Financial Time Series I
Topic 3: Analysis of Variance
Department of Mathematics National Taiwan University
1. One-Way Analysis of Variance – Introduction
– Examples – Basic Set-up – Data Example – Algebra
– Validality Conditions on F -test 2. Analysis of Variance of Ranks
3. Logistic Regression
4. Two-Way Analysis of Variance – Interaction
– Data Example
– Related R-command – Design
Comparison of Several Populations
• When some measurement, such as height or aptitude for a particular job, is made on sev- eral individuals, the values vary from person to person.
– The variability of a quantitative scale is measured by its variance.
– If the set of individuals is stratified into more homogeneous groups, the variance of the measurements within the more ho- mogeneous group will be less than that of the measurements in the entire group;
that is what “more homogeneous” means.
• As an example, consider the heights of pupils on an elementary school.
– Fact 1: The variance of the heights of pupils on an elementary school is usually greater than the variance of heights of pupils in just the first grade, the variance in the second grade, and the variance in each of the other grades.
The within-grade variances are less than the overall variance
– Fact 2: The average height of pupils also varies from grade to grade.
The averages vary between grades.
– The total variability (of heights) is made up of two components: the variability of individuals within groups (grades) and the variability of means between groups (grades).
V ar(Y ) = EV ar(Y |X)+V ar(E(Y |X)).
Y refers to the height of students and X refers to the grades.
– At the extreme, all of the variability of a measured variable may be within the groups and none of it between groups, that is, the means of the subgroups are equal. Mathematically, it refers to the case that E(Y |X) = E(Y ).
• The analysis of variance is a set of statisti- cal techniques for studying variability from different sources and comparing them to un- derstand the relative importance of each of the sources.
– It is also used to make inferences about the population through tests of signifi- cance, including the very important com- parison of the means of two or more sep- arate populations.
– The analysis of variance is the most straight- forward way to examine the association between a categorical variable (“groups”) and a numerical variable (the measure on which the means are based)?
– In the above example, we try to find the association between height of student and grade of student.
One-Way Analysis of Variance
Suppose that we wish to compare the means of three populations on some measured dependent variable.
• The population means are represented as µ1, µ2, and µ3.
• A hypothesis that may be tested is H0 : µ1 = µ2 = µ3
that is, that the means of the three popula- tions are equal.
• The hypothesis is rejected if the population means are different in any way, for example, if one of the means differs from the other two or if all three means are different.
• The procedure for testing H0 through the analysis of variance parallels that for other tests of significance.
• Samples of data are obtained and a test statis- tics is computed. The test statistic is pro- posed based on the following idea.
– The total variability of the data is allo- cated to two sources: variability among the group means and variability of the observations within the groups.
– Two measures of variability are computed, Mean Square between Groups and Mean Square within Groups, respectively.
– The hypothesis that the 3 population means are equal is tested by comparing the mag- nitude of these two mean squares.
It is the ratio of Mean Square Between Groups to Mean Square Within Groups.
– We need a reference, the F -distribution, to provide significance points for deciding if the ratio is large enough to reject the hypothesis of equal means.
• In this application the test statistic reflects the extent to which variation among the sam- ple means is greater than variation among observations within the groups.
• If the test statistic is large enough (that is, if it exceeds the corresponding significance point) then H0 is rejected.
• If the sample means are close together and hence the test statistic is small, then H0 is accepted.
• Can we apply this procedure for comparing the means of two populations?
• When only two groups are being compared, it is obvious from knowledge of the sample means which group has a significantly higher mean than the other.
– When there are three or more groups, however, rejection of H0 only means that at least one population mean is different from the others.
– It is necessary to follow the test of signif- icance by additional analysis-of-variance procedures to determine which if the pop- ulation means are different from which other means.
– What are these follow-up procedures?
• The “groups” compared through analysis of variance may be created in a number of dif- ferent ways.
– In many studies the data are already clas- sified into groups, such as states, coun- tries, gender groups, medical diagnostic categories, or religious affiliations.
– In other instances, the statistician may define the groups based on one or more measured characteristics.
For example, socioeconomic classifications may be created by combining individu- als’ educational levels, incomes, and em- ployment statuses according to some set of rules. The result might be, say, three or four or five separate socioeconomic cat- egories.
– A medical research might classify patients as “high risk” or “low risk” for some dis- ease based on a number of indicators, for example, age, family history, and health behaviors.
– In conducting surveys, groups are fre- quently formed from the respondents’ an- swers, for example, all people who sup- port a particular political issue, those who are opposed, and those who have no opin- ion.
– In an experiment, the groups are defined by the various experimental conditions or
“treatments,” such as instructional meth- ods, psychological or social interventions, or dosages or different forms of a drug.
In these situations, individuals are as- signed to the conditions by the researcher.
When possible, the assigned should be done randomly so that the only differ- ence between the groups is the explained by differences in the treatments received and not by other factors that are irrele- vant to the study.
A psychologist is interested in comparing the effects of three different informational “sets” in children’s ability to memorize words.
• Eighteen 7-year-old children were randomly assigned to one of three groups.
– In condition 1, the children were shown a list of 12 words and were asked to study them in preparation for a recall test.
– In condition 2, the children were told that the words comprised three global cate- gories, flowers, animals, and foods, and were asked to study the list.
– In condition 3, the children were told that the words comprised 6 more detailed cat- egories.
• After studying the word list for 10 minutes, each youngster was asked to list all of the words that he/she could remember.
• The number of correctly recalled words was the measured outcome variable of the study.
A study was performed to see if school grades are related to television viewing habits among high-school juniors.
• Respondents were monitored for 15 week- days during the year and classified into 4 groups according to the average amount of television they viewed on those days (0−0.5 hour; 0.5 − 1.5 hours; 1.5 − 3.0 hours; more than three hours).
• Each student’s grade-average (GPA) was recorded for all courses taken during the year.
• Mean GPAs were compared among the four television viewing groups.
In the study The Academic Mind by Lazarsfeld and Thielens, a total of 2451 social science fac- ulty members from 165 of the larger American colleges and universities were interviewed in or- der to assess the impact of the McCarthy era on social science faculties.
• At each college, the number of “academic freedom incidents” was counted.
• These were incidents mentioned by more than one respondent as an attack in the academic freedom of the faculty.
• They ranged from small-scale matters, such as a verbal denunciation of a professor by a student group, to large-scale matters, in- cluding a Congressional investigation.
• It was of interest to examine whether and how the institutional basis of a school’s sup- port and control affected the number of “in- cidents” occurring there.
• Hence, each college was classified as pub- licly controlled, privately controlled, or con- trolled by some other institution. (Teachers’
colleges and schools controlled by a religious institution were included in the “other” cat- egory.)
• The distributions of numbers of “incidents”
in the different types of institutions were studied.
A large manufacturing firm employs high-school
dropouts, high-school graduates, and individu- als who attended college as production-line work- ers.
• The company management speculated that job proficiency was related to educational attainment.
• If so, only high-school graduates, or perhaps only individuals who had attended college, would be hired in the future.
• To test their idea, the job performance of a sample of employees was rated by their su- pervisors on an extensive rating scale that yield possible proficiency scores ranging from 0 to 200.
• The mean ratings of the three education groups (high-school dropouts, high-school graduates, college attendees) were compared by the anal- ysis of variance.
A Complete Example with equal sample size The analysis of variance indicates whether pop- ulation means are different by comparing the variability among sample means with variability among individual observations within groups.
The following data gives the numbers of words memorized by 18 children given three different information sets as described in Example 1.
No information 3 categories 6 categories
2 9 4
4 10 5
3 10 6
4 7 3
5 8 7
6 10 5
Sums 24 54 30
Sample means 4 9 5
• The scores for three samples and pooled sam- ple can be graphed by box plot.
• The means of the three samples are 4, 9, and 5, respectively.
• There is clear variability from group to group.
• The variability in the entire pooled sample of 18 is much larger than the variability of these three groups.
• Variability within Groups: The measure of variability among observations within the groups is called the Mean Square Groups and is denoted M Sw.
– M Sw is the variance of all the scores, but computed separately for each of the three samples and then combined.
– Like the sample variance, M Sw has a sum of squared deviations in the numera- tor and the corresponding number of de- grees of freedom in the denominator.
– The numerator is called the Sum of Squares within Groups and the denominator is called the degrees of freedom within groups.
– The Sum of Squares within Groups (SSw) is the sum of squared deviations of indi- vidual scores from their subgroup means.
In this example, SSw = 10+8+10 = 28.
– The degrees of freedom within groups (dfw) is the total number of degrees of freedom
of the deviations within the groups.
In this example, dfw = (6 − 1) + 5 + 5 = 15.
• Variability between Groups
– The measure of variability among the group means is called the Mean Square between Groups and is denoted M SB.
– M SB is similar to a variance but is com- puted from the three subgroup means, 4, 9, and 5.
– M SB also has a sum of squared devia- tions in the numerator and a correspond- ing number of degrees of freedom in the denominator is called the Sum of Squares between Groups and the denominator is called the degrees of freedom between groups.
– The sum of Squares between Groups (SSB) is the sum of squared deviations of the subgroup means from the overall or “pooled”
∗ The pooled mean of all 18 observa- tions is 6.
∗ The squared deviations of the group
means from the pooled mean are (4 − 6)2 + (9 − 6)2 + (5 − 6)2 = 14.
∗ Before dividing by the degrees of free- dom between groups, one additional adjustment is necessary.
To test the null hypothesis of equal means, SSB is going to be compared to SSw by converting each to a mean square and then dividing one mean square by the other.
SSB cannot be meaningfully compared to SSw because SSB reflects variabil- ity among means while SSW reflects variability among individual measure- ments.
Put SSB in a comparable scale to SSw, it is multiplied by the number of ob- servations in each subgroup, that is, 6 in this example.
∗ SSB = 6 × 14 = 84.
– The degrees of freedom between groups (dfB) is the number of independent devi- ations summarized in SSB.
In this example, dfB = 3 − 1 = 2.
– In this example, M SB = 84/2 = 42.
• The Test Statistic:
– The test statistic to test H0 : µ1 = µ2 = µ3 is the ratio of M SB/M Sw, or 42/1.867 = 22.50.
How do you connect them with V ar(Y ) and V ar(E(Y |X))?
– This ratio seems to indicate that the vari- ability among groups is much greater than that within groups.
However, we know that different samples would give us different values of the ra- tio even of the three population means are equal.
– The calculated ratio of 22.50 is compared to values from the F -distribution to see if the test statistic is large enough for 3 samples and 18 observations to reject H0.
• Conclusion: If we are using a 1% significance level, the significance point from the table of the F -distribution with 2 and 15 degrees of freedom is 6.36.
Thus the calculated ratio of 22.50 is very sig-
nificance and we conclude that there are real differences in the average number of words memorized depending on the amount of or- ganizing information that is provided.
• The follow-up question remains, specifically:
Which of the three means are significantly different from which others?
The Algebra of ANOVA
• Suppose the number of groups being com- pared is k.
• The null hypothesis is
H0 : µ1 = µ2 = · · · = µk,
that is, the k population means are equal.
• Notation and Data:
1 2 · · · k Population Mean µ1 µ2 · · · µk
Variance σ12 σ22 · · · σk2 Sample Observations x11 x21 · · · xk1
x12 x22 · · · xk2 ... ... ...
Sample size n1 n2 · · · nk Mean x¯1 x¯2 · · · ¯xk Variance s21 s22 · · · s2k
• In R, try help.search(“anova”) and look at the help files of aov and manov.
• The number of sampled observations in a
“typical” group, group g, is denoted by ng
and the total number of observations is n = n1 + n2 + · · · + nk.
• Each observation has two subscripts, the first indicating the group to which the observa- tion belongs and the second indicating the observation number within that group. Thus, xgi represents the ith observation in group g.
– The sample mean of the ith subgroup is denoted by ¯xg = n−1g Pi xgi.
– The overall mean is denoted by ¯x = n−1 Pg Pi xgi.
• The ANOVA Decomposition: The “decom- position” of a single typical score xgi as
xgi − ¯x = (¯xg − ¯x) + (xgi − ¯xg).
– This expression shows that the deviation of a score from the overall mean ¯x (V ar(Y )) can be decomposed as the sum of two parts, the deviation of the group mean from the overall mean (V ar(E(Y |X))) and the deviation of an observation from its group mean (V ar(Y |X)).
– The terms on the right-hand side of the
above expression are the fundamental el- ements of “between-group” and “within- group” variability, respectively.
– By squaring and summing the terms in the expression we obtain exactly the sum- of-squares partition essential for the anal- ysis of variance.
• Total Sum of Squares:
SST = X
i (xgi − ¯x)2.
• By the property of least squares method, we have
SST = SSB + SSw.
• Between-group Sum of Squares and Mean Square:
SSB = X
g ng(¯xg − ¯x)2
and M SB = SSB/dfB where dfB = k − 1.
• Within-group Sum of Squares and Mean Square:
– SSw can be obtained by subtracting SSB from SST. (Recall the geometric interp- tation of least squares method.)
– dfw is the number of independent devia- tions of observed values from their sub- group means. There are ng observations in subgroup g, then the number of inde- pendent deviations in that group is ng − 1.
Summing these across the k groups, we have dfw = n − k.
– The Mean Square within Groups is M Sw = SSw/dfw.
• The F -ratio, F = M SB/M Sw.
Validity Conditions for the F -test
The analysis of variance rests on a set of as- sumptions that should be considered before the procedure is applied.
• The most inflexible assumption is that the n observations must be independent.
– Sampling observations at random is an important step in assuring that this con- dition is met.
– Steps should be taken to assure that the response on the measured outcome vari- able of no respondent is in any way af- fected by other respondents.
If observations are humans, they should not have the opportunity to hear, see, or otherwise be influenced by other sub- jects’ answers of behavior.
– If the assumption of independence is not met, then the analysis of variance tests of significance are not generally valid.
• The sampling distribution of the subgroup means should be nearly normal.
– This condition is usually met in analysis of variance, especially if the subgroup n’s are moderate to large, since the Central Limit Theorem assures us that means based on large sample sizes are nearly normally distributed.
– The normality condition will also be met if the distribution of the underlying mea- sured variable is normal, whether or not the sample sizes are large.
– It is always advisable to make histograms of the data to see whether there are any gross irregularities in the distributions, however.
– If the sample sizes are small and the dis- tribution of the measured variable is highly non-normal, then the analysis of variance of ranks should be used in place of the ANOVA methods presented so far.
• The F -test is based on an assumption that the population variances are all equal, that is σ12 = · · · = σk2.
– This condition is especially important if
the sample sizes (n1, n2, . . . , nk) are not equal.
– Sample variances should be computed prior to conducting the analysis of variance to see if they are in the same general range as one another.
If they appear to be very different, a for- mal test of equality of the σ2’s may be conducted.
– If the test indicates that the variances are not homogeneous, several options may be available.
The data may be transformed to a scale on which the variances are more equal.
For example, this might involve analyz- ing the logarithms of the original observed values, the square roots of the observed values, or some other function of the data.
Which Groups Differ from Which, and by How Much?
The F -test gives information about all the means µ1, µ2, . . . , µk simultaneously.
• If the hypothesis of equal means is accepted, the conclusion is that the data do not indi- cate differences among the population means.
• If the null hypothesis is rejected, the conclu- sion is that there are some differences; then the researcher may want to know which spe- cific means are significantly different from which other and the direction and the mag- nitudes of the differences.
What is the methods for comparing two means?
• Construct a confidence interval for the dif- ference of the two means as follows:
n−11 + n−12 , x¯1−¯x2+tn−2(α/2)s
n−11 + n−12 ).
Extend the above idea to the comparison of three or more groups.
• When three or more groups are compared by the analysis of variance, several specific
comparisons may be made after the overall hypothesis of equality is rejected.
– For example, if there are three groups and H0 is rejected, µ1 may be compared with µ2, µ1 may be compared with µ3, and µ2 may be compared with µ3.
– It is up to the researcher to decide which comparisons to make. The decision rests partially on the design of the research.
For example, if group 1 is a control group and groups 2 and 3 are two different ex- perimental conditions, then it would be sensible to compare µ2 with µ1 and µ3 with µ1, that is, both experimental con- ditions with the control.
– If students from four different universi- ties are being compared on mean scores on the Law School Admissions Test, ev- ery school’s mean might be routinely com- pared to every other school’s mean. In the latter case, the number of pairwise comparisons among k means is k × (k − 1)/2.
– The procedure for any one of these com- parisons is very much like comparing two groups as described before.
– But one additional factor needs to be considered: the probability of making a Type I error when performing several tests of significance from the same data set.
– If, indeed, all the µ’s were equal, so that there were no real differences, the proba- bility that any particular one of the pair- wise differences would exceed the corre- sponding t-value is α.
However, the probability of making at least one Type I error out of two or more pairwise comparisons is greater than this.
That is, when many differences are tested, the probability that some will appear to be “significant” when the corresponding population means are equal is greater than the nominal significance level α. The more comparisons that are made, the greater the probability of making at least one Type I error.
• How can a researcher protect against too
high a Type I error rate?
– One widely used approach is based on the Bonferroni inequality.
– The Bonferroni inequality states that the probability of making at least one Type I error out of a given set of comparisons is less than or equal to the sum of the α’s used for the separate comparisons.
– For example, if we make 2 comparisons among 3 group means and use an α level of 0.05 for each comparison, the proba- bility of making at least one Type I error is no greater than 0.10.
– This overall α level for the pair or “fam- ily” of tests is called the familywise (or experimentwise) Type I error rate.
– The Bonferroni inequality can be put to use to keep the familywise error rate ac- ceptably small.
– Suppose that we wish to make m spe- cific pairwise comparisons. We can then decide on a reasonable familywise error rate (α) and divide this value by m to
obtain a significance level to be used for each comparison separately; call this re- sult α∗.
– If α∗ is used for each of m comparisons, the probability of making at least one Type I error out of the set is not greater than m × α∗ = α.
Analysis of Variance of Ranks
• In ANOVA, its basic idea is to compare the distributions of a variable in several popula- tions by focusing on the means of the sam- ples.
– The ANOVA is based on the conditions that the sample means have a normal dis- tribution and that the populations being compared have the same variance σ2. – When these conditions are not met, or
when the raw data are ordinal, can we still use ANOVA technique to compare several populations?
– Consider the test procedure developed by Kruskal and Wallis (1952).
• Another approach to compare the distribu- tions of a variable in several populations:
– The locations of distribution can also be compared through an analysis of ranks.
– This approach can be applied even when the data are ordinal, and does not require the assumption of normality or equal vari- ances.
– The null hypothesis is that the locations of the populations are the same, and the alternative hypothesis is that they are not.
• Kruskal and Wallis (1952) considered the daily outputs of three bottle-cap machines.
– Use the following 12 values to test whether the three machines produce equal num- bers of bottle caps.
Machine A 340, 345, 330, 342, 338 Machine B 339, 333, 344
Machine C 347, 343, 349, 355
– Instead of using the above data, we use the following ranks instead.
Ranks Machine A 5, 9, 1, 6, 3 Machine B 4, 2, 8
Machine C 10, 7, 11, 12
– The test applies ANOVA formulas to these ranks. We have
SST = n(n + 1)(2n + 1)
6 − n
n + 1 2
= n(n + 1)(n − 1) 12
and M ST = n(n + 1)/12.
– The F -ratio in ANOVA is F = M SB/M Sw. – Unlike ANOVA, we consider H = M SB/M ST. – Observe that
SSB = X
ng − n
n + 1 2
. – In this example,
H = 12
12 × 13
5 + 142
3 + 402 4
−3×(12+1) = 5.66.
• The null hypothesis of no difference between the locations of the three populations is equiv- alent to tandom sampling, that is, that the ranks have been allocated at random to the k groups.
– When this hypothesis is true, the sam- pling distribution of H is approximately χ2 with k − 1 degrees of freedom if the sample sizes are large.
– Why? What is F -distribution?
Introduction to Logistic Regression
There are many situations in which we want to forecast events which cannot be modeled as a continuous variable. Examples include
• Whether a consumer will purchase a prod- uct from a menu of products, click on a web button, respond to a direct mail offer.
• Whether a firm will decide to repurchase stock, change accounting procedures, or write- off an asset.
• In these cases, the Y variable is not only dis- crete but takes only two values (1 “Event”
or 0 “No Event”).
This is the simplest type of random variable, a Bernoulli random variable. The classic ex- ample of this variable is the outcome of a coin toss.
• We want to make predictions about this dis- crete Y on the basis of some other explana- tory variables.
– For example, we want to study how prod- uct attributes and price influence con- sumer choice from an array of products
or how industry and firm financial con- ditions affect the decision to write-off as- sets.
– To make the conditional predictions, we need to incorporate the X variables into a prediction about Y .
– Unfortunately, the standard linear regres- sion of Y on the X variables is inappro- priate.
Our forecasts should be probabilities!
• If we think back to the introduction of re- gression, we viewed the regression model as a model for the conditional mean of Y given the X variables.
• For a Bernoulli random variable, the condi- tional mean will be a probability.
E(Y |X) = P (Y = 1|X)
• The logit model is one very convenient and useful way of forming probabilties from x variables.
P (Y = 1|x) = exp(β0 + β1X1 + · · · + βkXk) 1 + exp(β0 + β1X1 + · · · + βkXk)
– Think of V = β0 + β1X1 + · · · + βkXk as a “score.”
– As V gets large, the probability Y = 1 should increase (but never exceed 1).
– For example, if a firm is considering write- off an asset, then V might be the “desir- ability” of the write-off.
If V is negative, the write-off is not de- sired.
– The logit model is a particular form for how V gets mapped into a probability.
P (Y = 1) = exp(V ) 1 + exp(V ).
What does this probability curve or locus look like?
– Use R to plot it by yourself.
There are several keys aspects to notice about this curve:
i) As V increases to levels very much above or below 0 the probability of pur- chase goes very close to to 1 or 0.
ii) The sensitivity to changes in V varies depending on the level of utility and as- sociated probability. Around prob = 0.5,
the slope is at its maximum of 0.25. but at low or high probs the slope declines to very small numbers. This is because of the “Logistic” or S-shaped curve.
How do we estimate unknown parameters in such a model?
Use the method called Maximum Likelihood Es- timation (“MLE!L).
• Form a “likelihood function!L for each ob- servation.
exp(Vi) 1 + exp(Vi)
1 + exp(Vi)
where Vi = b0 + b1Xi1 + · · · + bkXik.
• Consider the logarithm of the likelihoods for all the observations.
• Find the values of b0, b1, b2, . . . that make this sum (log likelihood function) as large as possible.
• In R, try help(glm) and help(family) to find the MLE estimate.
– Logistic regression is a special case of gen- eralized linear model.
– We need to specify the error distribution and the link function.
– For logistic regression, the error distri- bution is binomial distribution and logit is being used to link the linear predictor and the mean of the error distribution.
– In glm, the “binomial” family admits the links “logit”, “probit”, “log, and “cloglog”
An Example of Logistic Regression
In “Causes and Effects of Discretionary Asset Write-Offs” by Francis, Hanna and Vincent (JAR 1996),
• The authors investigate the decision of firms to write-off assets by gathering a large sam- ple of firms that made and did not make write-offs in the 1989-92 period.
• The authors propose a set of independent variables which might be expected to influ- ence the write-off decision.
The set of explanatory variables are as fol- lows:
1. ryear1 - cumulative abnormal return over the preceding year
2. ryear5 - cumulative abnormal return 3. mtbdif - firm’s industry-adjusted Book
to Market ratio
4. mtbchg - change in firms B-t-M ratio 5. indmtb - change in industry’s B-t-M
6. roa - change in roa
7. indroa - change in industry roa
8. history - number of yrs in which firm reported special negative items
9. indhis - mean value of history for all other firms in industry
10. dmgmt - 1 if management changed in previsous year
11. poor - Unexpected earnings if U E <
0, 0 otherwise
12. good - UE- $amt of write-off, if > 0, 0 otherwise
13. lnsale - log of sales in yr preceding write-off
14. unless otherwise stated, all variables are averages of preceding 5 years
• They are interested in determining whether write-offs are the result of managers attempts to manipulate accounting performance or sim- ply the result of declines in the value of as- sets.
• The article relates the predicted probability of write-off to changes in stock price, but we will focus on the problem of predicting write-offs given on the basis of the financial
performance of the firm and the industry in which the firm operates.
Two-Way ANOVA; General Model
Two-way ANOVA has two factors of interest that occur in all combinations.
• Sometimes there is only one observation on each combination.
• Sometimes there are more (called “replica- tions”).
• The general model with replications has ob- servations labeled as Yijk.
The index i (i = 1, . . . , I) labels the value of the first factor, j (j = 1, . . . , J ) labels the value of the second factor, and k labels the replicated observations on the combination ij.
• When we always have the same number of replications for each ij combination, so that k = 1, . . . , K. Such a model is called “bal- anced”.
• There are two types of models - the “addi- tive model” and the “additive model with interactions”.
The second model needs K ≥ 2.
• The additive model decomposes the popula- tion means µij = E(Yij) as a sum of effects of the corresponding i and j factors.
µij = µ + αi + βj where Pi αi = 0 and Pj βj = 0..
• Without proper constraint on αi and βj, those unknown parameters are not identi- fiable.
• For the balanced model, µ, αs and βs can be estimated in the natural way (like in the one-way model) as ˆµ = ¯Y..., ˆαi = ¯Yi.. − ¯Y..., βˆj = ¯Y.j. − ¯Y.... These formulas correspond to the natural estimates of the βj. Thus
µij = ¯Yi.. + ¯Y.j. − ¯Y....
(For non-balanced models the correspond- ing formulas are more complicated, which is why we treat only balanced models.)
• There is an overall null hypothesis to be tested:
H0 : α1 = · · · = αI = 0 and β1 = · · · = βJ = 0 versus Ha : H0 is not true.
It includes two sub-hypotheses of interest:
H0A : α1 = · · · = αI = 0 versus HaA : H0A is not true.
H0B : β1 = · · · = βJ = 0 versus H0B : H0B is not true The Data:
Hourly arrivals of telephone calls to a telephone call center
• The data in this example involve telephone calls to a relatively small Israeli Bank tele- phone call center in 1999.
• The caller desires to speak to a telephone service agent.
• The call center managers want to be able to predict the number of calls that will arrive in any given hour.
– The working day at this center runs from 7am to 11:59pm.
– Look at data for all the full work-weeks in November and December 1999.
– Divide each day up into hourly intervals;
from 7 − 8, 8 − 9, etc.
Label these intervals as i = 1, 2, . . . , 17.
Thus interval i = 2 corresponds to 8am V 9am and I = 17.
– There are 5 regular working days each week (Sunday through Thursday in Is- rael.).
Label these as j = 1, . . . , 5.
Thus j = 2 corresponds to a Monday and J = 5.
– Let Nijk denote the number of calls arriv- ing during hourly interval i, day-of-week j, and week k. Note that K = 8.
• Use ANOVA to summarize this data set.
• How could we model Nijk?
– It is reasonable to conjecture that these arrival times are well modeled by an in- homogeneous Poisson process.
– The arrival rate for this process should depend only on time of day, and perhaps other calendar related covariates such as month or day of the week.
– Theory suggests that they may have a Poisson distribution with mean λij.
– If the arrival process for a given call cat- egory is as above then the number of ar- rivals each day within any given interval of time should be independent Poisson variables with a parameter that depends only on the given time interval.
If other covariates are involved (such as day of the week) then the Poisson param- eter may also depend on these.
– Poisson distribution an be approximated by a normal distribution when λij is large.
If so they would not be homoscedastic (since their variance would equal their mean).
– Anscombe’s (1948, Biometrika) variance stabilizing transformation suggests that the variables rNijk might be nearly ho- moscedastic with variance 1/4.
– Consider Yijk = rNijk + 1/4.
• Goal of the Manager:
– The manager of the call center would like to be able to predict the number of cus- tomers in any particular hour that will
call the center desiring to speak to an agent.
– Plan how many agents are needed at that time of day.
– (Other considerations also enter into this decision, such as the length of time that it takes an agent to serve a customer.)
• Next page gives a plot that tells the manager how to predict N for each hour on any day 2 (Monday).
• This plot also tells the manager what the 95% prediction intervals are for that predic- tion.
• Note that the prediction limits are pretty wide.
but it cant be helped if we only know about time-of-day and day-of-week.
• The 95% confidence intervals are also shown, but probably aren’t as important to the man- ager.
The tests of these hypotheses are summarized in an ANOVA table.
Summary of Fit
Root Mean Square Error 0.7645 Mean of Response 7.889
Analysis of Variance
Source DF SS MS F Ratio
Model (I − 1) + (J − 1) = 20 1671.63 83.581 143.00 Error n − I − J + 1 = 659 385.17 0.585 Prob > F C. Total n − 1 = 679 2056.80 < .0001
Source Npar. DF SS F Ratio Prob > F
Hour 16 I − 1 = 16 1658.40 177.34 < .0001 day of week 4 J − 1 = 4 13.23 5.66 0.0002
• Sum of Squares for “C Total” is SST =
Pi Pj Pk(Yijk − ¯Y...)2.
• Sum of Squares for “Error” is SSE = Pi Pj Pk(Yijk− ˆ
• In the model, we consider two factors: hour and day of the week.
SSM odel = SSA + SSB.
Residual by Predicted Plot
-2.0 -1.0 0.0 1.0 2.0 3.0
Y_ijk = "Root"(N_ijk) Residual
4 5 6 7 8 9 10 11 12
Y_ijk = "Root"(N_ijk) Predicted
0 50 100 150
No. of Customers/Hr
7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
time Y Pred N_ij
Lower Conf N_ij Upper Conf N_ij Lower Pred N_ij Upper Pred N_ij
Plot of Predicted Values of N for Mondays, as a function of Hour–of-Day
– SSM odel = SST − SSE.
– SS(Hour) is SSA = Pj Pk Pi( ¯Yi..− ¯Y...)2. – SS(day of week) is SSB = Pi Pk Pj( ¯Y.j.−
• F -ratios:
– The first F -ratio (= 143 = M SM/M SE) tests H0 and here has 20 and 659 df.
– The second F -ratio (= 177.34 = M SA/M SE) tests H0A and here has 16 and 659 df.
– The third F -ratio (= 5.66 = M SB/M SE) tests H0B and here has 4 and 659 df.
• The analysis tells us that knowing the day of the week makes a statistically significant dif- ference, but not a very important one; one could do almost as well just knowing the time of day. This means that there is not much point from the service point of view in the “manager” making different work sched- ules for different weekdays.
• CONCLUSION: All three null hypotheses are rejected, but the differences among the
factors B are much less striking than those among the factors A.
Model Checking: Checking the residuals
• Each RESIDUAL is the value of the OB- SERVATION V ITS PREDICTION. Sym- bolically, rijk = Yijk − ˆµij.
• When the basic ANOVA analysis is com- plete the residuals should then be examined for adherence to the basic assumptions V
– homoscedasticity – normality
• Standard graphical procedures: residual plot and quantile plot
• Residual plot:
– Look at a plot of the residuals against the values of Yijk or against the predicted values ˆµij.
This provides a check for homoscedastic- ity. Refer to next page.
– What you are looking for is to see that the vertical spreads of the data are ap- proximately the same at each value of
µij. (You need to allow for random vari- ation when evaluating such a plot V Our
basic assumption is that such an equality of spreads holds in the population.)
– CONCLUSION: So far as we can tell it
appears that the populations are homoscedas- tic (apart from that one outlier).
• Normal probability plot
– Once it has been decided that the data are acceptably homoscedastic it becomes of interest to check whether the within ij group residuals are also acceptably near normality.
– R can give you the values of the residuals and then form a Normal Quantile Plot to check normality.
– This residual plot shows startlingly good agreement with normality V except for the one annoying outlier.
– Usually we’re satisfied with considerably less convincing agreement to normality.
– What we want to avoid are heavily skewed residual distributions.
These would suggest that the analysis is invalid, and that some corrective ac-
tion (such as a transformation of the Y - variables) needs to be taken before re- analyzing the data.
Model Checking: Compare it to a model with interactions
• The ij population group means are modeled as
µij = µ + αi + βj + γij,
where Pi αi = 0, Pj βj = 0, Pi γij = 0, and
Pj γij = 0.
• The γij are the new part of this model.
They describe the difference between the µij here and their value in the additive model – µ + αi + βj.
• This model imposes no restrictions on the µij; they can take any value. Correspond- ingly, it turns out that
µij = ˆµ + ˆαi + ˆβj + ˆγij = ¯Yij.. Data Analysis
The tests of these hypotheses are summarized in an ANOVA table.
Summary of Fit
Root Mean Square Error 0.7617 Mean of Response 7.889
Analysis of Variance
Source DF SS MS F Ratio
Model 84 1711.59 20.3760 35.1193 Error 595 345.21 0.5802 Prob > F C. Total 679 2056.80 < .0001
Source Npar. DF SS F Ratio Prob > F
Hour 16 16 1658.40 177.34 < .0001
day of week 4 4 13.23 5.66 0.0002
Hour × day of week 64 64 39.96 1.08 0.3272
• The “new” part of this table is the line for Hour × day of week.
• This corresponds to a test of H0AB: all γij = 0 versus HaAB: not H0AB.
• It has df = (I − 1)(J − 1) = 64. (And, note that the “Model” now has df = IJ − 1 = 84, so that the Error df = n − IJ = 595, which is also not the same as in the additive model.)
• CONCLUSION: We DO NOT REJECT H0AB. We conclude that there is no statistically sig- nificant evidence of interaction effects. So far as we can determine the additive model fits this situation as well as the model with interactions.
• Profile Plots:
– Graphical method for the detection of in- teractions
– Plots of ˆµij as a function of i for each j are a good way to see the interaction estimates. (Or plots for each j as a func- tion of i.)
This produces a Profile Plot.
– If there were no interactions in the data then we would have
µij = ˆµ + ˆαi + ˆβj
so that the functions displayed on the profile plot would be exactly parallel.
– The degree to which they are not parallel indicates how much interaction there is in the data.
– Two-way Interaction Plot
– Plots the mean (or other summary) of the response for two-way combinations of factors, thereby illustrating possible in- teractions.
– Fitting Linear Models
– “lm” is used to fit linear models. It can be used to carry out regression, single stratum analysis of variance and anal- ysis of covariance (although “aov” may provide a more convenient interface for these).
– Models for lm are specified symbolically.
A typical model has the form “response
∼ terms” where response is the (numeric) response vector and terms is a series of terms which specifies a linear predictor for response.
– A terms specification of the form first + second indicates all the terms in first to- gether with all the terms in second with duplicates removed.
– A specification of the form first:second indicates the set of terms obtained by taking the interactions of all terms in first with all terms in second.
– The specification first*second indicates the cross of first and second. This is the same as first + second + first:second.
• g = lm(calltransf ∼ hour ∗ day, data) anova(g)
– anova: It gives Anova Tables
– Compute analysis of variance (or deviance) tables for one or more fitted model ob- jects.
• Diagnostics checking on the residual – qqnorm(g$res)
– plot(g$fitted,g$res,xlab=“Fitted”,ylab=“Residuals”, main=“Square root response”)
Blocking Designs and ANOVA
Suppose we want to compare 4 treatments and have 20 patients available.
• In completely randomized designs (CRD), the treatments are assigned to the experi- mental units at random.
The one-way ANOVA can be used to ana- lyze the resulting data to check on treatment effects.
– This design is appropriate when the units are homogenous.
– We may suspect that the units are het- erogenous, but we can’t describe the form it takes - for example, we may know a group of patients are not identical but we may have no further information about them.
– In this case, it is still appropriate to use a CRD.
– The randomization will tend to spread the heterogeneity around to reduce bias.
– Under the null hypothesis, there is no link between a factor and the response.
In other words, the responses have been assigned to the units in a way that is un- linked to the factor. This corresponds to the randomization used in assigning the levels of the factor to the units. This is why the randomization is crucial because it allows us to make this argument.
If the difference in the response between levels of the factor seems too unlikely to have occurred by chance, we can reject the null hypothesis.
• When the experimental units are heteroge- nous in a known way and can be arranged into blocks where the intrablock variation is ideally small but the interblock variation is large, a block design can be more efficient than a CRD.
– We might be able divide the patients in 5 blocks of 4 patients each where the pa- tients in each block have some relevant similarity.
– We would then randomly assign the treat- ments within each block.
– Suppose we want to test 3 crop varieties on 5 fields.
Divide each field into 3 strips and ran- domly assign the crop variety.
– We prefer to have block size equal to the number of treatments.
If this is not done or possible, an incom- plete block design must be used.
– We just motivate randomized block de- sign.
For the example we just described, we have one factor (or treatment) at 3 levels and one blocking variable at 5 levels.
We can use two-way ANOVA to check for interaction and check for a treatment effect.
We can also check the block effect but this is only useful for future reference.
• Notice that under the randomized block de- sign the randomization used in assigning the treatments to the units is restricted relative to the full randomization used in the CRD.
– Blocking is a feature of the experimen-
tal units and restricts the randomized as- signment of the treatments.
– This means that we cannot regain the degrees of freedom devoted to blocking even if the blocking effect turns out not to be significant.
– Only use blocking where there is some heterogeneity in the experimental units.
The decision to block is a matter of judg- ment prior to the experiment.
There is no guarantee that it will increase precision.
– How do we compare the efficiency of two designs?
• Latin Squares: These are useful when there are two blocking variables.
– Suppose, in a field used for agricultural experiments, the level of moisture may vary across the field in one direction and the fertility in another.
– In an industrial experiment, suppose we wish to compare 4 production methods (the treatment) - A, B, C, and D.
We have available 4 machines 1, 2, 3, and 4, and 4 operators, I, II, III, IV.
– A Latin square design is 1 2 3 4
I A B C D
II B D A C III C A D B IV D C B A
– Each treatment is assigned to each block once and only once.
The design and assignment of treatments and blocks should be random.
– Three-way ANOVA