intercept. As the intercept moves away from zero,
reagent grade water used for preparation of HPLC
eluents was obtained from a Millipore Milli-Q
the correlation coefficient (r) for the zero-inter-
Type 1 reagent grade water system.
cept model will decrease relative to the value for
the model with intercept, thereby giving an indi-
cation of the significance of the intercept.
Statistical analyses
For all on-site vs. laboratory comparisons, ex-
To see if there were significant concentration
cept location 6 (picrate), the sum of TNB, TNT
differences among sample positions at each sam-
and 2,4-DNT laboratory concentration estimates
pling location, analytical results from both meth-
were compared to on-site measurements. The
ods of analysis were subjected to one variable of
Janowsky ions produced for TNT and TNB both
classification, completely randomized Analysis of
have wavelengths of maximum absorption around
Variance (ANOVA) using CoStat version 1.03 soft-
540 nm and their molar absorptivities at that wave-
ware (CoHost Software, Inc.). For sampling loca-
length are nearly equal. (There is a peak with
tions 1, 2, 3, 6 and 8, where concentration varia-
higher absorptivity at lower wavelength but high
tions were extremely large, variances were not
humic background makes measurement at this
homogeneous (standard deviations were propor-
peak wavelength prone to interference.) In any
tional to concentrations, i.e., relative standard de-
case, the on-site TNT method will record the sum
viations [RSDs] were constant). In these instances,
of TNT and TNB (Jenkins and Walsh 1992). The
the concentrations were log-transformed prior to
absorptivity of the Janowsky ion from 2,4-DNT is
doing ANOVA. When the ANOVA demonstrated
not maximum at 540 nm but it is significant. How-
that there were significant differences among
ever, DNT reacts slower with the EnSys reagent
sample positions for a given sampling location,
than TNT and TNB, and the rate of color forma-
least significant differences (LSDs) were computed
tion varies with water concentration in the ex-
to identify specific differences.
tract. Since the contribution of DNT to the field
For sampling locations 4, 5, 7, 7R and 9, con-
centration ranges were less extreme and variances
TNT estimates will depend on analysis condi-
approached homogeneity. In these cases variances
tions, corrections are impractical, so we decided
were fractionated to yield estimates of the stan-
to use the total of these three analytes to represent
dard deviations for subsampling plus analysis (SA)
laboratory concentrations.
One further aspect of the statistical analysis
and for the field sampling (SS). Henceforth, all
requires mention. It has already been noted that
references to analytical error should be under-
total absolute variances for the seven sample po-
stood to include contributions from mixing and
sitions in some sampling locations were non-
subsampling, extraction, dilution, measurement
Gaussian. Furthermore, they were computed with-
and concentration computations, while sampling
out regard to the presence of variable amounts of
error refers to spatial heterogeneity at the sam-
spatial correlation between positions. We observed
pling location. CoStat software was also used to
that the spatial correlations were irregular in some
compute means and standard deviations of du-
cases, in contrast to a regular gradient such as the
plicates, overall means of the seven duplicates,
directional concentration change that one might
plus means and standard deviations of compos-
find on the edge of a plume of highly mobile
ites. Analytical precision of the seven duplicates
compounds. For example, see the pattern of TNT
for each sampling location and each analysis
concentrations observed for sampling location 1
method was expressed as the average of the seven
(Fig. 2). This spatial correlation undoubtedly in-
RSDs.
troduces some bias in the variance estimates, but
One-way ANOVA was also used to compare
we believe that the magnitude of this effect is
on-site vs. laboratory analyses of composites. A
insufficient to significantly affect the conclusions.
paired t-test and correlation analysis was used to
compare on-site vs. laboratory analysis for sets of
seven samples for a given sampling location. These
tests were done with Sigma Stat (Jandel Scien-
RESULTS AND DISCUSSION
tific). In addition to the linear least squares model
with intercept, correlations were also computed
Monite site
for the linear zero-intercept model on untrans-
Sampling location 1
formed data. When intercepts are close to zero,
Results for the on-site analysis and laboratory
the correlation coefficient for the zero-intercept
analyses for sampling location 1 are presented in
model approaches the value for the model with
8
Previous Page
