Non-Parametric Methods 17.1 INTRODUCTION Most of the hypothesis-testing procedures discussed so far, such as Z+test or Fest, are based on the assumption that the random samples are selected from a normal population, We have seen that most of these tests are still reasonably reliable for slight departures from normality, particularly, when the sample size is large. Traditionally, these testing procedures have been known as Parametric methods as they depend on the parameters, viz., mean or proportion or standard deviation etc. These parametric tests have used the parametric statistic of sample that comes from the population being tested. The need for technique that apply more broadly has led to the development of non- parametric methods. These do not require the assumption that the underlying population normal — or indced that they have any single mathematical form and some even apply to non-numerical data, The non-parametric methods are also called distribution-free methods that assume no knowledge, what so ever, about the distributions of underlying populations, except perhaps that they are continuous. 17.2 NON-PARAMETRIC OR DISTRIBUTION FREE METHODS Parametric Methods Before discussing non-parametric techniques, we should consider why the methods we usually use are called parametric, Parameters are indices. They index (or label) individual distributions within a particular family. For example, there are an infinite number of normal distributions. but each normal distribution is uniquely determined by its mean (1) and standard deviation (6) If you specify all of the purameters (here, \1 and 6), you've specified a unique normal distribution. Most common Wee ice ic ee ly used statistical techniques are properly called parametric because they ole estunating or testing values ic 4 eee sting the value(s) of parameter(s)—usually, population means OF proportions tishunll conc us uo suprise then, that won parametric methods are procedures work their magic without reference to specific purameters. 4Non-Parametric Methods 425 Parametric statistical methods are based on the stringent assumptions about the population from which the sample has been drawn, Particularly, the assumption like the form of probability distribution (whether it is normal or not), accuracy of observations, etc. are more common, Also the parametric methods are applicable primarily to the data which are measured in interval or ratio scale. Non-Parametric Methods If the assumptions (mentioned in parametric methods) do not hold good or the data donot meet the requirement of parametric statistical methods, non-parametric methods come to our rescue. Non-parametric methods entail very mild assumption of continuity and symmetry of the distribution. Non-parametric (or Distribution-free) statistical methods are mathematical procedures for statistical testing which unlike parametric statistics, make no assumptions about the frequency distributions of the variables being assessed Non-parametric tests are often used in place of their parametric counterparts when certain assumptions about the underlying population are questionable. For example, when comparing two independent samples, the Mann-Whitney (a non-parametric test) does not assume that the difference between the two samples is normally distributed, whereas its parametric counterpart, the two sample t-test does assume the condition of their being normally distributed Many non-parametric methods are applicable for ordered statistics. All tests involving ranked data (i.e., data that can be put in order) or involving ordinal data are non-parametric. Ordinal Data: A set of data is said to the ordinal if the values or observations belonging to it can be ranked (put in order) or have a rating scale attached. You can count and order, but not measure in ordinal data). Non-parametric tests may be, and often are, more powerful in detecting population differences when certain assumptions are not satisfied. Non-parametric models differ from parametric models in that the model structure is not specified a priori, but is instead determined from the data. The term non-parametric is not meant to imply that such models completely lack parameters, rather, the number and the nature of parameters is flexible and not fixed in advance. Non-parametric models are, therefore, also called Distribution free models. A histogram is a simple non-parametric estimate of a probability distribution. ‘The precise definition of non-parametric varies slightly among authors. You'll see the terms non-parametric and distribution-free. They have slightly different meanings, but are often used interchangeably. Non-parametric tests are often referred to as distribution-free tests. These tests have the obvious advantage of not requiring the assumption of normality or the assumption of homogeneity of variance. They compare medians rather than means and as a result. if the data have one ot two outliers, their influence is negated. Many non-parametric procedures are based on ranked data, Data are ranked by ordciing them from lowest to highest and assigning them, in order, the integer valucs from | to the sample size. Ties are resolved by assigning tied values the mean of the ranks they would haveze) Biostatistics received if there were no ties, e.g., 117, 119, 119, 125, 128, becomes 1, 2.5, 2.5, 4, 5. (IF the two 119s were not tied, they would have been assigned the ranks 2 and 3. The mean of 2 and 3 is. 2.5;). For large samples many non-parametric techniques can be viewed as the usual normal- theory-based procedures applied to ranks. The following table contains the names of some normal- theory-based procedures and their non-parametric counterparts. TABLE 17.1 : Some Commonly Used Statistical Tests Normal theory based test Corresponding Nonparametric test Purpose of test ‘test for independent samples | Mann-Whitney U test; Compares two independent Wilcoxon rank-sum test samples Paired ¢ test Wilcoxon matched pairs Examines a set of differences signed-rank test One way analysis of variance (Fest). Two way analysis of variance Kruskal-Wallis analysis of | Compares three or more groups variance by ranks Friedman Two way analysis | Compares groups classified by of variance two different factors Pearson correlation coefficient | Spearman rank correlation Assesses the linear association coefficient between two variables. For smaller sample sizes, the same statistic (or one mathematically equivalent to it) is used, but decisions regarding its significance are made by comparing the observed value to special tables of critical vales. 17.3 TYPES OF NON-PARAMETRIC TESTS A large number of non parametric tests exist, but in this chapter we shall discuss the following tests which are better known and widely used ones: J 1, Sign test for paired data and One Sample Sign Test. In these tests, positive or negative signs are substituted for quantitative values. y 2. Mann-Whitney U Test or a Rank Sum Test. This test is used to determine whether two independent samples have been drawn from the same population. / 3. Kruskal-Wallis Test. It is another rank sum test which generalizes the analysis of varian to enable us to dispense with the assumption that the population are normally distributed 4. One Sample Run Test. It is a method to determine the randomness with which the sample items have been selected. 5. Rank Correlation Test. It is a method for finding correlation coefficient when the data available is not in numerical form but the information is sufficient to rank the data first, second, third, and so on, 6. Kolmogorov-Smirnoy Test. It is another method of determining the goodness of fit between an observed and theoretical probability distributions.Non-Parametric Methods oa 7. Kendal Test of Concordance. This test is applicable to situations where we want to test the significance of more than two sets of rankings of individuals 8. Median Test for Two Independent Samples, / 9. Wilcoxon Signed-Rank Test. 17.4 ADVANTAGES OF NON-PARAMETRIC METHODS 1, Non-parametric tests are generally simple to understand, quicker and easier to apply when the sample sizes are small. 2. They do not require lengthy and laborious calculations and hence they are less time consuming. If significant results are obtained no further work is necessary. They do not require the assumption that a population is distributed in the shape of a normal curve or another specified shape 3. Non-parametric tests can be applied to all types of data — qualitative (nominal scaling) data in rank form (ordinary scaling) as well as data that have been measured more precisely (internal or ratio scaling). 5. The chief advantage of non-parametric tests is their inherently greater range of applicability because of milder assumptions. Non-parametric tests make fewer and less. stringent assumptions (that are more easily met) than the parametric tests. 7. Non-parametric test make less stringent demands of the data. 8. Non-parametric methods provide an air of objectivity when there is no reliable (universally recoganised) underlying scale for the original data and there is some concern that the results of standard parametric techniques would be criticized for their dependence on an artificial metric. 17.5 DISADVANTAGES OF NON-PARAMETRIC METHODS Distribution-free or non-parametric methods also have a number of important disadvantages which in practice often outweigh their advantages. 1. Non-parametric tests sometimes pay for freedom from assumptions by a marked lack of sensitivity. 2. Important effects can be entirely missed by a non-parametric technique because of its lack of power or because it produces confidence intervals that are too wide. To put it another way, some of these techniques are very blunt instruments. It may cause it unwise to substitute one of these techniques either because of computational ease or because the data do not quite satisfy some distributional assumption. ¥ The competing distribution-tied procedure may be quiet robust to departures from that assumption 4. An additional related disadvantage is that distribution-free techniques often disregard the actual scale of measurement and substitute either ranks or even a more gross comparison of relative magnitudes. This is not invariably bad, but is often the basis for the loss of power or sensitivity referred to in the preceding paragraph.428 Biostatistics 5. The another disadvantage is one that is gradually being corrected by research in st methodology. At present, we have available a much larger body of techniques in distribution. free hypothesis testing than in distribution-free estimations. To some extent, this is an inevitable accompanying feature of techniques that disregard or weaken the underlying scale of measurement, because estimation is often tied to that scale, whereas hypothesis testing is concerned with selecting one of the two competing decisions or actions. stical 6. Non-parametric tests are often wasteful of information and thus less efficient than the standard techniques which they replace. 7. The major disadvantage is that non-parametric procedures throw away information. Because information is discarded, non-parametric tests can never be as powerful as their parametric counter parts, when parametric tests can be used. For example sign test uses only the sign of the observations and discard the data as such 17.6 USES OF NON-PARAMETRIC METHODS There are four situations in which the use of a distribution-free or non-parametric technique is indicated: 1, When quick or preliminary data analysis is needed. May be you just want to see roughly how things are going with the data. May be somebody is about to leave for Atlantic City to present a paper tomorrow, and his or her chief thought it would be a nice idea to include some statistics in the paper, and you are feeling unusually tender hearted or browbeaten today. May be your computer is suffering an acute exacerbation of some rare electronic disease. May be your computer programmer is manifesting that chronic, epidemic, occupational disability: temporal disorientation with gross distortion of task-time relationships. May be you are doing research in a “primitive” setting without access to modern computational aids. 2. When the assumptions of a competing distribution-tied or parametric procedure are not satisfied and the consequences of this are either unknown or known to be serious. Notice that violation of assumptions per se is not a sufficient reason to reject one of the classical procedures. As we have emphasised throughout this book, these procedures are often insensitive to such violations. For example, the ¢-tests and intervals for normal means are relatively insensitive to non-normality. We think that one of the most common error made today is the under application of classical statistical methods because of the overzealous adherence to the letter of the law with respect to assumptions. Remember that assumptions will rarely be exactly satisfied. It is, therefore, as important to have a grasp of the consequences of violations of assumptions. A great deal of theoretical research in statistics today is directed toward investigating the robustness of statistical procedures. 3. ‘When data are only roughly scaled; for example, when only comparative rather than absolute magnitudes are available, In dealing with clinical data, perhaps patients can only be classified as better, unchanged, or worse. Perhaps only ranks, that is, largest, second largest, smallest, are available. 4. When the basic question of interest is distribution-free or non-parametric in’ nature. For example, do we have a random samples, or are these two samples drawn from populations with identical distributions?ai Non-Parametric Methods 47.7 THE SIGN TEST FOR PAIRED DATA The sign test is the easiest non-parametric test. Its name comes from the fact that it is base onthe direction (or sign for plus or minus) of a pair of obse ations and not on their numerical 429 magnitudes. In such problems, each pair of sample value is replaced b; between the paired observation is positive (that is, ifthe first value ex and by 2 minus sign if the difference between the paired observat it the first value is less than the second value) and it is disearded if the difference is zero. y a plus sign if the difference cds the second value) For applying sign test to solve a problem, we count: Number of +ve signs (+ signs) Number of —ve signs (- signs) Number of 0's (ie., which cannot be included either as positive or negative) We take the two hypothesis : Null Hypothesis: H, : p = 0.5 Alternative Hypothesis: H, : p #0. 5, It is a case of two tailed test. If you look carefully at the above two hypothesis, you will see that the situation is similar to a fair coin toss. If we toss a fait coin 30 times, then probability ‘p’ would be 0.5 and we would expect about fifteen heads and fifteen tails, In that case, we would use the binomial distribution as the appropriate sampling distribution. We also know that when mp and ng are each atleast 5, we ean use the normal distribution to approximate the binomial. Thus, we can apply normal distribution test. We calculate the standard error of the proportion p as given by: Standard error of proportion p : S.E. (p): 6, = f= . n The two limits of acceptance region at 0.05 level of significance are: Py, * 1966, and p,,~1.96 6, It is important to note that if the conditions that both mp and nq are not greater than 5, then we must use the binomial instead of normal distribution test. Example 1 : Use the sign test to see if there is a difference between the number of days required tocollect an account receivable before and after a new collection policy. Use the 0.05 significance level. - Before: 33 36 41 32 39 47 34 29 32 34 40 42 33 36° 27 Afier: 35 29 38 34 37 47 36 32 30 34 41 38 37 35 28 Solution : Table for Before and After Collection Policy Before (1) : 33 36 41 32 39 47 34 29 32 34 40 42 33 36 27 After (2) : 35 29 38 34 37 47 36 32 30 34 41 38 37 35 28 het 2 0 eS =e ()-@1—- 430 Biostatistics Ifwe count the bottom row of above table, we get Number of +? signs = 6 Number of signs = 7 Number of 0's = 2 Total sample size = 15. When we use the sign test, we exclude the evaluation 0°s as we are testing perceivable differences. We notice that we have 6 plus signs and 7 minus signs, for a total of 13 usable responses. If there is no difference between the two types of collections, p (the probability that the first collection exceeds the second collection) would be 0.5 and we would expect to get equal number of plus and minus signs. We set up the hypothesis as : 1, Null Hypothesi Hy: p=0.5 => — there is no difference between the two types of collections Alternative Hypothesis: H,:p #0.5 => there is a significant difference between the two types of collections => It is a case of two tailed test. Sample size : n = 13. Proportion : p = 2 = 0.46. . a ~ 046 = 0.54. Also q= 75 = 0.54. 7 = 6, and nq = 13 x 7 = 7 are both greater than 5, so we use the normal distribution to approximate the binomial. 2. Level of significance: « = 0.05 3. Z, for two tailed test = 1.96 (from the tables) 4, Standard error of the proportion p : S.E. (p) = 0, = \ 13 = 0.2484 = 0.498, > Vos 5. Limits of Acceptance Region. Upper Limit: = ((3)+ (1.96) x (0.498) = 1.476 Lower Limit: = 0.5 — (1.96) x (0.498) = — 0.476 - Decision. The two limits of acceptance region are 1.476 and — 0.476 The sample proportion P = 0.46 lies within these two limits, so the sample proportion falls within the acceptance region for this hypothesis. We accept the null hypothesis H, =? there is no difference between the two types of collections at 0.05 level significance.Non-Parametric Methods ant Example 2 : To determine the effectiveness of a new traffic-control system, the number of accidents that occurred at 12 dangerous intersections during four weeks before and four weeks after the installation of the new system was observed, and the following data was obtained. 3and1, Sand2, QandO, — 3and2,— 3.and2, — 3ando Oand2, 4and3, — land3, Gand 4, — dand 1 land 0 Use the paired — simple sign test at 0 = 0.05 level of significance to test the null hypothesis that (\1,), the new traffic-control system is only as effective as (jt,), the earlier system (the population samples are continuous, bu‘ this does not matter so long as zero differences are discarded). Solution : 1, Setting up of Hypothesis: Null Hypothesis: H, : 11, Alternative Hypothesis: H, ; 1, > 1, =9 It isa case of one-tailed test. . Level of significance: Here a = 0.05. . Test Statistic: Use the test statistic X= the observed number of plus signs. Replacing each positive difference of the two values by a plus sign and each negative difference by a minus sign, we get: wr Pr a a a ee ne ee eS Here, the number of paired values m = 12 and the number of plus signs X = 10 It is a case of Binomial distribution as the number of negative signs is only 2, which is less than 5. 1 From the table, we find that for n = 12, @ = 2 P(X > 10) = 0.0161 + 0.0029 + 0.0002 = 0.0192 [Note: For a = 0.05, for n P(X 2 10) = P (X= 10) + P (X= 11) + P (X= 12) .0161 + 0.0029 + 0.0002 = 0.0192] 4. Decision: Since P-value = 0.0192 is less than 0.05, so the null hypothesis must be rejected and we conclude that the new traffic-control system is not effective in reducing the number of accidents at dangerous crossings. w 17.8 ONE SAMPLE SIGN TEST One-sample sign test is a non parametric method which is used as an alternative to the one-sample ¢-test, where the null hypothesis is 1 =H, against a suitable alternative hypothesis. The only assumptions underlying the sign test are: the population sample is continuous and symmetrical. We assume that the population is continuous so that there is zero probability of getting a value equal to pi.) and we do not even need the assumption of symmetry if we change the null hypothesis to pt = Hg» where Ht is the population mean or median.432 Biostatistics In the sign test, we replace each sample value exceeding jt, with a plus sign and cach value less than j1, with a minus sign, Then we test the null hypothesis H, that X, the number of plus sign is a value of a random variable having Binomial distribution with parameters 1 n (the total number of plus or minus signs) and probability 0 = z The two-sided alternative hypothesis H, ; # Ht, thus becomes @ # >, and one-sided alternatives 1 w, I I . becomes 0 < 5 and @> > respectively. Ifa sample value equals ty, (which can happen when we deal with rounded data even though the population is continuous), we simply discard it. In order to perform a sign test, when the sample size is very small, we use the Table of Binomial probabilities and when the sample is large we use the normal approximation to the Binomial distribution with Mean = 18, and variance = nO (1 - 0) S.E. (0) = /n0 (1-0) fae X-n0 The test statistic in this case is, Z= 7————, [nd (1-0) with @ = Fyand X the number of plus signs. Example 3 : The following are the measurements of breaking strength of a certain kind of 2-inch cotton ribbon in pounds: 163 165° 160 «189 161 171 158151 169 162 163° 139-172 165 148 = «166172, «163 187 173 Use the sign test to test the null hypothesis = 160 against the hypothesis \. > 160 at the 0.05 level of significance. Solution : 1, Null Hypothesis: H, : 1 = 160. Alternative Hypothesis: H, : |. > 160 => It is a case of one tailed test. 2. Level of significance: 0. = 0.05. 3. Test statistic: X= The observed number of plus signs. Replacing cach value exceeding 160 with a plus sign, each value less than 160 with a minus sign, and discarding the value which equals 160, we get : Pe et ee Here m = the total number of plus and minus signs = 19. x = the total number of plus signs = 15.Non-Parametric Methods 433 It is a case of Binomial Distribution. From the Binomial Table given at the end, we find P(X2 15) = 0.0095 for @ = and n= 19, [For 8 = 0.05, P(X > 15) = P(X = 15) + P(X = 16) + P(X = 17) + P(X = 18) + P(X = 19) = 0.0074 + 0.0018 + 0.0003 + 0.0000 + 0.0000 = 0.0095] 4. Decision: Since P-value (= 0.0095) is less than 0.05, so the null hypothesis must be rejected and we conclude that the mean breaking strength of given kind of ribbon exceeds 160 pounds. Example 4 : The following data,( in tons) are the amounts of sulfur oxides emitted by a large industrial plant in 40 days 24 15 200 29191822 2S 9 7 2 17 6 m4 MW 15S 23 Mm 6 M9 238 16 19 10 23 18 31013 0 Use the sign test to test the null hypothesis = 21.5 against the alternative hypothesis [> 21.5 at the 0.01 level of significance. Solution : 1. Setting up the Hypothes Null Hypothesis: H, : = 21.5. Alternative Hypothesis: H,: 1 <21.5 => Itis a case of one-tailed test. nv . Test Statistic: Replacing each value exceeding 21.5 with a plus sign, each value le minus sign, and discarding the one value which equals 21.5, we get re ee ee s than 21.5 with a ee ee X= the number of plus signs = 16. n= total number of plus and minus signs = 16 + 24 = 40. As the sample size, 1 = 40 is very large, so we shall use the Normal approximation to Binomial distribution. X-n0 Test Statistic: 2-0-0) where @ = ; = 0.5, 1 = sample size. 10. (0.5) 4 _ 16 2° fro: | Z| = 1.26, [on 40, 0 © 0.5} )(0.5) 3.16434 Biostatistics 3. Level of significance: Here o = 0.01 4. Ctitical values: The critical value | Z, | for o = 0.01 for one-tailed test is 2.33. Decision: Since | Z| <| Z, | as 1.26 < 2.33, so we accept the null hypothesis. EXERCISE - 17.1 . Use the sign test to see if there is a difference between the number of days until collection of an account receivable before and after a collection policy. Use 0.05 significance level. Before : 30 28 34 35 40 42 33 38 34 45 28 27 25 41 36 Afier 2 34) 29" 33 328 478 49 240 242) 937 44027) 33) 302 382 36) [Ans, There is no significant difference before and after new collection policy in the accounts receivable.] Rv The following data show employees’ rates of defective work before and after a change in the wage incentive plan. Compare the two sets of data given below to see if the change lowered the defective units produced. Use the 0.10 level of significance. Before 8 7 6 9 7 10 8 6 5 8°10 8 Wee oo 6 Ob 6 GD 3. Use the sign test on the data given below to determine whether there is a statistical increase in the values produced by treatment B over those produced by treatment 4: Subject > 1 2 3 4 5 6 7 8 9 10 Treatment Alte tte 46ee 4\ge037)009200028)0043)00 42, 51128). 27) TreatmentB : 52 43 37 32 31 39 44 53 26 31. Use 0.05 level of significance. 4. When is the sign test used? The scores under two conditions X and Y obtained by the respondents are given below: xX : 122 6 8 6 4 8 You. 7 2 Wh S$) 2 Apply the sign test and comment on your findings at 0.05 level of significance. . Acompany claims that if its product is added to an automobile’s gasoline tank, the distance travelled in kilometres per litre will improve. To test the claim, 15 different automobiles are chosen and the distances with and without the additive are measured; the results are shown below. Assuming that the driving conditions are the same, determine whether there is a difference due to the additive at significance levels of (a) 0.05 and (5) 0.01 With additive + 17.3141 9.8 12.5 7.8 12.2 14.3 11.7 13.8 16.0 24.8 11.2 12.8 14.0 12.1 Without additive :15.713.6 10.2 12.3 7.4 11.1 13.4 12.0 13.1 15.7 14.4 11.5 12.0 13.6 114 [Ans. (a) There is a difference at the 0.05 level of significance (b) There is no difference at the 0.01 level of significance.] wnNon-Parametric Methods 435 6. The following: data represents the number of hours that a rechargeable hedge trimmer operates before a recharge is required. 1S 2.2 0.9 13 2.0 1.6 18 15 2.6 12 hs Use the sign test to test the hypothesis at 0.05 level of significance that this particular trimmer operates, on the average, 1.8 hours before requiring a recharge. (Hint: H, : p= 1.8, H, : 1 # 1.8, we have - +--+--+--, so, » = 10 and Y= 3. Apply Binomial Table] [Ans. Accept the null hypothesis H, : t= 1.8] : On 12 visits to a doctor’s clinic, a patient had to wait in minutes as under: 17; 32; 25; 15; 28; 25; 20; 12; 35; 20; 26; 24. before being seen by the doctor. Use the sign test with ot = 0.05 to test the doctor’s claim that, on the average, his patient dé not wait more thn 20 minutes before being examined by him. {Ans. H, : 1 = 20 is accepted] 17.9 RANK SUM TESTS ~ Rank sum tests are a whole family of tests. We shall concentrate only on the following two members of this family: /1. Mann-Whitney U Test _P> Kruskal-Wallis H Test. Mann-Whitney tests is used when there are only two populations whereas Kruskal- Wallis test is employed when more than two populations are involved. The use of these tests will enable us to determine whether independent samples have been drawn from the same population or different populations have the same distribution. 17.10 MANN-WHITNEY U-TEST It is a non-parametric method used to determine whether two independent samples have been drawn from populations with same distribution. This test is also known as U- Test. This test enables us to test the null hypothesis that both population medians are equal (or that the two samples are drawn from a single population). It requires the two samples to the independent samples of observations measured at least at an ordinal level, i.e., we can at least say, of any two observations, which is greater. This method does not require the assumption that the difference between the two samples are normally distributed. This method helps us to determine whether the two samples have come from identical populations. If it is true that the samples have come from the same population, it is reasonable to assume that the medians of ranks assigned to the values of two sampls are more or less the same, The alternative hypothesis H7, would be: That the medians of the populations are not equal. In this case, most of the smaller ranks will go to the values of one sample, while most of the higher ranks will go to the other sample. The test involves the calculation of a statistic usually ealled U, whose distribution under the null hypothesis is known. In case of small samples ize i ¢ distribution is tabulated but for samples above 8, i ze is less than 8), the distribution is p eal the normal distribution. there is good approximation using