252y0631 11/28/06 (Page layout view!) ECO252 QBA2 THIRD HOUR EXAM Nov 30 and Dec 1, 2006 Name KEY Hour of Class Registered (Circle) MWF 1 MWF 2 TR 12:30 TR 2 I. (8 points) Do all the following (2points each unless noted otherwise). Make Diagrams! Show your work! x ~ N 15, 9.3 20 15 Pz 0.54 Pz 0 P0 z 0.54 .5 .2054 .2946 1. Px 20 P z 9.3 For z make a diagram. Draw a Normal curve with a mean at 0. Indicate the mean by a vertical line! Shade the area above 0.54. Because this is on one side of zero we must subtract the area between 0 and 0.54 from the area above zero. If you wish, make a completely separate diagram for x . Draw a Normal curve with a mean at 15. Indicate the mean by a vertical line! Shade the area above 20. This area is totally above the mean so we must subtract the area between the mean and 20 from the area 50% above the mean. This is how we usually find a p-value when the distribution of the test ratio is Normal. 14 15 0 15 z P 1.61 z 0.11 P1.61 z 0 P0.11 z 0 2. P0 x 14 P 9.3 9.3 .4463 .0438 .4025 For z make a diagram. Draw a Normal curve with a mean at 0. Indicate the mean by a vertical line! Shade the area between -1.61 and -0.11. Because this is on one side of zero we must subtract the area between -0.11 and zero from the area between -1.61 and zero. If you wish, make a completely separate diagram for x . Draw a Normal curve with a mean at 15. Indicate the mean by a vertical line! Shade the area between zero and 14. This area is totally below the mean so we must subtract the area between 14 and the mean (15) from the area between zero and the mean. 16 15 16 15 z P 3.33 z 0.11 P3.33 z 0 P0 z 0.11 3. P16 x 16 P 9 . 3 9.3 .4996 .0438 .5434 For z make a diagram. Draw a Normal curve with a mean at 0. Indicate the mean by a vertical line! Shade the area between -3.33 and 0.11. Because this is on both sides of zero we must add together the area between -3.33 and zero and the area between zero and 0.11. If you wish, make a completely separate diagram for x . Draw a Normal curve with a mean at 15. Indicate the mean by a vertical line! Shade the area between -16 and 16. This area includes the mean (15), and areas to either side of it so we add together these two areas. 4. x.42 (Find z .42 first) Solution: (Do not try to use the t table to get this.) For z make a diagram. Draw a Normal curve with a mean at 0. z .42 is the value of z with 42% of the distribution above it. Since 100 – 42 = 58, it is also the 58th percentile. Since 50% of the standardized Normal distribution is below zero, your diagram should show that the probability between z .42 and zero is 58% - 50% = 8% or P0 z z.42 .0800 . The closest we can come to this is P0 z 0.20 .0783 . (0.21 is also acceptable here.) So z.42 0.20. To get from z .42 to x.42 , use the formula x z , which is the opposite of x . x 15 0.20 9.3 16 .86 . If you wish, make a completely separate diagram for x . Draw a Normal curve with a mean at 20. Show that 50% of the distribution is below the mean (7). If 42% of the distribution is above x.42 , it must be above the mean and have 8% of the distribution between it and the mean. Note that there will be some questions on the Final where you will need odd values of z like z .125 . z 16 .86 15 Pz 0.20 Pz 0 P0 z 0.20 .5 .0793 Check: Px 16 .86 P z 9.3 .4207 42% 1 252y0631 11/28/06 (Page layout view!) II. (22+ points) Do all the following (2points each unless noted otherwise). Do not answer question ‘yes’ or ‘no’ without giving reasons. Show your work in questions that are not multiple choice. Look them over first. The exam is normed on 50 points. Note the following: 1. This test is normed on 50 points, but there are more points possible including the take-home. You are unlikely to finish the exam and might want to skip some questions. 2. If you answer ‘None of the above’ in any question, you should provide an alternative answer and explain why. You may receive credit for this even if you are wrong. 3. Use a 5% significance level unless the question says otherwise. 4. Read problems carefully. A problem that looks like a problem on another exam may be quite different. 5. Use statistical tests! Just because two means or two proportions look different to you does not mean that they are significantly different unless you prove that the probability of getting the observed difference if the null hypothesis is true is very small. 1. Turn in your computer problems 2 and 3 marked to show the following: (5 points, 2 point penalty for not doing.) a) In computer problem 2 – what is tested and what are the results? b) In computer problem 3 – what coefficients are significant? What is your evidence? c) In the last graph in computer problem 3, where is the regression line? [5] 2. (Abronovic) The distance that a baseball travels after being hit is a function of the velocity (in mph) of the pitched ball. A ball is pitched to a batter with a 35 inch, 32 oz bat that is swung at 70mph from the waist and at an angle of 35%. The experiment is repeated 9 times. A partial Minitab printout appears below. Use .01 throughout this problem. DIST = ……… Predictor Constant VELOC s = 1.185 + ……… VELOC Coef StDev t-ratio p 31.311 0.999 …………… ………… 0.74667 0.01529 …………… ………… R-sq = ……… Rsq(adj) = ……… Analysis of Variance SOURCE DF SS Regression 1 3345.1 Error 7 9.8 Total 8 3354.9 MS 3345.1 1.4 F ……… p ………… a) The fastest pitchers can throw at about 100 mph. How far will such a pitch be hit? (2) b) What is the value of R-squared? (2) c) Fill in the F space in the ANOVA and explain specifically what is tested and what are the conclusions. (3) d) Is the constant (31.311) significant? Why? (2) [14] ˆ Solution: a) The equation given above is Y b0 b1 X or DIST = 31.311 + 0.74667 VELOC = 31.311 + 0.74667(100) = 105.98. b) The ANOVA tells us that SSR 3345 .1 and SSR SSY 3354 .9 . So it must be true that SSR 3345 .1 R2 .9971 SST 3354 .9 c) Analysis of Variance 2 252y0631 11/28/06 (Page layout view!) SOURCE Regression Error Total DF 1 7 8 SS 3345.1 9.8 3354.9 1,7 MS 3345.1 1.4 F 2389.36 The relevant table value of F is F.01 12.25 . Since our computed F is larger than the table value of F, we reject the null hypothesis that there is no relationship between the dependent variable and the independent variable(s). For a simple regression this null hypothesis is equivalent to saying that if Y 0 1 X , we have H 0 : 1 0 . Yˆ b0 b1 X H 0 : 0 00 b 00 d) The outline says that if we test , use t 0 . So if we test for s b0 H 1 : 0 00 H 0 : 0 0 significance the null hypothesis is that 0 is insignificant or , we use H 1 : 0 0 b 0 31 .311 t 0 2047 .8 . The degrees of freedom are n 2 7 . We cannot reject the null s b0 0.01529 7 7 3.449 and t .005 3.449 . hypothesis if the computed value of the t statistic is between t .005 Since the computed t is not between these values we must say that 0 is significant. 3. (Render) A firm renovates homes in upstate New York. Sales and payroll for the region for a random sample of years are given below. Sales are in $100,000 and Payroll is in $0.1 billions. Current sales are $210000.00 and because of the opening of several new plants payroll is anticipated to be about 0.6 billion dollars for the foreseeable future. Will average sales be significantly different from current sales? To answer this question. a) Complete the XY column (1) b) Find the regression equation (4) c) Find an appropriate interval for sales (3) d) State and justify your conclusion (2) [24] Y – Sales 2.0 3.0 2.5 2.0 2.0 3.5 15.0 X - Payroll 1 3 4 2 1 7 18 X2 1 9 16 4 1 49 80 Y2 4 9 6.25 4 4 12.25 39.50 XY Solution: a) Y – Sales 2.0 3.0 2.5 2.0 2.0 3.5 15.0 X - Payroll 1 3 4 2 1 7 18 X2 1 9 16 4 1 49 80 Y2 4 9 6.25 4 4 12.25 39.50 XY 2 9 10 4 2 24.5 51.5 b) n 6, X 18, Y 15, XY 51 .5, X 2 80 and Y 2 51 .5. 3 252y0631 11/28/06 Means: X (Page layout view!) X 18 3.00 n Spare Parts: SS x SS y 6 X Y 2 2 Y Y 15 2.5 n 6 nX 80 63 26 2 2 nY 2 39.5 62.52 2 SST (Total Sum of Squares) XY nXY 51.5 632.5 6.50 S XY nXY 6.5 0.25 Coefficients: b SS X nX 26 S xy xy 1 2 x 2 b0 Y b1 X 2.5 0.25 3 1.75 . So Yˆ 1.75 0.25 X The following computation may be useful, but is not needed: R2 S xy 2 6.52 .8125 SSR b1 S xy 0.256.5 .8125 SST SSy 2 SS x SS y 262 c) Since we are finding an interval for average sales, the confidence interval is appropriate and the SSE SST SSR SS y b1 S xy 2 0.256.5 significance interval is not. s e2 .09375 n2 n2 n2 4 s e .09375 .3062 The Confidence Interval is Y0 Yˆ0 t sYˆ , where 1 X X sY2ˆ s e2 0 n SS x 2 .09375 1 6 32 6 26 .09375 0.166667 0.346154 0.0480769 so 4 4.604 s ˆ 0.0480769 0.21926 t .005 Y Yˆ 1.75 0.256 3.25 y 3.25 4.604.21926 3.25 1.01 or 2.24 to 4.26. d) In hundreds of thousands, current sales are 2.1. since this is not on the confidence interval, sales will be significantly different from 2.1. 4 252y0631 11/28/06 (Page layout view!) 4. The following data may look familiar. It is the service method data from the last exam. 11 service methods were tried to see which is the fastest. I ran a multiple comparison of the 11 methods. Data were stacked as ‘time’ in column 12 with the column labels ‘tel’ in column 13. This was run on Minitab in three different ways, as a one-way ANOVA, using the Kruskal-Wallis test and using the Mood median test. I have no idea how to do a Mood median test but I know this much. The null hypothesis is equal medians. The assumptions are comparable to the Kruskal-Wallis test. The Mood median test is less powerful than the Kruskal-Wallis test but is less affected by outliers. The Mood median test produces a chi-squared statistic with degrees of freedom equal to the number of columns less 1. Row 1 2 3 4 5 6 7 8 9 10 11 12 13 14 1 2.9 3.9 2.6 3.1 3.9 2.6 3.3 3.0 3.5 3.1 3.2 2.4 4.0 4.3 2 2.8 2.6 2.6 2.9 2.9 2.8 2.3 2.4 2.0 2.5 2.4 2.0 4.1 4.3 3 2.6 2.7 3.2 2.8 3.6 2.1 2.3 2.6 2.6 2.9 2.4 2.3 0.5 4.7 4 7.7 2.9 4.3 2.7 3.4 4.4 5.5 3.4 3.4 3.5 4.0 3.4 4.1 3.8 5 2.4 13.4 5.8 1.5 9.8 2.7 2.7 4.5 2.3 5.8 4.8 4.2 5.8 6.1 6 6.6 3.7 9.7 1.9 10.1 4.5 2.9 9.9 3.0 31.5 3.5 5.3 9.8 5.3 7 3.5 8.4 4.3 3.3 11.9 3.7 3.0 2.9 3.6 5.4 4.4 3.0 4.3 5.4 8 3.4 8.3 4.2 3.2 11.0 3.6 2.9 2.8 3.5 5.3 4.3 2.9 4.2 5.3 9 3.5 3.4 3.8 3.4 3.6 3.5 4.8 3.5 5.3 3.7 3.4 3.6 3.8 3.8 10 2.3 6.9 3.3 5.3 3.0 3.3 6.1 3.1 2.6 4.4 15.0 6.9 2.1 10.4 11 3.4 4.4 3.1 3.6 4.4 3.1 3.8 3.5 4.0 3.6 3.7 2.9 4.5 4.8 The edited Minitab output follows. Note that the p-values are missing and for you to think about. Results for: 252x06031-01.MTW MTB > Name c72 "RESI1" #Residuals are stored for Normality test. MTB > Oneway c12 c13; SUBC> Residuals 'RESI1'; SUBC> GNormalplot; SUBC> NoDGraphs. One-way ANOVA: time versus tel Source DF SS MS F P tel 10 284.43 28.44 ……… ………… Error 143 1228.86 8.59 Total 153 1513.28 S = 2.931 R-Sq = 18.80% R-Sq(adj) = 13.12% Level tel 1 tel 10 tel 11 tel 2 tel 3 tel 4 tel 5 tel 6 tel 7 tel 8 tel 9 N 14 14 14 14 14 14 14 14 14 14 14 Mean 3.271 5.336 3.771 2.757 2.664 4.036 5.129 7.693 4.793 4.636 3.793 StDev 0.580 3.630 0.580 0.677 0.909 1.265 3.214 7.446 2.503 2.331 0.561 Individual 95% CIs For Mean Based on Pooled StDev ------+---------+---------+---------+--(-----*-----) (-----*------) (-----*-----) (-----*-----) (------*-----) (-----*-----) (------*-----) (-----*-----) (-----*-----) (------*-----) (-----*-----) ------+---------+---------+---------+--2.5 5.0 7.5 10.0 Pooled StDev = 2.931 Normplot of Residuals for time (This plot was identical to the one shown below) 5 252y0631 11/28/06 (Page layout view!) MTB > Kruskal-Wallis c12 c13. Kruskal-Wallis Test: time versus tel Kruskal-Wallis Test on time tel N Median Ave Rank tel 1 14 3.150 60.1 tel 10 14 3.850 86.8 tel 11 14 3.650 85.6 tel 2 14 2.600 33.3 tel 3 14 2.600 33.5 tel 4 14 3.650 85.4 tel 5 14 4.650 90.3 tel 6 14 5.300 107.2 tel 7 14 4.000 94.7 tel 8 14 3.900 89.5 tel 9 14 3.600 86.1 Overall 154 77.5 H = 41.97 Z -1.53 0.82 0.72 -3.89 -3.87 0.70 1.12 2.61 1.51 1.06 0.75 #H is the Kruskal-Wallis statistic that I #taught you about. MTB > Mood c12 c13. Mood Median Test: time versus tel Mood median test for time Chi-Square = 23.34 DF = 10 tel tel tel tel tel tel tel tel tel tel tel tel 1 10 11 2 3 4 5 6 7 8 9 N<= 10 7 5 12 12 7 5 4 5 6 6 N> 4 7 9 2 2 7 9 10 9 8 8 Median 3.15 3.85 3.65 2.60 2.60 3.65 4.65 5.30 4.00 3.90 3.60 Q3-Q1 1.08 4.00 1.08 0.53 0.68 0.93 3.25 6.45 2.18 2.18 0.33 P = ……… Individual 95.0% CIs -+---------+---------+---------+----(*--) (--*------------) (*--) *-) (*-) (*-) (-------*---) (------*-----------------) (--*-----) (--*----) *) -+---------+---------+---------+----2.5 5.0 7.5 10.0 Overall median = 3.50 MTB > NormTest c72; SUBC> KSTest. Probability Plot of RESI1 #This is a plot of the differences between means of #each columns and the actual data. 6 252y0631 11/28/06 (Page layout view!) MTB > Vartest c12 c13; SUBC> Confidence 95.0. Test for Equal Variances: time versus tel (This plot was not needed.) Bartlett's Test (normal distribution) Test statistic = 190.25, p-value = 0.000 Levene's Test (any continuous distribution) Test statistic = 3.29, p-value = 0.001 a) Are the means significantly different? Identify the test that answers this question, complete it and answer the question. Explain! (3) Solution: The null hypothesis of the ANOVA is that the means are all equal. The 28 .44 3.3108 . The F statistic has 10 and 143 degrees of computed F statistic is F 8.59 10,125 1.91 and F 10,200 1.88 . So F 10,143 freedom. According to the F table F.05 .05 .05 must be between these two values. Since the computer F statistic exceed the table value, reject the null hypothesis and say that the means are significantly different. b) Are the medians significantly different? Complete one or both of the remaining tests and answer the question. Explain! (4) Solution: The number of columns makes the Kruskal-Wallis table inapplicable. Under these circumstances the H statistic has the Chi-squared distribution with degrees of 10 freedom one less than the number of columns. According to our table 2 .05 18.3070. The printout says that H 41.97 . Since the computed value exceeds the table value, we can reject the null hypothesis of equal medians. The Chi-square statistic 2 23.34 produced by the Mood median test also has 10 degrees of freedom. Since the computed value exceeds the table value, we can reject the null hypothesis of equal medians. c) At the end of the printout there are a K-S (actually Lilliefors) test, a Bartlett and a Levene test. What do they tell us about the applicability of the ANOVA, Kruskal-Wallis or Mood test to the problem? What conclusion is the most reliable? (3) Solution: The p-value of less than 1% for the Lilliefors test means that we can reject the null hypothesis of Normality at the 5% or 1% level. This means that the ANOVA should not be used. The probability plot and a check of the numbers both imply the presence of outliers so that the Mood test is most applicable. d) On the basis of all this, which of the service methods are best? (1) [35] Solution: The Mood test printout says that the medians differ and tells us that the service methods with the lowest median time are methods 2 and 3. 7 252y0631 11/28/06 (Page layout view!) 5. Four experts rated 4 brands of Columbian coffee. By adding together ratings on a seven point scale for taste, aroma, richness and acidity, each coffee is given a rating on a 28 point scale. The following table gives these summed ratings. Assume that the underlying distribution is not Normal and test for a difference in the ratings of the four brands. (5) Note: .10 Row 1 2 3 4 Brand A Brand B Brand C Brand D x1 x2 x3 x4 24 27 19 24 26 27 22 27 25 26 20 25 22 24 16 23 Solution: (This is an abbreviated version of Problem 12.95) Since the underlying distribution is not Normal and the data is cross classified this is a Friedman test. Note that much time was wasted on this because people wanted to believe it was an ANOVA. ANOVA is not appropriate on data from non-Normal distributions unless they are very well-behaved. H0: 1 2 3 4 Where 1 is A, 2 is B, 3 is C and 4 is D. H1: At least one of the medians differs. First we rank the data within rows. The data appears below in columns marked x1 to x 4 and the ranks are in columns marked r1 to r4 . Row Brand A x1 1 2 3 4 SRi 24 27 19 24 Brand B r1 x2 2 3.5 2 2 9.5 26 27 22 27 Brand C r2 Brand D x3 4 3.5 4 4 15.5 25 26 20 25 r3 3 2 3 3 11 x4 22 24 16 23 r4 1 1 1 1 4 To check the ranking, note that the sum of the four rank sums is 9.5 + 15.5 + 11 + 4 = rcc 1 445 SRi 40 . 40, and that the sum of the rank sums should be 2 2 12 SRi2 3r c 1 Now compute the Friedman statistic F2 rc c 1 i 12 9.52 15 .52 112 42 345 12 467 .5 60 10 .125 . 80 4 45 Since the size of the problem is one shown in Table 8, we can use that table, which appears, in part, below. c4, r 4 F2 9.900 10.200 p value .006 .003 Since 10.125 lies between 9.9 and 10.2, the p-value is between .006 and .003, both of which are below .10, so reject the null hypothesis. 8 252y0631 11/28/06 (Page layout view!) ECO 252 QBA2 THIRD EXAM Nov 30 and Dec 1, 2006 TAKE HOME SECTION Name: _________________________ Student Number: _________________________ Class days and time: _________________________ Please Note: Computer problems 2 and 3 should be turned in with the exam (2). In problem 2, the 2 way ANOVA table should be checked. The three F tests should be done with a 5% significance level and you should note whether there was (i) a significant difference between drivers, (ii) a significant difference between cars and (iii) significant interaction. In problem 3, you should show on your third graph where the regression line is. Check what your text says about normal probability plots and analyze the plot you did. Explain the results of the t and F tests using a 5% significance level. (2) III Do the following. Note: Look at 252thngs (252thngs) on the syllabus supplement part of the website before you start (and before you take exams). Show your work! State H 0 and H 1 where appropriate. You have not done a hypothesis test unless you have stated your hypotheses, run the numbers and stated your conclusion. (Use a 95% confidence level unless another level is specified.) Answers without reasons or accompanying calculations usually are not acceptable. Neatness and clarity of explanation are expected. This must be turned in when you take the in-class exam. Note that from now on neatness means paper neatly trimmed on the left side if it has been torn, multiple pages stapled and paper written on only one side. Because so much of this exam is based on student numbers, there is a penalty for failing to state your correct student number. 1) Bassett et al. give the following numbers for the year and the number of pensioners in the United Kingdom. Pensioners are in millions. The 2000 number is a bit shaky, so subtract the last digit of your student number divided by 10 from the 12.00 that you see there. Label your answer to this problem with a version number. (Example: Good ol’ Seymour’s student number is 123456, so the 12.000 becomes 12.000 0.6 = 11.400 and he labels it Version 6.) 'Pensioners' is the dependent variable and 'Year' is the independent variable, so what you are going to get is a trend line. If you don’t know what dependent and independent variables are, stop work until you find out. 1 2 3 4 5 6 7 8 Year 1966 1971 1975 1976 1977 1978 1979 2000 Pensioners 6.679 7.677 8.321 8.510 8.637 8.785 8.937 12.000 Bassett et. al. strongly suggest that you change the base year to something other than the year zero. They recommend that you subtract 1970 from every number in the ‘Year’ column, so that 1966 becomes -4 and 2000 becomes 30. This will make your computations easier. a. Compute the regression equation Y b0 b1 x to predict the number of pensioners in each year. (3).You may check your results on the computer, but let me see real or simulated hand calculations. b. Compute R 2 . (2) c. Compute s e . (2) d. Compute s b1 and do a significance test on b1 (2) e. Use your equation to predict the number of pensioners in 2005 and 2006. Using the 2006 number, create a prediction interval for the number of pensioners for that year. Explain why a confidence interval for the number of pensioners is inappropriate. (3) 9 252y0631 11/28/06 (Page layout view!) f. Make a graph of the data. Show the trend line clearly. If you are not willing to do this neatly and accurately, don’t bother. (2) g. What percent rise in pensioners did the equation predict for 2006? What percent rise does it predict for 2050? The population of the United Kingdom grew at roughly 0.31% a year over the last quarter of the 20 th century. Can you intelligently guess what is wrong? (1) [15] Solution: Version 0 i X 1 2 3 4 5 6 7 8 -4 1 5 6 7 8 9 30 62 6.679 7.677 8.321 8.510 8.637 8.785 8.937 12.000 69.546 To summarize n 8, Y 2 16 1 25 36 49 64 81 900 1172 XY Y2 -26.716 7.677 41.605 51.060 60.459 70.280 80.433 360.000 644.798 44.609 58.936 69.239 72.420 74.598 77.176 79.870 144.000 620.848 X 62, Y 69 .546 , XY 644 .798 , X 2 1172 and 620 .848 . df n 2 8 2 6. X 62 7.75 Means: X n 8 X Spare Parts: SS x 2 Y Y 69.546 8.69325 n 8 nX 1172 87.752 691.50 2 Y nY 620.848 88.69325 16.2673 SST XY nXY 644 .798 87.75 8.69325 105 .8165 SS y S xy X2 Y 2 2 2 (Total Sum of Squares) Coefficients: b1 S xy SS x XY nXY X nX 2 2 105 .8165 0.1530 691 .500 b0 Y b1 X 8.69325 0.1530 7.75 7.5075 So our equation is Yˆ 7.5075 0.1539 X b. Compute R 2 . (2) Solution: SSR b1 Sxy 0.1530 105 .8165 16.1899 S xy 2 105 .8165 2 .9954 SSR b1 S xy 16 .1899 2 R .9952 or R SST SSy 16 .2673 SS x SS y 691 .516 .2673 2 c. Compute s e . (2) Solution: SSE SST SSR 16.2673 16.1899 .07740 s e2 SSE .07740 .01290 s e .01290 0.11358 n2 6 Note also that if R 2 is available s e2 SS y 1 R 2 n2 16.2673 1 .9954 .01247 6 and s e .01247 0.11168 . I will compromise and use s e2 .0127 10 252y0631 11/28/06 (Page layout view!) 1 0.0127 d. Compute s b1 and do a significance test on b1 . (2) Solution: s b21 s e2 0.00001837 SS x 691 .5 b 0 0.1539 s b 0.00001837 0.004286 . t 1 35 .908 . If we assume that .05 , compare this 1 s b1 0.004286 6 with t .025 2.447. Since the computed t is larger than the table t in absolute value, we reject the null hypothesis of no significance and say that the slope is significant. e. Use your equation to predict the number of pensioners in 2005 and 2006. Using the 2006 number, create a prediction interval for the number of pensioners for that year. Explain why a confidence interval for the number of pensioners is inappropriate. (3). Solution: Our base year was 1970 so the value of X for 2005 is 2005 – 1970 = 35. Our equation is Yˆ 7.5075 0.1539 X and for 2005 it gives Ŷ 7.5075 0.153935 12.894. Since the slope is 0.1539, add this to the 2005 value to get 13.048 for 2006. The formula for the 2 X0 X 2 2 1 ˆ Prediction Interval is Y0 Y0 t sY , where sY s e 1 . For 2006, X 0 36 . It’s time to n SS x remember that s e .01247 0.11168 , n 8, X 7.75 and SS x X 2 nX 2 691 .50 . 1 36 7.75 2 So sY2 .1247 1 .1247 0.1250 1.1541 1 .1247 2.2791 0.2842 8 691 .5 s 0.2842 0.5331 . We already know that t 6 2.447 and that for 2006 Yˆ 13 .048 . So that our Y .025 prediction interval is Y0 Yˆ0 t sY 13.048 2.4470.5331 13.048 1.304 , or between 11.7 and 14.3 million people. The confidence interval is inappropriate for this type of problem. To use the problem demonstrated in class, a confidence interval done when X 0 5 gives us a likely range in which the average number of children will fall for the average couple that wants 5 children. The prediction interval gives us a likely range in which the number of children will fall for one couple that wants 5 children. There is no average year 2006, since it will be 2006 only once, so an interval for an average makes no sense. f. Make a graph of the data. Show the trend line clearly. If you are not willing to do this neatly and accurately, don’t bother. (2) Suggestions: I don’t have the graphical power to do this but here’s what I would do. Since our years vary from 1966 to 2006 and the number of pensioners varies from 6.68 to a projected 13.0 million. Your x axis should be marked from 1966 to 2006, but I would probably only mark the years 1970 to 2010 by 5-year intervals. I might consider stretching the whole thing out to 2050 because of the next question. The y-axis might go from 6 to 14 million with marks for every two million. I can plot the regression line Yˆ 7.5075 0.1539 X by noting that for 1970 it gives us 7.5 million which grows to 13 million in 2006. these two points can be connected to give us the regression line and the 8 points that we used to get the regression equation can be plotted around it. g. What percent rise in pensioners did the equation predict for 2006? What percent rise does it predict for 2050? The population of the United Kingdom grew at roughly 0.31% a year over the last quarter of the 20 th century. Can you intelligently guess what is wrong? (1) Solution: Our base year was 1970 so the value of X for 2049 is 2049 – 1970 = 79. Our equation is Yˆ 7.5075 0.1539 X and for 2049 it gives Yˆ 7.5075 0.153979 19.666 . Since the slope is 0.1539, add this to the 2049 value to get 19.819 for 2050. We thus have the following. Year Number Per cent growth. 2005 12.894 2006 13.048 1.19 2049 19.666 2050 19.819 0. 78 11 252y0631 11/28/06 (Page layout view!) It is reasonable to expect the growth rate to fall as the absolute number of additional pensioners stays constant but the number of pensioners grows. However, we are predicting that for an 80 year period the number of pensioners will grow faster than population. Aside from the fact that this results in a substantial rise in the number of pensioners per worker that may be politically impossible, it’s hard to believe that the country will produce old people at a rate substantially higher than population grows for over 80 years. This is a basic problem with using a trend line to make predictions. A given slope may be appropriate for a long while, but it is no more appropriate to say it will last forever, than it is to say that Wal-Mart’s sales will continue to grow at a rate way above total retail sales for decades. Sooner or later Wal-Mart’s sales will be such a large part of retail sales that the only way to grow Wal-Mart’s sales at such a high rate is to grow all retail sales at a higher rate, which just won’t happen if we are all drawing our wages from Wal-Mart. 2) The Lees in their text ask whether experience makes a difference in student earnings and present the following data for student earnings versus years of work experience. To personalize these data, take the second to last digit of your student number call it a . Clearly label the problem with a version number based on your student number. Then take your a , multiply it by 0.5 and add it to the 13 in the lower left corner. . (Example: Good ol’ Seymour’s student number is 123456, so the 13 becomes 13 + 0.5(5) = 13 + 2.5 = 15.5 and he labels it Version 5.) Each column is to be regarded as an independent random sample. Years of Work Experience 1 2 3 16 19 24 21 20 21 18 21 22 13 20 25 a) State your null hypothesis and test it by doing a 1-way ANOVA on these data and explain whether the test tells us that experience matters or not. (4) b) Using your results from a) present two different confidence intervals for the difference between earnings for those with 1 and 3 years experience. Explain (i) under what circumstances you would use each interval and (ii) whether the intervals show a significant difference. (2) b) What other method could we use on these data to see if years of experience make a difference? Under what circumstances would we use it? Try it and tell what it tests and what it shows. (3) [24] c) (Extra Credit) Do a Levene test on these data and explain what it tests and shows. (4) Solution – Version 0: You should be able to do the calculations below. Only the three columns of numbers were given to us. 1 16 21 18 13 Years 2 19 20 21 20 3 24 20 22 25 Sum 68 + 80 + 92 240 nj 4+ 4+ 4 12 n x j 17 20 23 SS 1190 + 1602 + 2126 x 2j SST x nx 2 ij 289 2 400 + Sum 529 x ij 20 x x 1218 x 4918 2 ij 2 j 4918 12202 4918 4800 118 12 252y0631 11/28/06 SSB (Page layout view!) n x nx 417 2 j .j 2 2 Source SS Between 72 4202 4232 12202 41218 12202 4872 4800 72 DF MS F F.05 2 36 7.04 s F 2,9 4.26 H0 Column means equal Within 46 9 5.1111 Total 118 11 Because our computed F statistic exceeds the 5% table value, we reject the null hypothesis of equal means and conclude that experience matters. b) The following material is modified from the solution to the last graded assignment. Types of contrast between means. Assume 0.05 . m 4 is the number of columns. Individual Confidence Interval If we desire a single interval, we use the formula for the difference between two means when the variance is known. For example, if we want the difference between means of column 1 and column 2. 1 2 x1 x2 tn m s 2 1 1 , where s MSW . The within degrees of freedom are n1 n2 9 2.262 . n1 n 2 n3 n4 4 , so n m 9, , so we use t .025 s2 1 n1 1 n2 s 2 1 4 1 4 s 2 0.5 and our interval will be 1 3 x1 x3 2.262 s 2 .5 Scheffé Confidence Interval If we desire intervals that will simultaneously be valid for a given confidence level for all possible intervals 1 1 between column means, use 1 2 x1 x2 m 1Fm1,n m s where s MSW . n n 2 1 The degrees of freedom for columns are m 1 2 . The within degrees of freedom are n m 9, so we 2,9 4.26 . use F.05 n1 n 2 n3 n4 4 , so s2 1 n1 1 n2 s 2 18 18 s 2 0.25 and our interval will be 1 3 x1 x3 24.26 s 2 .5 x1 x2 2.9289 s 2 .5 Tukey Confidence Interval This also applies to all possible differences. 1 2 x1 x2 q m,n m s 1 1 . where s MSW . This gives rise to Tukey’s HSD n1 n 2 2 (Honestly Significant Difference) procedure. Two sample means x .1 and x .2 are significantly different if x.1 x.2 is greater than q m,n m s 2 1 1 n1 n 2 s2 2 s 2 1 1 3,9 3,9 3.95 . . We will need q .05 . The table says q.05 n1 n 2 1 1 2 2 .25 s and the interval will be 1 3 x1 x3 3.95 0.5s .5 4 4 x1 x2 2.7931 s 2 .25 13 252y0631 11/28/06 (Page layout view!) Contrasts for 3 and 1 . s 2 .5 .5.1111 .5 2.555556 1.5986 . Intervals for differences between Note that in all contrasts means that include zero show no significant difference. x1 x3 17 23 6. Individual – Used when you want only one interval. 1 3 x1 x3 2.262 s 2 .5 6 2.262 1.5986 6 3.61 Scheffé – Used when a collective confidence level is sought. 1 3 x1 x2 2.9289 s 2 .5 6 2.9289 1.5986 6 4.68 Tukey – More powerful, but similar to the Scheffé. 1 3 x1 x2 2.7931 s 2 .25 6 2.7931 1.5986 6 4.47 . All contrasts seem significant. c) The alternative to one-way ANOVA for situations in which the parent distribution is not Normal is the 2.0 4.0 11 .0 16 19 24 8 .0 5 .5 8 .0 21 20 21 Kruskal-Wallis test. The original data is replaced by ranks 3.0 8.0 10 .0 . Note that the 18 21 22 1.0 5.5 12 .0 13 20 25 14 .0 23 .0 41 .0 12 13 78 , so that we can check our ranking by noting that 14 + 23 + 2 SRi 2 3n 1 n i sum of the first twelve numbers is 12 41 = 78. H nn 1 i 12 14 2 23 2 412 313 1 1 196 529 1681 39 46 .2692 39 7.2692 . Since the 12 13 4 4 4 13 4 Kruskal-Wallis table for 4, 4, 4 says that 5.6923 has a p-value of .049 and 7.5385 has a p-value of .011, the p-value of 7.2692 must be below 5%, so that we can reject the null hypothesis of equal medians. 16 19 24 21 20 21 , which has column medians 18 21 22 13 20 25 1 1 1 4 0 2 of 17, 20 and 23, is replaced by the numbers with the column medians subtracted . Absolute 1 1 1 4 0 2 1 1 1 4 0 2 values are taken, so the columns become . An ANOVA is done on these 3 columns. 1 1 1 4 0 2 Source SS DF MS F.05 H0 F d) The Levene test is a test for equal variances. The original data Between 8 2 4 3.27 ns F 2,9 4.26 Column means equal Within 11 9 1.2222 Total 19 11 Because we cannot reject the null hypothesis, we cannot say that the variances are not equal. 14 252y0631 11/28/06 (Page layout view!) 3) (Abronovic) A group of 4 workers produces defective pieces at the rates shown below during different times of the day. Personalize the data by subtracting the last digit of your student number from the 14 in the lower right corner. Use the number subtracted to label this as a version number. (Example: Good ol’ Seymour’s student number is 123456, so the 14 becomes 14 - 6 =8 and he labels it Version 6.) Time Worker’s Name Apple Plum Pear Melon Early 10 11 8 12 Morning Late 9 8 7 10 Morning Early 12 13 11 11 Afternoon Late 13 14 10 14 Afternoon Sum of Row 1 = 41, SSQ of Row 1 = 429, Sum of Column 1 = 44, SSQ of Column 1 = 494, Sum of Row 2 = 34, SSQ of Row 2 = 294, Sum of Column 2 = 46, SSQ of Column 2 = 550, Sum of Row 3 = 47, SSQ of Row 3 = 555, Sum of Column 3 = 36, SSQ of Column 3 = 334, a) Do a 2-way ANOVA on these data and explain what hypotheses you test and what the conclusions are. (6) b) Using your results from a) present two different confidence intervals for the difference between numbers of defects for the best and worst worker and for the defects from the best and second best times. Explain which of the intervals show a significant difference and why. (3) c) What other method could we use on these data to see if time of day makes a difference while allowing for cross-classification? Under what circumstances would we use it? Try it and tell what it tests and what it shows. (3) [36] a) Time Apple Plum Pear Melon 1 2 3 4 Sum nj 10 9 12 13 44 4 11 8 13 14 46 4 8 7 11 10 36 4 12 10 11 14 47 4 41 34 47 51 173 16 x j 11 11.5 9 11.75 (10.8125) x SS 494 121 550 132.25 334 81 561 138.0625 1939 472.3125 2 xijk x j 2 Sum ni SS x i 4 4 4 4 16 n 10.25 8.50 11.75 12.75 (10.8125) x 429 294 555 661 1939 2 xijk xi 2 105.0625 72.25 138.0625 162.5625 477.9375 x i 2 x .2j . x nx 19391610.8125 19391870.5625 68.4375 SSR C x nx 4477.9375 1610.8175 1911.75 1870.5625 41.1875 SSC R x nx 4472.3125 1610.8125 1889.25 1870.5625 18.6875 SST 2 2 2 2 i. 2 2 2 .j 2 2 Source SS DF MS F F.05 F 3,9 3.86 F 3,9 3.86 Rows 41.1875 3 13.7292 14.43 s Columns 18.6875 3 6.2292 6.55 s H0 Row means equal Column means equal Within 8.5625 9 0.95139 Total 68.4375 15 As is shown by the computed F statistics and the table values, both computed Fs exceed the table values, meaning that we will reject the null hypotheses. A version of this problem has appeared on almost every Final. Using the format above is necessary if you want to use sums computed for you. 15 252y0631 11/28/06 (Page layout view!) b) The material on confidence intervals comes from the outline for 2-way ANOVA. Note that for columns Pear is the best worker (9) and Melon the Worst (11.75), so that the difference is 2.75. For rows Time 2 is the best time (8.50) and Time 1 (10.25) is the second best, so that the difference is 1.75. R 1C 1 9 20.95139 2 MSW 0.6897 , C 4 not significant. ‘s’ stands for significant. 2MSW R MSW 0.95139 , 20.95139 0.6897 . ‘ns’ stands for 4 i. A Single Confidence Interval If we desire a single interval we use the formula for a Bonferroni Confidence Interval with m 1 . . Note that since P 1 , we must replace RC P 1 with R 1C 1 . For row means 1 2 x1 x 2 t R1C 1 2 1.75 2.262 0.6897 1.75 1.56 . s 2MSW . So we have C For column means 1 2 x1 x2 t R1C 1 2 2.75 2.262 0.6897 2.75 1.56 . s t.9025 2.262 2MSW . So we have R ii. Scheffé Confidence Interval If we desire intervals that will simultaneously be valid for a given confidence level for all possible intervals between means, use the following formulas. Note that since P 1 , we must replace RC P 1 with R 1C 1 . 3,9 3.86 F.05 For row means, use 1 2 x1 x 2 x1 x 2 R 1FR 1,R 1C 1 R 1FR 1,R 1C 1 2MSW 2MSW So we have 1.75 33.86 0.6897 1.75 2.34 ns. C For column means, use 1 2 x1 x2 x1 x2 C 1FC 1,R 1C 1 C C 1FC 1,R 1C 1 2MSW R 2MSW 2.75 2.34 s R iii. Bonferroni Confidence Interval – not worth the effort. iv. Tukey Confidence Interval Note that since P 1 , we must replace RC P 1 with For row means, use 1 2 x1 x 2 qR ,R 1C 1 = x1 x 2 0.5qR ,R 1C 1 R 1C 1 . MSW C 2MSW . So we have 1.75 0.70711 4.410.6897 1.75 2.15 . ns C For column means, use 1 2 x1 x2 qC ,R 1C 1 x1 x2 0.5qC ,R 1C 1 4,9 q.05 4.41 MSW R 2MSW 2.75 2.15 s R 16 252y0631 11/28/06 (Page layout view!) c) The alternative to 2-way ANOVA with one measurement per cell is a Friedman test. Remember that it is 10 11 8 12 9 8 7 10 , time of day was represented by rows. 12 13 11 11 13 14 10 14 10 9 12 13 11 8 13 14 If we transpose the array, columns represent time of day. . Now replace the original data 8 7 11 10 12 10 11 14 2 1 3 4 2 1 3 4 with rank within rows and sum the columns. 2 1 4 3 . As a check note that the column sums should rows that we want to compare. In the original data 3 9 1 2 4 4 12 15 445 40 and that 9 + 4 + 12 + 15 = 40. The Friedman formula reads 2 12 SRi2 3r c 1 rc c 1 i add to F2 12 92 42 12 2 15 2 345 3 466 60 69 .90 60 9.90 20 445 The Friedman table says that 9.90 has a p-value of .006. Since this is less than 5% or 1% we reject the null hypothesis that the median number of defects is the same regardless of time of day. 17