Sx Solution See82 Comparsons—TWwe Independent Lge Samples 267 Pairing to eliminate a known source of variation A civil engineer needs to compare the durability of two different paints for marking lanes on a divided highway. One year after applying the paint, she will return and assign a number to the current quality of the marking. Ten widely separated I-mile sections of divided highway are available. From experience, she expects that the sections would carry substantially different traffic volumes. Describe an experiment for making the comparison, The two paints are the treatments which we call Treatment 1 and Treatment 2. The | -mile sections, in either direction, will carry approximately the same traffic volume for the whole mile, A I-mile section, with wailic going in one direction, is a unit The response is quality of marking after one year. The civil engineer expects trafic volume to heavily influence the response. The traffic volume could be eliminated from the comparison if we pair the two 1-mile sections with traffic going in opposite directions. Then a comparison will only be made within the pair. This is called a ‘matched pairs design. Still the road for one side may be subject to more shade, higher temperature, or other conditions different from the other. For each L-mile section, the engincer should flip a coin. If heads, the north or west direction receives Treatment 1. Ran- domization helps prevent these other uncontrolled variables from influencing the response in a systematic manner. ‘Note that by pairing like experimental units, we have eliminated traffic volume as an influencing variable. . ‘The term experimental design refers to the manner in which units are chosen and assigned to receive treatments. As introduced in the last wo examples, there are ‘ovo basic designs for comparing two treatments: 1. Independent samples (complete randomization) 2. Matched pairs sample (randomization within each matched pair) ‘We investigate the independent samples design in the next two sections and the ‘matched pairs sample design in the following section. 8.2 Comparisons—Two Independent Large Samples In this section, we consider the independent samples design when both sample sizes are large. To state the assumptions, we use X and Y for the observations and the subscript 1 or 2 for the mean and variance to distinguish the two populations. Assumptions—Large Samples 1X1. ...%qy isa random sample of size ny from population which has scan ~ jy and variance = 07 2. ¥4, Yo, ++ Yay 8 random sample of size ng from population 2 which has rican — jig and variance = oF 3. The two samples X1, Xp, ....Xq, and Yj, ¥a,.... Yn, are independent, Inferences will be made about the difference in means j1 — 113 = 5. Since, by Theorem 6.1 BF) =y, va) m nm, ER) =m, VarlX) 226B Chapter ® Comparing Two Treatments the mean of X —Y is EX 30, Chapter 5), ) = 11 — ne and, by independence (see Example Sy mm When the sample sizes m and ny are large, the central limit theorem implies that both Y and Y are approximately normal. Because they are independent, their dilference is also approximately normal and the two sample Z statistic van -F) = is approximately standard normal. Because the sample sizes m and nz are large— namely, both are greater than or equal to 30—the normal approximation remains valid when the sample variances replace the population variances, ‘When the sample sizes ny and ng are large—namely, ny, nz = 30 Statistic for large samples X_P_s inference concerning —— difference between two is approximately standard normal. Large Samples Confidence Intervals Large samples confidence intervals, for the difference of means 8 = j4y — 4; determined from the standard normal probability & m This last statement asserts tha, before we obtain the data, the probability is 1 that the random interval will cover the true unknown difference in the means 8 waa He Confidence limits fortarge | Limits of 100(1 — d) confidence interval for jy — Ha samples confidence fee interval for si — 4 Stanyote mthxu Solution See82 Comparsons—TWwe Independent Lge Samples 269) This confidence interval can also be obtained from the acceptance regions for the ‘owo-sided test on page 270, Large samples confidence interval for difference of means As a baseline for a study on the effects of changing electrical pricing for electricity during peak hours, July usage during peak hours was obtained for ny = 45 homes with air-conditioning and ny = 55 homes without.' The July on-peak usage (kWh) is summarized as Sample Population Size Mean Variance With 45 204.4 13,8253, Without 55 130.0 8,632.0 ‘Obtain a 95% confidence interval for § = [Ly — jp. For a 95% confidence interval, a = 0,05 and zo 9a5 = 1.96. We are given ny = 45, ny = 55, ¥ = 208.4, 2 = 13,825.3,F = 1300, and 53 = 8,632.0. Then the mits Of the confidence interval are + = = 204.4 — 130.0 + 1.96 3,825.3 8,632.0 ma Vas 73s = 144 + 1,96 Va6ETT so the 95% confidence interval is (32.17, 116.63 ). The mean on-peak usage for homes with air-conditioning is higher than for homes without, from 32.17 to 116.63 kWh per month, ‘The confidence interval not only reveals that the two population means are statis- tically different, because the confidence interval does not cover 0, but also quantifies the amount of difference. . Large Samples Tests for Differences of Means There are many statistical problems in which we are faced with decisions about the relative size of the means of two populations. For example, if two methods of welding ate being considered, we may take samples and decide which is better by comparing their mean strengths Formulating the problem more generally, we shall consider two populations having the means jz and jz and the variances of of and a3, We want to test the null hypothesis Hoi ay ~ 2 = 40 where fis a specified constant, on the basis of independent random samples of size zn and na. Analogous to the tests conceming one mean, we shall consider tests of this null hypothesis against each of the alternatives jy — #2 <6, shy — Ha > 40 and jy — 143 # 49, The test itself will depend on the distance, measured in esti mated standard deviation units, from the difference in sample means X — ¥ to the hypothesized value, 5p. When the sample sizes are large—namely, my and ny are Richard Johnson and Dean Wicher (2007), Applied Multivariate Static Analyt, 6h 8, page 29, Prniice Hall Upper Sale River.270 Chapter® Comparing Two Tresenens both greater than or equal to 30—we obtain this test statistic by specifying the null value y for the difference of means in the random Z defined above. When ny.ny > 30, test Hy: ay — 2 = ap using Test statistic for large 7 (EaP)=% samples concerning difference between two means ny ny which has, approximately, a standard normal distribution Analogous to the table of level o rejection region on page 252, the rejection regions for testing the null hypothesis 4 — j1z = 4p, using the two sample Z test Alternative hypothesis Reject null hypothesis if: MH <0 Zant Hy Ha > bo Z>t WH #50 Zs nto or Z> za) Although Sp can be any constant, it is worth noting that in the great majority of problems its value is zero and we test the null hypothesis of no difference, namely, the null hypothesis j2y = jy. A test for a the mean difference in driving performance With the goal of improving driving safety, engineers are quantifying the effects of such factors as drowsiness and alcohol on driver performance.* Volunteers drive a fixed course in a mid-sized car mounted in a sophisticated driving simulator. One performance measure isa standard deviation like score of the lateral deviation from center line ‘We consider the experiment where the first treatment specifies that the driver hhas a blood alcohol reading of 0 percent and the second treatment specifies that the driver imbibe and then be carefully monitored to reach a blood alcohol reading of 0.10 percent, The summary statistics for one segment of the drive are (courtesy of, Jobn Leo) ‘Treatment 1 ‘Treatment 2 OBlood Alcohol 1 % Blood Alcohol my = 54 ny = 54 F= 16 ys. s) = 0.177 sy = 0.183 as. $. Zhou, and John Lee, iflerentitngsleohokinduced diving behavior using testing whes! signals, IEEE Trans Tell Travap. Sys. 13,2012), 1355-1368.Solution Figure 8.1 ‘P-value for testing equality of sean lane deviation scores. Gz Solution See82 Comparsons—Twe Independent Lage Samples 271 ‘Conduct a test of hypotheses with the intent of establishing that the mean lateral deviation scores arc different. Take & = 0.02. ‘The test concerns 5 = jy — yz and the sample sizes m = 54 and ny = 54 are large. A. Null hypothesis: 8 = 0 Alternative hypothesis: | # 0 2. Level of significance: « = 0.02 3. Criterion: Reject the null hypothesis if Z < —2.33 or Z > 2.33 where Zis ‘given by the large sample formula above. 4, Calculations: The observed value of the test statistic is LO 177=0 ggg [0.177 , @.183)? 4 4 5. Decision: Since z= —4.04 is less than 2.33, the null hypothesis must be rejected at level of significance 0.02. The small P-value 0.000053 (see Figure 8.1) provides very strong evidence that the mean lateral deviation scores for drinkers is different from that of non-drinkers. 0000266 o.0000266 wsoe 235 -1 00 1 233 dou . Testing a difference in means with two large samples To test the claim that the resistance of electric wire can be reduced by more than 0.050 ohm by alloying, 32 values obtained for standard wire yielded % = 0.136 ohm. and s; = 0.004 ohm, and 32 values obtained for alloyed wire yielded ¥ = 0.083 ohm and s3 = 0.005 obm, At the 0.05 level of significance, does this support the claim? 1. Nuit hypothesis: 4y — Hq = 0.050 Alternative hypothesis: 1 ~ juz > 0.050 2. Level of significance: a = 0.05 3. Criterion: Reject the null hypothesis if Z > 1.645, where Z is given by the large samples formula above, Calculations: 0.136 ~ 0.083 0 9.050 _ 6s (0.004)? (0.005)? ED 5. Decision: Since z = 2.65 exceeds 1.645, the null hypothesis must be rejected: that is, the data substantiate the claim, From Table 3, the P-value is 0,004 (see Figure 8.2), so the evidence for alloying is very strong. Only 4 in 1,000 times would Z be at least 2.65, if the mean difference was 0.05. .272 Chapter® Comparing Two Treatments Figure 8.2 Large samples P-value = 0,004 for Z = 2.65 Solution 008 a. 0 Lois 2.65 To judge the strength of support for the null hypothesis when it is not rejected, ‘we consider Type I errors, for which the probabilities depend on the actual alterna- live differences 5’ = je; — 12. Fortunately, these can be determined using the single sample results (as long as we are sampling from normal populations with known standard deviations or both samples are large). ‘The calculation of Type II error is based on the results for a single sample on. page 260 and a fictitious single variance ? and sample size n, where They Also, replace 4g — 1 by 59 — 8 ‘The preferred method for calculation of Type If error is to use a computer soft- ware (see Exercise 7.72). Approximating the probability of Type Il error With reference to the preceding example, what is the probability of a Type Il error for 6" = 0.054 ohms? 0.0042 + 0.005? Refer to the formula on page 261. Since 8y = 0.050, 0? 0.000041, and n = 32 — (by — 5" 0.050 — 0.054 tat Vi =? 1645 4 vB 2 = 1,645—3.534 = -1.889 so the powers (0.054) = P(Z > —1.889) = 0.971. The Type Il exor probability 6 =1-0.971 = 0.029, . 8.3 Comparisons—Two Independent Small Samples ‘When ny and ng are both small and the population variances are unknown, we must impose additional assumptions, which we label 4 or 5 to emphasize that the three original assumptions still prevail. Additional Assumptions for Small Samples 4, Both populations are normal. 5. The two standard deviations have a common value 0) = 02 = 6 Because the populations are normal, X¥ and Y are normal and, because they are independent, their difference is also normal (see Example 41, Chapter 5). Recall. E(X—Y) = py — uz = 8. Under the assumption of common standardSee83 Comparsons—Two Independent Small Samples 273) deviation , the expression above for the variance of X — ¥ becomes 2 42 ey, 8 14 Var -F)= 442 2(c 42) mm mS in and the standardized version of ¥ — ovm has a normal distribution. The unknown ¢ must be estimated. Reasoning that each squared deviation (X, — X)? is an estimate of o? and so is each (Y, — ¥ )* from the second sample, ‘we estimate 0 by pooling the sums of squared deviations from the respective sam- ple means, That is, we estimate 0” by the pooled estimator Y (x) FP a x y _ Oy = SF + Om = 153 ” my bmg —2 m mg —2 where 3>(%; =F)? is the sum of the squared deviations from the mean forthe fist sample, while (1; — ¥ )? is the sum of the squared deviations from the mean for the second sample, We divide by nj + ng — 2, since there are n, — 1 independent eviations from the mean in the frst sample, n — 1 in the second, and we have zn +g ~2 independent deviations from their mean to estimate the popdlation More specifically, from the single sample results we know that both S? and SE are estimates of o? and that (m - DS} (m -1)S3 hhas a chi square distribution with n; — 1 degrees of freedom hhas a chi square distribution with ny — 1 degrees of freedom and these two random quantities are independent since the samples on which they are based are independent. By either the result on page 211 or Example 14, Chap- ter 6, the sum of the two chi square variables has a chi square distribution with degrees of freedom equal to the sum of the two degrees of freedom my + nz — 2. Farther, (m - DS} | (m -S$ (aa DSt (m=O _ (mi tm — 23S im =$f , 2-83 mtm—2 and we conclude that suave variable “a | degrees of feedom274 —Chapter® Comparing Two Tresenants and this can be shown to be independent ofthe standard normal based on X, ~ Xp. ‘Using the representation of fon page 211 as a standard normal over the square root of a chi square divided by its degrees of freedom, we obtain ¥-¥- (m= DS}+ (m= DS} my tng 2 Statistic for small sample test concerning dference Sp between two means + mm has a distribution with m +n — 2 degrees of freedom (4). [Note that by substituting Sp for o in the expression for Z on page 273, we ar- rive at the same statistic, With small sample sizes, the distribution is not standard normal but at ‘The criteria for the two sample ¢ test based on this statistic are like those for Z for testing the null hypothesis: Hp: y — 2 = bo, Level « Rejection Regions for Testing Cee keer ere ‘Alternative hypoth Reject null hypothesis it: M13 <8 tent =H > bo oh ni #50 tetas otf baa In the application of this test, my and ny may be small, yet ny ++ mg — 2 may be 30 ‘or more; in that case we use the normal critical value (also bottom line of Table 4.) A two sample # test to show a difference in strength ‘To reduce the amount of recycled construction materials entering landfills it is crushed for use inthe base of roadways. Green engineeting practices rege that thai stength, revlency modulus (MPa), be acessed. Measurements on m ng = 6 specimens of eecjeled materials om to diferent location produce the data (Courtesy of Tuncer Edil) Location 1: 107632604. 652.669 6TH. Location 2: 552 554484 630—GHB GLO Use the 0.05 level of significance to establish a difference in mean strength for ‘materials from the two locations. ‘The test concerns § = 4 — 42 and the sample sizes m) = nz = 6 are small ‘There are no obvious departure from normality 1. Null hypothesis: 8 = 0 Alternative hypothesis: 8 # 0 2. Level of significance: « = 0.05 3. Criterion: Reject the null hypothesis if ¢ < —to,o95 oF! > toons where toons = 2.228 for 6 + 6 — 2 = 10 degrees of freedom.Figure 8.3 P-value for Example 7. ‘Small sample confidence interval concerning difference between ‘two means See83 Comparions—Two Independent Small Samples 275 4, Calculations: The means and variances of the two samples are 656.3 3197 2 2 1 GPP Liane f= BO? 5.1399 so that sh = (6, 389.3 + 18, 699.3) /(5 + 5) = 2, 508.5, sp = $0.09 and the observed value of the test statistic is B_ _ 656.3 - 579.7 — 2.65 9 Lyd wleed 50.09% + 5. Decision: Since ¢ = 2.65 is greater than 2.228, the null hypothesis must be rejected at the 0.05 level of significance. The P-value 0.0243 (see Figure 8.3) provides stronger evidence that the mean strength of recycled materials is different at the two locations. oon oor -ms 10s [Using R: data x and y ttest(x,y,vanequal=T). ] . In the preceding example we went ahead and performed the two sample test, tacitly assuming that the population variances are equal. Fortunately, the test is not overly sensitive to small differences between the population variances, and the pro- cedure used in this instance is justifiable. As a rule of thumb, if one variance is four times the other, we should be concerned. A transformation will often improve the sit- uation. As another alternative there is the Smith-Satterthwaite test discussed below. Confidence intervals follow directly from the acceptance region for the tests. For two normal populations with equal variances, ‘The (1 — @)100% confidence interval for 8 = uy — jez has limits on [=D +0 - D5 : ae y ny tng —2 Vr me where fy /2 is based on v = my + nz 2 degrees of freedom. Graphics to accompany a two sample / test Example 7, Chapter 2, presents strength measurements on an aluminum alloy. A second alloy yielded measurements given in the following stem-and-leaf display. Find a 95% confidence interval for the difference in mean strength 5276 Chapter 8 Comparing Two Treatments ‘Solution We first place the observations on the two alloys in stem-and-leaf displays, Note that the observations from the first alloy appear normal, but those on the second alloy may deviate from a normal distribution, Since the sample sizes are relatively large, this will not cause any difficulty Alloy 1, N = $8 Alloy 2, = 27 Leaf unit = 0.10 Leaf unit = 0.10 66 | 4 66 67|7 67 o)o03 4689 68 olo12335567889 6 wloo1233456689 njo12335667889 71/28 mnl1234679 26 Bl1 35 nla79 m4)2 5 14/4 9 75|3 35|5 9 16 16/3579 1 ml13678 8 w|1 246 19 wlio 38 A computer calculation gives the sample means and standard deviations, N MEAN STDEV ALLOYT = 58 ~—70.70 1.80 ALLOY2 2776.13, 2.42 From another computer calculation (or by interpolation in Table 4), we find fg 025 = 1,99 for 83 degrees of freedom, so the 95% confidence limits are [m=Da+ 0m ny + —2 taj 77(1.80)7 + 26(2.42)? sO = 70.10 — 76.13 + 1.99 and 6.4 < py — pp < 45 We are 95% confident that the mean strength of alloy 2 is 4.5 to 6.4 thousand pounds per square inch higher than the mean strength of alloy 1 . It is good practice to show stem-and-leaf displays, boxplots, or histograms, Often they reveal more than a mean difference, For instance, in the last example, the first population is nearly symmetric but the second has a Tong tail to the left ‘The large sample confidence interval is obtained from the acceptance regions for the test on page 270.Solution Statistic for small samples inference, a, # o2,normal populations EXAMPLE 10) Ste83. Comparions—Two Independent Srall Samples 277 ‘Comparing the two confidence intervals Referring to the previous example, find the 95% large sample confidence interval. .80)7 | (2.42)? se) 10.10 — 76.13 + 1.96 6.5 < my Hp < 44 ‘There is not much difference between this 95% confidence interval and the one in. the previous example, where the variances were pooled . Small Sample Sizes but Unequal Standard Deviations—Normal Populations When we deal with independent random samples from normal populations whose variances seem to be unequal, we should not pool. As long as the populations are normal, an approximate ¢ distribution is available for making inferences, The statistic 1" is the same as the large samples statistic but, because sample sizes are small, its distribution is approximated as ar distribution For normal populations, when the sample sizes ny and np are not large and a1 Fo, 3 yo Fa) ie 3 m + mn is approximately distributed as a¢ with estimated degrees of freedom, ‘The estimated degrees of freedom for rare calculated from the observed values of 2 and #2 the sample variances sf and 53 4 8) my tay ny Gm (Gim)” | (Bima) Taya tg ‘The estimated degrees of freedom are often rounded down to an integer so a table can be consulted ‘The test based on ris called the Smith-Satterthwaite test. estimated degrees of freedom = Testing equality of mean product volume One process of making green gasoline takes sucrose, which can be derived from biomass, and converts it into gasoline using catalytic reactions. This is not a process for making a gasoline additive but fuel itself. so research is still atthe pilot plant stage. At one step in a pilot plant process, the product consists of carbon chains of length 3. Nine runs were made with each of two catalysts and the product volumes (gal) are Catalyst 1 0.63 2.64 1.85 1.68 1.09 1.67 0.73 Lot 0.68 Catalyst2 3.71 409 4.11 3.75 349 327 3.72 349 4.26278 Chapter® Comparing Two Treatments ‘The sample sizes ny =p = 9 andthe summary tatisis are Fa1gM, = 04548 52376 = 0.1089 chemical engineer wants o show that the mean product volumes are diferent Test witha = 008 Solution The test concems § = jz ~ jy and the sample sizes ny = ny = 9 are small. We note that there are no outliers and no obvious departure from normality. However, 57 = 0.4548 is more than four times s3 = 0.1089. We should not pool 1. Null hypothesis: 3 = 0 Alternative hypothesis: 5 #0 2. Level of significance: a = 0.05 3. Criterion: We choose the Smith-Satterthwaite test statistic with 5p = 0, X-7-5 ! = 3 Sy Se Vat 7 ‘The null hypothesis will be rejected if” < —I9 995 oF 1! > fp ans, but the value of 025 depends on the estimated degrees of freedom. 4. Calculations: As a first step, we estimate the degrees of freedom - o (im) (S/n? (0.4548/9 7 |, (0.1089/9 )* sre t= Se ee To use the # table, we round down to 11 and obtain fo 995 = 2.201 The observed value of the test statistic is ¥ 1,334 ~ 3,766 —0 /oasas_, 0.1089 yoo to 9.7 5. Decision: Since t' = ~9.71 is less than 2.201, the null hypothesis must be ejected at level of significance 0.05. The value of ris so small that the P-value {i 0,0000 when rounded. In other words, there is extremely strong evidence that the mean product volumes are different for the two catalysts [ Using R: data x and y ttest(xy) Use alt=“greater”, for one-sided upper tail test. | . Confidence intervals can tell us what differences in means are plausible, not just that the means are different. A 100 (1 — «)% confidence interval for 8 = uy — 2 Confidence interval for = 443 3 = 11 ~ Ha,normal uayor2 populations 01 #02 mm where fy 2 has the degrees of freedom estimated for 1!GzuaD Solution See83 Comparions—Twe Independent Small Samples 279 A confidence interval for the difference of mean yields when variances are unequal With reference to the previous example, obtain the 95% confidence interval for mH From the previous example we have ¥ = 1.334, s7 = 0.4548, = 3.76 0.1089, and tp o35 = 2.201 for 11 degrees of freedom. We get 6,3 = aSa89 0.1089 (1961-96-20 1.334 — 3.776 + 2.201, (Ca5a89 9 O:1089 o or (—2.982, ~1.880) gallons. The mean product volume for the second catalyst is greater than that of the first catalyst by 1.880 to 2.982 gallons. . ‘Although the computations for unequal standard deviations scem tedious by hhand, popular statistical software will allow this option in addition to pooling (see Exercise 8.35), Exercises 84 a2 83 a4 as Refer to Exercise 2.58, where m = 30 specimens of 2 x 4 lumber have ¥ = 1,908.8 and s, = 327.1 psi. A second sample of size mz = 40 specimens of larger dimension, 2 x 6, lumber yielded ¥ = 2,114.3, ands; = 472.3. Test, with a = 0.05, the null by- porhesis of equality of mean tensile stengths versus ‘he one-sided alternative thatthe mean tensile strength for the second population is greater than that of the first. Refer to Exercise 8.1 and obtain a 959 confidence in- terval forthe difference in mean tensile strength. ‘The dynamic modulus of concrete is obtained for two different concrete mixes. For the frst mix, my =33 (0.47 psi, For the second mix, 0.38, Test, with a =0.05 the null hypothesis of equality of mean dynamic mod- ulus versus the two-sided alternative, Rofer to Exercise 8,3 and obtain 2 959% confidence in- terval forthe difference in mean dynamic modulus, ‘An investigation of wo types of bulldozers showed that 50 failures of one type of bulldozer took on an average 6.8 hours to repair with a standard deviation fof 0.85 hours, while 50 failures of the other type of bulldozer took on an average 7.3 houts to repais with a standard deviation of 1.2 hours (9) Test the null hypothesis jy — p13 = 0 (namely, the hypothesis that on an average, ittakes an equal amount of time to repair either kind of bulldozer) against the alternative hypothesis 11 — 442 % 0 at the level of significance, a = 0.10. 86 87 (b) Using 0.85 and 1.2 as estimates of a and ap, find the probability of failing to reject the null hypoth- esis (uy — 43 = O with the etiterion of part (a) when actually sy — 3. Studying the flow of traffic at two busy interseetions between 4 Pa, and 6 Fat, (to determine the possible iced for tur signals), it was found that on 40 week- days there were on the average 247.3 ears approach ing the first intersection from the south that made left turns while on 30 weekdays there were onthe average 254.1 ears approaching the second intersection from ‘the south that made let turns. The corresponding saza- pile standard deviations are 51 = 15.2 and s2 = 18.7. {a) Test the null hypothesis 4 — 2 = O-against the altemative hypothesis 2; — jp # Oat the level of signticance a = 0.01, (b) Using 15.2 and 18,7 as estimates of oy and find the probability of failing to reject (accepting) the null hypothesis jt; — py = 0 when actually lua = 121 = 156 Given the my = 3 and np = 2 observations from Pop- ulation 1 and Population 2, respectively, Population! | 6 92 7 Population? | 14 10 {a) Calculate the three deviations x —¥ and two devi lions y — () Use your results from pat (a) to obtain the pooled280 Chapter ® Comparing Two Treatments 8.8 "Two methods for manufacturing a product are to be ‘compared. Given 12 units, six are manufactured using method M and six are manufactured using method N. (a) How would you assign manufacturing methods to the 12 units? (5) The response is the percent of finished product that did not meet quality standards. Suppose the Method M Method N 3 ° 8 2 1 1 4 5 6 1 2 9 Find a 99% confidence interval for the difference (©) What assumption did you make f to part ()? ‘8.9. Measuring specimens of nylon yarn taken from two spinning machines, i was found that 8 specimens from the ttst machine had a mean denier of 9.67 with astan- dard deviation of 1.81, while 10 specimens from the second machine had a mean denier of 7.43 with astan- dard deviation of 148. Assuming thatthe populations sampled are normal and have the same variance, test ‘the null hypothesis 2; ~ y42 = 1.5 against the alter- native hypothesis jz — 2 > 1.5 atthe 0.05 level of significance. 8.10 We know that silk fibers are very tough but in short supply: Breakthroughs by one research group result in the summa stalisties forthe stress (MPa) of synthetic silk fibers (Soures:F. Tele... (2012) Combining Sage {orm and agin spider sik mois to produce tunable systic biopolymer bers, Biopolymers, 916), $18-431.) ‘Small diameter Large diameter n=7 n=6 =1230 = 920 41 = 15.0 = 210 Use the 0.05 level of significance to test the claim that mean stress is largest for the small diameter fibers. Assume that both sampled populations have normal distributions with the same variance, 8.11 The following are the number of hydraulic pumps Which a sample of 10 industrial machines of Type A and a sample of 8 industrial machines of Type B man- ufactured over a certain fixed period of time’ Trea 8 6 79 4 8 6 9 Tee 43677 1 9 6 Assuming that the populations sampled can be approx imated closely with normal distributions having the same variance, test the null hypothesis py — yep = 0 ‘against the altemative hypothesis joy — 142 #0 atthe 0.05 level of significance 8.12. With reference to Example 5 construct 2 95% confi- ddence interval for the tue dilference between the av- erage resistance ofthe two kinds of wire 8.13 In each of the parts below, frst decide whether or not to use the pooled estimator of variance. Assume that the populations are normal (2) The following are the Brinell hardness values ob- tained for samples of two magnesium alloys be- ore testing! Alloy I: 66.3 63.5 64:9 618 64.3 64.7 65.1 64.5 68.4 63.2 Alloy 2: 71.3 60.4 62.6 63.9 68.8 70.1 64.8 68.9 65.8 66.2 Use the 0.05 level of significance to tet the null hypothesis soy ~ a = 0 against the alternative hypothesis jj — pp <0. (©) Tocompare two kinds of bumper guards, 6of cach kind, were mounted on a certain kind, of compact car, Then each cat was run into a concrete wall at S miles per hour, and the following are the costs ofthe repairs (in dollars): 407 448 423. 465 402 419 434 415 412. 451433 429 Bumper guard I Bumper guard 2: Use the 0.01 level of significance to test whether the difference between the two sample means is significant, 4. Matched Pairs Comparisons In the application of the two sample f test we need to watch that the samples are really independent. For instance, the test cannot be used when we deal with “before and after” data, the IQs of husbands and wives. and numerous other kinds of sit- uations where the data are naturally paired. Instead, comparisons are based on the matched pairs. A manufacturer is concerned about the loss of weight of ceramic parts during a baking step. The readings before and after baking, on the same specimen, are nat- ually paired, It would make no sense to compare the befor making weight of oneEXAMPLE 12, Solution See84 Matched Pats Comparvons 28 specimen with the after-baking weight of another specimen. Let the pair of ran- dom variables (X;,¥,) denote the weight before and weight after baking forthe ith specimen, for; = 1.2,...,n. A statistical analysis proceeds by considering the diferences Dj=X/-¥, fois 1,2... This collection of (signed) differences is then treated as a random sample of size n from a population having mean 1p. We interpret jz) = 0 as indicating that the means of the two responses are the same and Jip > 0 as indicating thal the mean sponse of the first is higher than that ofthe second Tests ofthe null hypothesis Mp: 1p) = sep,0 are based on the ratio _ yp, Yw-B? D- uo Solvn Unis small, andthe distribution ofa dlference is approximately normal, we teat this ratio a the one sampler statistic on page 253, Otherwise, we tea this ratio as the large sample statistic on page 252 where D= nT Conducting a paired test The following are the average weekly losses of worker-hours duc to accidents in 10 industrial plants before and after a certain safety program was put into operation: Before 45 73 46 124 33° 57 83 34 26 17 After 36 60 44 119 35 S177 29 24 11 Use the 0.05 level of significance to test whether the safety program is effective, We cannot apply the independent samples test because the before and after weekly losses of worker-hours in the same industrial plant are correlated, Here there is the obvious pairing of these two observations. 1, Null hypothesis: up = 0 Alternative hypothesis: xp > 0 2. Level of significance: « = 0.05 3. Criterion: Reject the null hypothesis if > 1.833, the value of tg 05 for 10 — 1 =9 degrees of freedom, where B-o Soi and D and Sp aze the mean and the standard deviation ofthe differences. 4, Calculations: The differences are 91325266526 thcir mean isd = 5.2, their standard deviation is sp = 4.08, so that 52-0 _ 4.08/10 8. Decision: Since t = 4.03 exceeds 1.833, the null hypothesis must be rejected at level « = 0.05, We conclude that the industrial safety program is effective. The evidence is very strong, since a computer calculation gives the P-value 0.0015 (see Figure 8.4). If zp = 0, only in 15 out of 10,000 times would we observe t greater than or equal to 4.03, . '282 —Chapter® Comparing Two Treaenents Figure 8.4 “The paired test: P-value = 0.0015 for t = 4.03 and degrees of freedom (df) = 9 EXAMPLE 13, Solution Grup 001s 400-20 188848 In conncetion with this kind of problem, the one sample testis referred to as a matched pairs f test or just the paired test. 95% confidence interval for the mean of a paired difference Scientists are making a major breakthrough by creating devices that can smell toxic chemicals.” An array of sites, each coated with different nanoporous pigments, change colors when exposed to various chemicals, Computer software prosduces the numer- ical value of the change, or difference, by subtracting an initial scanned image from, the image after exposure to the chemical, The red component of the difference of images, caused by exposure to a toxic level of formaldehyde, was measured seven times. (Courtesy of authors) 12% 134 «182058073078 1.10 Construct a 959% confidence interval for the mean change of the red color com- ponent at this site when exposed to a toxic level of formaldehyde ‘The sample size isn = T and fo.925 = 2.447 form — 1 = 6 degrees of freedom, We first calculate T= 1083 and 5 = 0.436 and the 95% confidence formula for jp becomes 0.436 0.436 1083 = 2.447, 286 uy < 1.083 + 2.447, 288 v7 MP Vi or 0.68 < up < 1.49, We are 95% conident that the interval from 0.68 to 1.49) contains the mean change in the red color component, The mean change is different from zero, ‘This site by itself contributes a substantial information for detecting formalde- hyde, By combining the measurements from all of the different sites in the array, scientists are actually able to identify many specitic toxic chemicals, These arrays can actually smell . ‘The next example illustrates some practical points when conducting a matched. pairs experiment, including randomization. Comparing measurements made at two laboratories A state law requires municipal wastewater treatment plants to monitor their dis- charges into rivers and streams. A treatment plant could choose to send its samples to a commercial laboratory of its choosing. Concer over this self-monitoring led a civil engineer to design a matched pairs experiment.‘ Exactly the same bottle of Liang Feng et al Colorimetric sensor aay fr determination and ientscation of toric industeal ciemieals. Anal Chem. €2 2010), 833-9440, R Sohason and D. Wicern, (2007), Applied MalivariateStatitical Analy, Prentice Hal, age 286Sec4 Matched Pats Comparions 283, effluent cannot be sent to two different laboratories. To match “identical” as closely as possible, she takes a sample of effluent in a large sample bottle and pours it back and forth over two open specimen bottles. When they are filled and capped, a coin is flipped to sce if the one on the right was sent to Commercial Laboratory A or the Wisconsin State Laboratory of Hygiene, This process was repeated 11 times. The results, for the response suspencied solids (SS) are Sample | 19203 4 5 6 7 8 9 1 I Commercial lab | 27 23 64 44 30 75 26 124 54 30 14 Statelab | 15 13 22 29 31 64 30 64 56 20 21 Difference xj — y; | 12 10 42 15 -1 11 -4 60 -2 10 ~7 Obtain a 95% confidence interval and look for any unusual features in the data. Solution The sample size is relatively small swe assume normality and base the confidence interval onthe distribution, We have n = 11 and calculate d = 13.27 and sf, = 418.61. Then, with n— 1 = 11 —1 = 10 degrees of freedom and fp ons = 2.228, the 95% confidence interval is 3 3 (o21-2ns (BH no vam BE) we com mm ‘This 95% confidence interval just covers 0, so no difference is indicated with this small sample size. But wait, look at the dot diagram of the differences in Figure 8.5 ‘There are two very large differences that would be unusual if the sample were taken from a normal population, ‘The validity of the confidence interval is, at least, under suspicion, In Exercise 8.17 you are asked to try the square root transformation to see if it improves the situation, . Figure 8.5 Dedagtam of diferencesin #2e@, fee ng suspendedsolids;ouliers = 30S present * Exercises 8.14 A civil engineer wants to compare two nuachines for grinding cement and sand. A sample of a fixed (quantity of cement and sand is taken and put in each machine, The machines are run and the fine- ress of each mixture is noted, This process is re- peated five times. The results, in microns, are as Tollows: Sample No. Machi 3 5 3 6 9284 817 8.21 Chapter 8 Comparing Two Tresments Find 2 99% confidence interval forthe mean difference Jn machine readings assuming the differences have a normal distribation, Refer to Exercise 8.14. Test with a = 0.01, thatthe mean difference is 0 versus a two-sided alternative. ‘The following data were obtained in an experiment de- signed to check whether there is a systematic dilfer- {ence in the weights obtained with two different scales: Weight in grams Seale | Seale It Rock Specimen 1 11.23 127 Rock Specimen2 14,36 1441 Rock Specimen 3 833 835 Rock Specimen 4 10.50 1052 Rock Specimen 23.42. 241 Rock Specimen 6 9.15 9.7 Rock Specimen? 13.47 Bs Rock Specimen 8 641 646 Rock Specimen9 12.40 124s Rock Specimen 1019.38 1935 ‘Use the paired 1 test at the 0.05 level of significance to tty to establish that the mean difference ofthe weights ‘obtained withthe two scales is nonzero. Refer to Example 14 concerning suspended solids in cfflaent from a treatment plant, Take the square root of ‘cach ofthe measurements and then take the dilference (a) Construct a 95% contidence interval fort. (b) Conduct a level « = 0.05 level test of Hy: kn = O against a two-sided alternative. Verify thatthe 8.18 aly 820 ‘conclusion is the same as that oblained from the ‘confidence interval, (©) Make a dot diagram of these differences and de- ‘ide ithe transformation has essentially removed the outliers. Refer to Example 14 conceming suspended solids in eiflint from treatment plant, Take the natural loga- ‘thm of each of the measurements and then take the difference. (a) Construct a 95% confidence interval for. (b) Conduct a level a = 0.05 level test of Ho: a 0 against a ewo-sided alternative, Verity that the conclusion is the same as that obtained from the confidence interval (©) Make a dot diagram of these diferences and de- tide ifthe transformation has essentially removed the outliers. A shoe manufacturer wants potential customers to ‘compare two types of shoes, one made of the current PVC material X and one made of a new PVC material ¥, Shoes made of both are available. Each person, in a sample of 52, is asked to Wear one pair of each type for a whole day. After a walk of 2 km, they are asked to score that day’s pair on a seale of I to 10, with higher scores being better. The differences in scares (New PVC Y) — (Current PVC X) have mean 2.6 and vatiance 3.9. Construct a 90% con- fidence interval for tae mean difference. Referring to Example 13, conduct a test to show that the mean change jp is different from 0. Take a = 0.05, In a study of the effectiveness of physical exercise in weight reduction, a group of 16 persons engaged in a prescribed program of physical exercise for one month showed the following results: Weight before Weight after | Weight before (pounds) (pounds) (pounds) 209 196 170 v8 m7 153 169 170 13 212 207 165 180 i 201 12 190 179 158 159 243 180 180 144 ‘Weight after (pounds) 164 152 179 162 199 13 BI 140 ‘Use the 0.01 level of signiticance to test whether the prescribed program of exercise is effective,