United States Environmental Protection Agency is EPA Research and Development Sheboygan Confined Treatment Facility (CTF) Bioremediation Workshop Prepared for Great Lakes National Program Office Region V U.S. Environmental Protection Agency Prepared by Environmental Research Laboratory Athens GA 30613 February 1992 ------- ¦ £36b pp(\ Sheboygan Confined Treatment Facility (CTF) yV Bioremediation Workshop (oOJ Cj- On May 30, 1991, a workshop was convened by the EPA q Environmental Research Laboratory, Athens, GA to review a 1 bioremediation pilot study to be conducted at the Sheboygan $Q\ Confined Treatment Facility (CTF). The objectives of the workshop were to : 1. Conduct a scientific review of the work plan for the pilot bioremediation study and provide recommendations for enhancing the experimental design. 2. Finalize a sampling plan for monitoring the CTF that would produce a statistically supportable documentation of the effectiveness of bioremediation. 3. Set up a plan for conducting a scientific evaluation of the monitoring data in order to determine loss of PCBs due to biodegradation as compared to losses resulting from other mechanisms. Cn) CO Participants comprised representatives from Blasland and Bouck (site contractors), Michigan State University, Wisconsin ^ Department of Natural Resources, EPA Region V, and the EPA Environmental Research Laboratory - Athens. A complete list of participants can be found in Appendix A. vO o Dr. Mark Brown of Blasland and Bouck Engineers started the meeting with a description of the bioremediation pilot study. The opening remarks were followed by an open discussion of the study. What follows is a summary of the workshop. A CTF has been constructed on the premises of the Tecumseh Products Company, Sheboygan Falls, WI by Blasland and Bouck Engineering. The facility is constructed of steel sheet piling (105 ft x 135 ft x 10 ft) lined with welded sheets of high- density polyethylene into which are inlaid pipe grids for detection of leaks and injection of amendment. The structure is divided in half to permit testing of two treatments, and within each half, approximately one-fourth the total area has been segregated and reserved as a treatment control. In addition, several materials are being tested as adsorbents for PCBs in the outflow water. Sediments were mechanically dredged from the Sheboygan River and transported to the facility in leak-proof containers. Although each load of sediment was not characterized before deposition in the facility, comparability of material in each treatment/control pair was attempted by delivering similar 1 ------- proportions of sediment from each of the dredging sites. Cells 1 and 2 were considered to have similar sediment/water compositions; cells 3 and 4 also were considered to be similar. However, cells 3 and 4 may be different from cells 1 and 2. Cells 2 and 3 are approximately one third the size of cells 1 and 4, respectively. Cells 2 and 3 were meant to serve as controls for cells 1 and 4, respectively. The requirements of mechanical dredging preclude mixing of the sediments before deposition in the CTF. Thus, the heterogeneous nature of the material in the facility introduces high variance into measurements of baseline levels of contaminants and responses to treatments. The sampling of sediments in the CTF cells has been limited to cell 4. Cell 4 was sampled by hand-held corers. A composite sampling approach was used to determine PCB concentrations. Composites were prepared from samples collected randomly from among a set of systematic samples taken from the cell. The composite samples were thoroughly mixed and then the mixture was subsampled for analysis. Material from seven cores were mixed to create seven composite samples. Only one subsample has been analyzed per composite sample thus far. The data so derived were utilized to estimate the mean PCB concentration in the cell, and the sample variance was calculated to measure variability. The mean PCB concentration of the composite samples was 325 +- 175 ppm. Cells 1, 2, and 3 were not sampled. This information was used to project the number of samples needed to detect 50% degradation in PCB levels. Several areas within the river have been armored by Blasland and Bouck Engineering to keep PCB contamination in place and eliminate transport and dissolution as loss mechanisms for sediment PCBs. Four areas of sediment having PCB concentrations ranging from 25 to 270 ppm were armored, and removable sections were installed for access to the underlying sediments. The rate of degradation in the armored area may serve as an indication of in situ rates and as a point of comparison for rates in the CTF. The CTF was constructed while bench-scale experiments were being conducted at the University of Michigan; the most promising treatments from the bench-scale work are to be implemented and tested in the pilot facility. Aerobic experiments were conducted on partially dechlorinated sediments and anaerobic experiments were performed with contaminated sediments having low (10 ppm) and high (500 ppm) PCB concentrations. Results to date suggest that the major component of the original contaminant was Aroclor 1248 and that dechlorination is occurring at this contaminated area of concern. Data collected from 1989 to 1991 support this suggestion since a . plot of the number of ortho- versus meta- and para-chlorines showed a shift away from the standard plot of Aroclor 1248. Dechlorination occurred at the meta- and para-positions of highly 2 ------- chlorinated congeners which resulted in the accumulation of less chlorinated ortho-substituted congeners. A faster rate of dechlorination was observed in sediments with higher concentrations of PCBs than the less contaminated sediments. Only mono- and dichlorinated congeners were degraded in the aerobic experiments, but the higher chlorinated congeners were degraded in the anaerobic experiments. The effect of amending anaerobic sediments (artificially contaminated with Aroclor 1242) with inorganic and organic nutrients also was investigated in bench scale experiments. Experimental treatments tested included mineral (RAMM medium) and secondary substrate (methanol and acetone) additions. There Was little conclusive evidence that addition of organic substrate plus RAMM medium enhanced rates of PCB (500 or 10 ppm) dechlorination at an incubation temperature of 30 °C. The rate observed of dechlorination was not enhanced significantly by these two amendments, and an equivalent rate of dechlorination was observed in the control vessels without addition of organic substrates. Data from non-RAMM controls were not available. Non-RAMM control experiments were initiated in May 1991. Future experiments are designed to increase the bioavailibility of sorbed PCBs. Surfactants, especially Triton X-705, will be used to increase the solubility of PCBs. Methods for transforming anaerobic sediments to aerobic conditions will be studied since the products of anaerobic dechlorination may serve as substrates for aerobic bacteria. In particular, the rate and amount of hydrogen peroxide needed to transform anaerobic to aerobic sediments will be investigated. In discussions that followed the introduction by Dr. Mark Brown, three major areas were addressed: the sampling design for evaluating treatment options, the treatment options to be tested this year, and the alternative treatments that could be pursued. Concern was voiced by an EPA statistician and others that composite samples obscured intersample variation and that the intracomposite variance had to be established (by subsampling) for valid comparisons of values. A composite sampling approach probably can be used advantageously for estimating mean PCB concentrations in the CTFs, but its use must be treated carefully in order that underestimation of variances or standard errors is avoided. Underestimation could lead to declaring significant differences where they do not really exist. It is important to realize that there are different sources of variability in this experiment and there has been no attempt to estimate or compare them. This means that such composites cannot be used to fine-tune a proposed design, as might be required when determining how many individual samples should be combined per composite sample, how many 3 ------- composite samples should be formed, how many subsamples should be collected from each, whether the mixing process is adequate, and whether the laboratory-based 10-gram subsampling is acceptable. A study to address all of these questions apparently would be prohibitively expensive. (A note on composite sampling can be found in Appendix B.) Because details of the computations were not available at the meeting, it is not possible to determine whether such calculations have been correctly implemented. This should be checked thoroughly inasmuch as small sample sizes have been suggested as being adequate. The number of proposed composite samples (five) does seem to be woefully small. The idea of using a 50% reduction in PCB levels as the tolerable difference (i.e., a difference to be detected with a given confidence level) seems to be too unrefined. It is a scientific question as to how long a 50% reduction is likely to take, and whether it is of interest to detect smaller changes. If smaller changes are of interest, there will be a direct impact on the number of samples required. If it is possible, available data should be utilized to estimate variance components and/or standard errors, and simulations should be conducted to investigate the proposed sampling strategies. This modification was subsequently incorporated into the sampling design. Subsequent to the meeting Dr. Mark Brown proposed (Appendix C, Dr. Rudy Parrish's response Appendix D) a sampling plan that would use 5 seven-core composites from each of the cells. For one composite in each cell, triplicate subsamples would be collected. The relative standard deviations would be compared to determine whether an alternative composite sampling design is needed. Appendix B contains applicable expressions for the variance of the estimate of the cell mean (Eq.4). This expression involves four variance components. Fortunately, it is not necessary to estimate these components individually in order to obtain an estimate of the variance of the estimator; however, the computation is facilitated by use of analysis of variance calculations whenever the data are balanced. Balanced, here, means that the same number of subsamples has been taken from each composite. In the unbalanced case, it is not clear exactly what the expressions should be, although they can probably be derived. If the cores are sufficiently large relative to the subsamples, then "S" will be large relative to "s", so that the quantity "(S- s)/(S-l)" is approximately the same as "S/(S-1)M. If this is, in fact, the case, then (EMSB - EMSJ can be used to produce an estimate of the second term of Eq.4 that is associated with the composite-related variance components. It seems that the proposal is to calculate the variances of the cell means for subsampling a single time versus subsampling three times, and then compare the results. While that might be appropriate from a general point of view, it might be better to attempt to estimate 4 ------- the variance components for: (1) between increments, (2) measurement error, and (3) the nonlinear combination of components due to imperfect compositing. The ANOVA-based approach might work well here if the design can be balanced. The unfortunate aspect of this approach is that design questions cannot be rigorously addressed without a knowledge of all of the individual variance components. Estimation of these requires a more complex (and expensive) design. Using the procedure above (with the simplifying assumption), one could investigate effects of differing numbers of subsamples, since ns" is a divisor in the second term of Eq.4. Similarly, one could investigate the effects of differing numbers of subsamples on the number of composite samples NrN. The effect of changing the number of samples per composite, however, cannot be examined in the same manner because "n" occurs nonlinearly. It would also be helpful to consider unbalanced method in terms of the variance components. Without that, it seems that a better approach would be to utilize an equal number of subsamples from each composite. Of course, it might also be possible to derive appropriate expressions for the unbalanced case. There are several features of the design of the CTFs that could have an impact on any experimental study that is conducted at the site. The CTFs are constructed asymmetrically to provide four cells, each of a different size. Statistical concerns include the following. (1) There might be a CTF-size effect that could interfere with treatment-control or treatment-treatment comparisons. (2) The composition of cells 1 and 2 and that of 3 and 4 might not be sufficiently similar to justify treating the smaller cells as controls for the corresponding larger cells. (3) The sediments in the cells were not mixed prior to being added, so that high spatial variability could be expected. There are also varying sediment depths. This impacts the level of sampling required for any given comparison or parameter estimation. (4) Having only four cells limits capability to use experimental designs that involve replication. A suggestion here is that perhaps cells 1 and 2 and cells 3 and 4 may be considered to form two blocks with two homogeneous experimental units (i.e., cells) in each. One further aspect of the design that should be mentioned is that the effectiveness of a given amendment option will be measured primarily on the basis of a change in concentration in a 5 ------- single cell over some time period, perhaps compared to similar observations in a corresponding control cell. There is only one experimental unit per option, so that any observed differences may, in fact, be confounded with the composition of the cell being used. There is no replication, which is needed to estimate the experimental error for testing differences among treatments and, thus, to make sound inferences. It might be possible to improve on the limitation of having only four compartments (cells) in the CTF. Cells 1 and 4 are large enough that they could be subdivided to form a total of eight cells in the facility. If the sediment composition in these could be distributed more uniformly among such cells, there would be wider latitude available for utilizing different experimental designs involving more replication (thus, higher power) and more treatments. Several concerns were raised about the study's capability to evaluate the potential for biodegradation of endogenous PCBs and also to develop new techniques that will enhance bioremediation of the Sheboygan River and Harbor area of concern. These studies are currently going on in the laboratory of Dr. Timothy Vogel at the University of Michigan. The laboratory experiments with Sheboygan sediments were conducted at 30aC. The justification for choosing this temperature was probably to obtain somewhat faster dechlorination rates to allow more rapid evaluation of the effects of different amendments, etc. The optimum temperature for dechlorination by the upper Hudson River microorganisms, however, is between 20 and 25*C, and para-dechlorinatinq activity is lost at 30*C. Thus, conducting laboratory experiments with the Sheboygan sediments at 30°C does not necessarily yield any advantage (in terms of more rapid assays), and has the obvious disadvantage that the results are not directly transferable to the field. While the Blasland and Bouck representatives did not have information on actual environmental temperatures, the sediments in the CTF are undoubtedly lower than 30°C. Average summer and winter temperatures for different depths in the CTF should be determined and future laboratory experiments conducted at comparable temperatures ranges. Blasland and Bouck Engneering has so far determined oil and grease levels for only a few of the Sheboygan sediment samples. Oil and grease have been shown to inhibit the aerobic degradation of PCBs by reducing their bioavailability. The same is likely true for anaerobic dechlorination. Results from Dr. John Quensen's laboratory at Michigan State University suggest that oil and grease levels above approximately 1,000 ug/g sediment (dry weight) begin to decrease the rate of dechlorination. Levels of 10,000 ug/g strongly inhibit dechlorination. 6 ------- Variations in oil and grease concentrations may account for the various extents of Is situ dechlorination observed in the Sheboygan sediments. The higher levels could limit dechlorination within the CTF. Sediments should be sampled routinely for oil and grease. If levels are, in fact, high enough to limit the dechlorination process, greater overall rates of PCB transformation may be achieved by cycling between aerobic and anaerobic conditions. The oil and grease should be degraded during the aerobic cycles. Several treatment options to address the question of availability of sediment-sorbed PCBs were discussed. Sediment organic matter, especially sediment oil and grease, strongly sorb PCBs, thus decreasing the rate of degradation. It was suggested that these organic materials might be decreased during an aerobic degradation period preceding the anaerobic treatment (e.g., 2 months aerobic followed by 8 months anaerobic). The sensitivity of the dechlorinating microorganisms to such cycling would need to be evaluated, but at least some of them are now known to be spore formers. Moreover, brief (30 minute) aeration of Hudson River inoculum did not diminish its subsequent activity under anaerobic conditions. The mechanics of altering the aeration status of the CTF sediments were discussed; the plumbing arrangement for the facility makes drainage difficult. Moreover, drainage water would have to be treated before discharge. The possibility of adding dissolved oxidants, including peroxides, as a means of shifting Oxidation status was also addressed. The final selection of treatment regimes for the CTF will be made pending further discussions among representatives of Blasland and Bouck Engineering, EPA scientists from the Office of Research and Development, and EPA Region V representatives. The aerobic degradation of PCBs is a co-metabolic process. The responsible microorganisms are induced with biphenyl and in some cases with mono-chlorobiphenyls. An attractive feature of an anaerobic-aerobic biotreatment sequence is that levels of mono-chlorobiphenyls may be achieved anaerobically that can induce the co-metabolism of other congeners when conditions become aerobic. The Sheboygan sediments, however, do not appear to accumulate the high levels of mono-chlorobiphenyls required. Thus aerobic degradation of the dechlorination products, in the absence of another inducer such as biphenyl, is likely to be quite limited. A possible solution is to inoculate the sediments with microorganisms that produce more monochlorobiphenyls, such as upper Hudson River microorganisms. This, however, may not be easily achieved< Drs. Vogel and Nies from the University of Michigan attempted such an inoculation by mixing Hudson River and Sheboygan sediments. They found this to have no effect on the 7 ------- extent of dechlorination. Similar results were obtained by Dr. John Quensen by mixing inocula from the upper Hudson River and Silver Lake. The Silver Lake microorganisms normally produce high levels of 2,4,2',4'-tetrachlorobiphenyl from the dechlorination of Aroclor 1260. Rather than dechlorinating the 2,4,2•,4'-PCB produced by the Silver Lake microorganisms, the upper Hudson River microorganisms apparently inhibited the activity of the Silver Lake microorganisms. Further attempts should be made at inoculating the Sheboygan sediments to achieve higher levels of mono-chlorobiphenyls. Sediments from sites other than the upper Hudson River should be tried. More work needs to be done to further the understanding of compatibility and incompatibility of different PCB dechlorinating activities. Blasland and Bouck engineers proposed several criteria besides total PCB concentration for assessing the progress of biotreatment in the CTF and armored areas in the river. These included toxic equivalency, total concentration of key dechlorination products, and average number of meta- and para- chlorines per biphenyl. These criteria emphasize different parameters, but are actually different ways of reducing the raw data consisting of congener-specific PCB analysis. The first two criteria are related to health and environmental hazards, whereas the last two are related to the extent of anaerobic dechlorination. All the above criteria should be considered (and all raw data retained for further possible evaluation after the experiment is finished). It appears that different criteria should be applied to the CTF and the armored areas. Armoring is essentially a technique for reducing risk (by decreasing the chance of exposure). Decreases in toxic and bioaccumulative congeners would further reduce risk, and should be the major criteria applied to the armored areas. In the case of the CTF, there is the additional goal of reducing total PCB concentrations. This means creating aerobic conditions at the most effective time. Therefore, criteria related to the progress of anaerobic dechlorination are also pertinent. The observed increase in the proportion of mono- chlorinated biphenyls is probably most important as these congeners may induce the aerobic co-metabolism of other PCBs still present as indicated previously. One could go a step further than Blasland and Bouck engineers proposed and, based on the relative degradabilities reported in the literature for the various congeners remaining, calculate the expected decrease in total concentration to be expected if a switch were made to aerobic conditions. PCB degradation rates appear to be slightly higher In situ than in the CTF, although, as mentioned, the values from the armored areas do not reflect the overall variability of sediment 8 ------- rates, which may not be statistically different from the CTF rates. Given the few openings and the small areas that are intensively sampled at the armored sites, the data from these areas must be interpreted cautiously. In summary the CTF provides a valuable outside laboratory for investigating technologies for the biological treatment of PCBs. One should be cautioned, however, that the selection of an appropriate sampling design is critical for the development of sufficient data to adequately evaluate each treatment option. Because of the size and construction of the facility, side-by- side testing of technologies would be difficult. Current research would suggest that several treatment options are available for testing. These options include addition of inorganic nutrients, organic nutrients, aerobic PCB-degrading bacteria, and the use of alternating aerobic/anaerobic systems. The proper selection of test conditions should result from the careful analysis of laboratory data. 9 ------- Appendix A. List of Participants Rebecca Adams Technology Applications, Inc. c/o Environmental Research Laboratory U s EPA Athens GA 30613-0801 404-546-3517 Rochelle Araujo Environmental Research Laboratory U. S. EPA Athens GA 30613-0801 404-546-3468 Mark Brown Blasland & Bouck Engineers, P.C. 6723 Towpath Road, Box 66 Syracuse NY 13214 315-446-9120 Dave Cowgill Great Lakes National Program Office 230 South Dearborn Chicago, IL 60604 312-353-3576 Bonnie Eleder Region V, US EPA 230 South Dearborn Street Chicago, IL 60604 312-886-4885 Terry Evanson WI Department of Natural Resources 101 S. Webster PO Box 7921 Madison, WI 53707 608-266-0941 Dawn Fester Blasland & Bouck Engineers, P.C. 6723 Towpath Road, Box 66 Syracuse NY 13214 325-446-9120 Jack Jones Environmental Research Laboratory U.S. EPA Athens GA 30613-0801 404-546-3468 10 ------- Rudy Parrish Computer Sciences Corp. c/o Environmental Research Laboratory U.S. EPA Athens GA 30613-0801 404-546-3604 John Quenson Department of Crop and Soil Sciences Michigan State University East Lansing MI 48823-1325 517-355-9285 John Rogers Environmental Research Laboratory U.S. EPA Athens 6A 30613-0801 404-546-3592 Robin Schmidt WI Department of Natural Resources 101 S. Webster PO BOX 7921 Madison, WI 53707 608-267-7569 Patricia Van Hoof University of Georgia c/o Environmental Research Laboratory U.S.EPA Athens GA 30613-0801 404-546-3349 Daniel Wubah University of Georgia Environmental Research Laboratory U.S. EPA Athens GA 30613-0801 404-546-3182 11 ------- Appendix B. A Note on Composite Sampling Suppose r x n equal-sized portions of sediment material (increments) are drawn randomly from a cell. From these, £ composites are formed by randomly partitioning the set of increments into £ subsets of n increments each and physically mixing the increments in each subset. Then, s subsamples are drawn randomly from each composite, and £ analyses are run on each subsample. Each composite is assumed to contain enough material for S subsamples. (If the subsample size were the same as the increment size, then £ = n*) To mathematically represent the variables involved in the composite sampling model, the following notation is defined: = the proportion of the jth subsample from the ith composite that comes from the 1th increment in that composite, ¦ the concentration value associated with the portion of the 1th increment that appears in the jth subsample from the 1th composite. Each variable a,^ has mean na and variance cra2. The variables 2£fjl have equal means and variances ax2 + a J, where ax2 is the between-increment variance and cth2 is the within-increment variance subsequent to compositing and subsampling. In different composites, the variables are uncorrelated, but they are correlated within the same increment, in the same subsample, and in different increments and subsamples within a composite. Similarly, the variables 2jji are correlated within the same increment, in the same subsample, and in different increments and subsamples within a composite. Elder et al. (1980) provided explicit expressions for these covariances. The concentration value associated with the jth subsample from the 1th composite is based on contributions from all increments in that composite. It can be expressed as the weighted mean The value observed in the laboratory from the Jcth analysis, however, is D (l) 12 ------- zijk ~ y±j + eijk (2) where is the measurement error associated with the Jcth test on the j.th subsample from the 1th composite. The variance of ejjk is defined as ae2. If the analysis procedure requires that additional subsampling be done, then an additional variance component should be incorporated. The estimator of px for composite sampling is * = ^ £ zij*/rst ¦ <3> As a result of the covariance structures for the 2{jL and X,- ,• t variables, the variance for this estimator involves four individual variance components. Elder et al. (1980) provided Vaz{z) =ol/(rn) + [ (S-s) / (5-1) ] [ol/n + no\ (o2x + oI)]/(ts) + al/ (rst) . Also, they established that the expected mean squares (between groups, within groups, and for error) associated with the analysis of variance for the composite sample groups are: EMSb = (8t/n)al * [t(S-s)/(S-l)] [a2w/n + na\(al + al)) * o| (5) EMS„ = [ tS/ (5-1) ] [a2w/n + nal(a2x + al) ] + o| (6) EMSb = al (7) Clearly, the variance of the estimator £ is equivalent to the value EMS,/(rst). Thus, the standard error of the estimate of nx can be obtained from a standard analysis of variance computation. References Brown, G. H. and N. I. Fisher. 1972. Subsampling a mixture of sampled material. Technometrics 14(3): 663-668. 13 ------- Elder, Robert S., William 0, Thompson, and Raymond H. Myers. 1980. Properties of composite sampling procedures. Technometrics 22(2): 179-186. Gilbert, Richard O. 1987. Statistical Methods for Environmental Pollution Monitoring. Van Nostrand Reinhold Company, New York. 320 pp. Parrish, R. S., G. 0. Ware, C. N. Smith, and P. A. Banks. 1990. Composite sampling techniques for determining pesticide concentrations. Proceedings of the KSU Conference on Applied Statistics in Agriculture. Manhattan, KS. Rhode, Charles A. 1976. Composite sampling. Biometrics 32, 273- 282. 14 ------- APPENDIX C 15 ------- BLASLAND & BOUCK ENGINEERS; P.C. ENGINEERS iOEOSCIENTISTS 6723 Towpath Road, Bpx 66, Syracuse, New York 13214 (315) 448-9120 FAX: (315) 449-0017 June 13, 1991 Rudy Parrish Environmental Research Laboratory U.S. EPA Athens, GA 30613 Re: Bioremediation Pilot Study Work Plan File: 176.06 #2 Dear Mr. Parrish: We appreciate your review and discussion of sampling design as contained in the Sheboygan River and Harbor Bioremediation Pilot Study Work Plan. At the fecent workshop in Atlanta it appeared that you and I reached an understanding that composite sampling was an acceptable approach. It was clear, however, that your acceptance of the approach was contingent upon an explicit assessment of the errors associated with subsampllng the composite samples. This assessment may be useful in revising our sampling approach if the errors are large relative to the errors associated with our current methods. If the errors are large, we would conduct a numerical analysis to reassess the sampling requirements balancing the number of seven-core composite samples against the number of composite subsamples tc minimize costs for a specified hypothesis-testing power. We are not sure that some of the other workshop attendees are as yet comfortable with use of composite samples for gauging the overall progress of the experiment^). A comment which was repeatedly offered Is that our compositing scheme obscured the real spatial variability of PCB concentrations within the cell and that this variability needs to be explicitly considered. We do not dispute the first contention. On the contrary, the intended purpose of compositing was to cope with the expected high degree ot spatial variability and the expense of analyzing large numbers of samples. For the benefit ot others, we are providing as an attachment, an example of how one might-approach a sampling design from error estimates for various contributors. The approach Illustrated by Neptune et al. can be compared, and contrasted to our approach in developing a sampling design. The fundamental objective of both approaches was to produce a sampling design with an acceptable power for hypothesis testing. The essential difference is that by estimating the individual errors, the EPA authors could optimize the number of individual samples per composite sample and the numbers of 16 Syracuse, NY • Rochester, NY • While Plains, NY • Syos.net, NY • Edison, NJ » Washington, DC ~ Columbus. OH Boca Raton, FL • Tampa. FL * Orlando. FL ------- Rudy Parrish June 13, 1991 Page 2 composite samples per area (note that sampling errors associated with subsampiing of composite samples Were ignored ior composites of live or (ewer samples) to produce the most cost-effective design. In contrast, we presumed that the spatial variability would be high and made an educated guess as to the number of cores required per composite sample. Based upon the actual distribution of PCB results for the composite samples, we then back calculated the requisite number of composite samples for a specilied power. Having obscured the actual spatial variability (in this case, the distribution of average PCB concentrations in individual cores), we have no basis for altering the number of cores within a composite sample. So, for the time being, we must stay with the seven-core composites. For the upcoming sampling, we propose to collect five seven-core composite samples from each of the four cells. In addition, we propose to submit triplicate subsamples of one composite sample from each of the three cells. The relative standard deviations of each of these composites would reflect not only our field errors in subsampiing composite samples, but also subsampiing errors by the laboratory and subsequent analytical errors. If these relative standard deviations are comparable to the relative standard deviations among composite samples within a cell, we would reconsider our sampling design based upon a numerical analysis (computer simulation ot error distributions using random normal deviates). We would potentially recommend a revised balance of composite samples and subsamples of ^composite samples to minimize the costs for an acceptable hypothesis-testing power. Please call me if you Wish to discuss this. We look forward to continued cooperation with EPA's ARCS program. Very truly yours, :rs, P.O. Manager MPB/nlg Attachment cc.: Bonnie L. Eieder, Environmental Protection Agency Robin R. Schmidt, Wisconsin Department o( Natural Resources Dawn S. Foster, P.E., Blasland & Bouck Engineers, P.C. John Rogers* Ph.D., Environmental Protection Agency Chad Jafvert, Ph.D., Environmental Protection Agency 17 * ii: I « fcotirf ENGINEERS, P.C. ------- APPENDIX D 18 ------- Computer Sciences Corporation c/o U.S. Environmental Protection Agency .* College Station Road 't'Athens, OA 30613-7799 • Phone: 404-546-3604 22 July 1991 Mark Brown, Ph.D. Blasland & Bouck Engineers, P.O. 6723 Towpath Road, Box 68 Syracuse, NY 13214 Dear Mark: Since talking with you a few weeks ago, I spoke briefly with John Rogers about the Sheboygan project and your letter to me of June 13. Earlier, I provided John With summary fcommerrts pertaining to the May 30th meeting In Atlanta; enclosed Is a copy of that memorandum tor your reference. It had been my understanding that John Would be Communicating those and other comments to you. The foflowtng remarks address your proposed Work, as described In your letter. I realize that you may already have begun this work. Your plan Indicates that you will use 6 seven-core composites from each of the cells, and for one composite In each cell, you wilt collect tripicate subsamples. Then, compare relative standard deviations to determine whether an alternative composite sampling design Is needed. The enclosed "Note on Composite Sampling* contains applicable expressions for the variance of the estimate of the cell mean (Eq.4). This expression Involves four variance components. Fortunately, It Is not necessary to estimate these components Individually In order to obtain an estimate of the variance of the estimator, however, the computation Is faclltated by use of analysis of variance Calculations whenever the data are balanced. Balanced, here, means that the same number of subsamples has been taken from each composite. In the unbalanced case, ft is not dear exactly what the expressions should be. although they can probably be derived. H the cores are sufficiently large relative to the subsamples, then *S* wttt be large relative to*s\ so that the quantity *(S-s)/(S-1)' Is approximately the same as *8/(S-1)\ If this Is, tot tad, the case, then (EMSB • EMSJ can be used to produce an estimate of the second term of Eq.4 that Is associated with the composite-related variance components. H seems that your proposal Is to calculate the variances of the cell means for subsampilng a single time versus subsampllng three times, and then compare. While that might be appropriate from a general point of view, it might be better to attempt to estimate the variance components for: (1) between increments, (2) measurement error, and (3) the nonlinear combination of components due to imperfect compositing. The ANOVA-based approach might work well here If the design can be made balanced. The unfortunate aspect of this is that design questions cannot be rigorously addressed without a knowledge of all of the individual variance components. Estimation of these requires a more complex (and expensive) design. Using the procedure above (With the simplifying assumption), one could Investigate effects of differing numbers of subsamples, since V is a divisor in the second term of Eq.4. SlmBariy, for the number of composite samples V. The affect of changing the number of samples per composite, however, cannot be examined in the same manner because "n* occurs honllnearty. Your procedure, I think, wotid be comparing two expressions based on Eq.4. it would be helpful to consider your method In terms of the variance components. Without that, it seems to me that a better 19 ------- approach would be to LJtBize an equal number of lubsamples from each composite. Of course. It might also be possible to derive appropriate expressions for the unbalanced case. Hopefully, these comments w>l be of Interest and of use to you. Please let me know If you would Rke to discuss this further. Cordially yours, Rudolph S. Parrlsh, Ph.D. Statistician 20 ------- June 10,1991 To: John Rogers Fm: Rudy Punish Re: Sheboygan Bioremediation Meeting Held in Atlanta May 30,1991 As you requested, this It to summarfce my impressions of the Sheboygan bioremediation work discussed at the meeting in Atlanta od May 30. these remarks are limited to statistical aspects of the experimental design proposed for the pilot study. Design of confined treatment facility (CTF) There are Several features of the design of the CTFs that could have an impact on any experimental study that is conducted at the site. The CTTFs art constructed asymmetrically to provide four cells, each of a different size. It was indicated that cells 1 and 2 have generally the same sediment/water composition, and similarly for cells 3 and 4. Cell 2 is of a size that is in approximately a 13 ratio to cell L Similarly, the size of cell 3 is in a 13 ratio to that of cell 4. Cells 2 and 3 were meant to serve as controls for cells 1 and 4, respectively. Statistical concerns include the following. (1) There might be a CTF-size effect that cotild interfere with treatment-control or treatment-treatment comparisons. (2) The composition of cells 1 and 2 and that of 3 and 4 might not be sufficiently similar to justify treating the smaller cells as controls for the corresponding larger cells. (3) The sediments in the cells were not mixed prior to being added, so that high spatial variability could be expected. There are also varying sediment depths. This impacts the level of sampling required for any given comparison or parameter estimation. (4) Having only 4 cells limits capability to use experimental designs that involve replication. A suggestion here is that perhaps cells 1 and 2 and cells 3 and 4 may be considered to form two blocks with two homogeneous experimental units (ie^ cells) in each. Sampling designs Tb* sampling of sediments in the OTP cells appears to have been limited to cell 4. The sampling approach has involved composite sampling techniques wherein several samples are collected randomly from among a set of systematic samples taken from the Cell, the samples are combined physically, and then the mixture is subsampled. Only one subsample has been utilized per composite sample thus far. The data so derived were utilized to estimate the mean concentration in the tell, and the sample variance was calculated to measure variability. This information was used to project sample sizes needed to detect 50% degradation in PCS levels. Because details of the computations Were not available at the meeting, it is not possible to determine whether such calculations have been Correctly implemented. This should be checked thoroughly inasmuch as small sample sizes have beefl Suggested as being adequate. The number of proposed composite samples (five) does seem to be Woefully too small. The idea of using a 50% reduction in PCB levels as the tolerable difference (Le* 1 difference to be detected with a given confidence level) 21 ------- seems to be too unrefined. It is", a scientific question as to how long a 50% (eduction is likely to take, and whether it is of interest to (feted smaller changes. If smaller changes are of interest, there will be a direct impact on the number of. samples required. A composite sampling approach pfpbably can be used advantageously for estimating means in the CTFs, but its use must be treated carefully in order that underestimation of variances or standard errors is avoided, which could lead to declaring significant differences where they do not really exist. It is important to realize that there are different sources of variability in this experiment and there has been bo attempt to estimate or compare them. This means that such components cannot be used to fine-tune a proposed design, as might be required when determining how many individual samples should be combined per composite sample, how many composite samples should be formed, how many subsamples should be collected from each, whether the mixing process is adequate, and whether the laboratory- based 10-gram subsampling is acceptable. A Study to address all of these questions apparently would be prohibitively expensive. (A note on composite sampling is attached.) If it is possible, available data should be Utilized to estimate Variance components and/or standard errors, and simulations should be conducted to investigate the proposed sampling strategies. Modification of the CTF One further aspect bf the design that Should be mentioned is that the effectiveness of a given amendment option will be measured primarily on the basis of a change in Concentration in a single cell over some time period, perhaps Compared to similar observations in a corresponding control cell. There essentially is only one experimental Unit per Option, so that any observed differences may, in fact, be confounded with the composition of the cell being used. There is no real replication, Which is needed to estimate the experimental error for testing differences among treatments and, thus, to make sound inferences. It might be possible to improve on the limitation of having only four compartments (cells) in the CTF. Cells 1 and 4 are large enough that they could be subdivided to form a total of eight cells in the facility. If the sediment composition in these could be distributed more uniformly among such cells, there would be wider latitude available for utilizing different experimental designs involving more replication (thus, higher power) and more treatments. 22 ------- A^Note on Composite Sampling Suppose rxn equal-sized portions of sediment material (increment*) are drawn randomly from a cell. From these, t composites are formed bytandomly partitioning the set of increments bto t subsets of n increments each and physically mixing the increments in each subset, then, I subsamples are drawn randomly from each composite, and (analyses are rati on eich subsample. Each composite is assumed to contain enough material for S subsamples. (If the Subsample size Were the same as the increment size, then 5 ¦ n.) To mathematically represent the variables Involved In the composite sampling model, the following notation is defined: 0p * the proportion tf the /th subsample from the Ith Composite that comes from the /th increment in font composite, X9 « the concentration Value associated with the portion of the /th increment that appears b the/th subsample from the Ith composite. Each variable has mean /*, and variance the variables have equal means p, and variances of 4 where o* is the between-ittaement variance and a, is the within-increment variance subsequent to compositing and subsfcmpling. In different composites, the variables fly ate ttncorrelated, but they are correlated within the same bcrement, b the same subsample, and b different increments and subsamples within a composite. Similarly, the variables X9 are correlated within the same bcrement, b the same subsample, and b different bcrements and subsamples within a composite. Elder et el. (1980) provided explicit expressions for these covariances. The concentration value associated with the /th subsample from the /th composite is based on contributions from all bcrements b that composite. It can be expressed as the weighted mean where e. is the measurement error associated with the Jtth test on the /th subsample from the Ith composite. The variance of e* is defined as If the analysis procedure requires that additional subsampling be done, then an additional variance component should be bcorporated. The estimator of ^ for composite sampling Is (1) The value observed b the laboratory from the tth analysis, however, is (2) f t t l-l M ft-i <3) 23 ------- As a result of the covariance structures lor the and X$ Variables, the variance for this estimator involves four individual variance components. Elder et ai. (1980) provided (to© . aim ~ l(.Ss)t(S-molln * no] (oj ~ ®i)V(rs) ~ o]l(rst). «> V S Also, they established that the expected mean squares (between groups, within groups, and for error) associated with the analysis of variance for the composite sample groups are: EMSt * (siln)*\ ~ [KS-sW-mol/n ~ no\(p\ + oj)] «¦ o] <*> EMSW m ItSKS-l)][olln 4- «o*(o* + a2J] + o] <«) EMSs - «rj <*> Clearly, the variance of the estimator I is equivalent to the value EMSB/(rst). Thus, the standard error of the estimate of /<, can be obtained from a standard analysis of variance computation. References Brown, O. H. and N. L Fisher. 1972. Subsampling a mixture of sampled material Technometrics 14(3): 663-668. Elder, Robert S., William O, Thompson, and Raymond H. Myers. 1980. Properties of composite sampling procedures. Technometrics 22(2): 179-186. Gilbert, Richard 0.1987. Statistical Methods for Environmental Pollution Monitoring, Van Nostrand Reinhold Company, New York. 320 pp. Parrish, R. S., O. O. Ware, C. N. Smith, and P. A. Banks. 1990. Composite templing techniques for determining pesticide concentrations. Proceedings of the KSU Conference on Applied Statistics in Agriculture! Manhattan, KS. Rhode, Charles A. 1976. Composite sampling. Biometrics 32,273-281 ,24 ------- |