1manly_b_f_j_statistics_for_environmental_science_and_managem
.pdf
10
Censored Data
10.1 Introduction
Censored values occur in environmental data most commonly when the level of a chemical in a sample of material is less than the limit of quantitation (LOQ) or the limit of detection (LOD), where the meaning of LOQ and LOD depends on the methods being used to measure the chemical (Keith 1991, chap. 10). Censored values are generally reported as being less than detectable (LTD), with the detection limit (DL) being specified.
Statisticians question why censoring is done just because a measurement falls below the reporting limit, arguing that an uncertain measurement is better than none at all (Lambert et al. 1991). Despite these arguments, it seems that censored values are inevitable in the foreseeable future in environmental data sets.
10.2 Single Sample Estimation
Suppose that there is a single random sample of observations, some of which are below the detection limit, DL. An obvious question then is how to estimate the mean and standard deviation of the population from which the sample was drawn. Some of the approaches that can be used are:
1.With the simple substitution method, the censored values are replaced by an assumed value. This might be zero, DL, DL/2, or a random value from a distribution over the range from zero to DL. After the censored values are replaced, the sample is treated as if it were complete to begin with. Obviously, replacing censored values by zero leads to a negative bias in estimating the mean, while replacing them with DL leads to a positive bias. Using random values from the uniform distribution over the range (0, DL) should give about the same estimated mean as is obtained from using DL/2, but it gives a better estimate of the population variance (Gilliom and Helsel 1986).
237
238Statistics for Environmental Science and Management, Second Edition
2.Direct maximum-likelihood methods are based on the original work of Cohen (1959). With these, some distribution is assumed for the data, and the likelihood function (which depends on both the observed and censored values) is maximized to estimate population parameters. Usually, a normal distribution is assumed, with the original data transformed to obtain this if necessary. These methods are well covered in the text by Cohen (1991).
3.Regression-on-order-statistics methods are alternatives to max- imum-likelihood methods, and these are easier to carry out in a spreadsheet, for example. One such approach works as follows for data from a normal distribution (Newman et al. 1995). First, the n data values are ranked from smallest to largest, with those below the DL treated as the smallest. A normal probability plot is then con-
structed, with the ith largest data value (xi) plotted against the normal score zi, such that the probability of a value less than or equal to zi is (i − 3/8)/(n + 1/4). Only the noncensored values can be plotted, but for these the plot should be approximately a straight line if the assumption of normality is correct. A line is fitted to the plot by ordi-
nary linear regression methods. If this fitted line is xi = a + bzi, then the mean and standard deviation of the uncensored normal distribution are estimated by a and b, respectively. It may be necessary to transform the data to normality before this method is used, in which case the estimates a and b will need to be converted to the mean and standard deviation for untransformed data.
4.With “fill in” methods, the complete data are used to estimate the mean and variance of the sampled distribution, which is assumed to be normal. The censored values are then set equal to their expected values based on the estimated mean and variance, and the resulting set of data is treated as if it were a full set to begin with. The process can be iterated if necessary (Gleit 1985).
5.The robust parametric method is also a type of fill-in method. A probability plot is constructed, assuming either a normal or lognormal distribution for the data. If the assumed distribution is correct, then the uncensored observations should plot approximately on a straight line. This line is fitted by a linear regression and extrapolated back to the censored observations, to give values for them. The censored values are then replaced by the values from the fitted regression line. If the detection limit varies, this method can allow for it (Helsel and Cohn 1988).
A Windows computer program called UNCENSOR is available on the Internet (www.vims.edu/env/research/software/vims_software.html) for carrying out eight different methods for estimating the censored data values in a sample, including versions of approaches 1 to 5 above. This is an updated version of an earlier program described by Newman et al. (1995). A
Censored Data |
239 |
program like this may be extremely useful, as standard statistical packages seldom have these types of calculations as a menu option.
It would be convenient if one method for handling censored data was always best. Unfortunately, this is not the case. A number of studies have compared different methods, and it appears that, in general, for estimating the population mean and variance from a single random sample, the robust parametric method is best when the underlying distribution of the data is uncertain, but if the distribution is known, then maximum-likelihood method performs well, with an adjustment for bias with a sample size less than or equal to about 20 (Akritas et al. 1994). In the manual for UNCENSOR, Newman et al. (1995) provide a flow chart for choosing a method that says more or less the same thing. On the other hand, in a manual on practical methods of data analysis, the U.S. Environmental Protection Agency (US EPA 2000a) suggests that (a) with less than 15% of values censored, replace these with DL, DL/2, or a small value; (b) with between 15 and 50% of censored values, use a trimmed mean, Cohen’s adjustment, or a Windsorized mean and standard deviation; and (c) with more than 50% of values censored, just base an analysis on the proportion of data values above a certain level.
See Akritas et al. (1994) for more information about methods for estimating means and standard deviations with multiple detection limits, and Helsel (2004) for a detailed discussion of the analysis of censored data in general.
Example 10.1: A Censored Sample of 1,2,3,4-Tetrachlorobenzene
Consider the data shown in Table 10.1 for a sample size of 75 values of 1,2,3,4-tetrachlorobenzene (TcCB) in parts per million from a possibly contaminated site. This sample has been used before in Example 1.7, and the original source was Gilbert and Simpson (1992, p. 6.22). For the present example, it is modified by censoring any values less than 0.25, which are shown in Table 10.1 as <0.25. In fact, this means that these values could be anywhere from 0.00 to 0.24 to two decimal places, so the detection limit is considered to be DL = 0.24.
For the uncensored data, the sample mean and standard deviation are 4.02 and 20.27. It is interesting to see how well these values can be recovered from the censored data with some of the methods in general use.
Table 10.1
Measurements of TcCB (parts per thousand million) from a Possibly Contaminated Site
1.33 |
<0.25 |
<0.25 |
0.28 |
<0.25 |
<0.25 |
<0.25 |
0.47 |
<0.25 |
<0.25 |
<0.25 |
<0.25 |
18.40 |
<0.25 |
<0.25 |
<0.25 |
<0.25 |
<0.25 |
<0.25 |
168.6 |
<0.25 |
0.25 |
0.25 |
<0.25 |
0.48 |
0.26 |
5.56 |
<0.25 |
0.29 |
0.31 |
0.33 |
3.29 |
0.33 |
0.34 |
0.37 |
0.25 |
2.59 |
0.39 |
0.40 |
0.28 |
0.43 |
6.61 |
0.48 |
<0.25 |
0.49 |
0.51 |
0.51 |
0.38 |
0.92 |
0.60 |
0.61 |
0.43 |
0.75 |
0.82 |
0.85 |
<0.25 |
0.94 |
1.05 |
1.10 |
0.54 |
1.53 |
1.19 |
1.22 |
0.62 |
1.39 |
1.39 |
1.52 |
0.33 |
1.73 |
2.35 |
2.46 |
1.10 |
51.972.61 3.06
Note: Values <0.25 are censored.
240 Statistics for Environmental Science and Management, Second Edition
First, consider the simple substitution methods. Replacing all of the censored values by zero, DL/2 = 0.12, DL = 0.24, and a uniform random value in the interval from 0.00 to 0.24 gave the following results for the sample mean and standard deviation (SD): replacement 0.00, mean = 3.97, SD = 20.28; replacement 0.12, mean = 4.00, SD = 20.28; replacement 0.24, mean = 4.03, SD = 20.27; and replacement uniform, mean = 4.00, SD = 20.28. Clearly in this example these simple substitution methods all work very well.
Newman et al.’s (1995) computer program UNCENSOR was used to calculate maximum-likelihood estimates of the population mean and standard deviation using Cohen’s (1959) method. The distribution was assumed to be lognormal because of the skewness indicated by three very large values. This gives the estimated mean and standard deviation to be 1.74 and 8.35, respectively. Using Schneider’s (1986, sec. 4.5) method for bias correction, the estimated mean and standard deviation change to 1.79 and 9.27, respectively. These maximum-likelihood estimates are rather poor, in the sense that they differ very much from the estimates from the uncensored sample.
The regression-on-order-statistics method can also be applied assuming a lognormal distribution, and it becomes apparent using this method that the assumption of a lognormal distribution is questionable. The calculations are shown in Table 10.2, and Figure 10.1 shows a normal probability plot for the logarithms of the uncensored values, i.e., the loge(X) values against the normal scores Z. The data should plot approximately
Table 10.2
Calculations for the Regression on Order Statistics
Order |
|
|
|
|
Order |
|
|
|
|
|
(i) |
Pia |
Zi |
Xi |
Loge(Xi) Fittedb |
(i) |
Pia |
Zi |
Xi |
Loge(Xi) |
Fittedb |
1 |
0.01 |
–2.40 |
<0.25 |
–5.01 |
39 |
0.51 |
0.03 |
0.47 |
–0.76 |
–0.77 |
2 |
0.02 |
–2.02 |
<0.25 |
–4.36 |
40 |
0.53 |
0.07 |
0.48 |
–0.73 |
–0.71 |
3 |
0.03 |
–1.81 |
<0.25 |
–3.99 |
41 |
0.54 |
0.10 |
0.48 |
–0.73 |
–0.65 |
4 |
0.05 |
–1.66 |
<0.25 |
–3.73 |
42 |
0.55 |
0.13 |
0.49 |
–0.71 |
–0.59 |
5 |
0.06 |
–1.54 |
<0.25 |
–3.52 |
43 |
0.57 |
0.17 |
0.51 |
–0.67 |
–0.53 |
6 |
0.07 |
–1.44 |
<0.25 |
–3.34 |
44 |
0.58 |
0.20 |
0.51 |
–0.67 |
–0.48 |
7 |
0.09 |
–1.35 |
<0.25 |
–3.19 |
45 |
0.59 |
0.24 |
0.54 |
–0.62 |
–0.42 |
8 |
0.10 |
–1.27 |
<0.25 |
–3.05 |
46 |
0.61 |
0.27 |
0.60 |
–0.51 |
–0.36 |
9 |
0.11 |
–1.20 |
<0.25 |
–2.93 |
47 |
0.62 |
0.30 |
0.61 |
–0.49 |
–0.30 |
10 |
0.13 |
–1.14 |
<0.25 |
–2.81 |
48 |
0.63 |
0.34 |
0.62 |
–0.48 |
–0.23 |
11 |
0.14 |
–1.07 |
<0.25 |
–2.70 |
49 |
0.65 |
0.38 |
0.75 |
–0.29 |
–0.17 |
12 |
0.15 |
–1.02 |
<0.25 |
–2.60 |
50 |
0.66 |
0.41 |
0.82 |
–0.20 |
–0.11 |
13 |
0.17 |
–0.96 |
<0.25 |
–2.51 |
51 |
0.67 |
0.45 |
0.85 |
–0.16 |
–0.05 |
14 |
0.18 |
–0.91 |
<0.25 |
–2.42 |
52 |
0.69 |
0.48 |
0.92 |
–0.08 |
0.02 |
15 |
0.19 |
–0.86 |
<0.25 |
–2.33 |
53 |
0.70 |
0.52 |
0.94 |
–0.06 |
0.09 |
16 |
0.21 |
–0.81 |
<0.25 |
–2.25 |
54 |
0.71 |
0.56 |
1.05 |
0.05 |
0.15 |
17 |
0.22 |
–0.77 |
<0.25 |
–2.17 |
55 |
0.73 |
0.60 |
1.10 |
0.10 |
0.22 |
18 |
0.23 |
–0.73 |
<0.25 |
–2.09 |
56 |
0.74 |
0.64 |
1.10 |
0.10 |
0.29 |
Censored Data |
|
|
|
|
|
|
|
|
241 |
||||
|
Table 10.2 (continued) |
|
|
|
|
|
|
|
|
||||
|
Calculations for the Regression on Order Statistics |
|
|
|
|
|
|||||||
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Order |
|
|
|
|
|
Order |
|
|
|
|
|
|
|
(i) |
Pia |
Zi |
Xi |
Loge(Xi) |
Fittedb |
(i) |
Pia |
Zi |
Xi |
Loge(Xi) |
Fittedb |
|
19 |
0.25 |
–0.68 |
<0.25 |
|
–2.02 |
57 |
0.75 |
0.68 |
1.19 |
0.17 |
0.36 |
|
|
20 |
0.26 |
–0.64 |
<0.25 |
|
–1.95 |
58 |
0.77 |
0.73 |
1.22 |
0.20 |
0.44 |
|
|
21 |
0.27 |
–0.60 |
0.25 |
–1.39 |
–1.88 |
59 |
0.78 |
0.77 |
1.33 |
0.29 |
0.52 |
|
|
22 |
0.29 |
–0.56 |
0.25 |
–1.39 |
–1.81 |
60 |
0.79 |
0.81 |
1.39 |
0.33 |
0.60 |
|
|
23 |
0.30 |
–0.52 |
0.25 |
–1.39 |
–1.74 |
61 |
0.81 |
0.86 |
1.39 |
0.33 |
0.68 |
|
|
24 |
0.31 |
–0.48 |
0.26 |
–1.35 |
–1.67 |
62 |
0.82 |
0.91 |
1.52 |
0.42 |
0.76 |
|
|
25 |
0.33 |
–0.45 |
0.28 |
–1.27 |
–1.61 |
63 |
0.83 |
0.96 |
1.53 |
0.43 |
0.86 |
|
|
26 |
0.34 |
–0.41 |
0.28 |
–1.27 |
–1.54 |
64 |
0.85 |
1.02 |
1.73 |
0.55 |
0.95 |
|
|
27 |
0.35 |
–0.38 |
0.29 |
–1.24 |
–1.48 |
65 |
0.86 |
1.07 |
2.35 |
0.85 |
1.05 |
|
|
28 |
0.37 |
–0.34 |
0.31 |
–1.17 |
–1.42 |
66 |
0.87 |
1.14 |
2.46 |
0.90 |
1.16 |
|
|
29 |
0.38 |
–0.30 |
0.33 |
–1.11 |
–1.36 |
67 |
0.89 |
1.20 |
2.59 |
0.95 |
1.27 |
|
|
30 |
0.39 |
–0.27 |
0.33 |
–1.11 |
–1.30 |
68 |
0.90 |
1.27 |
2.61 |
0.96 |
1.40 |
|
|
31 |
0.41 |
–0.24 |
0.33 |
–1.11 |
–1.24 |
69 |
0.91 |
1.35 |
3.06 |
1.12 |
1.54 |
|
|
32 |
0.42 |
–0.20 |
0.34 |
–1.08 |
–1.18 |
70 |
0.93 |
1.44 |
3.29 |
1.19 |
1.69 |
|
|
33 |
0.43 |
–0.17 |
0.37 |
–0.99 |
–1.12 |
71 |
0.94 |
1.54 |
5.56 |
1.72 |
1.87 |
|
|
34 |
0.45 |
–0.13 |
0.38 |
–0.97 |
–1.06 |
72 |
0.95 |
1.66 |
6.61 |
1.89 |
2.08 |
|
|
35 |
0.46 |
–0.10 |
0.39 |
–0.94 |
–1.00 |
73 |
0.97 |
1.81 |
18.40 |
2.91 |
2.34 |
|
|
36 |
0.47 |
–0.07 |
0.40 |
–0.92 |
–0.94 |
74 |
0.98 |
2.02 |
51.97 |
3.95 |
2.70 |
|
|
37 |
0.49 |
–0.03 |
0.43 |
–0.84 |
–0.89 |
75 |
0.99 |
2.40 |
168.6 |
5.13 |
3.36 |
|
|
|
38 |
0.50 |
0.00 |
0.43 |
–0.84 |
–0.83 |
|
|
|
|
|
|
|
Note: The censored TcCB data are arranged in order from the smallest values (the censored ones) to the largest values.
aPi = (i − 3/8)/(n + 1/4) is the probability used to calculate the Z scores, i.e., the probability of a value ≤Zi = Pi for the ith-order statistic.
bThe fitted values come from the fitted regression line shown in Figure 10.1. They are only used for the robust parametric method.
Log (X)
6
4
2
0
–2
–0.5 |
0.0 |
0.5 |
1.0 |
1.5 |
2.0 |
Normal Score (Z)
Figure 10.1
Normal probability plot for the logarithms of the uncensored TcCB concentrations, with a straight line fitted by ordinary regression methods.
242 Statistics for Environmental Science and Management, Second Edition
on a straight line if the logarithms of the TcCB concentrations are normally distributed. In fact, the plot appears to be curved, with the largest and smallest values being above the fitted straight line, showing that they are larger than expected for a normal distribution.
Ignoring the possible problem with the assumed type of distribution, the equation of the fitted line shown in Figure 10.1 is loge(X) = −0.83 + 1.75Z. The estimated mean and standard deviation for the logtransformed data are therefore −0.83 and 1.75, respectively. To produce estimates of the corresponding mean and variance for the original distribution of TcCB concentrations is now not all that straightforward. As a quick approximation, equations (3.15) and (3.16) can be used. Thus the estimated mean is
E(X) = exp(μ + ½σ2) ≈ exp(−0.83 + 0.5 × 1.752) = 2.01
and the estimated variance is
Var(X) = exp(2μ + σ2)[exp(σ2) − 1]
= exp[2x(−0.83) + 1.752][exp(1.752) − 1] = 81.58
so that the estimated standard deviation of TcCB concentrations is √81.58 = 9.03.
A better approach is to use the bias-corrected method that is incorporated into UNCENSOR, which is based on a series expansion due to Finney (1941), and takes into account the sample size. For the example data, this gives the estimated mean and standard deviation of TcCB concentrations to be 1.92 and 15.66, respectively. Compared with the mean and standard deviation for the uncensored sample of 4.02 and 20.27, respectively, the regression-on-order-statistics estimates without a bias correction are very poor, and not much better with a bias correction. Presumably this is because of the lack of fit of the lognormal distribution to the noncensored data (Figure 10.1).
Gleit’s (1985) iterative fill-in method is another option in UNCENSOR. This gives the estimated mean and variance of TcCB concentrations to be 1.92 and 15.66, respectively. These are the same as the estimates obtained from the bias-corrected regression-on-order-statistics method, and so they are, again, rather poor.
Finally, consider the robust parametric method. This starts off the same way as the regression-on-order-statistics method, with a probability plot of the data after a logarithmic transformation, with a fitted regression line (Figure 10.1). However, instead of using the regression line to estimate the mean and variance of the fitted distribution, this line is now extrapolated to obtain expected values for the censored data values, as shown in Figure 10.2. For example, the expected value for the smallest value in the sample is −5.0, corresponding to a normal score of −2.4, the second smallest value is −4.4, corresponding to a normal score of −2.0, and so on. The column headed “Fitted” in Table 10.2 gives these
Censored Data |
243 |
Log (X)
6 |
|
|
|
|
|
|
|
|
|
|
4 |
|
|
|
|
|
|
|
|
|
|
2 |
|
|
|
|
|
|
|
|
|
|
0 |
|
|
|
|
|
|
|
|
|
|
–2 |
|
|
|
|
|
|
|
|
|
|
–4 |
|
|
|
|
|
|
|
|
|
|
–6 |
|
|
|
|
|
|
|
|
|
|
–2.5 |
–2.0 |
–1.5 |
–1.0 |
–0.5 |
0.0 |
0.5 |
1.0 |
1.5 |
2.0 |
2.5 |
Normal Score (Z)
Figure 10.2
The regression line from Figure 10.1 extrapolated to estimate the censored values of the logarithm of TcCB values (● denotes an observed value of loge(X), and denotes an expected value from the regression line).
expected values for the order statistics. The robust parametric method simply consists of replacing the smallest 20 censored values for loge(X) with these expected values.
Having obtained values to fill in for the censored values of loge(X), these are untransformed to obtain values for X itself. The sample mean and variance can then be calculated in the normal way. The completed sample is shown in Table 10.3. The mean and variance are 3.99 and 20.28, respectively, which are almost exactly the same as the values for the real data without censoring.
Too much should not be concluded from just one example. However, the simple substitution methods and the robust parametric method have definitely worked better than the alternatives here for two reasons. First, the lognormal assumption is questionable for the methods that require this, other than the robust method. Second, the censored values are all very low, and as long as they are replaced by any value below the detection limit, the sample mean and standard deviation will be close to the values from the uncensored sample.
Table 10.3
Completed Sample for the Robust Parametric Method
1.33 |
0.04 |
0.09 |
0.28 |
0.08 |
0.11 |
0.07 |
0.47 |
0.14 |
0.12 |
0.07 |
0.04 |
18.40 |
0.02 |
0.02 |
0.01 |
0.01 |
0.03 |
0.05 |
168.64 |
0.11 |
0.25 |
0.25 |
0.06 |
0.48 |
0.26 |
5.56 |
0.05 |
0.29 |
0.31 |
0.33 |
3.29 |
0.33 |
0.34 |
0.37 |
0.25 |
2.59 |
0.39 |
0.40 |
0.28 |
0.43 |
6.61 |
0.48 |
0.10 |
0.49 |
0.51 |
0.51 |
0.38 |
0.92 |
0.60 |
0.61 |
0.43 |
0.75 |
0.82 |
0.85 |
0.13 |
0.94 |
1.05 |
1.10 |
0.54 |
1.53 |
1.19 |
1.22 |
0.62 |
1.39 |
1.39 |
1.52 |
0.33 |
1.73 |
2.35 |
2.46 |
1.10 |
51.972.61 3.06
Note: The filled-in values are underlined.
244 Statistics for Environmental Science and Management, Second Edition
10.3 Estimation of Quantiles
It may be better to describe highly skewed distributions with quantiles rather than using means and standard deviations. These quantiles are a set of values that divide the distribution into ranges covering equal percentages of the distribution. For example, the 0%, 25%, 50%, 75%, and 100% quantiles are the minimum value, the value that just equals or exceeds 25% of the distribution, the value that just equals or exceeds 50% of the distribution (i.e., the median), the value that just equals or exceeds 75% of the distribution, and the maximum value, respectively.
Sample quantiles can be used to estimate distribution quantiles that are above the detection limit, although Akritas et al. (1994) note that simulation studies indicate that this can lead to bias when the quantiles are close to this limit. It is therefore better to use a parametric maximum-likelihood approach when the distribution is known. When the distribution is uncertain, the robust parametric method can be used to fill in the censored data in the sample before evaluating the sample quantiles as estimates of those for the underlying distribution of the data.
Distribution quantiles can be estimated with multiple detection limits. See Akritas et al. (1994, sec. 2.6) for more details.
10.4 Comparing the Means of Two or More Samples
The comparison of the means of two or more samples is complicated with censored data, particularly if there is more than one detection limit. The simplest approach involves just replacing censored data by zero, DL, or DL/2, and then using standard methods either to test for a significant mean difference or to produce a confidence interval for the mean difference between the two sampled populations. In fact, this approach seems to work quite well and, based on a simulation study of 10 alternative ways for handling censoring, suggests that a good general strategy involves substituting DL for censored values when up to 40% of observations are censored, and substituting DL/2 when more than 40% of observations are censored (Clarke 1994). However, this strategy is not always the best, and the U.S. Environmental Protection Agency and U.S. Army Corps of Engineers (USEPA and US ACE 1998, Table D-12) give some more complicated rules that depend on the type of data, whether samples have equal variances, the coefficient of variation, and the type of data distribution.
When it can be assumed that the data come from a particular distribution, comparisons between groups can be based on the method of maximum likelihood, as described by Dixon (1998). One of the advantages of maximumlikelihood estimation is the approximate variances and covariances of the
Censored Data |
245 |
estimators that are available. Using these, it is possible to carry out a large sample test for whether the estimated population means are significantly different, or to find an approximate confidence interval for this difference.
For small samples, Dixon (1998) suggests the use of bootstrap methods for hypothesis testing and producing confidence intervals, as discussed further in the following example. This has obvious generalizations for use with other data distributions, and with more than two samples. Dixon also discusses the use of nonparametric methods for comparing samples, and the use of equivalence tests with data containing censored values.
Example 10.2: Upstream and Downstream Samples
The data from one of the examples considered by Dixon (1998) are shown in Table 10.4. The variable being considered is the dissolved orthophosphate concentration (DOP, mg/l) measured for water from the Savannah River in South Carolina. One sample is of 41 observations taken upstream of a potential contamination source, and the second sample is of 42 observations taken downstream. A higher general level of DOP downstream is clearly an indication that contamination has occurred. There are three DL values in this example, <1, <5, and <10, which occurred because the DL depends on dilution factors and other aspects of the chemical analysis that changed during the study.
The number of censored observations is high, consisting of 26 in each of the samples, and 63% of the values overall. Given the high detection limit of 10 for some of the data, simple substitution methods seem definitely questionable here, and an analysis assuming a parametric distribution seems like the only reasonable approach.
Dixon (1998) assumed that the data values X are lognormally distributed, with loge(X) having the same variance upstream and downstream of the potential source of contamination. On this basis, he obtained the following maximum-likelihood estimates: mean DOP upstream, 0.73 with standard error 0.19; mean DOP downstream, 1.02 with standard error 0.17; mean difference between downstream and upstream, 0.24
Table 10.4
Dissolved Orthophosphate Concentrations in Samples Upstream and Downstream of a Possible Source of Contamination, with Three Different Detection Limits
Sample 1: Upstream of Possible Contamination Source |
|
|
|
|
|
||||||
1 |
2 |
4 |
3 |
3 |
<10 |
2 |
<10 |
<5 |
<10 |
<5 |
3 |
<5 |
<5 |
<10 |
<5 |
<10 |
<1 |
<10 |
7 |
<5 |
<1 |
<5 |
2 |
<10 |
5 |
5 |
<5 |
<10 |
<1 |
<5 |
<10 |
<5 |
14 |
5 |
2 |
<10 |
<10 |
7 |
<1 |
<10 |
|
|
|
|
|
|
|
Sample 2: Downstream of Possible Contamination Source |
|
|
|
|
|||||||
4 |
<5 |
<1 |
4 |
3 |
9 |
<10 |
4 |
<5 |
<10 |
<10 |
8 |
<10 |
3 |
<5 |
<5 |
<10 |
5 |
<5 |
<10 |
6 |
<5 |
1 |
4 |
<10 |
<5 |
<5 |
<10 |
5 |
4 |
2 |
<5 |
<10 |
<5 |
<10 |
<5 |
<1 |
<10 |
4 |
<5 |
20 |
<10 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
