Precision and Bias of Graphite Furnace Analysis of Environmental Samples

* Determination of Cr(lIl)+Cr(VI). To 25 mL of sample, add 0.3 mL of 8% PHP, adjust pH to 3.5, and add 3 mL of 10% APDC and 5 mL of PHP-saturated MIBK. Extract for 20 min. Determine Cr(lll)+Cr(VI) in the MIBK layer. * Determination of Cr(VI). Proceed as above except for the addition of 3 mL of 8% PHP and 3 mL of 2% APDC and for the extraction time of 10 mm. * Determination of Cr(III). The value of Cr(III) is obtained as the difference between Cr(Ill) +Cr(VI) and Cr(VI). The concentration of Cr(111) is also determined by the SM-7 ion exchange procedure [1]. In all the above cases, the Cr atomic absorption signal was measured by using the dry/atomize program of 900 'C-30 s (ramp)-30 s (hold)/2500 0 C-O s (ramp)-6 s (hold).


Sample Preparation
The solid matrix samples were prepared using the nitric acid-hydrogen peroxide digestion originally proposed by the EPA EMSL/Cincinnati which is currently included in the EPA contract lab program protocol for low to medium concentration environmental samples.
The aqueous samples were not digested but were spiked to provide 50 lig/L of the elements of interest. Since the purpose of this study was to investigate instrumental precision and accuracy, samples were prepared only once and the four replicate determinations were performed on the same digestate. The precision estimates do not include any variability, due to that of digestion.

Instrumentation
Analysis was performed using a Perkin-Elmer Zeeman 5000 atomic absorption spectrophotometer equipped with an AS-40 auto-sampler and 3600 data station'.

Introduction
This study was undertaken to examine the effects of several operational modes of graphite furnace analysis on the precision and bias of environmental trace element analysis. The modes compared were:

Analysis Protocol
The study was designed so that all three desired comparisons could be done from a single run by storing each atomization peak on a computer disk.
' Disclaimer: Mention of trade names does not constitute endorsement by U.S. EPA.

Accuracy in Trace Analysis
Each sample matrix was injected in duplicate in the tion calculation. A separate least squares linear fit following manner: program provided the MSA values from cups 2, 3, Cup 1-Straight sample. and 4. This protocol was repeated on three addi-Cup 2-Sample diluted 50% with 0.5% HN0 3 . tional runs, usually on different days. Cup 3-Sample diluted 50% with 20 .vg/L standard.
All results were normalized to the appropriate A computer program calculates peak height and percentage of the reference or theoretical value. peak area for each atomization peak so that the first The concentration in solution in .g/L is also injection for cup I was used for the direct calibra-shown. Thirteen of the 32 required dilutions varytion single injection for both peak height and peak ing from 1-2 to 1-200 to bring them on scale for area. The average value obtained from the first and HGA analysis. Table 2 shows the mean and stansecond injections was used for the duplicate injec-dard deviation for each comparison. 6. Statistical Evaluation

Bias
Comparison of methods for bias (differences between the modes compared) was accomplished by pooling the mean recoveries of each method and using Student's t test in the form: t Sd where d=average difference between each "method mean" s= standard deviation of these differences n =number of differences If t>t, 9 5 , i.e., 3.18 in the case of four replicates, then the means of the two methods are considered to be statistically significantly different.

Precision
Equality of pooled within-sample variance for the modes compared was tested using the F test in the form: F=ST., where 5.>S." If F>F 9 , 5 , i.e., 15.4 in the case of four replicates, then there is a statistically significant difference between variances.

Precision
There were no significant differences in variance for any of the comparisons as summarized in table 3. Precision is not improved with peak area as compared to peak height or method of additions vs direct calibration.
Although not significant in this study possibly due to the small number of replications, we believe generally that precision is degraded in MSA analysis due to the multiple injections required. One of the primary objectives of this work was to investigate the value of duplicate injections. This practice is common in many, if not most, analytical laboratories. Duplicate injections essentially double the analysis time. If precision is not improved, then single injections would improve a laboratory's efficiency. This study indicated no significant improvement in precision. It has been our experience that poor precision is usually due to what we call "correctable problems." These include poor injection practices; improper furnace conditions; tube, platform, or contract ring conditions; inadequate background correction; or improper matrix modification. These and other such conditions should be corrected before continuing analysis.
Duplicate injections do not correct these conditions. As a result of this work and discussions with other analysts, we have discontinued the practice of duplicate injections.

Bias
In the context of this study, bias is concerned with the agreement of the two modes compared.
As was anticipated, direct calibration produced results that were generally lower than those by the method of additions. We have observed that interferences do sometimes tend to bias direct calibration results low. We have also noticed a high bias in the MSA results which we have not been able to explain.
For Se and Cd, the differences between direct calibration and MSA were not statistically significant. The 15% difference found for arsenic was significant. Both direct calibration and MSA gave only 26% recovery for sample type 5-the EPA Municipal Digested Sludge sample. We believe, as do other workers, that the true value may be less than that found by analysis Il]. Looking at the other seven sample types, the MSA mean recovery is 106% and the direct calibration is 88%, indicating that even though the MSA value is biased high, it is somewhat better than that resulting from direct calibration. The statistically significant difference found for lead is interesting in that the direct calibration results were somewhat better than the MSA results. The direct calibration averaged 96% recovery, and the MSA averaged 108%. Lead is a good example of the strides that have been made in eliminating interferences in graphite furnace. Only a few years ago, lead recoveries were often low. This was usually attributed to chloride interferences. Now accurate results may be obtained for lead even in hydrochloric acid. Selenium analysis is a good example of how statistical evaluation may sometimes be misleading. The differences in precision and bias found between MSA and direct calibration were not statistically significant. However, the difference for the NBS oyster tissue was the largest of any observed, the recoveries being 86% for MSA and only 41% by direct calibration. Since selenium is often present in tissue at significant levels, accurate analysis is important. These results indicate that one would not want to attempt selenium analysis for tissue using these conditions by direct calibration. This example demonstrates that, for complex and difficult matrices, more work needs to be done to remove interferences. But just as this is evident, we also believe that it is possible to overcome these interferences. For selenium and arsenic, perhaps palladium or a mixed palladium matrix modification may be the answer (see [2)). Although there are cases such as the above where the method of additions is required, we feel that there are disadvantages other than the time required that are often overlooked and hence we avoid MSAs whenever possible. These include errors when the sample to spike ratio is inappropriate, or small errors in calibration and blank (baseline) that can be magnified in MSAs. The sum of the sample and spike can put analysis out of the linear range. Small variations in the individual readings in an MSA determination can produce larger errors in the final result. One example which is quite typical from this study demonstrates this last problem. In the MSA value for arsenic in orchard leaves, the fourth replicate was biased positive 49%. Since the correlation coefficient of calibration was low, this replicate was repeated. The individual readings of the second run were very close to the first run with the largest difference for the highest spike level. The difference between the first and second run for this was 7.7% (33.8 to 36.5 gtg/L). In addition, the correlation coefficient improved and the bias was positive (only 15%). That is, a difference of 7.7% in one of the individual readings made a difference of 34% in the final result. This type of error is so common that we prefer to use direct calibration for the most accurate results for routine work such as verifying maximum contamination levels in drinking water or analyzing performance evaluation samples.
Errors introduced due to inappropriate spiking level and calibration error is discussed by Gaind and Odell [3]. They address the EPA contract lab "continuing calibration" criteria of ±+10% and its effect on the probability of successful spike recovery. This probability is low and they believe the 10% limit is too wide.

Conclusion
We believe that the data from this study demonstrate that reliable analytical data can be achieved, whether these data are obtained from peak height or peak area, single or double injections, or direct calibration or the method of additions. Certainly, there are exceptions to every rule, but we believe that advances have made the interferences the exception. These advances include such things as improved background correction, delayed atomization, automatic injection, and matrix modification. We are approaching the "standardless graphite furnace analysis" proposed by Walter Slavin and co-workers [4]. In the absence of interferences, absolute calibration by a "characteristic mass" could be a powerful concept in accurate trace analysis.