Carl E. Zipper

Department of Crop and Soil Environmental Sciences

Virginia Polytechnic Institute and State University

Golde I. Holtzman and Patrick Darken

Department of Statistics

Virginia Polytechnic Institute and State University

Pamela Thomas and Jason Gildea

Department of Crop and Soil Environmental Sciences

Virginia Polytechnic Institute and State University

Leonard Shabman

Virginia Water Resources Research Center

Virginia Polytechnic Institute and State University

The quality of water in Virginia’s rivers and streams affects the health and welfare of Virginia’s citizens, quality of life available in the Commonwealth’s communities, and the state’s economic development potential. Each year, large amounts of money are spent by both the state and the private sector to protect, improve, and monitor the quality of Virginia’s waters. Yet, little information is available on the long-term success of these water-quality-protection expenditures.

The research described in this report was the first-phase of a two-phased, multi-year effort to enhance Virginia Department of Environmental Quality’s (DEQ) capability to detect and interpret long-term water-quality trends. This report summarizes results of long-term trend analyses of Virginia water-quality monitoring data collected over varying periods at individual monitoring stations.

The objective of this research was to evaluate Virginia surface-water quality monitoring data statistically for trend.

We addressed the research objective by performing a statistical analysis of water quality data (9 variables) from 180 monitoring stations maintained by Department of Environmental Quality (DEQ), and 11 additional stations maintained by U.S. Geological Survey (USGS) using the seasonal Kendall technique. Four of the USGS stations are located out of state.

Water quality data (9 variables) for 191 water-quality monitoring stations were downloaded from the STORET data base by Virginia Tech personnel under the direction of Virginia DEQ. The stations included 180 Virginia monitoring stations maintained by Virginia DEQ, 7 Virginia monitoring stations maintained by USGS, and 4 out-of-state monitoring stations maintained by USGS.

DEQ personnel identified monitoring stations to be included in this study based on two primary criteria:

• availability of water quality data at a frequency sufficient to support the Seasonal Kendall analytical technique, and
• locations that provide statewide geographic representation.

Virginia DEQ monitoring station data extend from the earliest availability of data through early 1997. Most DEQ data sets begin in the early 1970s, although a few also contain data from the late 1960s; most extend through early 1997. Availability of some data for some USGS monitoring stations (notably, pH) extend as far back as the 1940s. In order to assure that the results of USGS monitoring-station analyses would be roughly comparable to those of the DEQ stations, data prior to 1966 were excluded from these analyses. Most stations were sampled monthly, although data sets for some stations show multiple-month and/or multiple-year gaps in the data record. A small number of stations lacked data for one or more variables.

DEQ personnel collected water samples from stream centers, generally using a bridge or boat and following EPA protocol. This activity occurred during the course of the agency’s ongoing surface water-quality monitoring program. Laboratory analyses were conducted by the Virginia Division of Consolidated Laboratory Services (DCLS) using methods based on U.S. Environmental Protection Agency’s "Methods for Chemical Analysis of Water and Wastes." Analytical methods used by DCLS (including detection limits and precision) changed during the period of analysis, as new instrumentation and more advanced analytical methods became available.

Precise flow data are not available for most monitoring stations. Therefore, statistical-analysis procedures did not include an adjustment for flow.

Water quality variables analyzed are listed below, with the STORET codes in parentheses.

• DO—Dissolved Oxygen Saturation (00301). DO saturation values were calculated from measured values of DO concentration and water temperature by Virginia DEQ.
• BOD—Biochemical Oxygen Demand (00310).
• PH—pH (00400). The pH data for the early years were listed with a precision of one-tenth pH unit (e.g., 7.1), whereas modern pH values were listed at a precision of one-one hundredth of a pH unit (e.g., pH 7.12). We rounded all pH values off to one-tenth of a pH unit prior to conducting seasonal-Kendall analyses.
• TR—Residue Total (00500).
• NFR—Non-Filterable Residue (00530).
• NN—Nitrate-Nitrite Nitrogen (00615 + 00620; 00630). NN values were determined using two separate methods. For some sampling dates at most sampling locations, the DEQ data included a single value representing nitrogen concentration in the nitrate and nitrite form (Storet Code 00630). In other cases, two separate analyses were performed, both nitrate (00615) and nitrite (00620); in these cases, the two values were summed to calculate NN. This process created some statistical complications when either one of these values was listed as a lower-detection limit.
• TKN—Total Kjeldahl Nitrogen (00625).
• TP—Total Phosphorous (00665).
• FC—Fecal Coliform (31616).

The statistical analysis (seasonal Kendall analysis) and data-screening procedures were automated by means of an original algorithm using SAS/IML (SAS/IML is a registered trademark of SAS Institute, Inc., Cary, NC) programming code. The programming code used in our earlier work (Rheem and Holtzman, 1990; Zipper et al., 1992) has been replaced by a more efficient program written by P. Darken and G. Holtzman called WQ1.

Values lying outside the analytical limits of the laboratory procedures, or beyond the boundaries of what would be reasonably expected to occur in natural waters even under extreme conditions, were discarded prior to statistical analysis. Limits used to identify values eliminated as outliers are written in the output for each station. These limits were identified by DEQ personnel.

Because of the wide range of conditions present, we were unable to automate the process of identifying erroneous observations. Erroneous values were identified by visual inspection on a station-specific basis. Virginia Tech personnel identified apparent outliers as suspect values through review of graphical and tabular output subsequent to a preliminary statistical analysis. DEQ personnel reviewed the raw data for each suspect variable; those values identified by DEQ personnel as erroneous were eliminated from the data set prior to statistical analysis, and identified as erroneous values in the program output.

The data consist of measurements of 9 water-quality variables made at 191 monitoring stations. The measurements were made over varying time periods between the late 1960s and early 1997 at intervals of varying lengths, but, generally, at three- to four-week intervals. The data were analyzed for trend using a seasonal Kendall Tau rank correlation test. Statistical procedures utilized are described in Zipper et al. (1998a and 1998b).

Some months had multiple observations. In each of those months, the median of the observations was chosen as a representative value for that month so that the data set would consist only of monthly observations.

STORET remark codes are single-character variables that accompany water-quality values entered in the STORET database. Non-blank remark codes typically indicate a condition that should be placed on interpretation of the associated data value. Data with non-blank remark codes to indicate conditions other than detection-limits or calculated values ($ - DO only) were eliminated from the trend analysis. Below each graphic display (Zipper et al., 1998a), those observations eliminated from the analysis due to remark codes are listed. Remark codes other than those indicating detection limits occurred only rarely.

For each variable at each monitoring station, all observations remark-coded as being at a lower detection limit were treated as tied. Similarly, all observations remarked as being at an upper detection limit were treated as tied. Values less than or equal to a lower detection limit (and similarly, greater than or equal to an upper detection limit) were treated as tied with values designated as being at the detection limit.

This procedure was complicated for the nitrate-nitrite nitrogen (NN) variable, because in some cases two values were added to create the NN value used in this analysis. This analysis required development of statistical procedure designated as the "epsilon method". Usage of the epsilon method is documented in the remark-code sections of the output for the NN variable at each monitoring station (Zipper et al, 1998a).

On a statewide basis, significant and apparent trends indicating water quality improvement outnumbered significant and apparent trends indicating water-quality deterioration for BOD, TP, FC, NFR, and DO. For BOD, NFR, and TP, trends representing water-quality improvement outnumbered trends representing water-quality deterioration by ratios exceeding 3:1. For BOD, declining trends representing water-quality improvement were predominant statewide.

For both NN and TKN, increasing trends outnumbered declining trends; increasing levels of nitrogen are generally interpreted to indicate deteriorating water quality. On a statewide basis, there is a tendency for increasing NN trends to occur at stations with relatively high medians.

Declining pH trends outnumbered increasing pH trends by a slight margin. Excluding coalfield stations (where acid-mine-drainage treatment may be responsible for the predominance of increasing pH trends), declining pH trends outnumbered increasing trends by a margin of nearly 2-to-1. Increasing TR trends and decreasing TR trends occurred in approximately equal numbers.

Because of several factors inherent in the data set, simple numerical comparisons of increasing and decreasing trends cannot be interpreted as a direct representation of general statewide change in water quality. For example, we have no basis for evaluating the extent to which the monitoring stations analyzed may or may not collectively represent statewide water quality. Whereas some monitoring stations may be located such that watershed characteristics are the predominant influence on water quality, others are located such that point-source discharges are the predominant influence. Similarly, because the time periods analyzed were based on data availability, rather than a fixed time period, not all trends represent change over the same time period. The fact that broad gaps of data coverage occurred for some variables at some stations between the starting and ending dates must also be considered in comparing findings of trend vs. no-trend among two or more stations.

We also confronted several other analytical issues during the course of this research. Many instances were present where significant or apparent trends were found to be present but the best-estimate of median-change per year (slope) is zero. This condition typically occurs in cases where many pairs of observations are interpreted as being tied (i.e., insufficient evidence is present to allow determination of which is the higher value), and was quite common for variables which yielded high numbers of observations at or below detection limits, including TP, BOD, FC, TKN, and NFR. Without flow data, we are unable to determine whether detected trends were the direct result of actual changes in water quality, or if they were an indirect result of differences in the distribution of high (or low) flow-volumes-at-sampling throughout the monitoring period. The accuracy of our analyses is dependent upon the assumption that variations of flow-volume are randomly distributed throughout the monitoring period.

More complete results for individual monitoring stations can be accessed on the internet (Zipper et al., 1998a).

On a statewide basis, significant and apparent trends indicating water quality improvement outnumbered significant and apparent trends indicating water-quality deterioration for BOD, TP, FC, NFR, and DO. For both NN and TKN, increasing trends outnumbered declining trends; increasing levels of nitrogen are generally interpreted to indicate deteriorating water quality. Declining pH trends outnumbered increasing pH trends by a slight margin. Increasing TR trends and decreasing TR trends occurred in approximately equal numbers. Because of uncertainties regarding the ability of the monitoring stations chosen to adequately represent statewide water quality and variations in periods of data coverage, simple numerical comparisons of increasing and decreasing trends cannot be interpreted as a direct representation of general statewide change in water quality.

This paper’s content is derived from a research report submitted to Virginia Department of Environmental Quality (Zipper et al., 1998b). This research was supported with funding provided by Virginia Department of Environmental Quality to the Virginia Water Resources Research Center. The authors express sincere thanks to all Virginia DEQ personnel who worked with us, including Ron Gregory, Roger Stewart, Stuart Torbeck, and David Lazarus.