Non-stationary approach to at-site flood frequency modelling II. Weighted least squares estimation
Introduction
For detecting a trend, standard statistical parametric and non-parametric techniques are used, which are based on the theory of hypothesis testing. The advantage of a non-parametric approach is that no assumption is made about the trend model and the functional form of the distribution function of random variables (e.g. Mitosek, 1992). If no tendency is detected with respect to the mean value, one can investigate a change in the variance by Mann's test (Mann, 1945). However, if the mean happens to be nonstationary there is neither a possibility of detecting trend in the variance nor investigating nonstationarity of the two first moments simultaneously. A great disadvantage of the non-parametric tests applied in hydrology is that they require much longer series than do parametric tests. Moreover, hydrologists expect that the information on statistically significantly increasing (or decreasing) tendency in the mean or the variance is supplemented by quantification of recorded trend and its prediction. The probability distribution function with time-variant parameters is the preferable output of trend investigation for further use in hydraulic design and water resources planning.
In the parametric approach, it is the form of a trend in the mean value that should be assumed. A linear or higher degree polynomial model is usually adopted for this purpose. Its coefficients are estimated by the least squares (LS) method and the parametric tests are used to verify the statistical significance of the trend (e.g. Giakoumakis and Baloutsos, 1997). However, most parametric tests require the characteristics under verification to be normally distributed, which rarely holds in hydrology. While estimating a trend in the mean by the LS method, no allowance is usually made for the time-variant variance. In this paper — the second of the three-part paper — a generalisation of the LS method for time-variable variance is presented which does not require as rigorous a distribution assumption as does the ML method. In fact, since water resources systems are supposed to perform satisfactorily during extreme hydrological conditions, such as exceptionally dry or wet years, a time-change of the variance may be even more important than that of the mean. For example, the difference ΔQ (m3/s) in the mean of flood frequency distribution results in the difference ΔQ in peak flow with a return period of 100 years (Q1%), while the same difference in the standard deviation (SD) makes the difference 2.33ΔQ in Q1%, if the normal distribution is assumed.
It should be stressed that even if an investigation of trend in only the mean is of concern, heteroscedasticity has to be taken into account. Therefore investigating a long time series of unknown character, one assumes a most general case, i.e. of functionally non-related trends in the first two statistical moments, and then may proceed to simpler cases ending at the stationary case, and finally identifying the model showing the best fit to the time series. The investigation can comprise various functional forms of trend, e.g. linear, parabolic or the periodic functions of time, while keeping in mind the need for parameter parsimony. Statistical significance of detected trends is not tested but the model showing the best fit of all competing models is considered here as the optimal one.
Section snippets
Method of moments as the LS method
The method of moments (MOM) is used to estimate the mean and other moments of time-series when all observations are considered to come from the same distribution, while the LS method is usually applied when this condition no longer holds. Since the subject of the study is the estimation of the time-variant moments, it is worth looking at the MOM from the point of view of the LS method.
Let the problem be to estimate, by the LS method, the moments of the probability distribution of the random
Non-identical parent distributions
So far, we have been considering the problem of estimation when all the observations come from the same underlying distribution. We will now examine the situation when this condition no longer holds. In general, the acceptability of the LS method depends on the properties of the estimators to which it leads. In a situation usually appearing in the investigation of trend in hydrological time series, it has such a property, even in small samples, that it provides unbiased estimators, linear in
Unequal known variances
In order to examine the influence of the time-variant variance on the properties of the estimator of the mean, a simple case will be considered first when elements of the sample differ in the variances which are known. It can be exemplified by the estimation of the mean from a sample with unequal accuracy of data. Let us place the weight γt(1) to every term St(1) of the sum of squares in Eq. (1)Then, from the LS method we get:For unbiased
Unknown variances
To use Eq. (19) (or Eq. (14a)) we need to know σ2(xt) for t=1,2,…,T. This rarely holds in practice. On the other hand, in order to estimate unknown variances, the functional form of a trend in variance should be assumed. If so and the time-dependent mean is also known, then the only problem left will be to estimate the parameters of time dependent variance (Eq. (10a)). To this end, the WLS method shall be applied this time in respect to the second central moment. We have:
Derivation of weighting factors
To solve , , the number of unknown variables would be reduced to those describing the time-dependent mean and variance. However, neither the variance of the second central moment nor the fourth central moment of xt in Eq. (22) is known. In order to express in terms of lower order moments, we have to presuppose the knowledge of the functional form of the distribution(s) generating the time series, which would contain less than four parameters. Similarly, to derive the coefficient of
The normality assumption
The LS method, while conceptually quite distinct from the ML method, coincides with the ML method in the case of normally distributed observations. If the distributions of xt about their expectations are independently normal with the same variance, as in the trend model A, the LS estimator will be identical with the ML estimator of time-variant mean mt. It will be shown here that the above identity holds for the WLS estimators.
The log likelihood function in a time series from normal
Concluding remarks
To deal with the trend in the first two moments, the LS method has been generalised to the case where the assumption of constant variance does not hold and functional form of a trend in the variance is given. Its generalisation is the WLS method where the parameters of trend in the mean and variance are estimated simultaneously. Each of the two weighted squares functions, i.e. for time-variant mean and variance, contains both variables and therefore they have to be minimised jointly. The first
Acknowledgements
The work reported in this study was supported in part by the Polish Committee of Scientific Research (KBN) grant No. PO4D 056 17, ‘Revision of applicability of the parametric methods for estimation of statistical characteristics of floods’. This support is gratefully acknowledged.
References (12)
- Aitchison, J., Brown, J.A.C., 1957. The Lognormal distribution. General Properties of Lognormal Distribution. Cambridge...
A new look at the statistical model identification
IEEE Trans. Automat. Contr.
(1974)Probability Theory and Mathematical Statistics
(1963)- et al.
Investigation of trend in hydrological time series of the Evinos river basin
Hydrol. Sci. J.
(1997) Statistical methods in Hydrology and Meteorology
(1977)- et al.
The advanced theory of statistics
(1969)
Cited by (90)
Analysis of long-term changes in inundation characteristics of near-natural temperate riparian habitats in the Lower Basin of the Biebrza Valley, Poland
2021, Journal of Hydrology: Regional StudiesCitation Excerpt :In our research, we adopted the statistical methodology from the second group (ii), called the weighted least squares (WLS) method. The WLS was developed by Strupczewski and Kaczmarek (1998, 2001b) to detect and estimate trends in parameters of the flood frequency models of peak annual river discharges. However, the simplicity and universality of this technique preordain it also to be used for trend detection in the area of residence of inundation waters and time of the inundation waters’ residence in the wetland pits.
Drought analysis framework based on copula and Poisson process with nonstationarity
2020, Journal of Hydrology