Download Statistical Analysis for Monotonic Trends and more Study notes Design in PDF only on Docsity!
Through the National Nonpoint Source Monitoring P rogram (NNPSMP) , states monitor and evaluate a subset of watershed projects funded by the Clean Water Act Section 319 Nonpoint Source Control Program. The program has two major objectives:
- To scientifically evaluate the effectiveness of watershed technologies designed to control nonpoint source pollution
- To improve our understanding of nonpoint source pollution NNPSMP Tech Notes is a series of publications that shares this unique research and monitoring effort. It offers guidance on data collection, implementation of pollution control technologies, and monitoring design, as well as case studies that illustrate principles in action.
Statistical Analysis for Monotonic Trends
Introduction
The purpose of this technical note is to present and demonstrate the basic analysis of long-term water quality data for trends. This publication is targeted toward persons involved in watershed nonpoint source monitoring and evaluation projects such as those in the National Nonpoint Source Monitoring Program (NNPSMP) and the Mississippi River Basin Initiative, where documentation of water quality response to the implementation of management measures is the objective. The relatively simple trend analysis techniques discussed below are applicable to water quality monitoring data collected at fixed stations over time. Data collected from multiple monitoring stations in programs intentionally designed to document response to treatment (e.g., paired-watershed studies or above/ below-before/after with control) or using probabilistic monitoring designs may need to apply other techniques not covered in this technical note.
Trend Analysis
For a series of observations over time—mean annual temperature, or weekly phosphorus concentrations in a river—it is natural to ask whether the values are going up, down, or staying the same. Trend analysis can be applied to all the water quality variables and all sampling locations in a project, not just the watershed outlet or the receiving water.
Broadly speaking, trends occur in two ways: a gradual change over time that is consistent in direction (monotonic^1 ) or an abrupt shift at a specific point in time (step trend). In watershed monitoring, the questions might be “Are streamflows increasing as urbanization increases?” [a monotonic trend] or “Did nonpoint source nutrient loads decrease after the TMDL was implemented in 2002?” [a step trend]. When a monitoring project involves widespread implementation of best management practices (BMPs), it is usually desirable
(^1) Linear trends are a subset of monotonic trends.
Trend analysis can
answer questions like:
“Are streamflows
increasing as
urbanization increases?”
or
“Have nutrient
loads decreased since
the TMDL was
implemented?”
Donald W. Meals, Jean Spooner, Steven A. Dressing, and Jon B. Harcum.
- Statistical analysis for monotonic trends, Tech Notes 6, November 2011. Developed for U.S. Environmental Protection Agency by Tetra Tech, Inc., Fairfax, VA, 23 p. Available online at https://www.epa. gov/polluted-runoff-nonpoint-source-pollution/nonpoint-source-monitoring- technical-notes.
November 2011
to know if water quality is improving: “Have suspended sediment concentrations gone down as conservation tillage adoption has gradually increased?” [a monotonic trend] or “Has the stream macroinvertebrate community improved after cows were excluded from the stream with fencing in 2005?” [a step trend]. If water quality is improving, it is also important to be able to state the degree of improvement.
Trend analysis has advantages and disadvantages for the evaluation of nonpoint source projects, depending on the specific situation (Table 1). Simple trend analysis may be the best — or only — approach to documenting response to treatment in situations where treatment was widespread, gradual, and inadequately documented, or where water quality data are collected only at a single watershed outlet station. For data from a short-term (e.g., 3 years) monitoring project operated according to a paired-watershed design (Clausen and Spooner 1993), analysis of covariance (ANCOVA) using data from the control watershed may be more appropriate than trend analysis to evaluate response to treatment because it directly accounts for the influences of climate and hydrology in a short-term data set. In contrast, for a long data record from a single watershed outlet station, trend analysis may be the best approach to evaluate gradual change resulting from widespread BMP implementation in the watershed in the absence of data from a control site.
Table 1. Advantages and disadvantages of simple trend analysis as the principal approach for evaluation of nonpoint source monitoring projects.
Advantages Disadvantages Can be done on data from a single monitoring station
Usually requires long, continuous data record
Does not require calibration period Difficult to account for variability in water quality data solely related to changes in land treatment or land management Applicable to large receiving waterbodies that may be subject to many influences
Not as powerful as other watershed monitoring designs that have baseline (or pre-BMP data) with controls (e.g., control watershed or upstream data), especially with small sample sizes Useful for BMPs that develop slowly or situations with long lag times
Provides no insight into cause(s) of trend
The application of trend analysis to evaluate the effects of a water quality project depends on the monitoring design. Data from a watershed project that uses an upstream/ downstream or before/after study design where intensive land treatment occurs over a short period generating an abrupt or step change may be evaluated for a step trend using a variety of parametric and nonparametric tests including the two sample t-test, paired t-test, sign test, analysis of (co)variance, or Kruskal-Wallis test. In general, these tests are most applicable when the data can be divided into logical groups.
precipitation, and flow. A snapshot of water quality data from a few months may suggest an increasing trend, while examination of an entire year shows this “trend” to be part of a regular cycle associated with temperature, precipitation, or cultural practices. Autocorrelation—the tendency for the value of an observation to be similar to the observation immediately before it—may also be mistaken for a trend over the short term. Changes in sampling schedules, field methods, personnel, or laboratory practices may also cause shifts in data that could be erroneously interpreted as trends. Characterization of project data through exploratory data analysis (see Tech Note #1 ) will help recognize and account for such features in a dataset.
General Considerations
Is a simple trend analysis appropriate?
The first step in trend analysis is to decide if it is an appropriate tool for answering the questions you have about project data. Effective trend analysis requires a fairly long sequence of data collected at a fixed location, collected by consistent methods, with few long gaps. It has been suggested that five years of monthly data are the minimum for monotonic trend (continuous rate of change, increasing or decreasing) analysis; for a step trend (abrupt shift up or down), at least two years of monthly data before and after treatment are required (Hirsch 1988). These time frames are only guidelines; longer periods of record and/or more intensive sampling frequency would generally provide a greater sensitivity to detect smaller changes. Trend analysis is best suited for a situation where the land treatment program has been successful in implementing BMPs over an extensive portion of the critical area, implementation occurs over several years, and water quality change is expected to be gradual.
The water resource type, project design, type of land treatment, and implementation schedule largely determine the type of trend to be expected. Most of the trend analysis techniques discussed in this publication apply to the evaluation of a monotonic trend, the kind of change that might be expected in response to gradual, widespread implementation of BMPs. Step trends may occur in response to an abrupt change in the watershed, such as the completion of a detention pond or a ban on winter manure application. To properly evaluate a step trend, it is critical to have a solid a priori hypothesis concerning when the step change took place; examination of the data themselves to search for the best place to locate a shift is inappropriate. Although techniques exist for testing for step trends, in many cases a two-sample test (e.g., t-test of before vs. after) may be a better choice when an abrupt change at a specific point in time is expected.
All increasing or decreasing patterns in water quality are not trends. Characterize your data to avoid misinterpreting seasonal cycles, autocorrelation, or changes in monitoring methods as significant trends.
Explore the data first
Before beginning trend analysis, define the question that needs to be answered and then conduct exploratory data analysis (EDA) on the data set (see Tech Note #1 ). EDA will often give preliminary indications of trends and set the stage for further trend analysis. Use EDA to evaluate how well the data satisfy assumptions of parametric statistical analysis (normal distribution, constant variance, and independence), evaluate the effectiveness of transformations, and characterize relationships between variables. EDA can reveal important explanatory variables (covariates) like flow or precipitation that drive dependent variables at this point. Some trend analysis techniques can account for covariates.
Evaluate the data set for significant missing observations, such as a year-long interruption in the middle of a 7-year program. Some techniques are sensitive to gaps in data collection. If a long gap exists in the data, step trend procedures (e.g., assessing the difference in sample means between the two periods using a two-sample t-test) may be more appropriate than the monotonic trend analysis techniques discussed below. Although there is no specific decision rule, Helsel and Hirsch (1992) advise using step trend rather than monotonic trend analysis if a data gap is greater than one-third of the total record.
Select variables
Trend analysis can be applied to all the water quality variables and all sampling locations in the project. In large projects tracking many variables at many stations, this can be a daunting task. If full analysis is not feasible, there are several options. First, a subset of monitored variables can be selected, focusing on those expected to be most responsive to land treatment or those that directly relate to water quality impairment. Alternatively, it may be possible to use an index that combines information from a number of variables, such as the Index of Biotic Integrity (IBI) for stream fish communities (Karr 1981), or the Oregon Water Quality Index (OWQI) that integrates measurements of temperature, dissolved oxygen, BOD, pH, ammonia+nitrate nitrogen, total phosphate, total solids, and fecal coliform (Cude 2005). Third, overall water quality trends have been efficiently assessed and presented by conducting trend analysis on principal components as surrogate variables for individual water quality constituents (Ye and Zou 1993).
Data reduction and flow adjustment
Before proceeding to trend analysis tests, it may be necessary or beneficial to perform some preliminary data reduction. Transformations may be necessary to satisfy assumptions for parametric analysis. If sampling has been collected regularly at very frequent intervals, the data can be aggregated to standard periods (e.g., from daily observations to monthly means or medians). Adjusting data because of changing sampling frequency (i.e., weekly
technique requires that a relationship exists between concentration and discharge. For this procedure to be valid, the streamflow distribution must be stationary, i.e., be itself free of trend. If the distribution of streamflow has changed over the period of record (e.g., because of diversions, detention ponds, or stormwater BMPs), then residuals analysis or any other flow-adjustment technique should not be used. Presence or absence of trend in flow can be verified through knowledge of changes in watershed hydrology or by independent analysis of trends in the streamflow record itself. Where streamflows are not stationary, it may be possible to remove the effects of varying hydrologic conditions on the concentration variable by using some appropriate measure of basin precipitation as a covariate or account for hydrologic changes by other trend analysis techniques.
Alternatively, because land treatment effects are generally expected to change the relationship between concentrations and flow, an analysis of covariance will usually be appropriate.
Graphing
Before proceeding to intensive numerical analysis, it is useful to re-examine the time series plots developed earlier in the process of exploratory data analysis. Visual inspection of a time series plot is the easiest way to look for a trend, but data variability may obscure a trend. Visualization of trends in noisy data can be clarified by various data smoothing techniques. Plotting moving averages or medians, for example, instead of raw data points, reduces apparent variation and may reveal general tendencies. Spreadsheets like Excel can display a moving-average trend line in time-series scatterplots with adjustable averaging periods. A more complex smoothing algorithm, such as LOWESS ( LO cally WE ighted S catterplot S moothing), can reveal patterns in very large datasets that would be difficult to resolve by eye. LOWESS is computationally intensive (see Helsel and Hirsch 1992), but computer programs exist that make the procedure relatively easy to accomplish.
Note, however, that visualization has limitations because people tend to focus on outliers, strong seasonal variation can mask trends in a variable of interest, and gradual trends are difficult to detect by eye alone. Additionally, simple visualization cannot adequately quantify the magnitude of a trend. Visualization is not a substitute for the hypothesis testing discussed below.
Monotonic Trend Analysis
A number of statistical tests are available to identify and quantify monotonic trends in a way that is defensible and repeatable. Statistical trend analysis is a hypothesis- testing process. The null hypothesis (HO ) is that there is no trend; each test has its own parameters for accepting or rejecting HO. Failure to reject HO does not prove that there
is not a trend, but indicates that the evidence is not sufficient to conclude with a specified level of confidence that a trend exists.
Table 2 lists some trend tests available for different circumstances, including adjustments for a covariate and the presence of seasonality. The tests are further divided into para- metric, nonparametric, and mixed types. Parametric tests are considered more powerful and/or sensitive to detect significant trends than are nonparametric tests, especially with a small sample number. However, unless the assumption of normal distribution for para- metric statistics is met, it is generally advisable to use a nonparametric test (Lettenmaier 1976, Hirsch et al. 1991, Thas et al. 1998). Both parametric and nonparametric tests require constant variance and independence. Methods for testing assumptions of distri- bution, constant variance, and independence required for parametric linear regressions are discussed in detail in USEPA (1997a). Nonparametric tests provide higher statistical power in case of nonnormality and are robust against outliers and large data gaps.
Table 2. Classification of tests for trend (adapted from Helsel and Hirsch 1992)
Type of test
Not Adjusted for covariate (X) Adjusted for covariate (X)
No seasonality
Parametric Linear regression of Y on t Multiple linear regression of Y on X and t Mixed – Mann-Kendall on residuals from regression of Y on X Nonparametric Mann-Kendall Mann-Kendall on residuals from LOWESS of Y on X
Seasonality
Parametric Linear regression of Y on t and periodic functions
Multiple linear regression of Y on X, t, and periodic functions Mixed Regression of deseasonalized Y on t
Seasonal Kendall on residuals from regression of Y on X
Nonparametric Seasonal Kendall on Y Seasonal Kendall on residuals from LOWESS of Y on X Y = dependent variable of interest; X = covariate, t = time
These tests will be discussed below, with emphasis on linear regression, Mann-Kendall, and seasonal Kendall procedures. For more detailed information, consult the references listed at the end of this technical note.
Tests without covariates (Y versus time)
Parametric test: Linear regression of Y on t (Example 1, p. 18).
If project data satisfy all the assumptions necessary for linear regression (Y is linearly related to t, residuals are normally distributed, residuals are independent, and variance of residuals is constant), a simple linear regression of Y on time is a test for linear trend:
Y = β 0 + β 1 t + ε
The null hypothesis is that the slope coefficient β 1 = 0. The t-statistic on β 1 is tested to determine if it is significantly different from zero. If the slope is nonzero, the null
which has a range of –1 to +1 and is analogous to the correlation coefficient in regression analysis. The null hypothesis of no trend is rejected when S and τ are significantly different from zero. If a significant trend is found, the rate of change can be calculated using the Sen slope estimator (Helsel and Hirsch 1992):
β 1 = median ( )
y (^) j - y (^) i x (^) j - x (^) i
for all i < j and i = 1, 2, …, n-1 and j = 2, 3,…, n; in other words, computing the slope for all pairs of data that were used to compute S. The median of those slopes is the Sen slope estimator.
Tests accounting for covariates
Variables other than time usually influence the behavior of water quality variables. These covariates are usually natural phenomena such as precipitation, temperature, or streamflow. By removing the variation caused by these explanatory variables, the noise may be reduced and a trend revealed. Correction for hydrologic and meteorologic variability is essential in both parametric and nonparametric trend tests to determine if the statistically significant trends are due to processes and transport changes such as land use changes, or to artifacts of system variability.
Selection of an appropriate covariate is critical. It should be a measure of the driving force behind the behavior of the variable of interest, but must not be subject to human manipulation during the course of the project, i.e., must not be changed by BMPs or the land treatment program. In nonpoint source monitoring, much of the variance in concentration data is usually a function of runoff and streamflow; thus, natural streamflow is a commonly used covariate in trend analysis. However, streamflow should not be used as a covariate if the land treatment program itself affects streamflow, such as with urban stormwater infiltration practices or conservation tillage. In such cases, precipitation may be a good choice for a covariate.
In deciding whether or not to remove the variation caused by flow from a data set, consider project objectives and the nature of the land treatment program. If a land treatment program has caused a measurable change in the watershed flow regime, such a change may in fact be a desired outcome and the resulting trend in both flow and pollutant concentration may be important to detect and quantify. Removing variation caused by flow may risk reducing the magnitude of any trend in concentration alone below detection level, considering other noise in the system. On the other hand, failure to account for a trend in flow that is not associated with the land treatment program may result in showing a trend in concentration where none exists. It is generally advisable to test the covariate data set independently for trend before proceeding.
Parametric: Multiple linear regression of Y on X and t
Multiple linear regression can be used to account for the effects of other variables such as flow, land management, or other water quality characteristics on a response variable. Multiple regression includes covariates in trend analysis in a single step. Appropriate covariates are those that are correlated with the water quality variable Y and adjust for changes in climate to better isolate trends due to BMPs. Consider multiple regression of concentration (Y) versus time (t) and flow (Q ):
Y = β 0 + β 1 t + β 2 Q + ε
The test accounts for the effects of the covariate by including them in the regression model. The null hypothesis for the trend test is β 1 = 0; the t-statistic for β 1 tests for trend. If the coefficient β 2 for the covariate is not significantly different from zero, the effect of the covariate is not significant and a simple regression model of Y on t should be used. An exception to this would be the case where flow is increasing over time and the effects of increasing flow are already accounted for in the time component; in such a case, flow might still be logically included in the regression model even if β 2 is not different from zero. It should be emphasized that as for simple linear regression, the assumptions that Y is linearly related to t and Q , that residuals are normally distributed and independent, and that variance of residuals is constant must be satisfied to use this test properly.
Mixed: Mann-Kendall on residuals from regression of Y on X
This is a hybrid test that includes removal of covariate effects by a parametric procedure, followed by a nonparametric test for trend. If a reasonable linear regression can be obtained (i.e., residuals have no extreme outliers, Y is approximately linear with X), the regression between Y and one or more Xs (i.e., Y = β 0 + β 1 X + ε) can remove the effect of X prior to performing the Mann-Kendall test for trend.
The residuals (R) from the regression model are computed as observed minus predicted values: R = Y - β 0 + β 1 X
Then the Kendall S statistic is computed on the R-time data pairs and tested to see if it differs significantly from zero. If assumptions for parametric statistics are seriously violated, a fully nonparametric alternative (e.g., using LOWESS) should be selected to estimate the relationship between Y and X as described in the next section.
Nonparametric: Mann-Kendall on residuals from LOWESS of Y on X
The LOWESS smoothing technique describes the relationship between Y and a covariate X without assuming linearity or normality of residuals. Applying LOWESS smoothing to a scatterplot of X and Y is roughly analogous to regression, without forcing a straight line. Given the LOWESS fitted value Y’, the residuals (R) are computed as:
R = Y - Y’
Then, the Kendall S statistic is computed on the R-t data pairs and tested to see if it differs significantly from zero.
Mixed: Seasonal Kendall on residuals from regression of Y on X and
Regression of deseasonalized Y on t
Two hybrid procedures may be used to account for seasonality. First, the seasonal Kendall test can be applied to residuals from a simple linear regression of Y versus X. This approach should only be used when the relationship of Y and X complies with the appropriate assumptions for parametric statistics.
Second, the data can be “deseasonalized” by subtracting seasonal medians or some other measure of seasonal effect from all the data within the season. The deseasonalized data is then regressed against time (Montgomery and Reckhow 1984). Although this technique has the advantage of producing a description of seasonality (seasonal medians), it has generally low statistical power.
Nonparametric: Seasonal Kendall on Y (Example 3, p. 21)
The seasonal Kendall test statistic is computed by performing a Mann-Kendall calculation for each season, then combining the results for each season. For monthly seasons, January observations are compared only to other January observations, etc. No comparisons are made across seasonal boundaries. The Seasonal Kendall test is highly robust and relatively powerful, and is often the recommended method for most water quality trend monitoring.
The Sk statistic is computed as the sum of the S from each season: m ∑ i = 1
S (^) k = S (^) i
where Si is the S from the ith^ season and m is the number of seasons.
The seasonal statistics are summed and a Z statistic is computed; consult other sources for the method of calculating Z (^) Sk (e.g., Helsel and Hirsch 1992, USEPA 1997b). If the number of seasons and years are sufficiently large (seasons * years > 25), the Z value may be compared to standard normal tables to test for a statistically significant trend. For fewer seasons/years, the applicability of standard normal tables has not been evaluated. An estimate of the trend slope for Y over time can be computed as the median of all slopes between data pairs within the same season using a generalized version of the Sen slope estimator described above. Consult other sources for the method of calculation (e.g., Helsel and Hirsch 1992, USEPA 1997b).
Emerging trend analysis techniques
A recent paper by Hirsch et al. (2010) called for a “next generation” of trend analysis techniques in response to the observations that new and longer monitoring data sets exist, new questions about the effectiveness of control efforts, and the availability of new
statistical tools. The authors identified seven critical attributes for the next generation of trend analysis:
l Focuses on revealing the nature and magnitude of change, rather than strict hypothesis testing; l Does not assume that the flow-concentration relationship is constant over time; l Makes no assumptions that seasonal patterns repeat exactly over the period of record, but allow the shape of seasonality to evolve over time; l Allows the shape of an estimated trend to be driven by the data and not constrained to follow a specific form such as linear or quadratic; trend patterns should be allowed to differ for different seasons or flow conditions; l Provides consistent results describing trends in both concentration and load; l Provides not only estimates of trends in concentration and flux but also trend estimates where the variation in water quality due to variation in streamflow has been statistically removed; and l Includes diagnostic tools to assist in understanding the nature of the changes that have taken place over time, e.g., to identify particular times of year or hydrologic conditions during which water quality changes are most pronounced.
The authors propose and demonstrate an experimental trend analysis technique called Weighted Regressions on Time, Discharge, and Season (WRTDS) that addresses these critical attributes. While a presentation of this approach is beyond the scope of this Tech Note, the reader is referred to the original paper for additional information.
Step Trends
Monotonic trend analysis may not be appropriate for all situations. Other statistical tests for discrete changes (step trends) should be applied where a known discrete event (like BMP implementation over a short period) has occurred. Testing for differences between the “before” and “after” conditions is done using two-sample procedures such as t-tests and analysis of covariance (parametric techniques) and nonparametric alternatives such as the rank-sum test, Mann-Whitney test, and the Hodges-Lehmann estimator of step- trend magnitude (Helsel and Hirsch 1992, Walker 1994).
Monitoring Program Design and Trend
Analysis
Trend analysis is effective with data sampled continuously at fixed-time intervals. If you are presently designing your watershed monitoring program, here are key points to consider if you plan to use trend analysis to evaluate your project: l Use consistent sampling locations throughout the monitoring period; l Operate the monitoring program continuously, starting before implementation and continuing after implementation;
References
Chesapeake Bay Program. 2007. Chesapeake Bay 2007 health & restoration assessment. http://www.chesapeakebay.net/content/publications/cbp 26038.pdf_ [Accessed 9-15-2011]
Clausen, J.C. and J. Spooner. 1993. Paired watershed study design. 841-F-93-009. U.S. Environmental Protection Agency, Office of Water, Washington, DC
Cude, C. 2005. Specific examples of trend analysis using the Oregon water quality index. Oregon Department of Environmental Quality, Laboratory Division, http://www.deq.state.or.us/lab/wqm/trendex.htm [Accessed 9-15-2011]
Helsel, D.R. and R.M. Hirsch. 1992. Statistical methods in water resources. Studies in Environmental Science 49. New York: Elsevier. (available on-line as a pdf file at: http://water.usgs.gov/pubs/twri/twri4a3/ ) [Accessed 9-15-2011].
Helsel, D.R., D.K. Mueller, and J.R. Slack. 2005. Computer program for the Kendall family of trend tests. USGS Scientific Investigations Report 2005-5275, U.S. Geological Survey, Reston, VA. http://pubs.usgs.gov/sir/2005/5275/ [Accessed 9-15-2011]
Hirsch, R.M. 1988. Statistical methods and sampling design for estimating step trends in surface water quality. Water Resour. Res. 24:493-503.
Hirsch, R.M., R.B. Alexander, and R.A. Smith.1991. Selection of methods for the detection and estimation of trends in water quality. Water Resour. Res. 27:803-813.
Hirsch, Robert M., Douglas L. Moyer, and Stacey A. Archfield, 2010. Weighted regressions on time, discharge, and season (WRTDS), with an application to Chesapeake Bay river inputs. J. Am. Water Resour. Assoc. 46(5):857-880.
Indiana Department of Environmental Management. 2011. Mann-Kendall spreadsheet. http://www.in.gov/idem/4213.htm [Accessed 9-15-2011].
Karr, J.R. 1981. Assessment of biotic integrity using fish communities. Fisheries 6(6):28-30.
Lettenmaier. D.P. 1976. Detection of trends in water quality data from records with dependent observations. Water Resour. Res. 12:1037-1046.
Meals, D.W. 2001. Lake Champlain Basin agricultural watersheds section 319 national monitoring program project, final project report: May, 1994-September, 2000. Vermont Dept. of Environmental Conservation, Waterbury, VT, 227 p.
Montgomery, R.H. and K.H. Reckhow. 1984. Techniques for detecting trends in lake water quality. Water Resour. Bull. 20(1):43-52.
Newbold, J.D., S. Herbert, B.W. Sweeney and P. Kiry. 2008. Water quality functions of a 15-year-old riparian forest buffer system. NWQEP Notes No. 130. NCSU Water Quality Group, NC State University, Raleigh, NC.
Richards, R.P. and D.B. Baker. 1993. Trends in nutrient and suspended sediment concentrations in Lake Erie tributaries, 1975-1990. J. Great Lakes Res. 19(2):200-211.
Thas O., L. Van Vooren, and J.P. Ottoy. 1998. Nonparametric test performance for trends in water quality with sampling design applications. J. American Water Resour. Assoc. 34(2):347-357.
US EPA. 1997a. Linear regression for nonpoint source pollution analysis. EPA- 841-B-97-007. Office of Water, Washington, DC.
US EPA. 1997b. Monitoring guidance for determining the effectiveness of nonpoint source control projects. EPA 841-B-96-004. Office of Water, Washington, DC.
Walker, J.F. 1994. Statistical techniques for assessing water quality effects of BMPs. J. Irr. Drainage Eng. ASCE. 120(2):334-347.
Ye, Y-S and S. Zou. 1993. Relating trends of principal components to trends of water- quality constituents. Water Resour. Bull. 29(5):797-806.
Further Reading and Resources
There is far more to trend analysis than is covered in this technical note. For more details on the techniques and calculations discussed, examples, and information on other approaches, consult these additional sources:
Berryman, D., B. Bobee, D. Cluis, and J. Haemmerli. 1988. Nonparametric tests for trend detection in water quality time series. Water Resour. Bull. 24(3):545-556.
Gilbert, R.O. 1987. Statistical methods for environmental pollution monitoring. New York: Van Nostrand Reinhold.
Grabow, G.L., L.A. Lombardo, D.E. Line, and J. Spooner. 1999. Detecting water quality improvement as BMP effectiveness changes over time: use of SAS for trend analysis. NWQEP Notes No. 95. NCSU Water Quality Group, NC State University, Raleigh, NC.
Hirsch, R.M, J.R. Slack, and R.A. Smith. 1982. Techniques of trend analysis for monthly water quality data. Water Resour. Res. 18:107-121.
Practical Stats , offering statistical training for environmental sciences and natural resources, with emphasis on applied environmental statistics, nondetects data analysis, and multivariate analysis. http://www.practicalstats.com/index.html [Accessed 9-15-2011].
Rheem, S. and G.I. Holtzman. 1990. A SAS program for seasonal Kendall trend analysis of monthly water quality data. Proceedings, Sixteenth Annual SAS Users Group International (SUGI) Conference, February 17-20, 1991, New Orleans.
Richards, R.P. and D.B. Baker. 2002. Trends in water quality in LEASEQ rivers and streams (northwestern Ohio), 1975-1995. J. Environ. Qual. 31(1): 90 - 96.
Taylor, C.H. and J. C. Loftis. 1989. Testing for trend in lake and ground water time series. Water Resour. Bull. 25(4):715-726.
Trend analysis example 2: Mann-Kendall
l Eight years of quarterly mean total phosphorus concentration data from Samsonville Brook, a stream draining a Vermont agricultural watershed l Data satisfy assumptions for constant variance and independence, but are not normally-distributed without transformation
The Mann-Kendall trend test for this example may be evaluated in two ways. First, in a manual calculation, use the formulas below. The value of S (sum of the signs of differences between all combinations of observations) can be determined either manually or by using a spreadsheet to compare combinations, create dummy variables (-1, 0, and +1), and sum for S. n - 1 ∑ i = 1
n ∑ j = i + 1
Mann – Kendall S = sign ( y (^) j - y (^) i ) = - 106
τ = S n(n - 1)/
= -^106
= - 0.353 decreasing trend
Calculating Z (^) S as (S ± 1) /σ s where σ s = (n/18) × (n - 1) × (2n + 5) = 42.
Z = -^105
= - 2.454 (^) (USEPA 1997b)
Month (n=25)
[TP] (mg/L) 1 0. 5 0. 9 0. 13 0. 17 0. 21 0. 25 0. 29 0. 33 0. 37 0. 41 0. 45 0. 49 0. 53 0. 57 0. 61 0. 65 0. 69 0. 73 0. 77 0. 81 0. 85 0. 89 0. 93 0. 97 0.
This Z statistic is significant at P = 0.014, indicating a significant trend, i.e., we are 98.6% confident there is a decreasing trend in TP. See USEPA (1997b) for the calculation of σS when there are ties among the data.
To estimate the rate of change, use the Sen slope estimator
β 1 = median ( )
y (^) j - y (^) i x (^) j - x (^) i
211 individual slopes - 0.00945 to +0. median slope = -0.0011 mg/L/month = -0.013 mg/L/yr
This result suggests that total P concentration decreased significantly over the period at a rate of about 0.013 mg/L/yr.
Alternatively, use a statistics computer program to run the Mann-Kendall procedure. For example, using the USGS program for the Kendall family of tests (Helsel et al. 2005), set up a text data input file specifying the Mann-Kendall test (test #4) without flow adjustment (“0”) or seasons (blanks) and name the data input file (“MKexample2.txt”) as:
4 0 MKexample 2 1 0. 5 0. 9 0. . . 97 0.
The output from the program gives the same results as shown above, including the estimated slope of the trend (-0.0011) computed by the Sen slope estimator above:
Kendall's tau Correlation Test US Geological Survey, 2005 Data set: MK Example 2 The tau correlation coefficient is -0. S = -106. z = -2. p = 0. The relation may be described by the equation: Y = 0.15412 + -0.1125E-02 * X
Note: data used in this example are taken from the Vermont NMP Project, Lake Champlain Basin agricultural watersheds section 319 national monitoring program project , 1993 – 2001 (Meals 2001). Helsel, D.R., D.K. Mueller, and J.R. Slack. 2005. Computer program for the Kendall family of trend tests. USGS Scientific Investigations Report 2005-5275, U.S. Geological Survey, Reston, VA. http://pubs.usgs.gov/sir/2005/5275/
