R. B. D'Agostino and M. A. Stephens, editors. Goodness-of-Fit Techniques. Marcel Dekker, Inc., 1986.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....weekend. On average, NYUHome received 1706 requests an hour: Figure 4(b) shows the minimum and maximum requests received during the same hour over the two week period. Figure 4(c) shows the cumulative distribution of the inter request arrival interval. Using the # method as the goodness of fit [15] measure, we found that this distribution is captured very well by an Exponential distribution with # = 0.526, suggesting a Poisson arrival process. This observation seemingly conflicts with that from previous studies of web servers [6] and telnet sessions [33] where it was found that the ....

R. B. D'Agostino and M. A. Stephens, editors. Goodness-of-Fit Techniques. Marcel Dekker, Inc, 1986.

....weekend. On average, NYUHome received 1706 requests an hour: Figure 4(b) shows the minimum and maximum requests received during the same hour over the two week period. Figure 4(c) shows the cumulative distribution of the inter request arrival interval. Using the X 2 method as the goodness of fit [15] measure, 3 we found that this distribution is captured very well by an Exponential distribution with ) 0.526, suggesting a Poisson arrival process. This observation seemingly conflicts with that from previous studies of web servers [6] and telnet sessions [33] where it was found that the ....

R.B. D'Agostino and M. A. Stephens, editors. Goodness-of-Fit Techniques. Marcel Dekker, Inc, 1986.

....approximations of the data. Unfortunately, because of the transformation used to convert the fractions of samples to fractions of events, we cannot apply standard goodness of t hypothesis testing techniques such as the ChiSquare goodness of t test or Anderson Darling goodness of t test [8] to these distributions. However, to put these ts in perspective, several observations are worth noting. First, the maximum positive and negative di erences from the data to the tted lines are D = 029, D = 003, D UB = 007, and D UB = 031. Second, the maximum di erence between the ....

R. D'Agostino and M. Stephens, editors. Goodness-of-Fit Techniques. Marcel Dekker, Inc., 1986.

....4. To develop a model for the object size distributions seen for different size limits, we use standard statistical methods similar to those used by Paxson et al. in [32] We used both Chi Square and Anderson Darling (A 2) empirical distribution functions (EDF) for estimation of goodness of fit [15]. By comparing with several distributions, such as Lognormal, Exponential, Weibull and Pareto, we found that the best 10 lOO lOOO 10000 lOO lOOO Objectsze Objectsze (a) cnn (b) yahoo 1.0 (c) nytimes 1 oo 1 ooo 1 oooo 1 oo 1 ooo Fig. 4. The cumulative distribution function (of ....

R. B. D'Agostino and M. A. Stephens, editors. Goodness-of-Fit Techniques. Marcel Dekker, Inc, 1986.

....its model parameters through one sweep of the data at each resolution. The error metric on the mean modeler is a variant of the root mean square error (RMSE) Our second model captures the normality of systematic partitions of the data by utilizing the Anderson Darling goodness of fit test [5]. This model is called the goodness of fit modeler. Similar to the mean modeler, the goodness of fit modeler is able to calculate its model parameters through one sweep of the data. However, this modeler attempts to fit the data to a normal distribution. The error on this model is the Type I error ....

....deviation, i . For the goodness of fit modeler, the partitioning step stops when the hypothesis test for normality is not rejected. We use the Anderson Darling test for normality (which is considered to be the most powerful goodness of fit test for normality) for our goodness of fit test [5]. The Anderson Darling test involves calculating the A metric for variable v N( i , i ) which is defined to be ( n j j n z j z j n A = 1 1 ln( ln( 1 2 1 2 where n = number of data points for v i and z j = i j x ) is the standard ....

[Article contains additional citation context not shown here]

D'Agostino, R.B., and Stephens, M.A. Goodness-of-fit Techniques, Marcel Dekker, Inc., 1986.

....methods similar to those used by Paxson in [22] Due to space restrictions, we discuss only the results for the cnn trace, which is representative of the others. Comparing several distributions, such as Lognormal, Exponential, Weibull and Pareto, using the Chi Square method as the goodness of fit [12] measure, we found that object sizes were best modeled using an Exponential distribution (with CDF F (x) 1 e #x ) The results of the Chi Square tests were in the range 0.1 to 2.5, showing a very close fit. Although not shown, we observed similar distribution fits for different settings of ....

R. B. D'Agostino and M. A. Stephens, editors. Goodness-of-Fit Techniques. Marcel Dekker, Inc, 1986.

....weekend. On average, NYUHome received 1706 requests an hour: Figure 4(b) shows the minimum and maximum requests received during the same hour over the two week period. Figure 4(c) shows the cumulative distribution of the interrequest arrival interval. Using the # method as the goodness of fit [15] measure, we found that this distribution is captured very well by an Exponential distribution with # = 0.526, suggesting a Poisson arrival process. This observation seemingly conflicts with that from previous studies of web servers [6] and telnet sessions [33] where it was found that the ....

R. B. D'Agostino and M. A. Stephens, editors. Goodness-of-Fit Techniques. Marcel Dekker, Inc, 1986.

.... Parametric techniques assume that knowledge about the underlying statistical distribution is available (often normal distribution is assumed) and that the task is to validate the assumption regarding the distribution, calculate the corresponding parameters, and establish intervals of confidence [D A86]. Nonparametric techniques do not make any assumptions about the statistical distribution. They aim to figure conceptually and quantitatively the simplest (and therefore best) model which fits the recorded data [Can99] Therefore, nonparametric techniques are significantly more computationally ....

R.B. D'Agostino, M. A. Stephens. "Goodness-of-fit techniques." New York: M. Dekker, 1986.

....reciprocal of the mean. A statistical test is that of Anderson Darling [6] This test has been found to be more powerful than either the Kolmogorov Smirnov or the # tests, i.e. its probability of correctly 136 rejecting the null hypothesis (that the distribution is exponential) is greater; see [15]. This is, in part, due to the fact that the Anderson Darling test employs the full empirical distribution (rather than binning, as in a # test) allowing it to give more weight to larger sample values whose presence can lead to a violation of the null hypothesis. For a set of n rank ordered ....

....is: A n (2i 1) log(1 e t i t t n 1 i t o where t = n 1 i=1 t i is the empirical mean inter event time. We reject the null hypothesis at significance level # if the test statistic exceeds the tabulated values appropriate for that level; see, e.g. Table 4. 11 in [15]. We note the importance of using the table appropriate to the present case in which the mean is estimated from the sample, rather than being specified in advance. Moreover, the table explicitly takes into account the e#ect of a finite sample size n. ....

R. B. D'Agostino and M. A. Stephens, editors. Goodness-of-Fit Techniques. Marcel Dekker, New York, 1986.

....approximations of the data. Unfortunately, because of the transformation used to convert the fractions of samples to fractions of events, we cannot apply standard goodness of fit hypothesis testing techniques such as the Chi Square goodness of fit test or Anderson Darling goodness of fit test [8] to these distributions. How ever, to put these fits in perspective, several observations are worth noting. First, the maximum positive and negative differences from the data to the fitted lines are Dr, s .029, DZs .003, Dv s = 007, and Ds = 031. Second, the maximum difference between the ....

R. D'Agostino and M. Stephens, editors. Goodness-of-Fit Techniques. Marcel Dekker, Inc., 1986.

....line. The Q Q plots in Figures 4 and 5 compare the log of the rate distribution to the normal distribution for two of the traces (Access1c and Regional2) The fit between the two is visually good. As in Reference [2] we further assess the goodness of fit using the Shapiro Wilk normality test [5]. For Access1c (Figure 4) we can not reject the null hypothesis that the log of rate comes from normal distribution at 25 significance level; for Regional2 (Figure 5) we can not reject normality at any level of significance. This suggests the fit for a normal distribution is indeed very good. ....

R. D'Agostino and M. Stephens, Eds., "Goodness-of-Fit Techniques," Marcel Dekker, New York, 1986.

....of data transferred, continuous in the non negative integers. As such the values of the variables do not naturally fall into a finite number of categories, which makes using the well known chi squared test less than ideal because it requires somewhat arbitrary choices regarding binning [Knuth81, DS86] The goodness of fit test commonly used with continuous data is the Kolmogorov Smirnov test. The authors of [DS86] however, recommend the Anderson Darling ( test [AD54] instead. They state that is often much more powerful than either Kolmogorov Smirnov or chi squared, and that ....

....fall into a finite number of categories, which makes using the well known chi squared test less than ideal because it requires somewhat arbitrary choices regarding binning [Knuth81, DS86] The goodness of fit test commonly used with continuous data is the Kolmogorov Smirnov test. The authors of [DS86] however, recommend the Anderson Darling ( test [AD54] instead. They state that is often much more powerful than either Kolmogorov Smirnov or chi squared, and that particularly good for detecting deviations in the tails of a distribution, often the most important to detect. We ....

[Article contains additional citation context not shown here]

R. B. D'Agostino and M. A. Stephens, editors, "Goodness-of-Fit Techniques", Marcel Dekker, Inc., 1986.

....4. To develop a model for the object size distributions seen for different size limits, we use standard statistical methods similar to those used by Paxson et al. in [27] We use both Chi Square and Anderson Darling (A 2) empirical distribution functions (EDF) for estimation of goodness of fit [12]. We found that most of object size distributions have a very good fit to the Exponential distribution, whose cumulative distribution function is F(x) II1, 1 I , I . o Object size Object size Object size (a) cnn (b) yahoo (c) nytimes o I1=1 I I I= ....

R.B. D'Agostino and M. A. Stephens, editors. Goodness-of-Fit Techniques. Marcel Dekker, Inc, 1986.

....behavior from Web client logs. To develop these models we use the client traces collected at Boston University in 1994 and 1995 by Cunha et al. 29] Our modeling methodology is similar to the techniques used by Paxson in [102] These techniques use goodness of fit tests described in detail in [22, 30, 109]. In this thesis, we present and use models extracted from the 94 and 95 client traces 1 . Workload generators are often used in a local area network when researchers do not have access to client systems deployed in the wide area. Use of workload generators in the local area means that effects ....

....models for each of these Web characteristics required the analysis of empirically measured Web workload traces. The most common way of specifying a statistical model for a set of data the is through the use of visual methods such as quantile quantile or cumulative distribution function (CDF) plots [30]. These methods, however, do not distinguish between two closely fitting distributions nor do they provide any level of confidence in the fit of the model. To address this drawback one can use goodness of fit tests [30, 102] However, these tests also present a number of problems. Methods which ....

[Article contains additional citation context not shown here]

R. D'Agostino and M. Stephens, editors. Goodness-of-Fit Techniques. Marcel Dekker, Inc., 1986.

....of the reciprocal of the mean. A statistical test is that of Anderson Darling. This test has been found to be more powerful than either the Kolmogorov Smirnov or the I tests, i.e. its probability of correctly rejecting the null hypothesis (that the distribution is exponential) is greater; see [DS86]. This is, in part, due to the fact that the Anderson Darling test employs the full empirical distribution (rather than binning, as in a I test) allowing it to give more weight to larger sample values whose presence can lead to a violation of the null hypothesis. For a set of rank ordered ....

....Anderson Darling statistic is: I u G yO L G where u G G is the empirical mean inter event time. We reject the null hypothesis at significance level if the test statistic exceeds the tabulated values appropriate for that level; see, e.g. Table 4. 11 in [DS86]. We note the importance of using the table appropriate to the present case in which the mean is estimated from the sample, rather than being specified in advance. Moreover, the table explicitly takes into account the effect of a finite sample size . ....

R.B. D'Agostino and M.A. Stephens, Eds., Goodness-of-Fit Techniques, Marcel Dekker, New York, 1986.

....of the reciprocal of the mean. A statistical test is that of Anderson Darling. This test has been found to be more powerful than either the Kolmogorov Smirnov or the 2 tests, i.e. its probability of correctly rejecting the null hypothesis (that the distribution is exponential) is greater; see [DS86]. This is, in part, due to the fact that the Anderson Darling test employs the full empirical distribution (rather than binning, as in a 2 test) allowing it to give more weight to larger sample values whose presence can lead to a violation of the null hypothesis. For a set of n rank ordered ....

....A 2 = n 1 n n X i=1 (2i 1) n log(1 e t i =t ) t n 1 i =t o where t = n 1 P n i=1 t i is the empirical mean inter event time. We reject the null hypothesis at significance level if the test statistic exceeds the tabulated values appropriate for that level; see, e.g. Table 4. 11 in [DS86]. We note the importance of using the table appropriate to the present case in which the mean is estimated from the sample, rather than being specified in advance. Moreover, the table explicitly takes into account the effect of a finite sample size n. ....

R.B. D'Agostino and M.A. Stephens, Eds., Goodness-of-Fit Techniques, Marcel Dekker, New York, 1986.

....uniformity The estimation approach outlined above depends on the spoofed source addresses being uniformly distributed across the entire IP address space. To check whether a sample of observed addresses are uniform in our monitored address range, we compute the Anderson Darling (A2) test statistic [9] to determine if the observations are consistent with a uniform distribution. In particular, we use the implementation of the A2 test as specified in RFC2330 [19] at a 0.05 significance level. 3.3 Analysis limitations There are three assumptions that underly our analysis: Address uniformity: ....

R. D'Agostino and M. Stephens. Goodness-of-Fit Techniques. Marcel Dekker, Inc., 1986.

....phase differences. An alternative approach performs complex addition of two band limited quadrature additive white Gaussian noise (AWGN) sources, a solution advocated for example by Arredondo et al. 35] and Goubran et al. 36] We carried out a rigorous Rayleigh distribution hypothesis testing [37] for the two methods and for a variety of reasons the latter quadrature noise source method was preferred. A convenient approach to understanding how the above parameters will affect the latency of an adaptive modulation scheme is to describe the transmission mode switching scheme with a Markov ....

....64 QAM frame were not negligible for the system parameters of interest, in particular for more rapidly fluctuating higher Doppler frequencies. Hence, we decided to determine the maximum Doppler frequency up which the Wang Moayeri [38] model is applicable to our adaptive regime using testing [24] [37]. The normalized Doppler frequencies, where the confidence levels became unacceptable are tabulated in Table II. Inspection of this table reveals that, for both sets of switching SNR s, the maximum normalized Doppler frequency, for which the Wang Moayeri model is applicable to our problem ....

R. B. D'Agostino and M. A. Stephens, Eds., Goodness-of-Fit Techniques. New York: Marcel Dekker, 1986.

....PDF that decreases exponentially over time. 4.4 Outlier Detection The theory and practice of identifying outliers in a dataset is another APPL application. The literature contains ample methods for identifying outliers in samples from a normal population. Regarding detection of outliers, D Agostino and Stephens (1986, p. 497) wrote We shall discuss here only the underlying assumption of normality since there is very little theory for any other case. Sarhan and Greenberg (1962, p. 302) and David (1981) proposed a number of test statistics based on standardized order statistics of normally distributed data. ....

D'Agostino, R. B., and Stephens, M. A. (1986), Goodness-of-Fit Techniques, New York: Marcel Dekker.

....have high variability Log log complementary distribution is a means for viewing tail characteristics: P [X x] log(Size) 0 1 2 3 4 5 6 # Web Server Performance Analysis Paul Barford Page 32 Analyzing Data Contd. ffl Modeling Fitting standard distributions to empirical distributions [21] Visual methods include CDF, PDF, LLCD and Quantile Quantile plots Sometimes its very hard to distinguish between standard distribution fits The reality is that most models are fit via visual methods Goodness of fit metrics include Chi Squared and 2 [49] methods Choice of bin size ....

R. B. D'Agostino and M. A. Stephens, editors. Goodness-of-Fit Techniques. Marcel Dekker, Inc., 1986.

....lobe tighter. For L shaped curved ( a 1 changing l has a less dramatic effect as the asymptotic behavior dominates. An alternate method for getting the parameters of the Weibull distribution is to detect a straight line on a probability plot of y x = ln( versus zFx= ln ( ln ( 1 [29]. From a straight line fit ymzc= one obtains a =1 m , and l = 1 e c . With these observations as guiding principles we estimated the parameters for the video sequences and superimposed the resulting distribution over the distribution derived from the trace data. 4.3 Fitting a ....

R. B. Agostino and M. A. Stephens, Goodness-of-fit Techniques, Marcel Dekker, 1986

....the course of the day. Figure 15.5 shows a histogram of the number of successful North American measurements made for each distinct hour of the day. The distribution appears fairly even, and, indeed, the measurement times pass the powerful Anderson Darling A 2 goodness of fit test for uniformity [DS86], using 5 significance (and, indeed, for higher significance) Figure 15.6 shows the same histogram for the European measurements. The bias towards the less busy early morning and late evening hours immediately stands out. The distribution fails A 2 at all significance levels, as one might ....

R. B. D' Agostino and M. A. Stephens, editors, Goodness-of-Fit Techniques, Marcel Dekker, Inc., 1986.

....(for the Pareto distribution only) and maximum likelihood estimation. The next step is to test whether the #t is adequate. This is usually done by comparing the #tted and empirical d.f. s, more precisely, by checking whether values of the #tted d.f. at sample points form a uniform distribution [11]. We applied the welland not so well known non parametric tests verifying the hypothesis of uniformity. The critical values C # of the tests, given a signi#cance level # (e.g. #=0:05) can be easily found in the literature [11,12] 274 K. Burnecki et al. Physica A 287 (2000) 269 278 A very ....

....values of the #tted d.f. at sample points form a uniform distribution [11] We applied the welland not so well known non parametric tests verifying the hypothesis of uniformity. The critical values C # of the tests, given a signi#cance level # (e.g. #=0:05) can be easily found in the literature [11,12]. 274 K. Burnecki et al. Physica A 287 (2000) 269 278 A very natural and well known is the # 2 statistics # 2 k = k k # i=1 (n i n=k) 2 n ; where n is the overall number of observations and n i is the number of observations which fall into the interval [ i 1) k; i=k] # 2 k ....

[Article contains additional citation context not shown here]

R.B. D'Agostino, M.A. Stephens, Goodness-of-Fit Techniques, Marcel Dekker, New York, 1986.

....are currently investigating networklevel congestion control protocols for CM, and the effect of client interactivity on the amount of CM data that should be transmitted in advance to the client. We are also currently integrating video into the MANIC system. More rigorous goodness of fits studies [2] can be performed to quantify the benefits of one analytic distribution over another. Finally, even though one distribution may more accurately model the workload than another, the overall effect of using different workload models in various performance studies is another topic for future ....

R. D'Agostino, Goodness of Fit Techniques, Marcel Decker, 1980.

....tests. We use KolmogorovSmirnov D statistics to perform these tests. For each stress level and distribution combination, we test the null hypothesis that the data obtained come from the corresponding random fatigue limit distribution. To do this, we adapt the methods discussed by D Agostino and Stephens (1986, chap. 4) Let w 1 = log(y 1 ) w n = log(y n ) be the ordered observations at log stress x. Let z i = FW (w i # x# ) for i = 1# : # n. Under the true value , Z i = FW (W i # x# ) are ordered uniform random variables. If one or more components of are unknown, these components are ....

....= FW (W i # x# ) are ordered uniform random variables. If one or more components of are unknown, these components are replaced by estimates, for example, the ML estimate b . However, z i = FW (w i # x# b ) will not be an ordered uniform sample even when the null hypothesis is true. D Agostino and Stephens (1986) suggested modifications of the test statistic to account for the use of b in place of . The modifications are functions of the test statistic and sample size. They give tables of percentage points for the modified test statistic. We shall replace unknown parameter values with ML estimates ....

[Article contains additional citation context not shown here]

D'Agostino, R., and Stephens, M. (1986), Goodness-of-Fit Techniques, New York: Marcel Dekker.

....) 1 2 (fa ;gsum ) h T ; T ; D 2 )i; h T ; T ; D T 2 )i) 5.2.2 Chi squared Goodness of Fit Statistic The computation of the chi squared statistic X 2 assumes that the entire space is partitioned into cells each of which is associated with expected and observed measures. See [DS86] for details. To apply the chi squared test to dt models, we use the regions associated with a decision tree T as the cells since these regions partition the entire attribute space. The expected and observed measures are: E( i ; D 2 ) i ; D 1 ) jD 2 j; O( i ; D 2 ) i ; D 2 ) jD ....

....The statistic X 2 can now be computed in a straightforward way except for two problems: 1) For the chi squared statistic to be well defined, E( i ; D 2 ) should not be zero. We follow the standard practice in Statistics and add a small constant c 0 (0. 5 is a common choice) to ensure this [DS86] 22 (2) At least 80 of the expected counts must be greater than 5 in order to use the standard X 2 tables. In a decision tree, this condition is often violated. For example, if all tuples in node n are of class i, the expected measures for regions n j ; j 6= i will be zero. The solution ....

[Article contains additional citation context not shown here]

Ralph B. D'Agostino and Michael A. Stephens. Goodness-of-fit techniques. New York: M.Dekker, 1986.

....tests. We use KolmogorovSmirnov D statistics to perform these tests. For each stress level and distribution combination, we test the null hypothesis that the data obtained come from the corresponding random fatigue limit distribution. To do this, we adapt the methods discussed by D Agostino and Stephens (1986, chap. 4) Let w 1 = log(y 1 ) w n = log(y n ) be the ordered observations at log stress x. Let z i = FW (w i ; x; for i = 1; n. Under the true value , Z i = FW (W i ; x; are ordered uniform random variables. If one or more components of are unknown, these components are ....

.... , Z i = FW (W i ; x; are ordered uniform random variables. If one or more components of are unknown, these components are replaced by estimates, for example, the ML estimate b . However, z i = FW (w i ; x; b ) will not be an ordered uniform sample even when the null hypothesis is true. D Agostino and Stephens (1986) suggested modifications of the test statistic to account for the use of b in place of . The modifications are functions of the test statistic and sample size. They give tables of percentage points for the modified test statistic. We shall replace unknown parameter values with ML estimates and ....

[Article contains additional citation context not shown here]

D'Agostino, R., and Stephens, M. (1986), Goodness-of-Fit Techniques, New York: Marcel Dekker.

....are currently investigating network level congestion control protocols for CM, and the effect of client interactivity on the amount of CM data that should be transmitted in advance to the client. We are also currently integrating video into the MANIC system. More rigorous goodness of fits studies [2] can be performed to quantify the benefits of one analytic distribution over another. Finally, even though one distribution may more accurately model the workload than another, the overall effect of using different workload models in various performance studies is another topic for future ....

R. D'Agostino, Goodness of Fit Techniques, Marcel Decker, 1980.

....T ; Sigma( Gamma T ; D 2 )i; h Gamma T ; Sigma( Gamma T ; D T 2 )i) 5.2.2 Chi squared Goodness of Fit Statistic The computation of the chi squared statistic X 2 assumes that the entire space is partitioned into cells each of which is associated with expected and observed measures. See [13] for details. To apply the chi squared test to dt models, we use the regions associated with a decision tree T as the cells since these regions partition the entire attribute space. The expected and observed measures are: E(fl i ; D 2 ) oe(fl i ; D 1 ) Delta jD 2 j; O(fl i ; D 2 ) oe(fl i ; ....

....f a,g sum:WOR;minSup=0.006 Figure 7: SD vs SF Figure 8: SD vs SF Figure 9: SD vs SF (1) For the chi squared statistic to be well defined, E(fl i ; D 2 ) should not be zero. We follow the standard practice in Statistics and add a small constant c 0 (0. 5 is a common choice) to ensure this [13]. 2) At least 80 of the expected counts must be greater than 5 in order to use the standard X 2 tables. In a decision tree, this condition is often violated. For example, if all tuples in node n are of class i, the expected measures for regions fl n j ; j 6= i will be zero. The solution to ....

[Article contains additional citation context not shown here]

Ralph B. D'Agostino and Michael A. Stephens. Goodness-of-fit techniques. New York: M.Dekker, 1986.

.... Miller [QM77, MQ79] made a thorough Monte Carlo simulation to compare a number of existing statistics for testing uniformity in [0; 1] and recommended the Watson s U 2 test [Wat62] and the Neyman s smooth test [Ney37] as general choices for testing uniformity in [0; 1] D Agostino and Stephens [DS86, Chapter 6] gave a comprehensive review on tests for uniformity in [0; 1] Testing uniformity of random samples in the multidimensional unit cube (d 2) C d = 0; 1] d = fx = x 1 ; Delta Delta Delta ; x d ) 0 2 R d : 0 x i 1; i = 1; dg; 1:1) seems to have received less attention in ....

R. B. D'Agostino and M. A. Stephens, Goodness-of-fit Techniques, Marcel Dekker, Inc., New York and Basel, 1986.

....in the case of data transferred, continuous in the non negative integers. As such the values of the variables do not naturally fall into a finite number of categories, which makes using the well known chi squared test less than ideal because it requires somewhat arbitrary choices regarding binning [Knuth81, DS86]. The goodness of fit test commonly used with continuous data is the Kolmogorov Smirnov test. The authors of [DS86] however, recommend the Anderson Darling (A 2 ) test [AD54] instead. They state that A 2 is often much more powerful than either Kolmogorov Smirnov or chi squared, and that A 2 ....

....fall into a finite number of categories, which makes using the well known chi squared test less than ideal because it requires somewhat arbitrary choices regarding binning [Knuth81, DS86] The goodness of fit test commonly used with continuous data is the Kolmogorov Smirnov test. The authors of [DS86], however, recommend the Anderson Darling (A 2 ) test [AD54] instead. They state that A 2 is often much more powerful than either Kolmogorov Smirnov or chi squared, and that A 2 is particularly good for detecting deviations in the tails of a distribution, often the most important. We ....

[Article contains additional citation context not shown here]

R. B. D'Agostino and M. A. Stephens, editors, "Goodness-of-Fit Techniques", Marcel Dekker, Inc., 1986.

.... (x) 2 dx = 1 3 s Gamma 2 N X z2P s Y j=1 1 Gamma z 2 j 2 1 N 2 X z;z 0 2P s Y j=1 [1 Gamma max(z j ; z 0 j ) This is known as the L 2 star discrepancy in the numerical quadrature literature, and also as the Cram er Von Mises goodness of fit statistic [1]. The averagecase analysis for this particular kernel was done in [8] and the worst case analysis in [9] In the next section the connection between discrepancies and goodness of fit statistics is generalized. 3. The Discrepancy is a Goodness of Fit Statistic From the definition in (4) it is ....

R. B. D'Agostino and M. A. Stephens (eds.), Goodness-of-fit techniques, Marcel Dekker, New York, 1986.

....and describing the random variables. Works such as Hogg and Craig (1995) Casella and Berger (1990) and David (1981) organize their efforts according to the second contribution, covering theoretical results that apply to random variables. Works such as Law and Kelton (1991) Lehmann (1986) and D Agostino and Stephens (1986) concentrate on the statistical applications of random variables, and tailor their explanations of probability theory to the portions of the field that have application in statistical analysis. In all these works, as well as countless others, one stark omission is apparent. There is no mention of ....

....a = 1, b = 0:5 and c = 10: 4.4 Outlier Detection The theory and practice of identifying outliers in a data set is another contribution provided by APPL. The literature contains ample methods for identifying outliers in samples related to a normal distribution. Regarding detection of outliers, D Agostino and Stephens (1986, p. 497) write We shall discuss here only the underlying assumption of normality since there is very little theory for any other case. Sarhan and Greenberg (1962, p. 302) and David (1981) propose a number of test statistics based on standardized order statistics of normally distributed data. ....

D'Agostino, R. B. and Stephens, M. A. (1986), Goodness-of-Fit Techniques, New York: Marcel Dekker.

No context found.

R. B. D'Agostino and M. A. Stephens, editors. Goodness-of-Fit Techniques. Marcel Dekker, Inc., 1986.

No context found.

M. A. Stephens, Goodness-of-fit techniques, New York : M. Dekker, 1986.

No context found.

R. B. D'Agostino and M. A. Stephens(editors), Goodness-of-fit Techniques. New York: Marcel Dekker, June, 1986, pp. 63-93, pp. 97-145, pp. 421-457.

No context found.

R. B. D'Agostino and M. A. Stephens. Goodness-Of-Fit Techniques. Marcel Dekker Inc., 1986.

No context found.

R. B. D'Agostino and M. A. Stephens (eds.), Goodness-of-Fit Techniques, 1986.

No context found.

R. D'Agostino and M. Stephens, editors, Goodness-of-Fit Techniques, Marcel Dekker, Inc., 1986.

No context found.

R. B. D'Agostino and M. A. Stephens, editors. Goodness of Fit Techniques. Marcel Dekker, 1986. (pp 68, 69)

No context found.

D' Agostino, R.B., and Stephens, M.A. Goodness-of-fit Techniques, Marcel Dekker, Inc., 1986.

No context found.

R.B. D'Agostino, M.A. Stephens, Goodness-of-Fit Techniques, Marcel Dekker, New York, 1986.

No context found.

D'Ag#U tino, R. and Stephens, M. (1986). Goodness-of-Fit Techniques. Marcel Dekker, New York.

No context found.

R. D'Agostino and M. Stephens, editors. Goodness-of-Fit Techniques. Marcel Dekker, Inc., 1986.

No context found.

D'Agostino, R. B. and Stephens, M. A., eds (1986), Goodness-of-Fit Techniques, Marcel Dekker, New York.

No context found.

D'Agostino, R. B. and Stephens, M. A. (1986), Goodness-of-fit Techniques, New York: Marcel Dekker, Inc.

No context found.

D'Agostino, R. B. and Stephens, M. A., eds (1986), Goodness-of-Fit Techniques, Marcel Dekker, New York.

No context found.

D'Agostino, R.B., and M.A. Stephens, Goodness-of-Fit Techniques (Marcel Dekker, New York, 1986).

No context found.

R. B. D'Agostino and M. A. Stephens, Goodness-of-Fit Techniques, Marcel-Dekker, Inc., 1986.

No context found.

D'Agostino, R. B. and Stephens, M. A. (1986) Goodness-of-Fit Techniques, Marcel Dekker, Inc., NY.

First 50 documents  Next 50

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC