Chapter 24 Analysis of Variance

24.
1 Testing Whether the Means 24.2 The ANOVA Table 24.3 Plot the Data . . . 24.4 Comparing Means
24
c hapter
Analysis of Variance*
of Several Groups Are Equal
Where are we going?

In Chapter 4, we looked at the performance of four brands of mugs to see which was the most effective at keeping coffee hot. Are all these brands equally good? How can we compare them all? We could run a t -test for each of the 6 head-to-head comparisons, but well learn a better way to compare more than two groups in this chapter.
Who What How
Hand washings by four different methods, assigned randomly and replicated 8 times each Number of bacteria colonies Sterile media plates incubated at 36C for 2 days
Figure 24.1
Boxplots of the bacteria colony counts for the four different washing methods suggest some differences between treatments.
id you wash your hands with soap before eating? Youve undoubtedly been asked that question a few times in your life. Mom knows that washing with soap eliminates most of the germs youve managed to collect on your hands. Or does it? A student decided to investigate just how effective washing with soap is in eliminating bacteria. To do this she tested four different methodswashing with water only, washing with regular soap, washing with antibacterial soap (ABS), and spraying hands with antibacterial spray (AS) (containing 65% ethanol as an active ingredient). Her experiment consisted of one experimental factor, the washing Method, at four levels. She suspected that the number of bacteria on her hands before washing might vary considerably from day to day. To help even out the effects of those changes, she generated random numbers to determine the order of the four treatments. Each morning, she washed her hands according to the treatment randomly chosen. Then she placed her right hand on a sterile media plate designed to encourage bacteria growth. She incubated each plate for 2days at 36C, after which she counted the bacteria colonies. She replicated this procedure 8 times for each of the four treatments. A side-by-side boxplot of the numbers of colonies seems to show some differences among the treatments:
200 Bacteria (# of colonies) 150 100 50
AS
ABS
Soap Method
Water
Copyright 2014 Pearson Education, Inc.
701
702
Part VII Inference When Variables Are Related
When we first looked at a quantitative variable measured for each of several groups in Chapter 4, we displayed the data this way with side-by-side boxplots. And when we compared the boxes, we asked whether the centers seemed to differ, using the spreads of the boxes to judge the size of the differences. Now we want to make this more formal by testing a hypothesis. Well make the same kind of comparison, comparing the variability among the means with the spreads of the boxes. It looks like the alcohol spray has lower bacteria counts, but as always, were skeptical. Could it be that the four methods really have the same mean counts and we just happened to get a difference like this because of natural sampling variability? What is the null hypothesis here? It seems natural to start with the hypothesis that all the group means are equal. That would say it doesnt matter what method you use to wash your hands because the mean bacteria count will be the same. We know that even if there were no differences at all in the means (for example, if someone replaced all the solutions with water) there would still be sample-to-sample differences. We want to see, statistically, whether differences as large as those observed in the experiment could naturally occur by chance in groups that have equal means. If we find that the differences in washing Methods are so large that they would occur only very infrequently in groups that actually have the same mean, then, as weve done with other hypothesis tests, well reject the null hypothesis and conclude that the washing Methods really have different means.1
For Example
Contrast baths are a treatment commonly used in hand clinics to reduce swelling and stiffness after surgery. Patients hands are immersed alternately in warm and cool water. (Thats the contrast in the name.) Sometimes, the treatment is combined with mild exercise. Although the treatment is widely used, it had never been verified that it would accomplish the stated outcome goal of reducing swelling. Researchers2 randomly assigned 59 patients who had received carpal tunnel release surgery to one of three treatments: contrast bath, contrast bath with exercise, and (as a control) exercise alone. Hand therapists who did not know how the subjects had been treated measured hand volumes before and after treatments in milliliters by measuring how much water the hand displaced when submerged. The change in hand volume (after treatment minus before) was reported as the outcome.
Question: Specify the details of the experiments design. Identify the subjects, the
sample size, the experiment factor, the treatment levels, and the response. What is the null hypothesis? Was randomization employed? Was the experiment blinded? Was it double-blinded?
ANSWER: Subjects were patients who received carpal tunnel release surgery. Sample size is 59 patients. The factor was contrast bath treatment with three levels: contrast baths alone, contrast baths with exercise, and exercise alone. The response variable is the change in hand volume. The null hypothesis is that the mean changes in hand volume will be the same for the three treatment levels. Patients were randomly assigned to treatments. The study was single-blind because the evaluators were blind to the treatments. It was not (and could not be) double-blind because the patients had to be aware of their treatments.
The alternative hypothesis is that the means are not all equal. Be careful not to confuse that with all the means are different. With 11 groups we could have 10 means equal to each other and 1 different. The null hypothesis would still be false. 2 Janssen, Robert G., Schwartz, Deborah A., and Velleman, Paul F., A Randomized Controlled Study of Contrast Baths on Patients with Carpal Tunnel Syndrome, Journal of Hand Therapy, 22:3, pp. 200207. The data reported here differ slightly from those in the original paper because they include some additional subjects and exclude some outliers. Copyright 2014 Pearson Education, Inc.
Chapter 24 Analysis of Variance*
703
24.1 Testing Whether the Means of Several Groups Are Equal

We saw in Chapter 20 how to use a t-test to see whether two groups have equal means. We compared the difference in the means to a standard error estimated from all the data. And when we were willing to assume that the underlying group variances were equal, we pooled the data from the two groups to find the standard error. Now we have more groups, so we cant just look at differences in the means.3 But all is not lost. Even if the null hypothesis were true, and the means of the populations underlying the groups were equal, wed still expect the sample means to vary a bit. We could measure that variation by finding the variance of the means. How much should they vary? Well, if we look at how much the data themselves vary, we can get a good idea of how much the means should vary. And if the underlying means are actually different, wed expect that variation to be larger. It turns out that we can build a hypothesis test to check whether the variation in the means is bigger than wed expect it to be just from random fluctuations. Well need a new sampling distribution model, called the F-model, but thats just a different table to look at (Table F can be found at the end of this chapter). To get an idea of how it works, lets start by looking at the following two sets of boxplots:
80 40.0
60
37.5 35.0
40
20
32.5 30.0
Figure 24.2
Its hard to see the difference in the means in these boxplots because the spreads are large relative to the differences in the means.
Figure 24.3
In contrast with Figure 24.2, the smaller variation makes it much easier to see the differences among the group means. (Notice also that the scale of the y-axis is considerably different from the plot on the left.)
Were trying to decide if the means are different enough for us to reject the null hypothesis. If theyre close, well attribute the differences to natural sampling variability. What do you think? Its easy to see that the means in the second set differ. Its hard to imagine that the means could be that far apart just from natural sampling variability alone. How about the first set? It looks like these observations could have occurred from treatments with the same means.4 This much variation among groups does seem consistent with equal group means. Believe it or not, the two sets of treatment means in both figures are the same. (They are 31, 36, 38, and 31, respectively.) Then why do the figures look so different? In the second figure, the variation within each group is so small that the differences between the means stand out. This is what we looked for when we compared boxplots by eye back in Chapter 4. And its the central idea of the F-test. We compare the differences between the means of the groups with the variation within the groups. When the differences between means are large compared with the variation within the groups, we reject the null hypothesis and conclude
3
You might think of testing all pairs, but that method generates too many Type I errors. Well see more about this later in the chapter. 4 Of course, with a large enough sample, we can detect any differences that we like. For experiments with the same sample size, its easier to detect the differences when the variation within each box is smaller. Copyright 2014 Pearson Education, Inc.
704
that the means are not equal. In the first figure, the differences among the means look as though they could have arisen just from natural sampling variability from groups with equal means, so theres not enough evidence to reject H0. How can we make this comparison more precise statistically? All the tests weve seen have compared differences of some kind with a ruler based on an estimate of variation. And weve always done that by looking at the ratio of the statistic to that variation estimate. Here, the differences among the means will show up in the numerator, and the ruler we compare them with will be based on the underlying standard deviationthat is, on the variability within the treatment groups.
For Example
Recap: Fifty-nine postsurgery patients were randomly assigned to one of three treatment levels. Changes in hand volume were measured. Here are the boxplots. The recorded values are volume after treatmentvolume before treatment, so positive values indicate swelling. Some swelling is to be expected.
15.0
Hand Volume
7.5
0.0
27.5
Bath
Bath1Exercise Exercise Treatment
Question: What do the boxplots say about the results? ANSWER: There doesnt seem to be much difference between the two contrast bath treatments. The exercise only treatment may result in less swelling.
Why variances?
How Different Are They?

The challenge here is that we cant take a simple difference as we did when comparing two groups. In the hand-washing experiment, we have differences in mean bacteria counts across four treatments. How should we measure how different the four group means are? With only two groups, we naturally took the difference between their means as the numerator for the t-test. Its hard to imagine what else we could have done. How can we generalize that to more than two groups? When weve wanted to know how different many observations were, we measured how much they vary, and thats what we do here. How much natural variation should we expect among the means if the null hypothesis were true? If the null hypothesis were true, then each of the treatment means would estimate the same underlying mean. If the washing methods are all the same, its as if were just estimating the mean bacteria count on hands that have been washed with plain water. And we have several (in our experiment, four) different, independent estimates of this mean. Here comes the clever part. We can treat these estimated means as if they were observations and simply calculate their (sample) variance. This variance is the measure well use to assess how different the group means are from each other. Its the generalization of the difference between means for only two groups.
Weve usually measured variability with standard deviations. Standard deviations have the advantage that theyre in the same units as the data. Variances have the advantage that for independent variables, the variances add. Because were talking about sums of variables, well stay with variances before we get back to standard deviations.
705
Level Alcohol Spray Antibacterial Soap Soap Water
n 8 8 8 8
Mean 37.5 92.5 106.0 117.0
The more the group means resemble each other, the smaller this variance will be. The more they differ (perhaps because the treatments actually have an effect), the larger this variance will be. For the bacteria counts, the four means are listed in the table to the left. If you took those four values, treated them as observations, and found their sample variance, youd get 1245.08. Thats fine, but how can we tell whether it is a big value? Now we need a model, and the model is based on our null hypothesis that all the group means are equal. Here, the null hypothesis is that it doesnt matter what washing method you use; the mean bacteria count will be about the same: H0: m1 = m2 = m3 = m4 = m. As always when testing a null hypothesis, well start by assuming that it is true. And if the group means are equal, then theres an overall mean, mthe bacteria count youd expect all the time after washing your hands in the morning. And each of the observed group means is just a sample-based estimate of that underlying mean. We know how sample means vary. The variance of a sample mean is s2 > n. With eight observations in a group, that would be s2 > 8. The estimate that weve just calculated, 1245.08, should estimate this quantity. If we want to get back to the variance of the observations, s2, we need to multiply it by 8. So 8 * 1245.08 = 9960.64 should estimate s2. Is 9960.64 large for this variance? How can we tell? Well need a hypothesis test. You wont be surprised to learn that there is just such a test. The details of the test, due to Sir Ronald Fisher in the early 20th century, are truly ingenious, and may be the most amazing statistical result of that century.
The Ruler Within

We need a suitable ruler for comparisonone based on the underlying variability in our measurements. That variability is due to the day-to-day differences in the bacteria count even when the same soap is used. Why would those counts be different? Maybe the experimenters hands were not equally dirty, or she washed less well some days, or the plate incubation conditions varied. We randomized just so we could see past such things. We need an independent estimate of s2, one that doesnt depend on the null hypothesis being true, one that wont change if the groups have different means. As in many quests, the secret is to look within. We could look in any of the treatment groups and find its variance. But which one should we use? The answer is, all of them! At the start of the experiment (when we randomly assigned experimental units to treatment groups), the units were drawn randomly from the same pool, so each treatment group had a sample variance that estimated the same s2. If the null hypothesis is true, then not much has happened to the experimental unitsor at least, their means have not moved apart. Its not much of a stretch to believe that their variances havent moved apart much either. (If the washing methods are equivalent, then the choice of method would not affect the mean or the variability.) So each group variance still estimates a common s2. Were assuming that the null hypothesis is true. If the group variances are equal, then the common variance they all estimate is just what weve been looking for. Since all the group variances estimate the same s2, we can pool them to get an overall estimate of s2. Recall that we pooled to estimate variances when we tested the null hypothesis that two proportions were equaland for the same reason. Its also exactly what we did in a pooled t-test. The variance estimate we get by pooling well denote, as before, by s2 p. For the bacteria counts, the standard deviations and variances are listed below.
Level Alcohol Spray Antibacterial Soap Soap Water n 8 8 8 8 Mean 37.5 92.5 106.0 117.0 Std Dev 26.56 41.96 46.96 31.13 Variance 705.43 1760.64 2205.24 969.08
706
If we pool the four variances (here we can just average them because all the sample sizes are equal), wed get s2 p = 1410.10. In the pooled variance, each variance is taken around its own treatment mean, so the pooled estimate doesnt depend on the treatment means being equal. But the estimate in which we took the four means as observations and took their variance does. That estimate gave 9960.64. That seems a lot bigger than 1410.10. Might this be evidence that the four means are not equal? Lets see what we have. We have an estimate of s2 from the variation within groups of 1410.10. Thats just the variance of the residuals pooled across all groups. Because its a pooled variance, we could write it as s2 p. Traditionally this quantity is also called the Error Mean Square, or sometimes the Within Mean Square and denoted by MSE. These names date back to the early 20th century when the methods were developed. If you think about it, the names do make sensevariances are means of squared differences.5 But we also have a separate estimate of s2 from the variation between the groups because we know how much means ought to vary. For the hand-washing data, when we took the variance of the four means and multiplied it by n we got 9960.64. We expect this to estimate s2 too, as long as we assume the null hypothesis is true. We call this quantity the Treatment Mean Square (or sometimes the Between Mean Square6) and denote by MST.
The F-Statistic
Now we have two different estimates of the underlying variance. The first one, the MST, is based on the differences between the group means. If the group means are equal, as the null hypothesis asserts, it will estimate s2. But, if they are not, it will give some bigger value. The other estimate, the MSE, is based only on the variation within the groups around each of their own means, and doesnt depend at all on the null hypothesis being true. So, how do we test the null hypothesis? When the null hypothesis is true, the treatment means are equal, and both MSE and MST estimate s2. Their ratio, then, should be close to 1.0. But, when the null hypothesis is false, the MST will be larger because the treatment means are not equal. The MSE is a pooled estimate in which the variation within each group is found around its own group mean, so differing means wont inflate it. That makes the ratio MST > MSE perfect for testing the null hypothesis. When the null hypothesis is true, the ratio should be near 1. If the treatment means really are different, the numerator will tend to be larger than the denominator, and the ratio will tend to be bigger than 1. Of course, even when the null hypothesis is true, the ratio will vary around 1 just due to natural sampling variability. How can we tell when its big enough to reject the null hypothesis? To be able to tell, we need a sampling distribution model for the ratio. Sir Ronald Fisher found the sampling distribution model of the ratio in the early 20th century. In his honor, we call the distribution of MST > MSE the F-distribution. And we call the ratio MST > MSE the F-statistic. By comparing this statistic with the appropriate F-distribution we (or the computer) can get a P-value. The F-test is simple. It is one-tailed because any differences in the means make the F-statistic larger. Larger differences in the treatments effects lead to the means being more variable, making the MST bigger. That makes the F-ratio grow. So the test is significant if the F-ratio is big enough. In practice, we find a P-value, and big F-statistic values go with small P-values. The entire analysis is called the Analysis of Variance, commonly abbreviated ANOVA (and pronounced uh-NO-va). You might think that it should be called the analysis of means, since its the equality of the means were testing. But we use the variances within and between the groups for the test.
Well, actually, theyre sums of squared differences divided by their degrees of freedom : ( n- 1 ) for the first variance we saw back in Chapter 3, and other degrees of freedom for each of the others weve seen. But even back in Chapter 3, we said this was a kind of mean, and indeed, it still is. 6 Grammarians would probably insist on calling it the Among Mean Square, since the variation is among all the group means. Traditionally, though, its called the Between Mean Square and we have to talk about the variation between all the groups (as bad as that sounds).
5
Notation Alert
Capital F is used only for this distribution model and statistic. Fortunately, Fishers name didnt start with a Z, a T, or an R.
707
Notation Alert
What, first little n and now big N? In an experiment, its standard to use N for all the cases and n for the number in each treatment group.
Like Students t-models, the F-models are a family. F-models depend on not one, but two, degrees of freedom parameters. The degrees of freedom come from the two variance estimates and are sometimes called the numerator df and the denominator df. The Treatment Mean Square, MST, is the sample variance of the observed treatment means. If we think of them as observations, then since there are k groups, this variance has k - 1 degrees of freedom. The Error Mean Square, MSE, is the pooled estimate of the variance within the groups. If there are n observations in each group, then we get n - 1 degrees of freedom from each, for a total of k ( n - 1 ) degrees of freedom. A simpler way of tracking the degrees of freedom is to start with all the cases. Well call that N. Each group has its own mean, costing us a degree of freedomk in all. So we have N - k degrees of freedom for the error. When the groups all have equal sample size, thats the same as k ( n - 1 ) , but this way works even if the group sizes differ. We say that the F-statistic, MST > MSE, has k - 1 and N - k degrees of freedom.
Back to Bacteria
For the hand-washing experiment, the MST = 9960.64. The MSE = 1410.14. If the treatment means were equal, the Treatment Mean Square should be about the same size as the Error Mean Square, about 1410. But its 9960.64, which is 7.06 times bigger. In other words, F = 7.06. This F-statistic has 4 - 1 = 3 and 32 - 4 = 28 degrees of freedom. An F-value of 7.06 is bigger than 1, but we cant tell for sure whether its big enough to reject the null hypothesis until we check the F3,28 model to find its P-value. (Usually, thats most easily done with technology, but we can use printed tables.) It turns out the P-value is 0.0011. In other words, if the treatment means were actually equal, we would expect the ratio MST > MSE to be 7.06 or larger about 11 times out of 10,000, just from natural sampling variability. Thats not very likely, so we reject the null hypothesis and conclude that the means are different. We have strong evidence that the four different methods of hand washing are not equally effective at eliminating germs.
24.2 The ANOVA Table

Youll often see the mean squares and other information put into a table called the ANOVA table. Heres the table for the washing experiment: Analysis of Variance Table Source Method Error Total
Calculating the ANOVA table This table has a long tradition stretching back to when ANOVA calculations were done by hand. Major research labs had rooms full of mechanical calculators operated by women. (Yes, always women; women were thought by the men in charge, at least to be more careful at such an exacting task.) Three women would perform each calculation, and if any two of them agreed on the answer, it was taken as the correct value.
Sum of Squares 29882 39484 69366
DF 3 28 31
Mean Square 9960.64 1410.14
F-Ratio 7.0636
P-Value 0.0011
The ANOVA table was originally designed to organize the calculations. With technology, we have much less use for that. Well show how to calculate the sums of squares later in the chapter, but the most important quantities in the table are the F-statistic and its associated P-value. When the F-statistic is large, the Treatment (here Method) Mean Square is large compared to the Error Mean Square ( MSE ) , and provides evidence that in fact the means of the groups are not all equal. Youll almost always see ANOVA results presented in a table like this. After nearly a century of writing the table this way, statisticians (and their technology) arent going to change. Even though the table was designed to facilitate hand calculation, computer programs that compute ANOVAs still present the results in this form. Usually the P-value is found next to the F-ratio. The P-value column may be labeled with a title such as Prob > F, sig, or Prob. Dont let that confuse you; its just the P-value. Youll sometimes see the two mean squares referred to as the Mean Square Between and the Mean Square Withinespecially when we test data from observational studies rather than experiments. ANOVA is often used for such observational data, and as long as certain conditions are satisfied, theres no problem with using it in that context.
708
For Example
Recap: An experiment to determine the effect of contrast bath treatments on swelling in postsurgical patients recorded hand volume changes for patients who had been randomly assigned to one of three treatments. Here is the Analysis of Variance for these data:
Analysis of Variance for Hand Volume Change Source Treatment Error Total df 2 56 58 Sum of Squares 716.159 2704.38 3420.54 Mean Square 358.080 48.2926 F-Ratio 7.4148 P-Value 0.0014
Question: What does the ANOVA say about the results of the experiment? Specifically, what does it say about the null hypothesis? ANSWER: The F-ratio of 7.4148 has a P-value that is quite small. We can reject the null hypothesis that the mean change in hand volume is the same for all three treatments.
The F-Table
Usually, youll get the P-value for the F-statistic from technology. Any software program performing an ANOVA will automatically look up the appropriate one-sided P-value for the F-statistic. If you want to do it yourself, youll need an F-table. F-tables are usually printed only for a few values of a, often 0.05, 0.01, and 0.001. They give the critical value of the F-statistic with the appropriate number of degrees of freedom determined by your data, for the a level that you select. If your F-statistic is greater than that value, you know that its P-value is less than that a level. So, youll be able to tell whether the P-value is greater or less than 0.05, 0.01, or 0.001, but to be more precise, youll need technology (or an interactive table like the one in the ActivStats program on the DVD). Heres an excerpt from an F-table for a = 0.05:
Figure 24.4
Part of an F-table showing critical values for a = 0.05 and highlighting the critical value, 2.947, for 3 and 28 degrees of freedom. We can see that only 5% of the values will be greater than 2.947 with this combination of degrees of freedom.
2.947 0.05
0
1 24 2 3.385 3.403
df (numerator)
3 2.991 3.009 4 2.759 2.776
2
5 2.603 2.621 6 2.490 2.508 7
df (denominator)
25 26 27 28 29 30 31 32
4.260 4.242 4.225 4.210 4.196 4.183 4.171 4.160 4.149
3.305 3.295
3.316
3.328
3.340
3.354
3.369
2.911
2.922
2.934
2.947
2.960
2.975
2.901
2.679 2.668
2.690
2.701
2.714
2.728
2.743
2.523 2.512
2.534
2.545
2.558
2.572
2.587
2.409 2.399
2.421
2.432
2.445
2.459
2.474
2.323 2.313
2.334
2.346
2.359
2.388 2.373
2.405
2.423
Notice that the critical value for 3 and 28 degrees of freedom at a = 0.05 is 2.947. Since our F-statistic of 7.06 is larger than this critical value, we know that the P-value is less than 0.05. We could also look up the critical value for a = 0.01 and find that its 4.568 and the critical value for a = 0.001 is 7.193. So our F-statistic sits between the two critical values 0.01 and 0.001, and our P-value is slightly greater than 0.001. Technology can find the value precisely. It turns out to be 0.0011.
709
Just Checking
A student conducted an experiment to see which, if any, of four different paper airplane designs results in the longest flights (measured in inches). The boxplots look like this (with the overall mean shown in red):
250
Distance
200
150
100 A B Design C D
The ANOVA table shows:
Analysis of Variance Source Design Error Total

DF Sum of Squares Mean Square F-Ratio Prob + F 3 51991.778 17330.6 37.4255 6 0.0001 32 14818.222 463.1 35 66810.000
1. What is the null hypothesis? 2. From the boxplots, do you think that there is evidence that the mean flight
distances of the four designs differ?

3. Does the F-test in the ANOVA table support your preliminary conclusion in (2)? 4. The researcher concluded that there is substantial evidence that all four of the
designs result in different mean flight distances. Do you agree?
The ANOVA Model

To understand the ANOVA table, lets start by writing a model for what we observe. We start with the simplest interesting model: one that says that the only differences of interest among the groups are the differences in their means. This one-way ANOVA model characterizes each observation in terms of its mean and assume that any variation around that mean is just random error: yij = mj + eij. That is, each observation is the sum of the mean for the treatment it received plus a random error. Our null hypothesis is that the treatments made no differencethat is, that the means are all equal: H0: m1 = m2 = g = mk. It will help our discussion if we think of the overall mean of the experiment and consider the treatments as adding or subtracting an effect to this overall mean. Thinking in this way, we could write m for the overall mean and tj for the deviation from this mean
710
to get to the jth treatment meanthe effect of the treatment (if any) in moving that group away from the overall mean: yij = m + tj + eij. Thinking in terms of the effects, we could also write the null hypothesis in terms of these treatment effects instead of the means: H0: t1 = t2 = g = tk = 0. We now have three different kinds of parameters: the overall mean, the treatment effects, and the errors. Well want to estimate them from the data. Fortunately, we can do that in a straightforward way. To estimate the overall mean, m, we use the mean of all the observations: y (called the grand mean.7) To estimate each treatment effect, we find the difference between the mean of that particular treatment and the grand mean: n j = yj - y. t Theres an error, eij, for each observation. We estimate those with the residuals from the treatment means: eij = yij - yj. we can write each observation as the sum of three quantities that correspond to our model: yij = y +
( yj - y ) + ( yij - yj ) .
What this says is simply that we can write each observation as the sum of

the grand mean, the effect of the treatment it received, and the residual Or: Observations = Grand mean + Treatment effect + Residual. If we look at the equivalent equation yij = y +
( yj - y ) + ( yij - yj )
closely, it doesnt really seem like weve done anything. In fact, collecting terms on the right-hand side will give back just the observation, yij again. But this decomposition is actually the secret of the Analysis of Variance. Weve split each observation into sourcesthe grand mean, the treatment effect, and error.
Where does the residual term come from? Think of the annual report from any Fortune 500 company. The company spends billions of dollars each year and at the end of the year, the accountants show where each penny goes. How do they do it? After accounting for salaries, bonuses, supplies, taxes, etc., etc., etc., whats the last line? Its always labeled other or miscellaneous. Using other as the difference between all the sources they know and the total they start with, they can always make it add up perfectly. The residual is just the statisticians other. It takes care of all the other sources we didnt think of or dont want to consider, and makes the decomposition work by adding (or subtracting) back in just what we need.
The father of your father is your grandfather. The mean of the group means should probably be the grandmean, but we usually spell it as two words. Copyright 2014 Pearson Education, Inc.
711
Lets see what this looks like for our hand-washing data. Here are the data again, displayed a little differently:
Alcohol AB Soap 51 70 5 164 19 88 18 111 58 73 50 119 82 20 17 95 37.5 92.5 Soap Water 84 74 51 135 110 102 67 124 119 105 108 139 207 170 102 87 106 117
Treatment Means
The grand mean of all observations is 88.25. Lets put that into a similar table:
Alcohol AB Soap 88.25 88.25 88.25 88.25 88.25 88.25 88.25 88.25 88.25 88.25 88.25 88.25 88.25 88.25 88.25 88.25 Soap 88.25 88.25 88.25 88.25 88.25 88.25 88.25 88.25 Water 88.25 88.25 88.25 88.25 88.25 88.25 88.25 88.25
The treatment means are 37.5, 92.5, 106, and 117, respectively, so the treatment effects are those minus the grand mean (88.25). Lets put the treatment effects into their table:
Alcohol AB Soap - 50.75 4.25 - 50.75 4.25 - 50.75 4.25 - 50.75 4.25 - 50.75 4.25 - 50.75 4.25 - 50.75 4.25 - 50.75 4.25 Soap 17.75 17.75 17.75 17.75 17.75 17.75 17.75 17.75 Water 28.75 28.75 28.75 28.75 28.75 28.75 28.75 28.75
Finally, we compute the residuals as the differences between each observation and its treatment mean:
Alcohol AB Soap 13.5 - 22.5 71.5 - 32.5 - 4.5 - 18.5 18.5 - 19.5 20.5 - 19.5 12.5 26.5 44.5 - 72.5 2.5 - 20.5
Soap Water - 22 - 43 18 - 55 4 - 15 7 - 39 13 - 12 2 22 101 53 - 4 - 30
712
Now we have four tables for which Observations = Grand Mean + Treatment Effect + Residual. Why do we want to think in this way? Think back to the boxplots in Figures 24.2 and 24.3. To test the hypothesis that the treatment effects are zero, we want to see whether the treatment effects are large compared to the errors. Our eye looks at the variation between the treatment means and compares it to the variation within each group. The ANOVA separates those two quantities into the Treatment Effects and the Residuals. Sir Ronald Fishers insight was how to turn those quantities into a statistical test. We want to see if the Treatment Effects are large compared with the Residuals. To do that, we first compute the Sums of Squares of each table. Fishers insight was that dividing these sums of squares by their respective degrees of freedom lets us test their ratio by a distribution that he found (which was later named the F in his honor). When we divide a sum of squares by its degrees of freedom we get the associated mean square. When the Treatment Mean Square is large compared to the Error Mean Square, this provides evidence that the treatment means are different. And we can use the F-distribution to see how large large is. The sums of squares for each table are easy to calculate. Just take every value in the table, square it, and add them all up. For the Methods, the Treatment Sum of Squares, SST = ( - 50.75 ) 2 + ( - 50.75 ) 2 + g + ( 28.75 ) 2 = 29882. There are four treatments, and so there are 3 degrees of freedom. So, In general, we could write the Treatment Sum of Squares as MST = SST > 3 = 29882 > 3 = 9960.64
( You can verify, for example, that the first observation, 51 = 88.25 + ( - 50.75 ) + 13.5 ) .
SST = a a ( yj - y ) 2. Be careful to note that the summation is over the whole table, rows and columns. Thats why there are two summation signs. And, The table of residuals shows the variation that remains after we remove the overall mean and the treatment effects. These are whats left over after we account for what were interested inin this case the treatments. Their variance is the variance within each group that we see in the boxplots of the four groups. To find its value, we first compute the Error Sum of Squares, SSE, by summing up the squares of every element in the residuals table. To get the Mean Square (the variance) we have to divide it by N - k rather than by N - 1 because we found them by subtracting each of the k treatment means. So, SSE = ( 13.5 ) 2 + ( - 32.5 ) 2 + g + ( - 30 ) 2 = 39484 and As equations: and Now where are we? To test the null hypothesis that the treatment means are all equal we find the F-statistic: and compare that value to the F-distribution with k - 1 and N - k degrees of freedom. When the F-statistic is large enough (and its associated P-value small) we reject the null hypothesis and conclude that at least one mean is different.
MST = SST > ( k - 1 ) .
MSE = SSE > ( 32 - 4 ) = 1410.14. SSE = a a ( yij - yj ) 2, MSE = SSE > ( N - k ) .
Fk - 1, N - k = MST > MSE
713
Theres another amazing result hiding in these tables. If we take each of these tables, square every observation, and add them up, the sums add as well! SSObservations = SSGrand Mean + SST + SSE The SSObservations is usually very large compared to SST and SSE, so when ANOVA was originally done by hand, or even by calculator, it was hard to check the calculations using this fact. The first sum of squares was just too big. So, usually the ANOVA table uses the Corrected Total sum of squares. If we write Observations = Grand Mean + Treatment Effect + Residual, we can naturally write Observations - Grand Mean = Treatment Effect + Residual. Mathematically, this is the same statement, but numerically this is more stable. Whats amazing is that the sums of the squares still add up. That is, if you make the first table of observations with the grand mean subtracted from each, square those, and add them up, youll have the SSTotal and SSTotal = SST + SSE. Thats what the ANOVA table shows. If you find this surprising, you must be following along. The tables add up, so sums of their elements must add up. But it is not at all obvious that the sums of the squares of their elements should add up, and this is another great insight of the Analysis of Variance.
Back to Standard Deviations

Weve been using the variances because theyre easier to work with. But when its time to think about the data, wed really rather have a standard deviation because its in the units of the response variable. The natural standard deviation to think about is the standard deviation of the residuals. The variance of the residuals is staring us in the face. Its the MSE. All we have to do to get the residual standard deviation is take the square root of MSE: sp = 2MSE = ae . D (N - k)
2
The p subscript is to remind us that this is a pooled standard deviation, combining residuals across all k groups. The denominator in the fraction shows that finding a mean for each of the k groups cost us one degree of freedom for each. This standard deviation should feel right. That is, it should reflect the kind of variation you expect to find in any of the experimental groups. For the hand-washing data, sp = 11410.14 = 37.6 bacteria colonies. Looking back at the boxplots of the groups, we see that 37.6 seems to be a reasonable compromise standard deviation for all four groups.
24.3 Plot the Data

Just as you would never find a linear regression without looking at the scatterplot of y vs. x, you should never embark on an ANOVA without first examining side-by-side boxplots of the data comparing the responses for all of the groups. You already know what to look forwe talked about that back in Chapter 4. Check for outliers within any of the groups and correct them if there are errors in the data. Get an idea of whether the groups have similar spreads (as well need) and whether the centers seem to be alike (as the null hypothesis claims) or different. If the spreads of the groups are very differentand especially if they seem to grow consistently as the means growconsider re-expressing the response variable to make the spreads more nearly equal. Doing so is likely to make the
714
analysis more powerful and more correct. Likewise, if the boxplots are skewed in the same direction, you may be able to make the distributions more symmetric with a re-expression. Dont ever carry out an Analysis of Variance without looking at the side-by-side boxplots first. The chance of missing an important pattern or violation is just too great.
Assumptions and Conditions

When we checked assumptions and conditions for regression we had to take care to perform our checks in order. Here we have a similar concern. For regression we found that displays of the residuals were often a good way to check the corresponding conditions. Thats true for ANOVA as well.
Independence Assumptions The groups must be independent of each other. No

test can verify this assumption. You have to think about how the data were collected. The assumption would be violated, for example, if we measured subjects performance before some treatment, again in the middle of the treatment period, and then again at the end.8 The data within each treatment group must be independent as well. The data must be drawn independently and at random from a homogeneous population, or generated by a randomized comparative experiment. We check the Randomization Condition: Were the data collected with suitable randomization? For surveys, are the data drawn from each group a representative random sample of that group? For experiments, were the treatments assigned to the experimental units at random? We were told that the hand-washing experiment was randomized.
Equal Variance Assumption The ANOVA requires that the variances of the treatment groups be equal. After all, we need to find a pooled variance for the MSE. To check this assumption, we can check that the groups have similar variances: Similar Spread Condition: There are some ways to see whether the variation in the treatment groups seems roughly equal:

Look at side-by-side boxplots of the groups to see whether they have roughly the same spread. It can be easier to compare spreads across groups when they have the same center, so consider making side-by-side boxplots of the residuals. If the groups have differing spreads, it can make the pooled variancethe MSElarger, reducing the F-statistic value and making it less likely that we can reject the null hypothesis. So the ANOVA will usually fail on the safe side, rejecting H0 less often than it should. Because of this, we usually require the spreads to be quite different from each other before we become concerned about the condition failing. If youve rejected the null hypothesis, this is especially true. Look at the original boxplots of the response values again. In general, do the spreads seem to change systematically with the centers? One common pattern is for the boxes with bigger centers to have bigger spreads. This kind of systematic trend in the variances is more of a problem than random differences in spread among the groups and should not be ignored. Fortunately, such systematic violations are often helped by reexpressing the data. (If, in addition to spreads that grow with the centers, the boxplots are skewed with the longer tail stretching off to the high end, then the data are pleading for a re-expression. Try taking logs of the dependent variable for a start. Youll likely end up with a much cleaner analysis.) Look at the residuals plotted against the predicted values. Often, larger predicted values lead to larger magnitude residuals. This is another sign that the condition is violated. (This may remind you of the Does the Plot Thicken? Condition of
8 There is a modification of ANOVA, called repeated measures ANOVA, that deals with such data. (If the design reminds you of a paired-t situation, youre on the right track, and the lack of independence is the same kind of issue we discussed in Chapter 21.)
715
regression. And it should.) When the plot thickens (to one side or the other), its usually a good idea to consider re-expressing the response variable. Such a systematic change in the spread is a more serious violation of the equal variance assumption than slight variations of the spreads across groups. Lets check the conditions for the hand-washing data. Heres a boxplot of residuals by group and a scatterplot of residuals by predicted value:
Figure 24.5
Residuals (# of colonies) Residuals (# of colonies) AS AB Soap Soap Method Water Boxplots of residuals for the four washing methods and a plot of residuals vs. predicted values. Theres no evidence of a systematic change in variance from one group to the other or by predicted value. 80 40 0 40 80 40 0
40
50
75 100 Predicted (# of colonies)
Neither plot shows a violation of the condition. The IQRs (the box heights) are quite similar and the plot of residuals vs. predicted values does not show a pronounced widening to one end. The pooled estimate of 37.6 colonies for the error standard deviation seems reasonable for all four groups.
Normal Population Assumption Like Students t-tests, the F-test requires the underlying errors to follow a Normal model. As before when weve faced this assumption, well check a corresponding Nearly Normal Condition. Technically, we need to assume that the Normal model is reasonable for the populations underlying each treatment group. We can (and should) look at the side-by-side boxplots for indications of skewness. Certainly, if they are all (or mostly) skewed in the same direction, the Nearly Normal Condition fails (and re-expression is likely to help). In experiments, we often work with fairly small groups for each treatment, and its nearly impossible to assess whether the distribution of only six or eight numbers is Normal (though sometimes its so skewed or has such an extreme outlier that we can see that its not). Here we are saved by the Equal Variance Assumption (which weve already checked). The residuals have their group means subtracted, so the mean residual for each group is 0. If their variances are equal, we can group all the residuals together for the purpose of checking the Nearly Normal Condition. Check Normality with a histogram or a Normal probability plot of all the residuals together. The hand-washing residuals look nearly Normal in the Normal probability plot, although, as the boxplots showed, theres a possible outlier in the Soap group. Because we really care about the Normal model within each group, the Normal Population Assumption is violated if there are outliers in any of the groups. Check for outliers in the boxplots of the values for each treatment group. The Soap group of the hand-washing data shows an outlier, so we might want to compute the analysis again without that observation. (For these data, it turns out to make little difference.)
One-way ANOVA F-test We test the null hypothesis H0: m1 = m2 = g = mk against the alternative that the group means are not all equal. We test the hypothesis with the MST , where MST is the Treatment Mean Square, found from the variance of F-statistic, F = MSE the means of the treatment groups, and MSE is the Error Mean Square, found by pooling the variances within each of the treatment groups. If the F-statistic is large enough, we reject the null hypothesis.
Residuals (# of colonies)
80 40 0 40 1.25 0.00 1.25 Normal Scores
Figure 24.6
The hand-washing residuals look nearly Normal in this Normal probability plot.
716
Step-by-Step Example Analysis of Variance

In Chapter 4, we looked at side-by-side boxplots of four different containers for holding hot beverages. The experimenter wanted to know which type of container would keep his hot beverages hot longest. To test it, he heated water to a temperature of 180F, placed it in the container, and then measured the temperature of the water again 30 minutes later. He randomized the order of the trials and tested each container 8 times. His response variable was the difference in temperature (in F) between the initial water temperature and the temperature after 30 minutes. Question: Do the four containers maintain temperature equally well?
data.
Temperature Change (F)
Think Plot Plot the side-by-side boxplots of the
25 20 15 10 5 0 CUPPS Nissan SIGG Starbucks Container
Plan State what you want to know and the
null hypothesis you wish to test. For ANOVA, the null hypothesis is that all the treatment groups have the same mean. The alternative is that at least one mean is different.
I want to test whether there is any difference among the four containers in their ability to maintain the temperature of a hot liquid for 30 minutes. Ill write mk for the mean temperature difference for container k, so the null hypothesis is that these means are all the same: H0: m1 = m2 = m3 = m4. The alternative is that the group means are not all equal.
Think about the assumptions and check the conditions.
Randomization Condition: The experimenter performed the trials in random order, so its reasonable to assume that the performance of one tested cup is independent of other cups. Similar Spread Condition: The Nissan mug variation seems to be a bit smaller than the others. Ill look later at the plot of residuals vs. predicted values to see if the plot thickens.
SHOW Mechanics Fit the ANOVA model.

Source Container Error Total DF 3 28 31
Analysis of Variance Sum of Squares 714.1875 622.1875 1336.3750 Mean Square 238.063 22.221 F-Ratio 10.713 P-Value 6 0.0001
717
Nearly Normal Condition, Outlier Condition: The Normal probability plot is not very straight, but there are no outliers.
8 Residuals (F) 4 0 4 1.25 0.00 1.25 Normal Scores
The histogram shows that the distribution of the residuals is skewed to the right:
8 6 Counts 4 2
0 Residuals
The table of means and SDs (below) shows that the standard deviations grow along with the means. Possibly a re-expression of the data would improve matters. Under these circumstances, I cautiously find the P-value for the F-statistic from the F-model with 3 and 28 degrees of freedom. The ratio of the mean squares gives an F-ratio of 10.7134 with a P-value of 6 0.0001.
SHOW Show the table of means.
This seems like a reasonable value for the error standard deviation in the four treatments (with the possible exception of the Nissan mug).
Level CUPPS Nissan SIGG Starbucks n 8 8 8 8 Mean 10.1875 2.7500 16.0625 10.2500 Std Dev 5.20259 2.50713 5.90059 4.55129
From the ANOVA table, the Error Mean Square, MSE, is 22.22, which means that the standard deviation of all the errors is estimated to be 122.22 = 4.71 degrees F.
718
Interpretation Tell what the F-test TELL means.
An F-ratio this large would be very unlikely if the containers all had the same mean temperature difference.
Think State your conclusions.

(You should be more worried about the changing variance if you fail to reject the null hypothesis.) More specific conclusions might require a re-expression of the data.
Conclusions: Even though some of the conditions are mildly violated, I still conclude that the means are not all equal and that the four cups do not maintain temperature equally well.
The Balancing Act

The two examples weve looked at so far share a special feature. Each treatment group has the same number of experimental units. For the hand-washing experiment, each washing method was tested 8 times. For the cups, there were also 8 trials for each cup. This feature (the equal numbers of cases in each group, not the number 8) is called balance, and experiments that have equal numbers of experimental units in each treatment are said to be balanced or to have balanced designs. Balanced designs are a bit easier to analyze because the calculations are simpler, so we usually try for balance. But in the real world, we often encounter unbalanced data. Participants drop out or become unsuitable, plants die, or maybe we just cant find enough experimental units to fit a particular criterion. Everything weve done so far works just fine for unbalanced designs except that the calculations get a bit more complicated. Where once we could write n for the number of experimental units in a treatment, now we have to write nk and sum more carefully. Where once we could pool variances with a simple average, now we have to adjust for the different ns. Technology clears these hurdles easily, so youre safe thinking about the analysis in terms of the simpler balanced formulas and trusting that the technology will make the necessary adjustments.
For Example
Recap: An ANOVA for the contrast baths experiment had a statistically significant
F-value. Here are summary statistics for the three treatment groups: Group Bath Bath+ Exercise Exercise Count 22 23 14 Mean 4.54545 8 - 1.07143 StdDev 7.76271 7.03885 5.18080
Question: What can you conclude about these results? ANSWER: We can be confident that there is a difference. However, it is the exercise treatment that appears to reduce swelling and not the contrast bath treatments. We might conclude (as the researchers did) that contrast bath treatments are of limited value.
719
24.4 Comparing Means

When we reject H0, its natural to ask which means are different. No one would be happy with an experiment to test 10 cancer treatments that concluded only with We can reject H0the treatments are different! Wed like to know more, but the F-statistic doesnt offer that information. What can we do? If we cant reject the null, weve got to stop. Theres no point in further testing. If weve rejected the simple null hypothesis, however, we can do more. In particular, we can test whether any pairs or combinations of group means differ. For example, we might want to compare treatments against a control or a placebo, or against the current standard treatment. In the hand-washing experiment, we could consider plain water to be a control. Nobody would be impressed with (or want to pay for) a soap that did no better than water alone. A test of whether the antibacterial soap (for example) was different from plain water would be a simple test of the difference between two group means. To be able to perform an ANOVA, we first check the Similar Variance Condition. If things look OK, we assume that the variances are equal. If the variances are equal then a pooled t-test is appropriate. Even better (this is the special part), we already have a pooled estimate of the standard deviation based on all of the tested washing methods. Thats sp, which, for the hand-washing experiment, was equal to 37.55 bacteria colonies. The null hypothesis is that there is no difference between water and the antibacterial soap. As we did in Chapter 20, well write that as a hypothesis about the difference in the means: H0: mW - mABS = 0. The alternative is H0: mW - mABS 0. The natural test statistic is yW - yABS, and the (pooled) standard error is SE ( mW - mABS ) = sp
Level Alcohol Spray Antibacterial Soap Soap Water n 8 8 8 8 Mean Std Dev 37.5 26.56 92.5 41.96 106.0 46.96 117.0 31.13
The difference in the observed means is 117.0 - 92.5 = 24.5 colonies. The standard 24.5 error comes out to 18.775. The t-statistic, then, is t = = 1.31. To find the P-value 18.775 we consult the Students t-distribution on N - k = 32 - 4 = 28 degrees of freedom. The P-value is about 0.2not small enough to impress us. So we cant discern a significant difference between washing with the antibacterial soap and just using water. Our t-test asks about a simple difference. We could also ask a more complicated question about groups of differences. Does the average of the two soaps differ from the average of three sprays, for example? Complex combinations like these are called contrasts. Finding the standard errors for contrasts is straightforward but beyond the scope of this book. Well restrict our attention to the common question of comparing pairs of treatments after H0 has been rejected.
1 1 + . nABS B nW
*Bonferroni Multiple Comparisons

Our hand-washing experimenter was pretty sure that alcohol would kill the germs even before she started the experiment. But alcohol dries the skin and leaves an unpleasant smell. She was hoping that one of the antibacterial soaps would work as well as alcohol so she could use that instead. That means she really wanted to compare each of the other treatments against the alcohol spray. We know how to compare two of the means with a t-test. But now we want to do several tests, and each test poses the risk of a Type I error. As we do more and more tests, the risk that we might make a Type I error grows bigger than the a level of each individual test. With each additional test, the risk of making an error grows. If we do enough tests, were almost sure to reject one of the null hypotheses by mistakeand well never know which one. There is a defense against this problem. In fact, there are several defenses. As a class, they are called methods for multiple comparisons. All multiple comparisons methods
720
Carlo Bonferroni (18921960) was a mathematician who taught in Florence. He wrote two papers in 1935 and 1936 setting forth the mathematics behind the method that bears his name.
require that we first be able to reject the overall null hypothesis with the ANOVAs F-test. Once weve rejected the overall null, then we can think about comparing severalor even allpairs of group means. Lets look again at our test of the water treatment against the antibacterial soap treatment. This time well look at a confidence interval instead of the pooled t-test. We did a test at significance level a = 0.05. The corresponding confidence level is 1 - a = 95%. For any pair of means, a confidence interval for their difference is ( y1 - y2 ) { ME, where the margin of error is ME = t* * sp 1 1 + . n2 B n1
As we did in the previous section, we get sp as the pooled standard deviation found from all the groups in our analysis. Because sp uses the information about the standard deviation from all the groups its a better estimate than we would get by combining the standard deviation of just two of the groups. This uses the Equal Variance Assumption and borrows strength in estimating the common standard deviation of all the groups. We find the critical value t* from the Students t-model corresponding to the specified confidence level found with N - k degrees of freedom, and the nk s are the number of experimental units in each of the treatments. To reject the null hypothesis that the two group means are equal, the difference between them must be larger than the ME. That way 0 wont be in the confidence interval for the difference. When we use it in this way, we call the margin of error the least significant difference (LSD for short). If two group means differ by more than this amount, then they are significantly different at level a for each individual test. For our hand-washing experiment, each group has n = 8, sp = 37.55, and df = 32 - 4 = 28. From technology or Table T, we can find that t* with 28 df (for a 95% confidence interval) is 2.048. So LSD = 2.048 * 37.55 * 1 1 + = 38.45 colonies, B8 8
and we could use this margin of error to make a 95% confidence interval for any difference between group means. Any two washing methods whose means differ by more than 38.45 colonies could be said to differ at a = 0.05 by this method. Of course, were still just examining individual pairs. If we want to examine many pairs simultaneously, there are several methods that adjust the critical t*-value so that the resulting confidence intervals provide appropriate tests for all the pairs. And, in spite of making many such intervals, the overall Type I error rate stays at (or below) a. One such method is called the Bonferroni method. This method adjusts the LSD to allow for making many comparisons. The result is a wider margin of error called the minimum significant difference, or MSD. The MSD is found by replacing t* with a slightly larger number. That makes the confidence intervals wider for each contrast and the corresponding Type I error rates lower for each test. And it keeps the overall Type I error rate at or below a. The Bonferroni method distributes the error rate equally among the confidence intervals. It divides the error rate among J confidence intervals, finding each interval at a confidence level 1 - instead of the original 1 - a. To signal this adjustment, we label J the critical value t** rather than t*. For example, to make the three confidence intervals comparing the alcohol spray with the other three washing methods, and preserve our overall a risk at 5%, wed construct each with a confidence level of 1 0.05 = 1 - 0.01667 = 0.98333. 3
The only problem with this is that t-tables dont have a column for 98.33% confidence (or, correspondingly, for a = 0.01667). Fortunately, technology has no such constraints.
721
For the hand-washing data, if we want to examine the three confidence intervals comparing each of the other methods with the alcohol spray, the t**-value (on 28 degrees of freedom) turns out to be 2.546. Thats somewhat larger than the individual t*-value of 2.048 that we would have used for a single confidence interval. And the corresponding ME is 47.69 colonies (rather than 38.45 for a single comparison). The larger critical value along with correspondingly wider intervals is the price we pay for making multiple comparisons. Many statistics packages assume that youd like to compare all pairs of means. Some will display the result of these comparisons in a table like this:
Level Alcohol Spray Antibacterial Soap Soap Water n 8 8 8 8 Mean Groups 37.5 A 92.5 B 106.0 B 117.0 B
This table shows that the alcohol spray is in a class by itself and that the other three hand-washing methods are indistinguishable from one another.
ANOVA on Observational Data

So far weve applied ANOVA only to data from designed experiments. Thats natural for several reasons. The primary one is that, as we saw in Chapter 11, randomized comparative experiments are specifically designed to compare the results for different treatments. The overall null hypothesis, and the subsequent tests on pairs of treatments in ANOVA, address such comparisons directly. In addition, as we discussed earlier, the Equal Variance Assumption (which we need for all of the ANOVA analyses) is often plausible in a randomized experiment because the treatment groups start out with sample variances that all estimate the same underlying variance of the collection of experimental units. Sometimes, though, we just cant perform an experiment. When ANOVA is used to test equality of group means from observational data, theres no a priori reason to think the group variances might be equal at all. Even if the null hypothesis of equal means were true, the groups might easily have different variances. But if the side-by-side boxplots of responses for each group show roughly equal spreads and symmetric, outlier-free distributions, you can use ANOVA on observational data. Observational data tend to be messier than experimental data. They are much more likely to be unbalanced. If you arent assigning subjects to treatment groups, its harder to guarantee the same number of subjects in each group. And because you are not controlling conditions as you would in an experiment, things tend to be, well, less controlled. The only way we know to avoid the effects of possible lurking variables is with control and randomized assignment to treatment groups, and for observational data, we have neither. ANOVA is often applied to observational data when an experiment would be impossible or unethical. (We cant randomly break some subjects legs, but we can compare pain perception among those with broken legs, those with sprained ankles, and those with stubbed toes by collecting data on subjects who have already suffered those injuries.) In such data, subjects are already in groups, but not by random assignment. Be careful; if you have not assigned subjects to treatments randomly, you cant draw causal conclusions even when the F-test is significant. You have no way to control for lurking variables or confounding, so you cant be sure whether any differences you see among groups are due to the grouping variable or to some other unobserved variable that may be related to the grouping variable. Because observational studies often are intended to estimate parameters, there is a temptation to use pooled confidence intervals for the group means for this purpose. Although these confidence intervals are statistically correct, be sure to think carefully about the population that the inference is about. The relatively few subjects that happen to be in a group may not be a simple random sample of any interesting population, so their true mean may have only limited meaning.
722
Step-by-Step Example One More Example

Heres an example that exhibits many of the features weve been discussing. It gives a fair idea of the kinds of challenges often raised by real data. A study at a liberal arts college attempted to find out who watches more TV at college. Men or women? Varsity athletes or non-athletes? Student researchers asked 200 randomly selected students questions about their backgrounds and about their television-viewing habits and received 197 legitimate responses. The researchers found that men watch, on average, about 2.5 hours per week more TV than women, and that varsity athletes watch about 3.5 hours per week more than those who are not varsity athletes. But is this the whole story? To investigate further, they divided the students into four groups: male athletes (MA), male non-athletes (MNA), female athletes (FA), and female non-athletes (FNA). Question: Do these four groups of students spend about the same amount of time watching TV?
Think Variables
Name the variables, report the Ws, and specify the questions of interest.
I have the number of hours spent watching TV in a week for 197 randomly selected students. We know their sex and whether they are varsity athletes or not. I wonder whether TV watching differs according to sex and athletic status. Here are the side-by-side boxplots of the data:
25 20 TV Time (hrs) 15 10 5 0 FNA FA MNA MA / /
Plot Always start an ANOVA with side-byside boxplots of the responses in each of the groups. Always.
These data offer a good example why.
This plot suggests problems with the data. Each box shows a distribution skewed to the high end, and outliers pepper the display, including some extreme outliers. The box with the highest center (MA) also has the largest spread. These data just dont pass our first screening for suitability. This sort of pattern calls for a re-expression.
The responses are countsnumbers of TV hours. You may recall from Chapter 6 that a good re-expression to try first for counts is the square root.
Here are the boxplots for the square root of TV hours.

5
TV Time (hrs)
4 3 2 1 0 FNA FA MNA MA
723
The spreads in the four groups are now more similar and the individual distributions more symmetric. And now there are no outliers.
Think about the assumptions and check the conditions.
Randomization Condition: The data come from a random sample of students. The responses should be independent. It might be a good idea to see if the number of athletes and men are representative of the campus population. Similar Spread Condition: The boxplots show similar spreads. I may want to check the residuals later.
Fit the ANOVA model.
The ANOVA table looks like this:

Source Group Error Total DF 3 193 196 Sum of Squares 47.24733 237.26114 284.50847 Mean Square 15.7491 1.2293 F-Ratio 12.8111 P-Value 6 0.0001
Nearly Normal Condition, Outlier Condition: A histogram of the residuals looks reasonably Normal:
60
Counts
40
20
3.00
0.00 Residuals
3.00
Interestingly, the few cases that seem to stick out on the low end are male athletes who watched no TV, making them different from all the other male athletes. Under these conditions, its appropriate to use Analysis of Variance.
Tell Interpretation
The F-statistic is large and the corresponding P-value small. I conclude that the TV-watching behavior is not the same among these groups.
724
*So Do Male Athletes Watch More TV?

Heres a Bonferroni comparison of all pairs of groups:
Differing Standard Errors?
In case you were wondering The standard errors are different because this isnt a balanced design. Differing numbers of experimental units in the groups generate differing standard errors.
FAFNA MNAFNA MNAFA MAFNA MAFA MAMNA
Difference 0.049 0.205 0.156 1.497 1.449 1.292
Std. Err. 0.270 0.182 0.268 0.250 0.318 0.248
P-Value 0.9999 0.8383 0.9929 6 0.0001 6 0.0001 6 0.0001
Three of the differences are very significant. It seems that among women theres little difference in TV watching between varsity athletes and others. Among men, though, the corresponding difference is large. And among varsity athletes, men watch significantly more TV than women. But wait. How far can we extend the inference that male athletes watch more TV than other groups? The data came from a random sample of students made during the week of March 21. If the students carried out the survey correctly using a simple random sample, we should be able to make the inference that the generalization is true for the entire student body during that week. Is it true for other colleges? Is it true throughout the year? The students conducting the survey followed up the survey by collecting anecdotal information about TV watching of male athletes. It turned out that during the week of the survey, the NCAA mens basketball tournament was televised. This could explain the increase in TV watching for the male athletes. It could be that the increase extends to other students at other times, but we dont know that. Always be cautious in drawing conclusions too broadly. Dont generalize from one population to another.
What Can Go Wrong?

Watch out for outliers. One outlier in a group can change both the mean and the spread of that group. It will also inflate the Error Mean Square, which can influence the F-test. The good news is that ANOVA fails on the safe side by losing power when there are outliers. That is, you are less likely to reject the overall null hypothesis if you have (and leave) outliers in your data. But they are not likely to cause you to make a Type I error. Watch out for changing variances. The conclusions of the ANOVA depend crucially on the assumptions of independence and constant variance, and (somewhat less seriously as n increases) on Normality. If the conditions on the residuals are violated, it may be necessary to re-express the response variable to approximate these conditions more closely. ANOVA benefits so greatly from a judiciously chosen re-expression that the choice of a re-expression might be considered a standard part of the analysis. Be wary of drawing conclusions about causality from observational studies. ANOVA is often applied to data from randomized experiments for which causal conclusions are appropriate. If the data are not from a designed experiment, however, the Analysis of Variance provides no more evidence for causality than any other method we have studied. Dont get into the habit of assuming that ANOVA results have causal interpretations. Be wary of generalizing to situations other than the one at hand. Think hard about how the data were generated to understand the breadth of conclusions you are entitled to draw. Watch for multiple comparisons. When rejecting the null hypothesis, you can conclude that the means are not all equal. But you cant start comparing every pair of treatments in your study with a t-test. Youll run the risk of inflating your Type I error rate. Use a multiple comparisons method when you want to test many pairs.
725
Connections
We first learned about side-by-side boxplots in Chapter 4. There we made general statements about the shape, center, and spread of each group. When we compared groups, we asked whether their centers looked different compared with how spread out the distributions were. Now weve made that kind of thinking precise. Weve added confidence intervals for the difference and tests of whether the means are the same. We pooled data to find a standard deviation when we tested the hypothesis of equal proportions. For that test, the assumption of equal variances was a consequence of the null hypothesis that the proportions were equal, so it didnt require an extra assumption. Means dont have a linkage with their corresponding variances, so to use pooled methods we must make the additional assumption of equal variances. In a randomized experiment, thats a plausible assumption. Chapter 11 offered a variety of designs for randomized comparative experiments. Each of those designs can be analyzed with a variant or extension of the ANOVA methods discussed in this chapter. Entire books and courses deal with these extensions, but all follow the same fundamental ideas presented here. ANOVA is closely related to the regression analyses we saw in Chapter 23. (In fact, most statistics packages offer an ANOVA table as part of their regression output.) The assumptions are similarand for good reason. The analyses are, in fact, related at a deep conceptual (and computational) level, but those details are beyond the scope of this book. The pooled two-sample t-test for means is a special case of the ANOVA F-test. If you perform an ANOVA comparing only two groups, youll find that the P-value of the F-statistic is exactly the same as the P-value of the corresponding pooled t-statistic. Thats because in this special case the F-statistic is just the square of the t-statistic. The F-test is more general. It can test the hypothesis that several group means are equal.
What Have We Learned?

Learning Objectives
Recognize when to use an Analysis of Variance (ANOVA) to compare the means of several groups. Know how to read an ANOVA table to locate the degrees of freedom, the Mean Squares, and the resulting F-statistic. Know how to check the three conditions required for an ANOVA: Independence of the groups from each other and of the individual cases within each group. Equal variance of the groups. Normal error distribution. Know how to create and interpret confidence intervals for the differences between each pair of group means, recognizing the need to adjust the confidence interval for the number of comparisons made.
Review of Terms
Error (or Within) Mean Square (MSE) Treatment (or Between) Mean Square (MST) The Error Mean Square (MSE) is the estimate of the error variance obtained by pooling the variances of each treatment group. The square root of the (MSE) is the estimate of the error standard deviation, sp (p. 706). The Treatment Mean Square (MST) is the estimate of the error variance under the assumption that the treatment means are all equal. If the (null) assumption is not true, the MST will be larger than the error variance (p. 706).
726
F-distribution
The F-distribution is the sampling distribution of the F-statistic when the null hypothesis that the treatment means are equal is true. It has two degrees of freedom parameters, one for the numerator, ( k - 1 ) , and one for the denominator, ( N - k ) , where N is the total number of observations and k is the number of groups (p. 706). The F-statistic is the ratio MST > MSE. When the F-statistic is sufficiently large, we reject the null hypothesis that the group means are equal (p. 706).
F-statistic F-test
The F-test tests the null hypothesis that all the group means are equal against the onesided alternative that they are not all equal. We reject the hypothesis of equal means if the F-statistic exceeds the critical value from the F-distribution corresponding to the specified significance level and degrees of freedom (p. 706). An analysis method for testing equality of means across treatment groups (p. 706). The ANOVA table is convenient for showing the degrees of freedom, the Treatment Mean Square, the Error Mean Square, their ratio, the F-statistic, and its P-value. There are usually other quantities of lesser interest included as well (p. 707). The model for a one-way (one response, one factor) ANOVA is yij = mj + eij. nij = y j and residuals eij = yij - y j Estimating with yij = y j + eij gives predicted values y (p. 709).
ANOVA ANOVA table
One-way ANOVA model
Residual standard deviation
The residual standard deviation, sp = 2MSE =
gives an idea of the underlying variability of the response values (p. 703). Balance An experiments design is balanced if each treatment level has the same number of experimental units. Balanced designs make calculations simpler and are generally more powerful (p. 718). If we reject the null hypothesis of equal means, we often then want to investigate further and compare pairs of treatment group means to see if they differ. If we want to test several such pairs, we must adjust for performing several tests to keep the overall risk of a Type I error from growing too large. Such adjustments are called methods for multiple comparisons (p. 719). The standard margin of error in the confidence interval for the difference of two means is called the least significant difference. It has the correct Type I error rate for a single test, but not when performing more than one comparison (p. 720). One of many methods for adjusting the length of the margin of error when testing the differences between several group means (p. 720). The Bonferroni methods margin of error for the confidence interval for the difference of two group means is called the minimum significant difference. This can be used to test differences of several pairs of group means. If their difference exceeds the MSD, they are different at the overall rate (p. 720).
ae , DN - k
2
Methods for multiple comparisons
Least significant difference (LSD) *Bonferroni method Minimum significant difference (MSD)
727
On the Computer ANOVA

Most analyses of variance are found with computers. And all statistics packages present the results in an ANOVA table much like the one we discussed. Technology also makes it easy to examine the side-by-side boxplots and check the residuals for violations of the assumptions and conditions. Statistics packages offer different choices among possible multiple comparisons methods (although Bonferroni is quite common). This is a specialized area. Get advice or read further if you need to choose a multiple comparisons method. As we saw in Chapter 4, there are two ways to organize data recorded for several groups. We can put all the response values in a single variable and use a second, factor, variable to hold the group identities. This is sometimes called stacked format. The alternative is to place the data for each group in its own column or variable. Then the variable identities become the group identifiers. Most statistics packages expect the data to be in stacked format because this form also works for more complicated experimental designs. Some packages can work with either form, and some use one form for some things and the other for others. (Be careful, for example, when you make side-by-side boxplots; be sure to give the appropriate version of the command to correspond to the structure of your data.) Most packages offer to save residuals and predicted values and make them available for further tests of conditions. In some packages you may have to request them specifically.
Data Desk
Select
the response variable as Y and the factor variable as X. the Calc menu, choose ANOVA. Desk displays the ANOVA table.
Comments
Data Desk expects data in stacked format. You can change the ANOVA by dragging the icon of another variable over either the Y or X variable name in the table and dropping it there. The analysis will recompute automatically.
From Data
Select
plots of residuals from the ANOVA tables HyperView menu.
Excel
In
Excel 2003 and earlier, select Data Analysis from the Tools menu. Excel 2007 , select Data Analysis from the Analysis Group on the Data Tab. Anova Single Factor from the list of analysis tools. the OK button. the data range in the box provided. the Labels in First Row box, if applicable. an alpha level for the F-test in the box provided.
Comments
The data range should include two or more columns of data to compare. Unlike all other statistics packages, Excel expects each column of the data to represent a different level of the factor. However, it offers no way to label these levels. The columns need not have the same number of data values, but the selected cells must make up a rectangle large enough to hold the column with the most data values.
In
Select Click
Enter
Check Enter Click
the OK button.
JMP
From
the Analyze menu select Fit Y by X.
From
Select
variables: a quantitative Y, Response variable, and a categorical X, Factor variable. opens the Oneway window. on the red triangle beside the heading, select Display Options, and choose Boxplots.
the same menu choose the Means/ANOVA.t-test command. opens the oneway ANOVA output.
JMP
JMP Click
Comments
JMP expects data in stacked format with one response and one factor variable.
728
MINITAB
Choose Choose In
ANOVA from the Stat menu. One-way . . . from the ANOVA submenu.
Click Click
the OK button to return to the ANOVA dialog. the OK button to compute the ANOVA.
the One-way Anova dialog, assign a quantitative Y variable to the Response box and assign a categorical X variable to the Factor box. the Store Residuals check box. the Graphs button.
Comments
If your data are in unstacked format, with separate columns for each treatment level, choose One-way (unstacked) from the ANOVA submenu.
Check Click In
the ANOVA-Graphs dialog, select Standardized residuals, and check Normal plot of residuals and Residuals versus fits.
R
To perform an analysis of variance of a variable y on a categorical variable (factor) x:
myaov
To get confidence or prediction intervals use:

predict(myaov,interval
= aov(yx) summary(myaov) # gives the ANOVA table
= confidence)
or
predict(myaov,
interval = prediction)
SPSS
Choose Choose
Compare Means from the Analyze menu.
One-way ANOVA from the Compare Means submenu. the One-Way ANOVA dialog, select the Y-variable and move it to the dependent target. Then move the X-variable to the independent target. the OK button.
Comments
SPSS expects data in stacked format. The Contrasts and Post Hoc buttons offer ways to test contrasts and perform multiple comparisons. See your SPSS manual for details.
In
Click
Statcrunch
To compute an ANOVA:
Click
on Stat. ANOVA One Way.
Choose the single column containing the data (Responses) and the column containing the Factors.
Click
Choose Choose
on Calculate.
the Columns for all groups. (After the first one, you may hold down the ctrl or command key to choose more.) OR
Exercises
Section 24.1
1. Popcorn A student runs an experiment to test four different popcorn brands, recording the number of kernels left unpopped. She pops measured batches of each brand 4 times, using the same popcorn popper and randomizing the order of the brands. After collecting her data and analyzing the results, she reports that the F-ratio is 13.56. a) What are the null and alternative hypotheses? b) How many degrees of freedom does the treatment sum of squares have? How about the error sum of squares?
729
c) Assuming that the conditions required for ANOVA are satisfied, what is the P-value? What would you conclude? d) What else about the data would you like to see in order to check the assumptions and conditions?
measured how many seconds it took for the same amount of dough to rise to the top of a bowl. He randomized the order of the recipes and replicated each treatment 4 times. Here are the boxplots of activation times from the four recipes:
800 700 Activation Times (sec) 600 500 400 300 200 100 A B Recipe C D
2. Skating A figure skater tried various approaches to her Salchow jump in a designed experiment using 5 different places for her focus (arms, free leg, midsection, takeoff leg, and free). She tried each jump 6 times in random order, using two of her skating partners to judge the jumps on a scale from 0 to 6. After collecting the data and analyzing the results, she reports that the F-ratio is 7.43. a) What are the null and alternative hypotheses? b) How many degrees of freedom does the treatment sum of squares have? How about the error sum of squares? c) Assuming that the conditions are satisfied, what is the P-value? What would you conclude? d) What else about the data would you like to see in order to check the assumptions and conditions? 3. Gas mileage A student runs an experiment to study the effect of three different mufflers on gas mileage. He devises a system so that his Jeep Wagoneer uses gasoline from a one-liter container. He tests each muffler 8 times, carefully recording the number of miles he can go in his Jeep Wagoneer on one liter of gas. After analyzing his data, he reports that the F-ratio is 2.35 with a P-value of 0.1199. a) What are the null and alternative hypotheses? b) How many degrees of freedom does the treatment sum of squares have? How about the error sum of squares? c) What would you conclude? d) What else about the data would you like to see in order to check the assumptions and conditions? e) If your conclusion in part c is wrong, what type of error have you made? 4. Darts A student interested in improving her dart-throwing technique designs an experiment to test 4 different stances to see whether they affect her accuracy. After warming up for several minutes, she randomizes the order of the 4 stances, throws a dart at a target using each stance, and measures the distance of the dart in centimeters from the center of the bulls-eye. She replicates this procedure 10 times. After analyzing the data she reports that the F-ratio is 1.41. a) What are the null and alternative hypotheses? b) How many degrees of freedom does the treatment sum of squares have? How about the error sum of squares? c) What would you conclude? d) What else about the data would you like to see in order to check the assumptions and conditions? e) If your conclusion in part c is wrong, what type of error have you made?
T
The ANOVA table follows:

Source Recipe Error Total DF 3 12 15 Sum of Squares 638967.69 57128.25 696095.94 Mean Square 212989 4761 F-Ratio 44.7392 P-Value 6 0.0001
a) State the hypotheses about the recipes (both numerically and in words). b) Assuming that the assumptions for inference are satisfied, perform the hypothesis test and state your conclusion. Be sure to state it in terms of activation times and recipes. c) Would it be appropriate to follow up this study with multiple comparisons to see which recipes differ in their mean activation times? Explain.
6. Frisbee throws A student performed an experiment with three different grips to see what effect it might have on the distance of a backhanded Frisbee throw. She tried it with her normal grip, with one finger out, and with the Frisbee inverted. She measured in paces how far her throw went. The boxplots and the ANOVA table for the three grips are shown below:
40
Distance (paces)
35
30
Section 24.2
T
25 Finger Out Inverted Grip Normal
5. Activating baking yeast To shorten the time it takes him to make his favorite pizza, a student designed an experiment to test the effect of sugar and milk on the activation times for baking yeast. Specifically, he tested four different recipes and
(continued)
730
Source Grip Error Total
DF 2 21 23
Sum of Squares 58.58333 300.75000 359.33333
Mean Square 29.2917 14.3214

F-Ratio 2.0453 P-Value 0.1543
T
b) Do the conditions for an ANOVA seem to be met here? Why or why not?
Section 24.4
9. Tellers A bank is studying the time that it takes 6 of its tellers to serve an average customer. Customers line up in the queue and then go to the next available teller. Here is a boxplot of the last 200 customers and the times it took each teller:
120 Time (min) 90 60 30 1 2 3 4 Teller 5 6
a) State the hypotheses about the grips. b) Assuming that the assumptions for inference are satisfied, perform the hypothesis test and state your conclusion. Be sure to state it in terms of Frisbee grips and distance thrown. c) Would it be appropriate to follow up this study with multiple comparisons to see which grips differ in their mean distance thrown? Explain.
Section 24.3
T
7. Fuel economy Here are boxplots that show the relationship between the number of cylinders a cars engine has and the cars fuel economy.
35
Miles per Gallon
30
Source Teller Error Total
DF 5 134 139
ANOVA Table Sum of Mean Squares Square 3315.32 663.064 58919.1 439.695 62234.4
F-Ratio 1.508
P-Value 0.1914
25
20 T 4 6 5 Cylinders 8
a) What are the null and alternative hypotheses? b) What do you conclude? c) Would it be appropriate to run a multiple comparisons test (for example, a Bonferroni test) to see which tellers differ from each other? Explain.
a) State the null and alternative hypotheses that you might consider for these data. b) Do the conditions for an ANOVA seem to be met here? Why or why not?
10. Hearing A researcher investigated four different word lists for use in hearing assessment. She wanted to know whether the lists were equally difficult to understand in the presence of a noisy background. To find out, she tested 96 subjects with normal hearing randomly assigning 24 to each of the four word lists and measured the number of words perceived correctly in the presence of background noise. Here are the boxplots of the four lists:
50 45 40 35 30 25 20 15 10 1 2 List 3 4 Hearing (% of words)
8. Finger Lakes Wines Here are case prices (in dollars) of wines produced by wineries along three of the Finger Lakes.
150 125 100 75
Case Price ($)
ANOVA Table
Cayuga Keuka Location Seneca
a) What null and alternative hypotheses would you test for these data? Talk about prices and location, not symbols.
Source List Error Total
DF 3 92 95
Sum of Squares 920.4583 5738.1667 6658.6250
Mean Square 306.819 62.371
F-Ratio 4.9192
P-Value 0.0033
731
a) What are the null and alternative hypotheses? b) What do you conclude? c) Would it be appropriate to run a multiple comparisons test (for example, a Bonferroni test) to see which lists differ from each other in terms of mean percent correct? Explain.
peoples ZIP codes vary by the last product they bought. They have 16 different products, and the ANOVA table of ZIP code by product showed the following:
Source Product Error Total DF 15 475 490 Sum of Squares 3.836e10 2.45787e11 2.84147e11 Mean Square 2.55734e9 517445573 F-Ratio 4.9422 P-Value 6 0.0001
Chapter Exercises
11. Eye and hair color In Chapter 4, Exercise 32, we saw a survey of 1021 school-age children conducted by randomly selecting children from several large urban elementary schools. Two of the questions concerned eye and hair color. In the survey, the following codes were used:
Hair Color 1 = Blond 2 = Brown 3 = Black 4 = Red 5 = Other Eye Color 1 = Blue 2 = Green 3 = Brown 4 = Grey 5 = Other
(Nine customers were not included because of missing ZIP code or product information.) What criticisms of the analysis might you make? What alternative analysis might you suggest? 13. Yogurt An experiment to determine the effect of several methods of preparing cultures for use in commercial yogurt was conducted by a food science research group. Three batches of yogurt were prepared using each of three methods: traditional, ultrafiltration, and reverse osmosis. A trained expert then tasted each of the 9 samples, presented in random order, and judged them on a scale from 1 to 10. A partially completed Analysis of Variance table of the data follows.
Source Treatment Residual Total Sum of Squares 17.300 0.460 17.769 Degrees of Freedom Mean Square F-Ratio
The students analyzing the data were asked to study the relationship between eye and hair color. They produced this plot:
5 4 Eye Color 3 2 1 0 1 2 3 Hair Color 4 5
They then ran an Analysis of Variance with Eye Color as the response and Hair Color as the factor. The ANOVA table they produced follows:
Source Hair Color Error Total DF 4 1016 1020 Sum of Squares 1.46946 927.45317 928.92263 Mean Square 0.367365 0.912848 F-Ratio 0.4024 P-Value 0.8070
a) Calculate the mean square of the treatments and the mean square of the error. b) Form the F-statistic by dividing the two mean squares. c) The P-value of this F-statistic turns out to be 0.000017. What does this say about the null hypothesis of equal means? d) What assumptions have you made in order to answer part c? e) What would you like to see in order to justify the conclusions of the F-test? f) What is the average size of the error standard deviation in the judges assessment?
What suggestions do you have for the Statistics students? What alternative analysis might you suggest? 12. ZIP codes, revisited The intern from the marketing de partment at the Holes R Us online piercing salon (Chapter 3, Exercise 55) has recently finished a study of the companys 500 customers. He wanted to know whether
14. Smokestack scrubbers Particulate matter is a serious form of air pollution often arising from industrial production. One way to reduce the pollution is to put a filter, or scrubber, at the end of the smokestack to trap the particulates. An experiment to determine which smokestack scrubber design is best was run by placing four scrubbers of different designs on an industrial stack in random order. Each scrubber was tested 5 times. For each run, the same material was produced, and the particulate emissions coming out of the scrubber were measured (in parts per billion). A partially completed Analysis of Variance table of the data follows on the next page. (continued)
732
a) Calculate the mean square of the treatments and the mean square of the error. b) Form the F-statistic by dividing the two mean squares. c) The P-value of this F-statistic turns out to be 0.0000949. What does this say about the null hypothesis of equal means? d) What assumptions have you made in order to answer part c? e) What would you like to see in order to justify the conclusions of the F-test? f) What is the average size of the error standard deviation in particulate emissions?
Noise (decibels)
Source Treatment Residual Total
Sum of Squares 81.2 30.8 112.0
Degrees of Freedom
Mean Square
F-Ratio
boxplots from the data on noise reduction (in decibels) of the two filters. Type 1 = standard; Type 2 = Octel.
86 85 84 83 82 81 80 79 78 77 76 75 1 Type 2
ANOVA Table Source Type Error Total Level Standard Octel DF 1 33 34 n 18 17 Sum of Squares 6.31 271.47 2.77 Mean 81.5556 80.7059 Mean Square 6.31 8.22 F-Ratio 0.7673 P-Value 0.3874
15. Eggs A student wants to investigate the effects of real vs. substitute eggs on his favorite brownie recipe. He enlists the help of 10 friends and asks them to rank each of 8 batches on a scale from 1 to 10. Four of the batches were made with real eggs, four with substitute eggs. The judges tasted the brownies in random order. Here is a boxplot of the data:
8 7 Scores 6 5 4 Real Eggs Substitute T
Means and Std Deviations StdDev 3.2166 2.43708
a) What are the null and alternative hypotheses? b) What do you conclude from the ANOVA table? c) Do the assumptions for the test seem to be reasonable? d) Perform a two-sample pooled t-test of the difference. What P-value do you get? Show that the square of the t-statistic is the same (to rounding error) as the F-ratio.
ANOVA Table Source Eggs Error Total DF 1 6 7 Sum of Squares 9.010013 1.739875 10.749883 Mean Square 9.01001 0.28998 F-Ratio 31.0712 P-Value 0.0014
17. School system A school district superintendent wants to test a new method of teaching arithmetic in the fourth grade at his 15 schools. He plans to select 8 students from each school to take part in the experiment, but to make sure they are roughly of the same ability, he first gives a test to all 120 students. Here are the scores of the test by school:
27 25 Test Scores 23 21 19 17 15 13 A B C D E F G H I School J K L M N O
The mean score for the real eggs was 6.78 with a standard deviation of 0.651. The mean score for the substitute eggs was 4.66 with a standard deviation of 0.395. a) What are the null and alternative hypotheses? b) What do you conclude from the ANOVA table? c) Do the assumptions for the test seem to be reasonable? d) Perform a two-sample pooled t-test of the difference. What P-value do you get? Show that the square of the t-statistic is the same (to rounding error) as the F-ratio.
T
The ANOVA table shows:

Source School Error Total DF 14 105 119 Sum of Squares 108.800 760.125 868.925 Mean Square 7.7714 7.2392 F-Ratio 1.0735 P-Value 0.3899
16. Auto noise filters In a statement to a Senate Public Works Committee, a senior executive of Texaco, Inc., cited a study on the effectiveness of auto filters on reducing noise. Because of concerns about performance, two types of filters were studied, a standard silencer and a new device developed by the Associated Octel Company. Here are the
733
a) What are the null and alternative hypotheses? b) What does the ANOVA table say about the null hypothesis? (Be sure to report this in terms of scores and schools.) c) An intern reports that he has done t-tests of every school against every other school and finds that several of the schools seem to differ in mean score. Does this match your finding in part b? Give an explanation for the difference, if any, of the two results.
15 10 5 0 1 2 Shelf 3
18. Fertilizers A biology student is studying the effect of 10 different fertilizers on the growth of mung bean sprouts. She sprouts 12 beans in each of 10 different petri dishes, and adds the same amount of fertilizer to each dish. After one week she measures the heights of the 120 sprouts in millimeters. Here are boxplots and an ANOVA table of the data:
140 130 120 110 100 90 80 70 60 50 A B C D E F Fertilizer G H I J
Sugars (g)
Source Shelf Error Total Level 1 2 3
DF 2 74 76 n 20 21 36
Sum of Squares 248.4079 1253.1246 1501.5325 Mean 4.80000 9.61905 6.52778
Mean Square 124.204 16.934
F-Ratio 7.3345
P-Value 0.0012
Means and Std Deviations StdDev 4.57223 4.12888 3.83582
Heights (mm)
Source Fertilizer Error Total
DF 9 110 119
Sum of Squares 2073.708 21331.083 23404.791
Mean Square 230.412 193.919
F-Ratio 1.1882
P-Value 0.3097
a) What are the null and alternative hypotheses? b) What does the ANOVA table say about the null hypothesis? (Be sure to report this in terms of heights and fertilizers). c) Her lab partner looks at the same data and says that he did t-tests of every fertilizer against every other fertilizer and finds that several of the fertilizers seem to have significantly higher mean heights. Does this match your finding in part b? Give an explanation for the difference, if any, between the two results.
a) What are the null and alternative hypotheses? b) What does the ANOVA table say about the null hypothesis? (Be sure to report this in terms of Sugars and Shelves.) c) Can we conclude that cereals on shelf 2 have a higher mean sugar content than cereals on shelf 3? Can we conclude that cereals on shelf 2 have a higher mean sugar content than cereals on shelf 1? What can we conclude? d) To check for significant differences between the shelf means, we can use a Bonferroni test, whose results are shown below. For each pair of shelves, the difference is shown along with its standard error and significance level. What does it say about the questions in part c? Dependent Variable: SUGARS
Mean (I) (J) Difference SHELF SHELF (I-J) Bonferroni 1 2 3 2 3 1 95% Confidence P-Value Interval Lower Upper Bound Bound - 4.819(*) 1.2857 0.001 - 7.969 - 1.670 - 1.728 1.1476 0.409 - 4.539 1.084 4.819(*) 1.2857 0.001 1.670 7.969 Std. Error 0.323 5.859 - 1.084 4.539 - 5.859 - 0.323
19. Cereals Supermarkets often place similar types of cereal on the same supermarket shelf. We have data on the shelf as well as the sugar, sodium, and calorie content of 77 cereals. Does sugar content vary by shelf? At the top of the next column is a boxplot and an ANOVA table for the 77 cereals.
3 3.091(*) 1.1299 0.023 1 1.728 1.1476 0.409 - 3.091(*) 1.1299 0.023 2 *The mean difference is significant at the 0.05 level.
734
20. Cereals, redux We also have data on the protein content of the cereals in Exercise 19 by their shelf number. Here are the boxplot and ANOVA table:
6 5 Protein (g) 4 3 2 1 1 2 Shelf 3
conclude that cereals on shelf 2 have a lower mean protein content than cereals on shelf 1? What can we conclude? d) To check for significant differences between the shelf means we can use a Bonferroni test, whose results are shown below. For each pair of shelves, the difference is shown along with its standard error and significance level. What does it say about the questions in part c? Dependent Variable: PROTEIN
(I) (J) SHELF SHELF Mean Difference (I-J) Std. Error 95% Confidence Interval Lower Upper Bound Bound 1.53 - 0.04 0.49 - 0.92 - 1.53 0.04 - 1.65 - 0.26 - 0.49 0.92 0.26 1.65
P-Value
Bonferroni
2 0.75 0.322 0.288 - 0.21 3 - 0.75 0.322 2 1 0.283 - 0.96(*) 3 3 1 0.21 0.288 2 0.96(*) 0.283 *The mean difference is significant at the 0.05 level.
T
Source Shelf Error Total Level 1 2 3
DF 2 74 76 n 20 21 36
Sum of Squares 12.4258 78.6650 91.0909 Mean 2.65000 1.90476 2.86111
Mean Square 6.2129 1.0630
F-Ratio 5.8445
P-Value 0.0044
0.070 1.000 0.070 0.004 1.000 0.004
Means and Std Deviations StdDev 1.46089 0.99523 0.72320
a) What are the null and alternative hypotheses? b) What does the ANOVA table say about the null hypothesis? (Be sure to report this in terms of protein content and shelves.) c) Can we conclude that cereals on shelf 2 have a lower mean protein content than cereals on shelf 3? Can we
21. Downloading To see how much of a difference time of day made on the speed at which he could download files, a college sophomore performed an experiment. He placed a file on a remote server and then proceeded to download it at three different time periods of the day. He downloaded the file 48 times in all, 16 times at each Time of Day, and recorded the Time in seconds that the download took. a) State the null and alternative hypotheses, being careful to talk about download Time and Time of Day as well as parameters.
Time of Day Early (7 a.m.) Early (7 a.m.) Early (7 a.m.) Early (7 a.m.) Early (7 a.m.) Early (7 a.m.) Early (7 a.m.) Early (7 a.m.) Early (7 a.m.) Early (7 a.m.) Early (7 a.m.) Early (7 a.m.) Early (7 a.m.) Early (7 a.m.) Early (7 a.m.) Early (7 a.m.)
Time (sec) 68 138 75 186 68 217 93 90 71 154 166 130 72 81 76 129
Time of Day Evening (5 p.m.) Evening (5 p.m.) Evening (5 p.m.) Evening (5 p.m.) Evening (5 p.m.) Evening (5 p.m.) Evening (5 p.m.) Evening (5 p.m.) Evening (5 p.m.) Evening (5 p.m.) Evening (5 p.m.) Evening (5 p.m.) Evening (5 p.m.) Evening (5 p.m.) Evening (5 p.m.) Evening (5 p.m.)
Time (sec) 299 367 331 257 260 269 252 200 296 204 190 240 350 256 282 320
Time of Day Late Night (12 a.m.) Late Night (12 a.m.) Late Night (12 a.m.) Late Night (12 a.m.) Late Night (12 a.m.) Late Night (12 a.m.) Late Night (12 a.m.) Late Night (12 a.m.) Late Night (12 a.m.) Late Night (12 a.m.) Late Night (12 a.m.) Late Night (12 a.m.) Late Night (12 a.m.) Late Night (12 a.m.) Late Night (12 a.m.) Late Night (12 a.m.)
Time (sec) 216 175 274 171 187 213 221 139 226 128 236 128 217 196 201 161
735
b) Perform an ANOVA on these data. What can you conclude? c) Check the assumptions and conditions for an ANOVA. Do you have any concerns about the experimental design or the analysis? d) (Optional) Perform a multiple comparisons test to determine which times of day differ in terms of mean download time.
22. Analgesics A pharmaceutical company tested three formulations of a pain relief medicine for migraine headache sufferers. For the experiment, 27 volunteers were selected and 9 were randomly assigned to one of three drug formulations. The subjects were instructed to take the drug during their next migraine headache episode and to report their pain on a scale of 1 = no pain to 10 = extreme pain 30 minutes after taking the drug.
Drug A A A A A A A A A Pain 4 5 4 3 2 4 3 4 4 Drug B B B B B B B B B Pain 6 8 4 5 4 6 5 8 6 Drug C C C C C C C C C Pain 6 7 6 6 7 5 6 5 5
a) State the null and alternative hypotheses, being careful to talk about Drug and Pain levels as well as parameters. b) Perform an ANOVA on these data. What can you conclude? c) Check the assumptions and conditions for an ANOVA. Do you have any concerns about the experimental design or the analysis? d) (Optional) Perform a multiple comparisons test to determine which drugs differ in terms of mean pain level reported.
Just Checking
Answers
1. The null hypothesis is that the mean flight distance
for all four designs is the same.

2. Yes, it looks as if the variation between the means is
greater than the variation within each boxplot.

3. Yes, the F-test rejects the null hypothesis with a
P-value 6 0.0001.
4. No. The alternative hypothesis is that at least one
mean is different from the other three. Rejecting the null hypothesis does not imply that all four means are different.
Denominator df
736
3 4 5 6 7 8 9 10 Numerator df 11 12 13 14 15 16 17 18 19 20 21 22
Table F A 0.01
1 4052.2 4999.3 5403.5 5624.3 5764.0 5859.0 5928.3 5981.0 6022.4 6055.9 6083.4 6106.7 6125.8 6143.0 6157.0 6170.0 6181.2 6191.4 6200.7 6208.7 6216.1 6223.1 2 98.50 99.00 99.16 99.25 99.30 99.33 99.36 99.38 99.39 99.40 99.41 99.42 99.42 99.43 99.43 99.44 99.44 99.44 99.45 99.45 99.45 99.46 3 34.12 30.82 29.46 28.71 28.24 27.91 27.67 27.49 27.34 27.23 27.13 27.05 26.98 26.92 26.87 26.83 26.79 26.75 26.72 26.69 26.66 26.64 4 21.20 18.00 16.69 15.98 15.52 15.21 14.98 14.80 14.66 14.55 14.45 14.37 14.31 14.25 14.20 14.15 14.11 14.08 14.05 14.02 13.99 13.97 5 16.26 13.27 12.06 11.39 10.97 10.67 10.46 10.29 10.16 10.05 9.96 9.89 9.82 9.77 9.72 9.68 9.64 9.61 9.58 9.55 9.53 9.51 6 13.75 10.92 9.78 9.15 8.75 8.47 8.26 8.10 7.98 7.87 7.79 7.72 7.66 7.60 7.56 7.52 7.48 7.45 7.42 7.40 7.37 7.35 7 12.25 9.55 8.45 7.85 7.46 7.19 6.99 6.84 6.72 6.62 6.54 6.47 6.41 6.36 6.31 6.28 6.24 6.21 6.18 6.16 6.13 6.11 8 11.26 8.65 7.59 7.01 6.63 6.37 6.18 6.03 5.91 5.81 5.73 5.67 5.61 5.56 5.52 5.48 5.44 5.41 5.38 5.36 5.34 5.32 9 10.56 8.02 6.99 6.42 6.06 5.80 5.61 5.47 5.35 5.26 5.18 5.11 5.05 5.01 4.96 4.92 4.89 4.86 4.83 4.81 4.79 4.77 10 10.04 7.56 6.55 5.99 5.64 5.39 5.20 5.06 4.94 4.85 4.77 4.71 4.65 4.60 4.56 4.52 4.49 4.46 4.43 4.41 4.38 4.36 11 9.65 7.21 6.22 5.67 5.32 5.07 4.89 4.74 4.63 4.54 4.46 4.40 4.34 4.29 4.25 4.21 4.18 4.15 4.12 4.10 4.08 4.06 12 9.33 6.93 5.95 5.41 5.06 4.82 4.64 4.50 4.39 4.30 4.22 4.16 4.10 4.05 4.01 3.97 3.94 3.91 3.88 3.86 3.84 3.82 13 9.07 6.70 5.74 5.21 4.86 4.62 4.44 4.30 4.19 4.10 4.02 3.96 3.91 3.86 3.82 3.78 3.75 3.72 3.69 3.66 3.64 3.62 14 8.86 6.51 5.56 5.04 4.69 4.46 4.28 4.14 4.03 3.94 3.86 3.80 3.75 3.70 3.66 3.62 3.59 3.56 3.53 3.51 3.48 3.46 15 8.68 6.36 5.42 4.89 4.56 4.32 4.14 4.00 3.89 3.80 3.73 3.67 3.61 3.56 3.52 3.49 3.45 3.42 3.40 3.37 3.35 3.33 16 8.53 6.23 5.29 4.77 4.44 4.20 4.03 3.89 3.78 3.69 3.62 3.55 3.50 3.45 3.41 3.37 3.34 3.31 3.28 3.26 3.24 3.22 17 8.40 6.11 5.19 4.67 4.34 4.10 3.93 3.79 3.68 3.59 3.52 3.46 3.40 3.35 3.31 3.27 3.24 3.21 3.19 3.16 3.14 3.12 18 8.29 6.01 5.09 4.58 4.25 4.01 3.84 3.71 3.60 3.51 3.43 3.37 3.32 3.27 3.23 3.19 3.16 3.13 3.10 3.08 3.05 3.03 19 8.18 5.93 5.01 4.50 4.17 3.94 3.77 3.63 3.52 3.43 3.36 3.30 3.24 3.19 3.15 3.12 3.08 3.05 3.03 3.00 2.98 2.96 20 8.10 5.85 4.94 4.43 4.10 3.87 3.70 3.56 3.46 3.37 3.29 3.23 3.18 3.13 3.09 3.05 3.02 2.99 2.96 2.94 2.92 2.90 21 8.02 5.78 4.87 4.37 4.04 3.81 3.64 3.51 3.40 3.31 3.24 3.17 3.12 3.07 3.03 2.99 2.96 2.93 2.90 2.88 2.86 2.84 22 7.95 5.72 4.82 4.31 3.99 3.76 3.59 3.45 3.35 3.26 3.18 3.12 3.07 3.02 2.98 2.94 2.91 2.88 2.85 2.83 2.81 2.78 23 7.88 5.66 4.76 4.26 3.94 3.71 3.54 3.41 3.30 3.21 3.14 3.07 3.02 2.97 2.93 2.89 2.86 2.83 2.80 2.78 2.76 2.74 24 7.82 5.61 4.72 4.22 3.90 3.67 3.50 3.36 3.26 3.17 3.09 3.03 2.98 2.93 2.89 2.85 2.82 2.79 2.76 2.74 2.72 2.70 25 7.77 5.57 4.68 4.18 3.85 3.63 3.46 3.32 3.22 3.13 3.06 2.99 2.94 2.89 2.85 2.81 2.78 2.75 2.72 2.70 2.68 2.66 26 7.72 5.53 4.64 4.14 3.82 3.59 3.42 3.29 3.18 3.09 3.02 2.96 2.90 2.86 2.81 2.78 2.75 2.72 2.69 2.66 2.64 2.62 27 7.68 5.49 4.60 4.11 3.78 3.56 3.39 3.26 3.15 3.06 2.99 2.93 2.87 2.82 2.78 2.75 2.71 2.68 2.66 2.63 2.61 2.59 28 7.64 5.45 4.57 4.07 3.75 3.53 3.36 3.23 3.12 3.03 2.96 2.90 2.84 2.79 2.75 2.72 2.68 2.65 2.63 2.60 2.58 2.56 29 7.60 5.42 4.54 4.04 3.73 3.50 3.33 3.20 3.09 3.00 2.93 2.87 2.81 2.77 2.73 2.69 2.66 2.63 2.60 2.57 2.55 2.53 30 7.56 5.39 4.51 4.02 3.70 3.47 3.30 3.17 3.07 2.98 2.91 2.84 2.79 2.74 2.70 2.66 2.63 2.60 2.57 2.55 2.53 2.51 32 7.50 5.34 4.46 3.97 3.65 3.43 3.26 3.13 3.02 2.93 2.86 2.80 2.74 2.70 2.65 2.62 2.58 2.55 2.53 2.50 2.48 2.46 35 7.42 5.27 4.40 3.91 3.59 3.37 3.20 3.07 2.96 2.88 2.80 2.74 2.69 2.64 2.60 2.56 2.53 2.50 2.47 2.44 2.42 2.40 40 7.31 5.18 4.31 3.83 3.51 3.29 3.12 2.99 2.89 2.80 2.73 2.66 2.61 2.56 2.52 2.48 2.45 2.42 2.39 2.37 2.35 2.33 45 7.23 5.11 4.25 3.77 3.45 3.23 3.07 2.94 2.83 2.74 2.67 2.61 2.55 2.51 2.46 2.43 2.39 2.36 2.34 2.31 2.29 2.27 50 7.17 5.06 4.20 3.72 3.41 3.19 3.02 2.89 2.78 2.70 2.63 2.56 2.51 2.46 2.42 2.38 2.35 2.32 2.29 2.27 2.24 2.22 60 7.08 4.98 4.13 3.65 3.34 3.12 2.95 2.82 2.72 2.63 2.56 2.50 2.44 2.39 2.35 2.31 2.28 2.25 2.22 2.20 2.17 2.15 75 6.99 4.90 4.05 3.58 3.27 3.05 2.89 2.76 2.65 2.57 2.49 2.43 2.38 2.33 2.29 2.25 2.22 2.18 2.16 2.13 2.11 2.09 100 6.90 4.82 3.98 3.51 3.21 2.99 2.82 2.69 2.59 2.50 2.43 2.37 2.31 2.27 2.22 2.19 2.15 2.12 2.09 2.07 2.04 2.02 120 6.85 4.79 3.95 3.48 3.17 2.96 2.79 2.66 2.56 2.47 2.40 2.34 2.28 2.23 2.19 2.15 2.12 2.09 2.06 2.03 2.01 1.99 140 6.82 4.76 3.92 3.46 3.15 2.93 2.77 2.64 2.54 2.45 2.38 2.31 2.26 2.21 2.17 2.13 2.10 2.07 2.04 2.01 1.99 1.97 180 6.78 4.73 3.89 3.43 3.12 2.90 2.74 2.61 2.51 2.42 2.35 2.28 2.23 2.18 2.14 2.10 2.07 2.04 2.01 1.98 1.96 1.94 250 6.74 4.69 3.86 3.40 3.09 2.87 2.71 2.58 2.48 2.39 2.32 2.26 2.20 2.15 2.11 2.07 2.04 2.01 1.98 1.95 1.93 1.91 400 6.70 4.66 3.83 3.37 3.06 2.85 2.68 2.56 2.45 2.37 2.29 2.23 2.17 2.13 2.08 2.05 2.01 1.98 1.95 1.92 1.90 1.88 1000 6.66 4.63 3.80 3.34 3.04 2.82 2.66 2.53 2.43 2.34 2.27 2.20 2.15 2.10 2.06 2.02 1.98 1.95 1.92 1.90 1.87 1.85
Table F (cont.) A 0.01 25 26 27 28 29 30 32 35 50 60 75 100 120 140 180 250 400
23
24
Numerator df 40 45
1000
Denominator df
1 6228.7 6234.3 6239.9 6244.5 6249.2 6252.9 6257.1 6260.4 6266.9 6275.3 6286.4 6295.7 6302.3 6313.0 6323.7 6333.9 6339.5 6343.2 6347.9 6353.5 6358.1 6362.8 2 99.46 99.46 99.46 99.46 99.46 99.46 99.46 99.47 99.47 99.47 99.48 99.48 99.48 99.48 99.48 99.49 99.49 99.49 99.49 99.50 99.50 99.50 3 26.62 26.60 26.58 26.56 26.55 26.53 26.52 26.50 26.48 26.45 26.41 26.38 26.35 26.32 26.28 26.24 26.22 26.21 26.19 26.17 26.15 26.14 4 13.95 13.93 13.91 13.89 13.88 13.86 13.85 13.84 13.81 13.79 13.75 13.71 13.69 13.65 13.61 13.58 13.56 13.54 13.53 13.51 13.49 13.47 5 9.49 9.47 9.45 9.43 9.42 9.40 9.39 9.38 9.36 9.33 9.29 9.26 9.24 9.20 9.17 9.13 9.11 9.10 9.08 9.06 9.05 9.03 6 7.33 7.31 7.30 7.28 7.27 7.25 7.24 7.23 7.21 7.18 7.14 7.11 7.09 7.06 7.02 6.99 6.97 6.96 6.94 6.92 6.91 6.89 7 6.09 6.07 6.06 6.04 6.03 6.02 6.00 5.99 5.97 5.94 5.91 5.88 5.86 5.82 5.79 5.75 5.74 5.72 5.71 5.69 5.68 5.66 8 5.30 5.28 5.26 5.25 5.23 5.22 5.21 5.20 5.18 5.15 5.12 5.09 5.07 5.03 5.00 4.96 4.95 4.93 4.92 4.90 4.89 4.87 9 4.75 4.73 4.71 4.70 4.68 4.67 4.66 4.65 4.63 4.60 4.57 4.54 4.52 4.48 4.45 4.41 4.40 4.39 4.37 4.35 4.34 4.32 10 4.34 4.33 4.31 4.30 4.28 4.27 4.26 4.25 4.23 4.20 4.17 4.14 4.12 4.08 4.05 4.01 4.00 3.98 3.97 3.95 3.94 3.92 11 4.04 4.02 4.01 3.99 3.98 3.96 3.95 3.94 3.92 3.89 3.86 3.83 3.81 3.78 3.74 3.71 3.69 3.68 3.66 3.64 3.63 3.61 12 3.80 3.78 3.76 3.75 3.74 3.72 3.71 3.70 3.68 3.65 3.62 3.59 3.57 3.54 3.50 3.47 3.45 3.44 3.42 3.40 3.39 3.37 13 3.60 3.59 3.57 3.56 3.54 3.53 3.52 3.51 3.49 3.46 3.43 3.40 3.38 3.34 3.31 3.27 3.25 3.24 3.23 3.21 3.19 3.18 14 3.44 3.43 3.41 3.40 3.38 3.37 3.36 3.35 3.33 3.30 3.27 3.24 3.22 3.18 3.15 3.11 3.09 3.08 3.06 3.05 3.03 3.02 15 3.31 3.29 3.28 3.26 3.25 3.24 3.23 3.21 3.19 3.17 3.13 3.10 3.08 3.05 3.01 2.98 2.96 2.95 2.93 2.91 2.90 2.88 16 3.20 3.18 3.16 3.15 3.14 3.12 3.11 3.10 3.08 3.05 3.02 2.99 2.97 2.93 2.90 2.86 2.84 2.83 2.81 2.80 2.78 2.76 17 3.10 3.08 3.07 3.05 3.04 3.03 3.01 3.00 2.98 2.96 2.92 2.89 2.87 2.83 2.80 2.76 2.75 2.73 2.72 2.70 2.68 2.66 18 3.02 3.00 2.98 2.97 2.95 2.94 2.93 2.92 2.90 2.87 2.84 2.81 2.78 2.75 2.71 2.68 2.66 2.65 2.63 2.61 2.59 2.58 19 2.94 2.92 2.91 2.89 2.88 2.87 2.86 2.84 2.82 2.80 2.76 2.73 2.71 2.67 2.64 2.60 2.58 2.57 2.55 2.54 2.52 2.50 20 2.88 2.86 2.84 2.83 2.81 2.80 2.79 2.78 2.76 2.73 2.69 2.67 2.64 2.61 2.57 2.54 2.52 2.50 2.49 2.47 2.45 2.43 21 2.82 2.80 2.79 2.77 2.76 2.74 2.73 2.72 2.70 2.67 2.64 2.61 2.58 2.55 2.51 2.48 2.46 2.44 2.43 2.41 2.39 2.37 22 2.77 2.75 2.73 2.72 2.70 2.69 2.68 2.67 2.65 2.62 2.58 2.55 2.53 2.50 2.46 2.42 2.40 2.39 2.37 2.35 2.34 2.32 23 2.72 2.70 2.69 2.67 2.66 2.64 2.63 2.62 2.60 2.57 2.54 2.51 2.48 2.45 2.41 2.37 2.35 2.34 2.32 2.30 2.29 2.27 24 2.68 2.66 2.64 2.63 2.61 2.60 2.59 2.58 2.56 2.53 2.49 2.46 2.44 2.40 2.37 2.33 2.31 2.30 2.28 2.26 2.24 2.22 25 2.64 2.62 2.60 2.59 2.58 2.56 2.55 2.54 2.52 2.49 2.45 2.42 2.40 2.36 2.33 2.29 2.27 2.26 2.24 2.22 2.20 2.18 26 2.60 2.58 2.57 2.55 2.54 2.53 2.51 2.50 2.48 2.45 2.42 2.39 2.36 2.33 2.29 2.25 2.23 2.22 2.20 2.18 2.16 2.14 27 2.57 2.55 2.54 2.52 2.51 2.49 2.48 2.47 2.45 2.42 2.38 2.35 2.33 2.29 2.26 2.22 2.20 2.18 2.17 2.15 2.13 2.11 28 2.54 2.52 2.51 2.49 2.48 2.46 2.45 2.44 2.42 2.39 2.35 2.32 2.30 2.26 2.23 2.19 2.17 2.15 2.13 2.11 2.10 2.08 29 2.51 2.49 2.48 2.46 2.45 2.44 2.42 2.41 2.39 2.36 2.33 2.30 2.27 2.23 2.20 2.16 2.14 2.12 2.10 2.08 2.07 2.05 30 2.49 2.47 2.45 2.44 2.42 2.41 2.40 2.39 2.36 2.34 2.30 2.27 2.25 2.21 2.17 2.13 2.11 2.10 2.08 2.06 2.04 2.02 32 2.44 2.42 2.41 2.39 2.38 2.36 2.35 2.34 2.32 2.29 2.25 2.22 2.20 2.16 2.12 2.08 2.06 2.05 2.03 2.01 1.99 1.97 35 2.38 2.36 2.35 2.33 2.32 2.30 2.29 2.28 2.26 2.23 2.19 2.16 2.14 2.10 2.06 2.02 2.00 1.98 1.96 1.94 1.92 1.90 40 2.31 2.29 2.27 2.26 2.24 2.23 2.22 2.20 2.18 2.15 2.11 2.08 2.06 2.02 1.98 1.94 1.92 1.90 1.88 1.86 1.84 1.82 45 2.25 2.23 2.21 2.20 2.18 2.17 2.16 2.14 2.12 2.09 2.05 2.02 2.00 1.96 1.92 1.88 1.85 1.84 1.82 1.79 1.77 1.75 50 2.20 2.18 2.17 2.15 2.14 2.12 2.11 2.10 2.08 2.05 2.01 1.97 1.95 1.91 1.87 1.82 1.80 1.79 1.76 1.74 1.72 1.70 60 2.13 2.12 2.10 2.08 2.07 2.05 2.04 2.03 2.01 1.98 1.94 1.90 1.88 1.84 1.79 1.75 1.73 1.71 1.69 1.66 1.64 1.62 75 2.07 2.05 2.03 2.02 2.00 1.99 1.97 1.96 1.94 1.91 1.87 1.83 1.81 1.76 1.72 1.67 1.65 1.63 1.61 1.58 1.56 1.53 100 2.00 1.98 1.97 1.95 1.93 1.92 1.91 1.89 1.87 1.84 1.80 1.76 1.74 1.69 1.65 1.60 1.57 1.55 1.53 1.50 1.47 1.45 120 1.97 1.95 1.93 1.92 1.90 1.89 1.87 1.86 1.84 1.81 1.76 1.73 1.70 1.66 1.61 1.56 1.53 1.51 1.49 1.46 1.43 1.40 140 1.95 1.93 1.91 1.89 1.88 1.86 1.85 1.84 1.81 1.78 1.74 1.70 1.67 1.63 1.58 1.53 1.50 1.48 1.46 1.43 1.40 1.37 180 1.92 1.90 1.88 1.86 1.85 1.83 1.82 1.81 1.78 1.75 1.71 1.67 1.64 1.60 1.55 1.49 1.47 1.45 1.42 1.39 1.35 1.32 250 1.89 1.87 1.85 1.83 1.82 1.80 1.79 1.77 1.75 1.72 1.67 1.64 1.61 1.56 1.51 1.46 1.43 1.41 1.38 1.34 1.31 1.27 400 1.86 1.84 1.82 1.80 1.79 1.77 1.76 1.75 1.72 1.69 1.64 1.61 1.58 1.53 1.48 1.42 1.39 1.37 1.33 1.30 1.26 1.22 1000 1.83 1.81 1.79 1.77 1.76 1.74 1.73 1.72 1.69 1.66 1.61 1.58 1.54 1.50 1.44 1.38 1.35 1.33 1.29 1.25 1.21 1.16
737
Denominator df
738
3 224.6 19.25 9.12 6.39 5.19 4.53 4.12 3.84 3.63 3.48 3.36 3.26 3.18 3.11 3.06 3.01 2.96 2.93 2.90 2.87 2.84 2.82 2.80 2.78 2.76 2.74 2.73 2.71 2.70 2.69 2.67 2.64 2.61 2.58 2.56 2.53 2.49 2.46 2.45 2.44 2.42 2.41 2.39 2.38 230.2 19.30 9.01 6.26 5.05 4.39 3.97 3.69 3.48 3.33 3.20 3.11 3.03 2.96 2.90 2.85 2.81 2.77 2.74 2.71 2.68 2.66 2.64 2.62 2.60 2.59 2.57 2.56 2.55 2.53 2.51 2.49 2.45 2.42 2.40 2.37 2.34 2.31 2.29 2.28 2.26 2.25 2.24 2.22 234.0 19.33 8.94 6.16 4.95 4.28 3.87 3.58 3.37 3.22 3.09 3.00 2.92 2.85 2.79 2.74 2.70 2.66 2.63 2.60 2.57 2.55 2.53 2.51 2.49 2.47 2.46 2.45 2.43 2.42 2.40 2.37 2.34 2.31 2.29 2.25 2.22 2.19 2.18 2.16 2.15 2.13 2.12 2.11 236.8 19.35 8.89 6.09 4.88 4.21 3.79 3.50 3.29 3.14 3.01 2.91 2.83 2.76 2.71 2.66 2.61 2.58 2.54 2.51 2.49 2.46 2.44 2.42 2.40 2.39 2.37 2.36 2.35 2.33 2.31 2.29 2.25 2.22 2.20 2.17 2.13 2.10 2.09 2.08 2.06 2.05 2.03 2.02 238.9 19.37 8.85 6.04 4.82 4.15 3.73 3.44 3.23 3.07 2.95 2.85 2.77 2.70 2.64 2.59 2.55 2.51 2.48 2.45 2.42 2.40 2.37 2.36 2.34 2.32 2.31 2.29 2.28 2.27 2.24 2.22 2.18 2.15 2.13 2.10 2.06 2.03 2.02 2.01 1.99 1.98 1.96 1.95 240.5 19.38 8.81 6.00 4.77 4.10 3.68 3.39 3.18 3.02 2.90 2.80 2.71 2.65 2.59 2.54 2.49 2.46 2.42 2.39 2.37 2.34 2.32 2.30 2.28 2.27 2.25 2.24 2.22 2.21 2.19 2.16 2.12 2.10 2.07 2.04 2.01 1.97 1.96 1.95 1.93 1.92 1.90 1.89 241.9 19.40 8.79 5.96 4.74 4.06 3.64 3.35 3.14 2.98 2.85 2.75 2.67 2.60 2.54 2.49 2.45 2.41 2.38 2.35 2.32 2.30 2.27 2.25 2.24 2.22 2.20 2.19 2.18 2.16 2.14 2.11 2.08 2.05 2.03 1.99 1.96 1.93 1.91 1.90 1.88 1.87 1.85 1.84 243.0 19.40 8.76 5.94 4.70 4.03 3.60 3.31 3.10 2.94 2.82 2.72 2.63 2.57 2.51 2.46 2.41 2.37 2.34 2.31 2.28 2.26 2.24 2.22 2.20 2.18 2.17 2.15 2.14 2.13 2.10 2.07 2.04 2.01 1.99 1.95 1.92 1.89 1.87 1.86 1.84 1.83 1.81 1.80 243.9 19.41 8.74 5.91 4.68 4.00 3.57 3.28 3.07 2.91 2.79 2.69 2.60 2.53 2.48 2.42 2.38 2.34 2.31 2.28 2.25 2.23 2.20 2.18 2.16 2.15 2.13 2.12 2.10 2.09 2.07 2.04 2.00 1.97 1.95 1.92 1.88 1.85 1.83 1.82 1.81 1.79 1.78 1.76 244.7 19.42 8.73 5.89 4.66 3.98 3.55 3.26 3.05 2.89 2.76 2.66 2.58 2.51 2.45 2.40 2.35 2.31 2.28 2.25 2.22 2.20 2.18 2.15 2.14 2.12 2.10 2.09 2.08 2.06 2.04 2.01 1.97 1.94 1.92 1.89 1.85 1.82 1.80 1.79 1.77 1.76 1.74 1.73 245.4 19.42 8.71 5.87 4.64 3.96 3.53 3.24 3.03 2.86 2.74 2.64 2.55 2.48 2.42 2.37 2.33 2.29 2.26 2.22 2.20 2.17 2.15 2.13 2.11 2.09 2.08 2.06 2.05 2.04 2.01 1.99 1.95 1.92 1.89 1.86 1.83 1.79 1.78 1.76 1.75 1.73 1.72 1.70 245.9 19.43 8.70 5.86 4.62 3.94 3.51 3.22 3.01 2.85 2.72 2.62 2.53 2.46 2.40 2.35 2.31 2.27 2.23 2.20 2.18 2.15 2.13 2.11 2.09 2.07 2.06 2.04 2.03 2.01 1.99 1.96 1.92 1.89 1.87 1.84 1.80 1.77 1.75 1.74 1.72 1.71 1.69 1.68 246.5 19.43 8.69 5.84 4.60 3.92 3.49 3.20 2.99 2.83 2.70 2.60 2.51 2.44 2.38 2.33 2.29 2.25 2.21 2.18 2.16 2.13 2.11 2.09 2.07 2.05 2.04 2.02 2.01 1.99 1.97 1.94 1.90 1.87 1.85 1.82 1.78 1.75 1.73 1.72 1.70 1.68 1.67 1.65 246.9 19.44 8.68 5.83 4.59 3.91 3.48 3.19 2.97 2.81 2.69 2.58 2.50 2.43 2.37 2.32 2.27 2.23 2.20 2.17 2.14 2.11 2.09 2.07 2.05 2.03 2.02 2.00 1.99 1.98 1.95 1.92 1.89 1.86 1.83 1.80 1.76 1.73 1.71 1.70 1.68 1.66 1.65 1.63 247.3 19.44 8.67 5.82 4.58 3.90 3.47 3.17 2.96 2.80 2.67 2.57 2.48 2.41 2.35 2.30 2.26 2.22 2.18 2.15 2.12 2.10 2.08 2.05 2.04 2.02 2.00 1.99 1.97 1.96 1.94 1.91 1.87 1.84 1.81 1.78 1.74 1.71 1.69 1.68 1.66 1.65 1.63 1.61 247.7 19.44 8.67 5.81 4.57 3.88 3.46 3.16 2.95 2.79 2.66 2.56 2.47 2.40 2.34 2.29 2.24 2.20 2.17 2.14 2.11 2.08 2.06 2.04 2.02 2.00 1.99 1.97 1.96 1.95 1.92 1.89 1.85 1.82 1.80 1.76 1.73 1.69 1.67 1.66 1.64 1.63 1.61 1.60 248.0 19.45 8.66 5.80 4.56 3.87 3.44 3.15 2.94 2.77 2.65 2.54 2.46 2.39 2.33 2.28 2.23 2.19 2.16 2.12 2.10 2.07 2.05 2.03 2.01 1.99 1.97 1.96 1.94 1.93 1.91 1.88 1.84 1.81 1.78 1.75 1.71 1.68 1.66 1.65 1.63 1.61 1.60 1.58 248.3 19.45 8.65 5.79 4.55 3.86 3.43 3.14 2.93 2.76 2.64 2.53 2.45 2.38 2.32 2.26 2.22 2.18 2.14 2.11 2.08 2.06 2.04 2.01 2.00 1.98 1.96 1.95 1.93 1.92 1.90 1.87 1.83 1.80 1.77 1.73 1.70 1.66 1.64 1.63 1.61 1.60 1.58 1.57 4 5 6 7 8 9 10 Numerator df 11 12 13 14 15 16 17 18 19 20 21 22 248.6 19.45 8.65 5.79 4.54 3.86 3.43 3.13 2.92 2.75 2.63 2.52 2.44 2.37 2.31 2.25 2.21 2.17 2.13 2.10 2.07 2.05 2.02 2.00 1.98 1.97 1.95 1.93 1.92 1.91 1.88 1.85 1.81 1.78 1.76 1.72 1.69 1.65 1.63 1.62 1.60 1.58 1.57 1.55
Table F (cont.) A 0.05
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 32 35 40 45 50 60 75 100 120 140 180 250 400 1000
161.4 18.51 10.13 7.71 6.61 5.99 5.59 5.32 5.12 4.96 4.84 4.75 4.67 4.60 4.54 4.49 4.45 4.41 4.38 4.35 4.32 4.30 4.28 4.26 4.24 4.23 4.21 4.20 4.18 4.17 4.15 4.12 4.08 4.06 4.03 4.00 3.97 3.94 3.92 3.91 3.89 3.88 3.86 3.85
199.5 19.00 9.55 6.94 5.79 5.14 4.74 4.46 4.26 4.10 3.98 3.89 3.81 3.74 3.68 3.63 3.59 3.55 3.52 3.49 3.47 3.44 3.42 3.40 3.39 3.37 3.35 3.34 3.33 3.32 3.29 3.27 3.23 3.20 3.18 3.15 3.12 3.09 3.07 3.06 3.05 3.03 3.02 3.00
215.7 19.16 9.28 6.59 5.41 4.76 4.35 4.07 3.86 3.71 3.59 3.49 3.41 3.34 3.29 3.24 3.20 3.16 3.13 3.10 3.07 3.05 3.03 3.01 2.99 2.98 2.96 2.95 2.93 2.92 2.90 2.87 2.84 2.81 2.79 2.76 2.73 2.70 2.68 2.67 2.65 2.64 2.63 2.61
Table F (cont.) A 0.05 23 25 26 27 28 29 30 32 35 50 60 75 100 120 140 180 250 400
24
Numerator df 40 45
1000
Denominator df
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 32 35 40 45 50 60 75 100 120 140 180 250 400 1000 249.5 19.46 8.63 5.76 4.52 3.83 3.40 3.10 2.89 2.72 2.59 2.49 2.41 2.33 2.27 2.22 2.17 2.13 2.10 2.07 2.04 2.01 1.99 1.97 1.95 1.93 1.91 1.90 1.88 1.87 1.85 1.82 1.77 1.74 1.72 1.68 1.64 1.61 1.59 1.57 1.56 1.54 1.52 1.51 249.6 19.46 8.63 5.76 4.51 3.82 3.39 3.10 2.88 2.72 2.59 2.48 2.40 2.33 2.27 2.21 2.17 2.13 2.09 2.06 2.03 2.00 1.98 1.96 1.94 1.92 1.90 1.89 1.88 1.86 1.84 1.81 1.77 1.73 1.71 1.67 1.63 1.60 1.58 1.57 1.55 1.53 1.51 1.50 249.8 19.46 8.62 5.75 4.50 3.82 3.39 3.09 2.87 2.71 2.58 2.48 2.39 2.32 2.26 2.21 2.16 2.12 2.08 2.05 2.02 2.00 1.97 1.95 1.93 1.91 1.90 1.88 1.87 1.85 1.83 1.80 1.76 1.73 1.70 1.66 1.63 1.59 1.57 1.56 1.54 1.52 1.50 1.49 250.0 19.46 8.62 5.75 4.50 3.81 3.38 3.08 2.87 2.70 2.58 2.47 2.39 2.31 2.25 2.20 2.15 2.11 2.08 2.05 2.02 1.99 1.97 1.95 1.93 1.91 1.89 1.88 1.86 1.85 1.82 1.79 1.75 1.72 1.69 1.66 1.62 1.58 1.56 1.55 1.53 1.51 1.50 1.48 250.1 19.46 8.62 5.75 4.50 3.81 3.38 3.08 2.86 2.70 2.57 2.47 2.38 2.31 2.25 2.19 2.15 2.11 2.07 2.04 2.01 1.98 1.96 1.94 1.92 1.90 1.88 1.87 1.85 1.84 1.82 1.79 1.74 1.71 1.69 1.65 1.61 1.57 1.55 1.54 1.52 1.50 1.49 1.47 250.4 19.46 8.61 5.74 4.49 3.80 3.37 3.07 2.85 2.69 2.56 2.46 2.37 2.30 2.24 2.18 2.14 2.10 2.06 2.03 2.00 1.97 1.95 1.93 1.91 1.89 1.87 1.86 1.84 1.83 1.80 1.77 1.73 1.70 1.67 1.64 1.60 1.56 1.54 1.53 1.51 1.49 1.47 1.46 250.7 19.47 8.60 5.73 4.48 3.79 3.36 3.06 2.84 2.68 2.55 2.44 2.36 2.28 2.22 2.17 2.12 2.08 2.05 2.01 1.98 1.96 1.93 1.91 1.89 1.87 1.86 1.84 1.83 1.81 1.79 1.76 1.72 1.68 1.66 1.62 1.58 1.54 1.52 1.51 1.49 1.47 1.45 1.43 251.1 19.47 8.59 5.72 4.46 3.77 3.34 3.04 2.83 2.66 2.53 2.43 2.34 2.27 2.20 2.15 2.10 2.06 2.03 1.99 1.96 1.94 1.91 1.89 1.87 1.85 1.84 1.82 1.81 1.79 1.77 1.74 1.69 1.66 1.63 1.59 1.55 1.52 1.50 1.48 1.46 1.44 1.42 1.41 251.5 19.47 8.59 5.71 4.45 3.76 3.33 3.03 2.81 2.65 2.52 2.41 2.33 2.25 2.19 2.14 2.09 2.05 2.01 1.98 1.95 1.92 1.90 1.88 1.86 1.84 1.82 1.80 1.79 1.77 1.75 1.72 1.67 1.64 1.61 1.57 1.53 1.49 1.47 1.46 1.44 1.42 1.40 1.38 251.8 19.48 8.58 5.70 4.44 3.75 3.32 3.02 2.80 2.64 2.51 2.40 2.31 2.24 2.18 2.12 2.08 2.04 2.00 1.97 1.94 1.91 1.88 1.86 1.84 1.82 1.81 1.79 1.77 1.76 1.74 1.70 1.66 1.63 1.60 1.56 1.52 1.48 1.46 1.44 1.42 1.40 1.38 1.36 252.2 19.48 8.57 5.69 4.43 3.74 3.30 3.01 2.79 2.62 2.49 2.38 2.30 2.22 2.16 2.11 2.06 2.02 1.98 1.95 1.92 1.89 1.86 1.84 1.82 1.80 1.79 1.77 1.75 1.74 1.71 1.68 1.64 1.60 1.58 1.53 1.49 1.45 1.43 1.41 1.39 1.37 1.35 1.33 252.6 19.48 8.56 5.68 4.42 3.73 3.29 2.99 2.77 2.60 2.47 2.37 2.28 2.21 2.14 2.09 2.04 2.00 1.96 1.93 1.90 1.87 1.84 1.82 1.80 1.78 1.76 1.75 1.73 1.72 1.69 1.66 1.61 1.58 1.55 1.51 1.47 1.42 1.40 1.38 1.36 1.34 1.32 1.30 253.0 19.49 8.55 5.66 4.41 3.71 3.27 2.97 2.76 2.59 2.46 2.35 2.26 2.19 2.12 2.07 2.02 1.98 1.94 1.91 1.88 1.85 1.82 1.80 1.78 1.76 1.74 1.73 1.71 1.70 1.67 1.63 1.59 1.55 1.52 1.48 1.44 1.39 1.37 1.35 1.33 1.31 1.28 1.26 253.3 19.49 8.55 5.66 4.40 3.70 3.27 2.97 2.75 2.58 2.45 2.34 2.25 2.18 2.11 2.06 2.01 1.97 1.93 1.90 1.87 1.84 1.81 1.79 1.77 1.75 1.73 1.71 1.70 1.68 1.66 1.62 1.58 1.54 1.51 1.47 1.42 1.38 1.35 1.33 1.31 1.29 1.26 1.24 253.4 19.49 8.55 5.65 4.39 3.70 3.26 2.96 2.74 2.57 2.44 2.33 2.25 2.17 2.11 2.05 2.00 1.96 1.92 1.89 1.86 1.83 1.81 1.78 1.76 1.74 1.72 1.71 1.69 1.68 1.65 1.61 1.57 1.53 1.50 1.46 1.41 1.36 1.34 1.32 1.30 1.27 1.25 1.22 253.6 19.49 8.54 5.65 4.39 3.69 3.25 2.95 2.73 2.57 2.43 2.33 2.24 2.16 2.10 2.04 1.99 1.95 1.91 1.88 1.85 1.82 1.79 1.77 1.75 1.73 1.71 1.69 1.68 1.66 1.64 1.60 1.55 1.52 1.49 1.44 1.40 1.35 1.32 1.30 1.28 1.25 1.23 1.20 253.8 19.49 8.54 5.64 4.38 3.69 3.25 2.95 2.73 2.56 2.43 2.32 2.23 2.15 2.09 2.03 1.98 1.94 1.90 1.87 1.84 1.81 1.78 1.76 1.74 1.72 1.70 1.68 1.67 1.65 1.63 1.59 1.54 1.51 1.47 1.43 1.38 1.33 1.30 1.29 1.26 1.23 1.20 1.17
248.8 19.45 8.64 5.78 4.53 3.85 3.42 3.12 2.91 2.75 2.62 2.51 2.43 2.36 2.30 2.24 2.20 2.16 2.12 2.09 2.06 2.04 2.01 1.99 1.97 1.96 1.94 1.92 1.91 1.90 1.87 1.84 1.80 1.77 1.75 1.71 1.67 1.64 1.62 1.61 1.59 1.57 1.56 1.54
249.1 19.45 8.64 5.77 4.53 3.84 3.41 3.12 2.90 2.74 2.61 2.51 2.42 2.35 2.29 2.24 2.19 2.15 2.11 2.08 2.05 2.03 2.01 1.98 1.96 1.95 1.93 1.91 1.90 1.89 1.86 1.83 1.79 1.76 1.74 1.70 1.66 1.63 1.61 1.60 1.58 1.56 1.54 1.53
249.3 19.46 8.63 5.77 4.52 3.83 3.40 3.11 2.89 2.73 2.60 2.50 2.41 2.34 2.28 2.23 2.18 2.14 2.11 2.07 2.05 2.02 2.00 1.97 1.96 1.94 1.92 1.91 1.89 1.88 1.85 1.82 1.78 1.75 1.73 1.69 1.65 1.62 1.60 1.58 1.57 1.55 1.53 1.52
254.0 19.49 8.53 5.64 4.38 3.68 3.24 2.94 2.72 2.55 2.42 2.31 2.22 2.15 2.08 2.02 1.98 1.93 1.89 1.86 1.83 1.80 1.77 1.75 1.73 1.71 1.69 1.67 1.66 1.64 1.61 1.58 1.53 1.49 1.46 1.41 1.37 1.31 1.29 1.27 1.24 1.21 1.18 1.14
254.2 19.49 8.53 5.63 4.37 3.67 3.23 2.93 2.71 2.54 2.41 2.30 2.21 2.14 2.07 2.02 1.97 1.92 1.88 1.85 1.82 1.79 1.76 1.74 1.72 1.70 1.68 1.66 1.65 1.63 1.60 1.57 1.52 1.48 1.45 1.40 1.35 1.30 1.27 1.25 1.22 1.18 1.15 1.11
739
Denominator df
740
3 55.8 9.24 5.34 4.11 3.52 3.18 2.96 2.81 2.69 2.61 2.54 2.48 2.43 2.39 2.36 2.33 2.31 2.29 2.27 2.25 2.23 2.22 2.21 2.19 2.18 2.17 2.17 2.16 2.15 2.14 2.13 2.11 2.09 2.07 2.06 2.04 2.02 2.00 1.99 1.99 1.98 1.97 1.96 1.95 57.2 9.29 5.31 4.05 3.45 3.11 2.88 2.73 2.61 2.52 2.45 2.39 2.35 2.31 2.27 2.24 2.22 2.20 2.18 2.16 2.14 2.13 2.11 2.10 2.09 2.08 2.07 2.06 2.06 2.05 2.04 2.02 2.00 1.98 1.97 1.95 1.93 1.91 1.90 1.89 1.88 1.87 1.86 1.85 58.2 9.33 5.28 4.01 3.40 3.05 2.83 2.67 2.55 2.46 2.39 2.33 2.28 2.24 2.21 2.18 2.15 2.13 2.11 2.09 2.08 2.06 2.05 2.04 2.02 2.01 2.00 2.00 1.99 1.98 1.97 1.95 1.93 1.91 1.90 1.87 1.85 1.83 1.82 1.82 1.81 1.80 1.79 1.78 58.9 9.35 5.27 3.98 3.37 3.01 2.78 2.62 2.51 2.41 2.34 2.28 2.23 2.19 2.16 2.13 2.10 2.08 2.06 2.04 2.02 2.01 1.99 1.98 1.97 1.96 1.95 1.94 1.93 1.93 1.91 1.90 1.87 1.85 1.84 1.82 1.80 1.78 1.77 1.76 1.75 1.74 1.73 1.72 59.4 9.37 5.25 3.95 3.34 2.98 2.75 2.59 2.47 2.38 2.30 2.24 2.20 2.15 2.12 2.09 2.06 2.04 2.02 2.00 1.98 1.97 1.95 1.94 1.93 1.92 1.91 1.90 1.89 1.88 1.87 1.85 1.83 1.81 1.80 1.77 1.75 1.73 1.72 1.71 1.70 1.69 1.69 1.68 59.9 9.38 5.24 3.94 3.32 2.96 2.72 2.56 2.44 2.35 2.27 2.21 2.16 2.12 2.09 2.06 2.03 2.00 1.98 1.96 1.95 1.93 1.92 1.91 1.89 1.88 1.87 1.87 1.86 1.85 1.83 1.82 1.79 1.77 1.76 1.74 1.72 1.69 1.68 1.68 1.67 1.66 1.65 1.64 60.2 9.39 5.23 3.92 3.30 2.94 2.70 2.54 2.42 2.32 2.25 2.19 2.14 2.10 2.06 2.03 2.00 1.98 1.96 1.94 1.92 1.90 1.89 1.88 1.87 1.86 1.85 1.84 1.83 1.82 1.81 1.79 1.76 1.74 1.73 1.71 1.69 1.66 1.65 1.64 1.63 1.62 1.61 1.61 60.5 9.40 5.22 3.91 3.28 2.92 2.68 2.52 2.40 2.30 2.23 2.17 2.12 2.07 2.04 2.01 1.98 1.95 1.93 1.91 1.90 1.88 1.87 1.85 1.84 1.83 1.82 1.81 1.80 1.79 1.78 1.76 1.74 1.72 1.70 1.68 1.66 1.64 1.63 1.62 1.61 1.60 1.59 1.58 60.7 9.41 5.22 3.90 3.27 2.90 2.67 2.50 2.38 2.28 2.21 2.15 2.10 2.05 2.02 1.99 1.96 1.93 1.91 1.89 1.87 1.86 1.84 1.83 1.82 1.81 1.80 1.79 1.78 1.77 1.76 1.74 1.71 1.70 1.68 1.66 1.63 1.61 1.60 1.59 1.58 1.57 1.56 1.55 60.9 9.41 5.21 3.89 3.26 2.89 2.65 2.49 2.36 2.27 2.19 2.13 2.08 2.04 2.00 1.97 1.94 1.92 1.89 1.87 1.86 1.84 1.83 1.81 1.80 1.79 1.78 1.77 1.76 1.75 1.74 1.72 1.70 1.68 1.66 1.64 1.61 1.59 1.58 1.57 1.56 1.55 1.54 1.53 61.1 9.42 5.20 3.88 3.25 2.88 2.64 2.48 2.35 2.26 2.18 2.12 2.07 2.02 1.99 1.95 1.93 1.90 1.88 1.86 1.84 1.83 1.81 1.80 1.79 1.77 1.76 1.75 1.75 1.74 1.72 1.70 1.68 1.66 1.64 1.62 1.60 1.57 1.56 1.55 1.54 1.53 1.52 1.51 61.2 9.42 5.20 3.87 3.24 2.87 2.63 2.46 2.34 2.24 2.17 2.10 2.05 2.01 1.97 1.94 1.91 1.89 1.86 1.84 1.83 1.81 1.80 1.78 1.77 1.76 1.75 1.74 1.73 1.72 1.71 1.69 1.66 1.64 1.63 1.60 1.58 1.56 1.55 1.54 1.53 1.51 1.50 1.49 61.3 9.43 5.20 3.86 3.23 2.86 2.62 2.45 2.33 2.23 2.16 2.09 2.04 2.00 1.96 1.93 1.90 1.87 1.85 1.83 1.81 1.80 1.78 1.77 1.76 1.75 1.74 1.73 1.72 1.71 1.69 1.67 1.65 1.63 1.61 1.59 1.57 1.54 1.53 1.52 1.51 1.50 1.49 1.48 61.5 9.43 5.19 3.86 3.22 2.85 2.61 2.45 2.32 2.22 2.15 2.08 2.03 1.99 1.95 1.92 1.89 1.86 1.84 1.82 1.80 1.79 1.77 1.76 1.75 1.73 1.72 1.71 1.71 1.70 1.68 1.66 1.64 1.62 1.60 1.58 1.55 1.53 1.52 1.51 1.50 1.49 1.47 1.46 61.6 9.44 5.19 3.85 3.22 2.85 2.61 2.44 2.31 2.22 2.14 2.08 2.02 1.98 1.94 1.91 1.88 1.85 1.83 1.81 1.79 1.78 1.76 1.75 1.74 1.72 1.71 1.70 1.69 1.69 1.67 1.65 1.62 1.60 1.59 1.56 1.54 1.52 1.50 1.50 1.48 1.47 1.46 1.45 61.7 9.44 5.19 3.85 3.21 2.84 2.60 2.43 2.30 2.21 2.13 2.07 2.01 1.97 1.93 1.90 1.87 1.84 1.82 1.80 1.78 1.77 1.75 1.74 1.73 1.71 1.70 1.69 1.68 1.68 1.66 1.64 1.61 1.59 1.58 1.55 1.53 1.50 1.49 1.48 1.47 1.46 1.45 1.44 61.7 9.44 5.18 3.84 3.21 2.84 2.59 2.42 2.30 2.20 2.12 2.06 2.01 1.96 1.92 1.89 1.86 1.84 1.81 1.79 1.78 1.76 1.74 1.73 1.72 1.71 1.70 1.69 1.68 1.67 1.65 1.63 1.61 1.58 1.57 1.54 1.52 1.49 1.48 1.47 1.46 1.45 1.44 1.43 61.8 9.44 5.18 3.84 3.20 2.83 2.59 2.42 2.29 2.19 2.12 2.05 2.00 1.96 1.92 1.88 1.86 1.83 1.81 1.79 1.77 1.75 1.74 1.72 1.71 1.70 1.69 1.68 1.67 1.66 1.64 1.62 1.60 1.58 1.56 1.53 1.51 1.48 1.47 1.46 1.45 1.44 1.43 1.42 4 5 6 7 8 9 10 Numerator df 11 12 13 14 15 16 17 18 19 20 21
Table F (cont.) A 0.1 1
22 61.9 9.45 5.18 3.84 3.20 2.83 2.58 2.41 2.29 2.19 2.11 2.05 1.99 1.95 1.91 1.88 1.85 1.82 1.80 1.78 1.76 1.74 1.73 1.71 1.70 1.69 1.68 1.67 1.66 1.65 1.64 1.62 1.59 1.57 1.55 1.53 1.50 1.48 1.46 1.45 1.44 1.43 1.42 1.41
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 32 35 40 45 50 60 75 100 120 140 180 250 400 1000
39.9 8.53 5.54 4.54 4.06 3.78 3.59 3.46 3.36 3.29 3.23 3.18 3.14 3.10 3.07 3.05 3.03 3.01 2.99 2.97 2.96 2.95 2.94 2.93 2.92 2.91 2.90 2.89 2.89 2.88 2.87 2.85 2.84 2.82 2.81 2.79 2.77 2.76 2.75 2.74 2.73 2.73 2.72 2.71
49.5 9.00 5.46 4.32 3.78 3.46 3.26 3.11 3.01 2.92 2.86 2.81 2.76 2.73 2.70 2.67 2.64 2.62 2.61 2.59 2.57 2.56 2.55 2.54 2.53 2.52 2.51 2.50 2.50 2.49 2.48 2.46 2.44 2.42 2.41 2.39 2.37 2.36 2.35 2.34 2.33 2.32 2.32 2.31
53.6 9.16 5.39 4.19 3.62 3.29 3.07 2.92 2.81 2.73 2.66 2.61 2.56 2.52 2.49 2.46 2.44 2.42 2.40 2.38 2.36 2.35 2.34 2.33 2.32 2.31 2.30 2.29 2.28 2.28 2.26 2.25 2.23 2.21 2.20 2.18 2.16 2.14 2.13 2.12 2.11 2.11 2.10 2.09
Table F (cont.) 23 A 0.1 26 27 28 29 30 32 35 50 60 75 100 120 140 180 250 400
24
25
Numerator df 40 45
1000
Denominator df
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 32 35 40 45 50 60 75 100 120 140 180 250 400 1000 62.1 9.45 5.17 3.83 3.18 2.81 2.57 2.40 2.27 2.17 2.09 2.03 1.97 1.93 1.89 1.86 1.83 1.80 1.78 1.76 1.74 1.72 1.70 1.69 1.68 1.67 1.65 1.64 1.63 1.63 1.61 1.59 1.56 1.54 1.52 1.50 1.47 1.45 1.43 1.42 1.41 1.40 1.39 1.38 62.1 9.45 5.17 3.82 3.18 2.81 2.56 2.39 2.26 2.17 2.09 2.02 1.97 1.92 1.88 1.85 1.82 1.80 1.77 1.75 1.73 1.72 1.70 1.69 1.67 1.66 1.65 1.64 1.63 1.62 1.60 1.58 1.56 1.53 1.52 1.49 1.47 1.44 1.43 1.42 1.40 1.39 1.38 1.37 62.2 9.46 5.17 3.82 3.18 2.81 2.56 2.39 2.26 2.16 2.08 2.02 1.96 1.92 1.88 1.85 1.82 1.79 1.77 1.75 1.73 1.71 1.69 1.68 1.67 1.66 1.64 1.63 1.62 1.62 1.60 1.58 1.55 1.53 1.51 1.49 1.46 1.43 1.42 1.41 1.40 1.39 1.37 1.36 62.2 9.46 5.17 3.82 3.18 2.80 2.56 2.39 2.26 2.16 2.08 2.01 1.96 1.92 1.88 1.84 1.81 1.79 1.76 1.74 1.72 1.71 1.69 1.68 1.66 1.65 1.64 1.63 1.62 1.61 1.59 1.57 1.55 1.52 1.51 1.48 1.45 1.43 1.41 1.41 1.39 1.38 1.37 1.36 62.3 9.46 5.17 3.82 3.17 2.80 2.56 2.38 2.25 2.16 2.08 2.01 1.96 1.91 1.87 1.84 1.81 1.78 1.76 1.74 1.72 1.70 1.69 1.67 1.66 1.65 1.64 1.63 1.62 1.61 1.59 1.57 1.54 1.52 1.50 1.48 1.45 1.42 1.41 1.40 1.39 1.37 1.36 1.35 62.3 9.46 5.17 3.81 3.17 2.80 2.55 2.38 2.25 2.15 2.07 2.01 1.95 1.91 1.87 1.83 1.80 1.78 1.75 1.73 1.71 1.69 1.68 1.66 1.65 1.64 1.63 1.62 1.61 1.60 1.58 1.56 1.53 1.51 1.49 1.47 1.44 1.41 1.40 1.39 1.38 1.36 1.35 1.34 62.4 9.46 5.16 3.81 3.16 2.79 2.54 2.37 2.24 2.14 2.06 2.00 1.94 1.90 1.86 1.82 1.79 1.77 1.74 1.72 1.70 1.68 1.67 1.65 1.64 1.63 1.62 1.61 1.60 1.59 1.57 1.55 1.52 1.50 1.48 1.45 1.43 1.40 1.39 1.38 1.36 1.35 1.34 1.32 62.5 9.47 5.16 3.80 3.16 2.78 2.54 2.36 2.23 2.13 2.05 1.99 1.93 1.89 1.85 1.81 1.78 1.75 1.73 1.71 1.69 1.67 1.66 1.64 1.63 1.61 1.60 1.59 1.58 1.57 1.56 1.53 1.51 1.48 1.46 1.44 1.41 1.38 1.37 1.36 1.34 1.33 1.32 1.30 62.6 9.47 5.16 3.80 3.15 2.77 2.53 2.35 2.22 2.12 2.04 1.98 1.92 1.88 1.84 1.80 1.77 1.74 1.72 1.70 1.68 1.66 1.64 1.63 1.62 1.60 1.59 1.58 1.57 1.56 1.54 1.52 1.49 1.47 1.45 1.42 1.40 1.37 1.35 1.34 1.33 1.31 1.30 1.29 62.7 9.47 5.15 3.80 3.15 2.77 2.52 2.35 2.22 2.12 2.04 1.97 1.92 1.87 1.83 1.79 1.76 1.74 1.71 1.69 1.67 1.65 1.64 1.62 1.61 1.59 1.58 1.57 1.56 1.55 1.53 1.51 1.48 1.46 1.44 1.41 1.38 1.35 1.34 1.33 1.32 1.30 1.29 1.27 62.8 9.47 5.15 3.79 3.14 2.76 2.51 2.34 2.21 2.11 2.03 1.96 1.90 1.86 1.82 1.78 1.75 1.72 1.70 1.68 1.66 1.64 1.62 1.61 1.59 1.58 1.57 1.56 1.55 1.54 1.52 1.50 1.47 1.44 1.42 1.40 1.37 1.34 1.32 1.31 1.29 1.28 1.26 1.25 62.9 9.48 5.15 3.78 3.13 2.75 2.51 2.33 2.20 2.10 2.02 1.95 1.89 1.85 1.80 1.77 1.74 1.71 1.69 1.66 1.64 1.63 1.61 1.59 1.58 1.57 1.55 1.54 1.53 1.52 1.50 1.48 1.45 1.43 1.41 1.38 1.35 1.32 1.30 1.29 1.27 1.26 1.24 1.23 63.0 9.48 5.14 3.78 3.13 2.75 2.50 2.32 2.19 2.09 2.01 1.94 1.88 1.83 1.79 1.76 1.73 1.70 1.67 1.65 1.63 1.61 1.59 1.58 1.56 1.55 1.54 1.53 1.52 1.51 1.49 1.47 1.43 1.41 1.39 1.36 1.33 1.29 1.28 1.26 1.25 1.23 1.21 1.20 63.1 9.48 5.14 3.78 3.12 2.74 2.49 2.32 2.18 2.08 2.00 1.93 1.88 1.83 1.79 1.75 1.72 1.69 1.67 1.64 1.62 1.60 1.59 1.57 1.56 1.54 1.53 1.52 1.51 1.50 1.48 1.46 1.42 1.40 1.38 1.35 1.32 1.28 1.26 1.25 1.23 1.22 1.20 1.18 63.1 9.48 5.14 3.77 3.12 2.74 2.49 2.31 2.18 2.08 2.00 1.93 1.87 1.82 1.78 1.75 1.71 1.69 1.66 1.64 1.62 1.60 1.58 1.57 1.55 1.54 1.53 1.51 1.50 1.49 1.47 1.45 1.42 1.39 1.37 1.34 1.31 1.27 1.26 1.24 1.22 1.21 1.19 1.17 63.1 9.49 5.14 3.77 3.12 2.74 2.49 2.31 2.18 2.07 1.99 1.92 1.87 1.82 1.78 1.74 1.71 1.68 1.65 1.63 1.61 1.59 1.57 1.56 1.54 1.53 1.52 1.51 1.50 1.49 1.47 1.44 1.41 1.38 1.36 1.33 1.30 1.26 1.24 1.23 1.21 1.19 1.17 1.15 63.2 9.49 5.14 3.77 3.11 2.73 2.48 2.30 2.17 2.07 1.99 1.92 1.86 1.81 1.77 1.73 1.70 1.67 1.65 1.62 1.60 1.59 1.57 1.55 1.54 1.52 1.51 1.50 1.49 1.48 1.46 1.43 1.40 1.37 1.35 1.32 1.29 1.25 1.23 1.22 1.20 1.18 1.16 1.13
61.9 9.45 5.18 3.83 3.19 2.82 2.58 2.41 2.28 2.18 2.11 2.04 1.99 1.94 1.90 1.87 1.84 1.82 1.79 1.77 1.75 1.74 1.72 1.71 1.70 1.68 1.67 1.66 1.65 1.64 1.63 1.61 1.58 1.56 1.54 1.52 1.49 1.47 1.46 1.45 1.43 1.42 1.41 1.40
62.0 9.45 5.18 3.83 3.19 2.82 2.58 2.40 2.28 2.18 2.10 2.04 1.98 1.94 1.90 1.87 1.84 1.81 1.79 1.77 1.75 1.73 1.72 1.70 1.69 1.68 1.67 1.66 1.65 1.64 1.62 1.60 1.57 1.55 1.54 1.51 1.49 1.46 1.45 1.44 1.43 1.41 1.40 1.39
62.1 9.45 5.17 3.83 3.19 2.81 2.57 2.40 2.27 2.17 2.10 2.03 1.98 1.93 1.89 1.86 1.83 1.80 1.78 1.76 1.74 1.73 1.71 1.70 1.68 1.67 1.66 1.65 1.64 1.63 1.62 1.60 1.57 1.55 1.53 1.50 1.48 1.45 1.44 1.43 1.42 1.41 1.39 1.38
63.2 9.49 5.14 3.77 3.11 2.73 2.48 2.30 2.17 2.06 1.98 1.91 1.86 1.81 1.76 1.73 1.70 1.67 1.64 1.62 1.60 1.58 1.56 1.54 1.53 1.52 1.50 1.49 1.48 1.47 1.45 1.43 1.39 1.37 1.34 1.31 1.27 1.24 1.22 1.20 1.18 1.16 1.14 1.11
63.3 9.49 5.1 3.76 3.11 2.72 2.47 2.30 2.16 2.06 1.98 1.91 1.85 1.80 1.76 1.72 1.69 1.66 1.64 1.61 1.59 1.57 1.55 1.54 1.52 1.51 1.50 1.48 1.47 1.46 1.44 1.42 1.38 1.36 1.33 1.30 1.26 1.22 1.20 1.19 1.16 1.14 1.12 1.08
741

Chapter 24 Analysis of Variance

Uploaded by

Copyright:

Available Formats

Chapter 24 Analysis of Variance

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Chapter 24 Analysis of Variance

Uploaded by

Copyright:

Available Formats

24.

of Several Groups Are Equal

Where are we going?

Who What How

Copyright 2014 Pearson Education, Inc.

Part VII Inference When Variables Are Related