 11.45 5 Steel 0.1696 0.1049 0.0252 1.4585 5.78 6 Glass 0.4449 0.3286 0.0000 3.4304 3.36 7 Newspaper 1.0920 0.6921 0.0000 6.1633 2.04 8 Corrugated 0.4764 0.3866 0.0187 3.0007 2.24 9 Boxboard / Cores 0.3467 0.1450 0.0822 0.8771 1.11 10 PET Bottles & Jars 0.1576 0.1236 0.0130 2.0103 9.34 11 HDPE Bottles & Jugs 0.0713 0.0480 0.0000 0.5803 5.11 12 Aluminum Cans 0.0632 0.0583 0.0000 0.8613 7.62 13 Steel Food & Beverage Cans 0.1487 0.0837 0.0175 0.8923 3.26 14 Clear Glass Alcoholic 0.0977 0.1386 0.0000 1.4063 4.85 15 Coloured Glass Alcoholic 0.1339 0.1633 0.0000 1.4326 3.02 16 Clear Glass Food & Beverage 0.1888 0.1066 0.0000 0.8544 1.75 17 Coloured Glass Food & Bev. 0.0245 0.0369 0.0000 0.3149 3.50 Outlier Identification Probability plots identified a small number of potential outliers in the data. In one case, several audits from seasonal households were excluded from further analysis because they were clearly very different from all other audit data. However, in all other cases, there was no pattern in the data that would suggest a systematic error or any other reason to specifically exclude any data points. This suggests that no one audit sample was an outlier (due to a consistent problem in sorting procedures for example). Some outliers may be due to errors in recording data or they may be legitimate data points. For example, an abnormally large sample weight could be caused by a homeowner cleaning out a garage or basement. Since there was no clear pattern to these outliers and because including or excluding them did not have a significant impact on estimates of the mean generation rate the points were left in the database. Confidence Intervals for Mean Values The mean values reported in Tables 1 and 2 are sample means, based on the data collected. If the data were Normally distributed, putting a confidence interval on this mean would be a straightforward exercise. However, since the data are distinctly non-normal, establishing a reasonable confidence interval on the mean is slightly problematic. Since the data are skewed, the confidence interval will be non-symmetrical about the mean. There are procedures for establishing approximate confidence intervals for the sample mean if the data can be represented by a lognormal distribution.
Probability Plot of Corrugated 3-Parameter Lognormal - 90% CI Percent 99.9 99 95 90 80 70 60 50 40 30 20 10 5 Loc -0.8395 Scale 0.6593 Thresh -0.05915 N 374 AD 0.292 P-Value * 1 0.1 0.1 1.0 Corrugated Packaging - kgs/household/week 10.0 Figure 1 Log-normal Probability Plot for Corrugated Percent 99.9 99 95 90 80 70 60 50 40 30 20 10 5 1 Probability Plot of Food and Beverage Glass Coloured 3-Parameter Lognormal - 90% CI Loc -6.119 Scale 3.597 Thresh -0.00001 N 374 AD 39.914 P-Value * 0.1 0.000000 0.000000 0.000000 0.000001 0.000010 0.000100 0.001000 0.010000 0.100000 1.000000 10.000000 100.000000 Coloured Food and Beverage Glass - kg/hhld/week 1.0000E+03 Figure 2 Log-normal Probability Plot for Coloured Food and Beverage Glass
In this paper, the modified Cox method reported by Olsson (2005) was used to determine 95% confidence intervals for the mean generation rate for each material category. The results are presented in Table 3 for single family household audits and Table 4 for Multi-family household audits. These tables show, for example, that the mean weekly household generation rate for Printed Paper from Single Family Audits was 2.5179 kgs and that the mean lies between 2.4573 and 2.6632 kgs with 95% confidence. Note that because the data for categories 14, 15, and 17 are not lognormal, no confidence interval can be established for the mean generation rate for these categories. Differences in the Data Due to Different Classifications Analysis of Variance (ANOVA) procedures were used to identify differences in the data based on differences in Area Type (Large Urban or Small Urban and Rural) and Season. The analysis was conducted in MINITAB 15. The results are presented in Table 5 and show, for example, that Large Urban areas generated 0.389 kgs/hhld/week more Printed Paper than Small Urban & Rural areas and that the difference is statistically significant with 95% confidence. Conversely, Large Urban areas generated less steel than Small Urban & Rural areas. Table 6 summarizes the seasonal differences in the data. It shows, for example, generation of Printed Paper is significantly higher in the Fall & Spring than in the Summer & Winter. Table 3: 95% Confidence Intervals on the Mean by Material Single Family Samples Number Category Lower Mean Upper Kilograms per household per week 1 Printed Paper 2.4573 2.5179 2.6632 2 Paper Packaging 1.5328 1.5908 1.6446 3 Plastic 1.0494 1.0727 1.1037 4 Aluminum 0.1241 0.1279 0.1345 5 Steel 0.2277 0.2328 0.2465 6 Glass 0.7031 0.7182 0.7777 7 Newspaper 1.5814 1.6050 1.7501 8 Corrugated 0.6026 0.6191 0.6822 9 Boxboard / Cores 0.4914 0.5021 0.5233 10 PET Bottles & Jars 0.1968 0.2015 0.2119 11 HDPE Bottles & Jugs 0.1090 0.1115 0.1181 12 Aluminum Food & Beverage Cans 0.0950 0.0968 0.1049 13 Steel Food & Beverage Cans 0.2020 0.2047 0.2203 14 Clear Glass Alcoholic N/A 0.1631 N/A 15 Coloured Glass Alcoholic N/A 0.2176 N/A 16 Clear Glass Food & Beverage 0.2932 0.2968 0.3213 17 Coloured Glass Food & Beverage N/A 0.0407 N/A Seasonal Variations Analysis of Variance (ANOVA) procedures were used to examine potential differences in the data between seasons. No consistent annual trends could be identified in the data, in part because of changes in the location of the samples from year to year and because of the short period of record. The annual trends that did appear in the data were contradictory, based on small sample sizes, and could be due to factors other than long term trends in waste generation rates (Wilson 2008). ANOVA analysis suggested that there were differences in the mean total waste generation by season for all audit types. Waste generation was lowest in the Winter and highest in the Fall, but none of the differences in the data were statistically significant. Table 6 summarizes the seasonal differences in the single family data. It shows, for example, that generation of Paper Packaging is
significantly higher in the Fall and the generation of PET is higher in the summer. Overall, seasonal variations in the Multi-Family data were smaller than those in the Single Family data. The seasonal differences apparent in the data were smaller than expected. Given the importance and size of seasonal variations discussed in the waste management literature, further investigation of seasonal effects in the data is recommended. It may be that the variations in the waste management literature are not applicable to these data or that the waste samples are collected in such a way that they are missing time periods with higher or lower average waste generation (Wilson 2008). Table 4: 95% Confidence Intervals on the Mean by Material Multi-Family Samples Number Category Lower Mean Upper Kilograms per household per week 1 Printed Paper 1.6658 1.7436 1.8889 2 Paper Packaging 1.1840 1.2546 1.3269 3 Plastic 0.7906 0.8382 0.8732 4 Aluminum 0.0801 0.0865 0.0908 5 Steel 0.1601 0.1696 0.1768 6 Glass 0.4188 0.4449 0.4820 7 Newspaper 1.0474 1.0920 1.2185 8 Corrugated 0.4501 0.4764 0.5461 9 Boxboard / Cores 0.3325 0.3467 0.3624 10 PET Bottles & Jars 0.1475 0.1576 0.1654 11 HDPE Bottles & Jugs 0.0680 0.0713 0.0767 12 Aluminum Food & Beverage Cans 0.0588 0.0632 0.0684 13 Steel Food & Beverage Cans 0.1410 0.1487 0.1566 14 Clear Glass Alcoholic N/A 0.0977 N/A 15 Coloured Glass Alcoholic N/A 0.1339 N/A 16 Clear Glass Food & Beverage 0.1824 0.1888 0.2050 17 Coloured Glass Food & Beverage N/A 0.0245 N/A Program Variables Different municipalities have different systems for managing solid wastes, have different restrictions on what can and cannot be disposed of, spend different amounts on educating and informing citizens about waste management, and run recycling programs of varying effectiveness. All of these factors can be expected to have an impact on overall waste generation within a municipality. To investigate the potential impact of differences in waste management programs, the single family data records in the database were classified according to whether or not the municipality had a Pay As You Throw (PAYT) program in place. Municipalities were classified according to the Waste Diversion Organization Data Call as either having a PAYT or not. The results of this ANOVA analysis suggested that municipalities with a PAYT program have lower mean waste generation rates than those without (12.318 kgs/hhld/yr vs 13.890 kgs/hhld/yr). The results were significant at a 95% confidence interval. Similar results were noted for many individual materials or material categories. For example, PAYT households generate less old newsprint (1.6180 vs 1.3027 kgs/hhld/yr) and more PET (0.17389 vs 0.15147 kgs/hhld/yr). Differences in other waste categories exist, but are not presented here. A similar analysis was conducted for waste management programs with a limit on the number of bags or waste containers that residents are allowed to set out for collection. No significant effects based on bag limits were detected in the data (Wilson 2008).
Table 5: ANOVA Analysis by Area Type Single Family Samples Category Large Urban Small Urban/Rural Difference Significant at 95% Kilograms per household per week Confidence? 1 Printed Paper 2.7180 2.3290 0.3890 Yes 2 Paper Packaging 1.5756 1.6050-0.0294 No 3 Plastic 1.0450 1.0989-0.0539 No 4 Aluminum 0.1223 0.1332-0.0109 No 5 Steel 0.2135 0.2510-0.0375 Yes 6 Glass 0.7148 0.7214-0.0066 No 7 Newspaper 1.7497 1.4688 0.2809 Yes 8 Corrugated 0.6227 0.6157 0.0070 No 9 Boxboard / Cores 0.4872 0.5161-0.0289 No 10 PET Bottles & Jars 0.1941 0.2085-0.0144 No 11 HDPE Bottles & Jugs 0.1109 0.1121-0.0012 No 12 Aluminum Cans 0.0899 0.1034-0.0135 Yes 13 Steel Food & Bev. Cans 0.1836 0.2246-0.0410 Yes 14 Clear Glass Alcoholic 0.1491 0.1763-0.0272 No 15 Coloured Glass Alcoholic 0.2372 0.1992 0.0380 No 16 Clear Glass Food & Bev. 0.2837 0.3091-0.0254 No 17 Coloured Glass Food/Bev. 0.0449 0.0368 0.0081 No Table 6: ANOVA Analysis by Season Single Family Samples Category Differences Observed Significant at 95% C.I.? 1 Printed Paper Fall & Spring higher than Summer & Winter Yes 2 Paper Packaging No significant differences No 3 Plastic Spring & Summer higher than Winter & Fall Yes 4 Aluminum Summer & Winter higher than Spring & Fall Yes 5 Steel Winter highest, Summer lowest Yes 6 Glass No significant differences No 7 Newspaper Fall & Spring higher than Summer & Winter Yes 8 Corrugated Spring highest Yes 9 Boxboard / Cores No significant differences No 10 PET Bottles & Jars Summer highest Yes 11 HDPE Bottles & Jugs No significant differences No 12 Aluminum Food/Bev. Cans Summer highest Yes 13 Steel Food & Bev. Cans Winter highest, Summer lowest Yes 14 Clear Glass Alcoholic Summer highest Yes 15 Coloured Glass Alcoholic No significant differences No 16 Clear Glass Food & Bev. No significant differences No 17 Coloured Glass Food/Bev. No significant differences No Differences Between and Among Municipalities ANOVA techniques also identified significant differences in average waste generation rates between municipalities and within different neighbourhoods in the same municipality. For example, within one city the samples on one street generated 65% more total waste than the samples on a different street. This variation is to be expected, since sample areas were selected with a view to
obtaining sample areas which represent a cross-section of socio-economic backgrounds within each municipality. The difference in generation rate among different areas of the same city was found to be as large as 7 kg/hhld/wk, which is larger than the differences found in the data due to audit type, housing type, season, or annual trend. This table demonstrates that socio-economic differences are very important in determining waste generation and suggests that differences in the data could be better explained by relating it to demographic information on the sample area. Data Correlations As a final check on the data, a simple correlation analysis was conducted. Correlation coefficients were calculated between all waste categories in the database. The results, some of which are shown in Table 7, indicated that there are significant correlations between many of the waste categories. This means that, for example, households that generate large amounts of PET are also likely to generate large amounts of aluminium cans or corrugated cartons. Table 7 shows that although not all of the correlations are strong, most of them are statistically significant. The most likely explanation for this apparent correlation is that many waste generation categories are strongly correlated to the number of people in the household. Interestingly, many of the post-consumer packaging categories did not correlate to waste categories for durable goods such as furniture or electronic equipment. Table 7: Correlation Coefficients between Waste Categories All Audit Types ONP Corr PET HDPE Alum Steel Clear Glass Corr 0.188 PET 0.149 0.425 HDPE 0.163 0.171 0.253 Alum 0.074 0.291 0.486 0.243 Steel - 0.300 0.411 0.197 0.492 Clear Glass 0.193 0.182 0.300 0.334 0.307 0.261 Colour Glass 0.206 0.116 0.184 0.121 0.068 0.085 0.395 RELATING WASTE AUDIT DATA TO ADDITIONAL DATA SOURCES This project also examined the potential for linking the Stewardship Ontario waste audit data to other existing or potential sources of data with a view to improving estimates of waste generation rates. This task was limited to a description of potential sources of additional data for improving province wide waste generation estimates and the development of a framework for an improved waste generation model. The project identified a considerable volume of data that could be related to the Stewardship Ontario waste audit data, usually by linking the sample data to a specific civic address, postal code, or geographic feature (Wilson 2008). Specifically, the project identified online census data from Statistics Canada ( e.g. http://www12.statcan.ca/english/census06/data/profiles/ct/), information from municipal taxation databases and Geographic Information Systems (e.g. mapping and taxation available online from municipalities such as the City of Hamilton, Ontario at http://map.hamilton.ca/interactivemaps.asp), weigh scale records available from municipal waste management departments, and data collected by equipping waste collection vehicles with Global Positioning System (GPS) data recorders (Wilson et al. 2007).
CONCLUSIONS An analysis of the Stewardship Ontario Waste Audit Database and a review of the waste auditing literature resulted in the following findings: 1. Although some minor errors were identified, it was possible to conclude that the data set appears to be complete, robust, and consistent and that Stewardship Ontario's current sample size recommendations are consistent with the available literature and with practise in other jurisdictions. 2. None of the data in the database can reasonably be described by a Normal probability distribution, but a Log Normal probability distribution fits most of the data reasonably well. 3. Although significant seasonal differences were not evident in the data, they were expected. This may be due to a bias in sampling procedures during each season. Some significant seasonal differences were found for specific material categories. 4. Significant differences in the data can be attributed to the classifications currently used by Stewardship Ontario such as Audit Type, Housing Type, and Building Type. 5. Significant differences in the data were also identified using additional classifications of the data. The data support segregation according to details regarding the waste collection system employed by the municipality and according to demographic information about the households and areas sampled. ACKNOWLEDGEMENTS This work was conducted under contract to Stewardship Ontario. Throughout the project I worked closely with John Dixie, Mustan Lalani, Liz Parry, and Guy Perry of StewardEdge Inc.. a company under contract to Stewardship Ontario to provide management services for the Blue Box Program. The assistance of those individuals throughout the project and in the preparation of this paper is gratefully acknowledged. REFERENCES Olsson, U. (2005). Confidence Intervals for the Mean of a Log-Normal Distribution, Journal of Statistics Education Volume 13, Number 1, available online at: www.amstat.org/publications/jse/v13n1/olsson.html. Stewardship Ontario (2005). Guide for Single-Family Waste Audits, available online at: http://www.stewardshipontario.ca/bluebox/pdf/eefund/waste_audit_guide2005_sf.pdf. Wilson, B.G., Agar, B.J., Baetz, B.W., and Winning, A. (2007). Practical Applications for GPS Data from Solid Waste Collection Vehicles, Canadian Journal for Civil Engineering 34(5), 678-681. Wilson, B.G. (2008). Improving the Stewardship Ontario Residential Waste Audit Program: Report 1: Preliminary Findings; Report 2: Audit Methodology and Waste Generation Estimates; Report 3: Results for Specific Material Categories; Reports submitted to Stewardship Ontario by the University of New Brunswick.