Chapter 4. Quality Evaluation and Intercalibration
1.0 Introduction
The measurements described in the next chapters provide part of the core set of data for the scientists of JGOFS and the U.S.JGOFS Bermuda Atlantic Time-series Study (BATS). The continuous CTD data are calibrated by the bottle-collected salinity and oxygen data. Most of the techniques are standard and widely used. However, there are also numerous ways that the data can be inaccurate, from mechanical failure of the Niskin bottles to accidents in the laboratory. Since these kinds of problems are unavoidable, a lab must set
up a series of procedures for checking the data both internally (consistency with the other similar data) and externally (consistency with historical data for the area and intercalibrations with other labs). These quality control methods are used primarily to evaluate the salinity, dissolved oxygen, dissolved inorganic carbon, and nutrient data, and to a lesser extent the particulate and rate measurements. The methods used in the BATS
program are presented here as an illustration of a procedure that might be applicable to similar datasets.

The measures that BATS employs are a combination of formal and informal examinations of the data for inconsistencies and errors. The technicians who are making the measurements are well trained and make the same measurements month to month. They often spot an error in the data set as the number is being generated or as the data are entered into the computer. They know the values that they usually get at each depth and can spot many of the outliers. Such points are not automatically discarded. The identification of an aberrant result, either at this step or in the subsequent examinations, is only cause for rechecking the previous steps in the data generation process (sampling, analysis, data entry and calculation, etc.) for inadvertent errors. If no inadvertent error can be found, then a decision must be made. If the datum is out of the bounds of possibility the datum is likely discarded (see below).

The next step in data inspection is to graph the data with depth and visually examine the profile. At this step, aberrant points can also become evident as deviations from the continuity of the profile. These deviations are checked as above. The other analyses of samples from the same Niskin bottle are also examined to see if they all are aberrant, indicating that the bottle misfired or leaked. If a bottle appears to have leaked, all the
measurements from that bottle are discarded, even if some of them appear to fall within the correct range.

Other graphical methods are also employed to examine the data. T-S diagrams are plotted and compared with historical data. Nutrients are plotted against temperature and density and against each other. Contour plots of a measurement on axes of potential density and time are particularly useful in identifying anomalous data and calibration errors. Nitratephosphate plots have proved very useful in identifying both individual and systematic problems in those nutrient data.

The final examination procedure is the comparison with a carefully selected set of data called our QC windows. In our case, this is a data set compiled by G. Heimerdinger (National Oceanic Data Center) from a number of cruises to within 200 miles of Bermuda between 1975 and 1985. These are data that he believes are of high quality and also reflect the kinds of variation that would be seen at the BATS station. Salinity and oxygen are well represented in this data set, while nutrients are present for only four cruises. G.
Heimerdinger is constantly expanding this QC data set. As the BATS data grows, we have compiled a second set of QC windows from BATS data to compliment G. Heimerdinger's. The BATS data are graphically overlaid on both sets of the QC data and both systematic and individual variations noted and checked carefully as above. Similar data can be compiled to construct QC windows for other ocean regions. This may not be helpful in coastal areas with great variability.

The most difficult problems to resolve are small systematic deviations from the QC envelopes. We are unwilling to automatically discard every deviation from the existing data, especially when they can find no reason that a previously reliable analysis should show the deviation. If the measurements were meant to come out invariant, there would be no reason to collect new data. Therefore, some of the data that are reported deviate from the QC envelope and it is left to others to decide whether they agree with the values. These
deviations are noted in the cruise summaries that accompany each data report. BATS does not flag individual values. In the WOCE program the data reporting system is different. All of the measurements are reported and each is accompanied by a quality flag (see WOCE Manual cited previously).

Finally, one must constantly expand the methods used to check data quality. For many measurements, BATS has added internal standards, sample carry-overs between months and other procedures to prevent accuracy and standardization biases from giving false temporal change. They are currently involved in a number of intercalibration/intercomparison efforts between the BATS lab and other laboratories that regularly make
these kinds of analyses. The results of these intercalibrations (and other types of methods checks) are reported in regular data reports.