What is a General Linear Model

download What is a General Linear Model

of 70

description

Details about linear models

Transcript of What is a General Linear Model

  • What is a general linear model?

    Use General Linear Model to determine whether the means of two or more groups differ. You can include random factors, covariates, or a mix of crossed and nested factors. You can also use stepwise regression to help determine the model. You can then use the model to predict values for new observations, identify the combination of predictor values that jointly optimize one or more fitted values, and create surface plots, contour plots, and factorial plots.

    GLM is an ANOVA procedure in which the calculations are performed using a least squares regression approach to describe the statistical relationship between one or more predictors and a continuous response variable. Predictors can be factors and covariates. GLM codes factor levels as indicator variables using a 1, 0, - 1 coding scheme, although you can choose to change this to a binary coding scheme (0, 1). Factors may be crossed or nested, fixed or random. Covariates may be crossed with each other or with factors, or nested within factors. The design may be balanced or unbalanced. GLM can perform multiple comparisons between factor level means to find significant differences.

    Example of a general linear model

    Suppose you are studying the affect of an additive (factor with three levels) and temperature (covariate) on the coating thickness of your product. You collect your data and fit a general linear model. The following output is a portion of the results from Minitab:

    Factor Information

    Factor Type Levels Values

    Additive fixed 3 1, 2, 3

    Analysis of Variance

    Source F P

    Temperature 719.21 0.000

    Additive 56.65 0.000

    Additive*Temperature 69.94 0.000

    Model Summary

    S R-Sq R-Sq(adj) R-sq(pred)

    19.1185 99.73% 99.61% 99.39%

    Coefficients

    Term Coef T P

    Constant -4968 -25.97 0.000

    Temperature 83.87 26.82 0.000

  • Additive*Temperature -0.2852 -22.83 0.000

    Additive

    1 -24.40 -5.52 0.000

    2 -27.87 -6.30 0.000

    Because the p-values are less than any reasonable alpha level, evidence exists that your two predictors and their interaction have a significant affect on strength. In addition, your model explains 99.73% of the variance. The coefficient for the covariate, temperature, indicates that the mean strength increases by 83.87 units per one degree increase in temperature when all other predictors are held constant. For the additive factor, the mean for level 1 is 24.40 units below the overall mean while level 2 is 27.87 units below the overall mean. Level 3 is the baseline value so it is not displayed. You can calculate the baseline factor level mean by adding all the level coefficients for a factor (excluding the intercept) and multiplying by - 1. In this case, it is 52.27 ((-24.40-27.87) * -1) units above the overall mean.

    Perform a fully nested ANOVA

    Use Fully Nested ANOVA to determine whether the means of two or more groups differ when all the factors are nested. For example, compare the production rates of two machines that each have unique operators. To perform ANOVA with nested factors in Minitab you can use either Fully Nested ANOVA or General Linear Model. Fully Nested ANOVA does not display F and p values when the data are unbalanced while General Linear Model does.

    The following options show how to perform fully nested ANOVA using both methods using an example. Suppose you want to understand the sources of variability in the manufacture of glass jars. You do an experiment and measure furnace temperature three times during a work shift for each of four operators from each plant on four different shifts. Using the Minitab sample data set FURNTEMP.MTW where Temp is the response and the four nested factors are Plant, Operator, Shift and Batch.

    IN THIS TOPIC

    Option 1: Use Fully Nested ANOVA

    Option 2: Use General Linear Model

    Option 1: Use Fully Nested ANOVA

    1. Choose Stat > ANOVA > Fully Nested ANOVA.

    2. In Responses, enter Temp.

    3. In Factors, enter Plant Operator Shift Batch. Click OK.

    Note

    In step 3, the factors are listed in hierarchical order.

  • Return to top

    Option 2: Use General Linear Model

    1. Choose Stat > ANOVA > General Linear Model > Fit General Linear Model

    2. In Responses, enter Temp.

    3. In Factors, enter Plant Operator Shift Batch.

    4. Click Random/Nest. Complete the Nesting Table as follows:

    Factor/Covariate Nested in specified factors

    Plant

    Operator Plant

    Shift Operator

    Batch Shift

    5. In the Factor type table, change all of the factors to Random.

    6. Click OK in each dialog box.

    The general linear model output will not be identical to the fully nested ANOVA output.

    What is analysis of means?

    Analysis of means is a graphical analog to ANOVA that tests the equality of population means. The graph displays each factor level mean, the overall mean, and the decision limits. If a point falls outside the decision limits, then evidence exists that the factor level mean represented by that point is significantly different from the overall mean.

  • For example, you are investigating how temperature and additive settings affect the rating of your product. After your experiment, you use ANOM to generate the following

    graph. The top plot shows that the interaction effects are well within the decision limits, signifying no evidence of interaction. The lower two plots show the means for the levels of the two factors, with the main effect being the difference between the mean and the center line. In the lower left plot, the point representing the third mean of the factor Temperature is displayed by a red symbol, indicating that there is evidence that the Temperature 200 mean is significantly different from the overall mean at = 0.05. The main effects for levels 1 and 3 of the Additive factor are well outside the decision limits of the lower right plot, signifying that there is evidence that these means are different from the overall mean.

    Return to top

    Comparison of ANOM (analysis of means) and ANOVA

    ANOVA tests whether the treatment means differ from each other. ANOM tests whether the treatment means differ from the overall mean (also called grand mean).

    Often, both analyses yield similar results. However, there are some scenarios in which the results can differ:

    If one group of means is above the overall mean and a different group of means is below the overall mean, ANOVA might indicate evidence for differences where ANOM might not.

    If the mean of one group is separated from the other means, the ANOVA F-test might not indicate evidence for differences whereas ANOM might flag this group as being different from the overall mean.

  • One more important difference is that ANOVA assumes that your data follow a normal distribution, while ANOM can be used with data that follows a normal, binomial, or Poisson distribution.

    What is MANOVA (multivariate analysis of variance)?

    A test that analyzes the relationship between several response variables and a common set of predictors at the same time. Like ANOVA, MANOVA requires continuous response variables and categorical predictors. MANOVA has several important advantages over doing multiple ANOVAs, one response variable at a time. Increased power

    You can use the covariance structure of the data between the response variables to test the equality of means at the same time. If the response variables are correlated, then this additional information can help detect differences too small to be detected through individual ANOVAs.

    Detects multivariate response patterns

    The factors may affect the relationship between responses instead of affecting a single response. ANOVAs will not detect these multivariate patterns as the following figures show.

    Controls the family error rate

    Your chance of incorrectly rejecting the null hypothesis increases with each successive ANOVA. Doing one MANOVA to test all response variables at the same time keeps the family error rate equal to your alpha level.

    For example, you are studying the affects of different alloys (1, 2, and 3) on the strength and flexibility of your company's building products. You first perform two separate ANOVAs but the results are not significant. Surprised, you plot the raw data for both response variables using individual value plots. These plots visually confirm the insignificant ANOVA results.

  • Because the response variables are correlated, you perform a MANOVA. This time the results are significant with p-values less than 0.05. You create a scatterplot to better understand the results.

    The individual value plots show, from a univariate perspective, that the alloys do not significantly affect either strength or flexibility. However, the scatterplot of the same data shows that the different alloys change the relationship between the

  • two response variables. That is, for a specified flexibility score, Alloy 3 usually has a higher strength score than Alloys 1 and 2. MANOVA can detect this type of multivariate response whereas ANOVA cannot.

    Note

    Usually, you should graph the data before conducting any analyses because it will help you decide what approach is appropriate.

    Return to top

    Which multivariate tests are included in MANOVA?

    Minitab automatically does four multivariate tests for each term in the model and for specially requested terms:

    Wilk's test

    Lawley-Hotelling test

    Pillai's test

    Roy's largest root test

    All four tests are based on two SSCP (sums of squares and cross products) matrices:

    An H (hypothesis) matrix associated with each term; also called between sample sums of squares

    An E (error) matrix associated with the error for the test; also called within sample sums of squares

    The SSCP matrices are displayed when you request the hypothesis matrices.

    You can express the test statistics either as H, E, or H and E, or as the eigenvalues of E-1 H. You can request to have these eigenvalues displayed. (If the eigenvalues are repeated, corresponding eigenvectors are not unique and in this case, the eigenvectors Minitab displays and those in books or other software may not agree. The MANOVA tests, however, are always unique.)

    Analyzing a repeated measures design

    You can use Fit General Linear Model to analyze a repeated measures design in Minitab. To use Fit General Linear Model, choose Stat > ANOVA > General Linear Model > Fit General Linear Model.

    In all cases, you must arrange the data in the Minitab worksheet so the response values are in one column, subject IDs are in a different column, and each factor has its own separate column.

  • The following examples show analyses of several different repeated measures designs. You can find the data and more information on these examples in J. Neter, M.H. Kutner, C.J. Nachtsheim, and W. Wasserman (1996). Applied Linear Statistical Models, 4th edition. WCB/McGraw-Hill.

    IN THIS TOPIC

    Example of a single-factor experiment with repeated measures on all treatments

    Example of a two-factor experiment with repeated measures on both factors

    Example of a two-factor experiment with repeated measures on one factor

    Example of a single-factor experiment with repeated measures on all

    treatments

    In this designed experiment each subject receives each treatment in succession. Create three columns in the Minitab worksheet: one column for the measurements, one column identifying which subject corresponds to that measurement, and one column identifying the treatment applied to that subject. Each row represents a single measurement.

    For more information, see page 1166, model 29.1 in Neter, Kutner, Nachtsheim, and Wwasserman (1996).

    C1 C2 C3

    Subject Dosage Measurement

    A low 1.33

    A medium 0.27

    B medium 0.49

    B low 0.99

    C medium 0.41

    C low 1.12

    1. Choose Stat > ANOVA > General Linear Model > Fit General Linear Model

    2. In Responses, enter Measurement.

    3. In Factors, enter Subject Dosage.

    4. Click Random/Nest.

    5. Under Factor type, choose Random in the field beside Subject.

    6. Click OK in each dialog box.

    Return to top

    Example of a two-factor experiment with repeated measures on both factors

  • In this designed experiment each subject is measured after receiving, successively, every combination of the levels of the two factors A and B. For example, suppose there are three subjects, and factors A and B each have two levels. For more information, see page 1177, model 29.10 in Neter, Kutner, Nachtsheim, and Wwasserman (1996). The designed experiment continues as follows:

    1 2 3 4

    Subject 1 A1B2 A2B2 A1B1 A2B1

    Subject 2 A2B1 A1B2 A2B2 A1B1

    Subject 3 A1B1 A2B1 A1B2 A2B2

    1. Create four columns in the Minitab worksheet: one column for the measurements, one column identifying which subject corresponds to that measurement, one column for Factor A, and one column for Factor B.

    C1 C2 C3 C4

    Subject Temperature Fabric Measurement

    A High Old 10.4

    A High New 9.5

    A Low New 7.6

    A Low Old 6.9

    B High New 9.1

    B High Old 7.9

    B Low New 10.0

    B Low Old 8.1

    2. Choose Stat > ANOVA > General Linear Model > Fit General Linear Model.

    3. In Responses, enter Measurement.

    4. In Factors, enter Subject Temperature Fabric.

    5. Click Random/Nest.

    6. Under Factor type, choose Random in the field beside Subject.

    7. Click OK.

  • 8. Click Model.

    9. Use the dialog box to add interactions to the model. For example, to add the interaction between Temperature and Fabric:

    1. In the field under Factors and covariates, select both Temperature and Fabric.

    2. Verify that 2 is selected beside Interactions through order.

    3. Click Add beside the field that has 2 selected.

    4. Click OK in each dialog box.

    Return to top

    Example of a two-factor experiment with repeated measures on one factor

    In this designed experiment each subject is measured after receiving, successively, all levels of Factor B in combination with only one level of Factor A. For more information, see page 1186, model 29.16 in Neter, Kutner, Nachtsheim, and Wwasserman (1996). This designed experiment continues as follows:

    Factor A Factor B Treatment Order 1 Treatment Order 2

    A1 1

    ...

    n

    A1B1

    ...

    A1B2

    A1B2

    ...

    A1B1

    A2 n+1

    ...

    2n

    A2B2

    ...

    A2B1

    A2B1

    ...

    A2B2

    1. Create four columns in the Minitab worksheet: one column for the measurement, one column identifying which subject corresponds to that measurement, one column for Factor A, and one column for Factor B.

    C1 C2 C3 C4

    Subject Temperature Fabric Measurement

    A High Old 1.1

    A High New 2.2

    B High New 1.9

    B High Old 1.2

  • C1 C2 C3 C4

    Subject Temperature Fabric Measurement

    C Low Old 0.8

    C Low New 1.1

    D Low Old 0.9

    D Low New 1.3

    2. Choose Stat > ANOVA > General Linear Model > Fit General Linear Model.

    3. In Responses, enter Measurement.

    4. In Factors, enter Subject Temperature Fabric.

    5. Click Random/Nest.

    6. Under Nesting in specified factors, enter Temperature beside Subject.

    7. Under Factor type, choose Random in the field beside Subject.

    Note

    If any factors besides Subject are random, choose Random for them too.

    8. Click OK.

    9. Click Model.

    10. Use the dialog box to add interactions to the model. For example, to add the interaction between Temperature and Fabric:

    1. In the field under Factors and covariates, select both Temperature and Fabric.

    2. Verify that 2 is selected beside Interactions through order.

    3. Click Add beside the field that has 2 selected.

    4. Click OK in each dialog box.

    11. How variability can affect your ANOVA

    12. The data sets in the following two individual value plots have exactly the same factor level means. Therefore, the variability in the data because of the factor is the same for both data sets. When you examine the plots, you might be tempted to conclude that the means are different in both cases. Notice, however, that the variability within factor levels is much greater in the second data set than in the first.

    13. To assess the differences between means, you must compare these differences with the spread of the observations about the means. This is exactly what an analysis of variance does. Using analysis of variance, the p-value corresponding to the first plot is 0.000, whereas the p-value corresponding to the second plot is 0.109.

  • 14. Therefore, using an of 0.05, the test says the means in the first data set are significantly different. The differences in the sample means for the second data set, however, could very well be a random result of the large overall variability in the data.

    15.

    16. Plot with low variability

    17.

    18. Plot with high variability

    Perform a two-way ANOVA

    To perform a two-way ANOVA in MInitab, use Stat > ANOVA > General Linear Model > Fit General Linear Model. Suppose your response is called A and your factors are B and C.

    1. Choose Stat > ANOVA > General Linear Model > Fit General Linear Model.

    2. In Responses, enter A.

    3. In Factors, enter B C.

    4. Click Model.

    5. In Models and covariates, select both B and C. To the right of Interactions through order, choose 2 and click Add.

    6. Click OK in each dialog box.

    What is a main effects plot?

    Use a main effects plot to examine differences between level means for one or more factors. There is a main effect when different levels of a factor affect the response differently. A main effects plot graphs the response mean for each factor level connected by a line.

  • When you choose Stat > ANOVA > Main Effects Plot Minitab creates a plot that uses data means. After you have fit a model, you can use the stored model to generate plots that use fitted means.

    Example

    For example, fertilizer company B is comparing the plant growth rate measured in plants treated with their product compared to plants treated by company A's fertilizer. They tested the two fertilizers in two locations. The following are the main effects plots of these two factors.

    Fertilizer seems to affect the plant growth rate because the line is not horizontal. Fertilizer B has a higher plant growth rate mean than fertilizer A. Location also affects the plant growth rate. Location 1 had a higher plant growth rate mean than location 2. The reference line represents the overall mean.

    General patterns to look for:

    When the line is horizontal (parallel to the x-axis), then there is no main effect. Each level of the factor affects the response in the same way, and the response mean is the same across all factor levels.

    When the line is not horizontal, then there is a main effect. Different levels of the factor affect the response differently. The steeper the slope of the line, the greater the magnitude of the main effect.

    Main effects plots will not show interactions. To view interactions between factors, use an interaction plot.

    Important

    To determine whether a pattern is statistically significant, you must do an appropriate test.

  • What is sum of squares?

    The sum of squares represents a measure of variation or deviation from the mean. It is calculated as a summation of the squares of the differences from the mean. The calculation of the total sum of squares considers both the sum of squares from the factors and from randomness or error.

    Return to top

    Sum of squares in ANOVA

    In analysis of variance (ANOVA), the total sum of squares helps express the total variation that can be attributed to various factors. For example, you do an experiment to test the effectiveness of three laundry detergents.

    The total sum of squares = treatment sum of squares (SST) + sum of squares of the residual error (SSE)

    The treatment sum of squares is the variation attributed to, or in this case between, the laundry detergents. The sum of squares of the residual error is the variation attributed to the error.

    Converting the sum of squares into mean squares by dividing by the degrees of freedom lets you compare these ratios and determine whether there is a significant difference due to detergent. The larger this ratio is, the more the treatments affect the outcome.

    Return to top

    Sum of squares in regression

    In regression, the total sum of squares helps express the total variation of the y's. For example, you collect data to determine a model explaining overall sales as a function of your advertising budget.

    The total sum of squares = regression sum of squares (SSR) + sum of squares of the residual error (SSE)

    The regression sum of squares is the variation attributed to the relationship between the x's and y's, or in this case between the advertising budget and your sales. The sum of squares of the residual error is the variation attributed to the error.

    By comparing the regression sum of squares to the total sum of squares, you determine the proportion of the total variation that is explained by the regression model (R2, the coefficient of determination). The larger this value is, the better the relationship explaining sales as a function of advertising budget.

  • Return to top

    Comparison of sequential sums of squares and adjusted sums of squares

    Minitab breaks down the SS Regression or Treatments component of variance into sums of squares for each factor. Sequential sums of squares

    Sequential sums of squares depend on the order the factors are entered into the model. It is the unique portion of SS Regression explained by a factor, given any previously entered factors.

    For example, if you have a model with three factors, X1, X2, and X3, the sequential sums of squares for X2 shows how much of the remaining variation X2 explains, given that X1 is already in the model. To obtain a different sequence of factors, repeat the regression procedure entering the factors in a different order.

    Adjusted sums of squares

    Adjusted sums of squares does not depend on the order the factors are entered into the model. It is the unique portion of SS Regression explained by a factor, given all other factors in the model, regardless of the order they were entered into the model.

    For example, if you have a model with three factors, X1, X2, and X3, the adjusted sum of squares for X2 shows how much of the remaining variation X2 explains, given that X1 and X3 are also in the model.

    When will the sequential and adjusted sums of squares be the same?

    The sequential and adjusted sums of squares are always the same for the last term in the model. For example, if your model contains the terms A, B, and C (in that order), then both sums of squares for C represent the reduction in the sum of squares of the residual error that occurs when C is added to a model containing both A and B.

    The sequential and adjusted sums of squares will be the same for all terms if the design matrix is orthogonal. The most common case where this occurs is with factorial and fractional factorial designs (with no covariates) when analyzed in coded units. In these designs, the columns in the design matrix for all main effects and interactions are orthogonal to each other. Plackett-Burman designs have orthogonal columns for main effects (usually the only terms in the model) but interactions terms, if any, may be partially confounded with other terms (that is, not orthogonal). In response surface designs, the columns for squared terms are not orthogonal to each other.

    For any design, if the design matrix is in uncoded units then there may be columns that are not orthogonal unless the factor levels are still centered at zero.

  • Can the adjusted sums of squares be less than, equal to, or greater than the

    sequential sums of squares?

    The adjusted sums of squares can be less than, equal to, or greater than the sequential sums of squares.

    Suppose you fit a model with terms A, B, C, and A*B. Let SS (A,B,C, A*B) be the sum of squares when A, B, C, and A*B are in the model. Let SS (A, B, C) be the sum of squares when A, B, and C are included in the model. Then, the adjusted sum of squares for A*B, is:

    SS(A, B, C, A*B) - SS(A, B, C)

    However, with the same terms A, B, C, A*B in the model, the sequential sums of squares for A*B depends on the order the terms are specified in the model.

    Using similar notation, if the order is A, B, A*B, C, then the sequential sums of squares for A*B is:

    SS(A, B, A*B) - SS(A, B)

    Depending on the data set and the order in which the terms are entered, all the following cases are possible:

    SS(A, B, C, A*B) - SS(A, B, C) < SS(A, B, A*B) - SS(A, B), or

    SS(A, B, C, A*B) - SS(A, B, C) = SS(A, B, A*B) - SS(A, B), or

    SS(A, B, C, A*B) - SS(A, B, C) > SS(A, B, A*B) - SS(A, B)

    Return to top

    What is uncorrected sum of squares?

    Squares each value in the column, and calculates the sum of those squared values. That is, if the column contains x1, x2, ... , xn, then sum of squares calculates (x12 + x22+ ... + xn2). Unlike the corrected sum of squares, the uncorrected sum of squares includes error. The data values are squared without first subtracting the mean.

    In Minitab, you can use descriptive statistics to display the uncorrected sum of squares (choose Stat > Basic Statistics > Display Descriptive Statistics). You can also use the sum of squares (SSQ) function in the Calculator to calculate the uncorrected sum of squares for a column or row. For example, you are calculating a formula manually and you want to obtain the sum of the squares for a set of response (y) variables.

    Choose Calc > Calculator and enter the expression: SSQ (C1)

  • Store the results in C2 to see the sum of the squares, uncorrected. The following worksheet shows the results from using the calculator to calculate the sum of squares of column y.

    C1 C2

    y Sum of Squares

    2.40 41.5304

    4.60

    2.50

    1.60

    2.20

    0.98

    Note

    Minitab omits missing values from the calculation of this function.

    What are mean squares?

    Mean squares represent an estimate of population variance. It is calculated by dividing the corresponding sum of squares by the degrees of freedom.

    Regression

    In regression, mean squares are used to determine whether terms in the model are significant.

    The term mean square is obtained by dividing the term sum of squares by the degrees of freedom.

    The mean square of the error (MSE) is obtained by dividing the sum of squares of the residual error by the degrees of freedom. The MSE is the variance (s2) around the fitted regression line.

    Dividing the MS (term) by the MSE gives F, which follows the F-distribution with degrees of freedom for the term and degrees of freedom for error.

    ANOVA

    In ANOVA, mean squares are used to determine whether factors (treatments) are significant.

  • The treatment mean square is obtained by dividing the treatment sum of squares by the degrees of freedom. The treatment mean square represents the variation between the sample means.

    The mean square of the error (MSE) is obtained by dividing the sum of squares of the residual error by the degrees of freedom. The MSE represents the variation within the samples.

    For example, you do an experiment to test the effectiveness of three laundry detergents. You collect 20 observations for each detergent. The variation in means between Detergent 1, Detergent 2, and Detergent 3 is represented by the treatment mean square. The variation within the samples is represented by the mean square of the error.

    Return to top

    What are adjusted mean squares?

    Adjusted mean squares are calculated by dividing the adjusted sum of squares by the degrees of freedom. The adjusted sum of squares does not depend on the order the factors are entered into the model. It is the unique portion of SS Regression explained by a factor, assuming all other factors in the model, regardless of the order they were entered into the model.

    For example, if you have a model with three factors, X1, X2, and X3, the adjusted sum of squares for X2 shows how much of the remaining variation X2 explains, assuming that X1 and X3 are also in the model.

    Return to top

    What are expected mean squares?

    If you do not specify any factors to be random, Minitab assumes that they are fixed. In this case, the denominator for F-statistics will be the MSE. However, for models which include random terms, the MSE is not always the correct error term. You can examine the expected means squares to determine the error term that was used in the F-test.

    When you perform General Linear Model, Minitab displays a table of expected mean squares, estimated variance components, and the error term (the denominator mean squares) used in each F-test by default. The expected mean squares are the expected values of these terms with the specified model. If there is no exact F-test for a term, Minitab solves for the appropriate error term in order to construct an approximate F-test. This test is called a synthesized test.

    The estimates of variance components are the unbiased ANOVA estimates. They are obtained by setting each calculated mean square equal to its expected mean square, which gives a system of linear equations in the unknown variance components that is then solved. Unfortunately, this approach can cause negative estimates, which should be set to zero. Minitab, however, displays the negative estimates because they

  • sometimes indicate that the model being fit is inappropriate for the data. Variance components are not estimated for fixed terms.

    How the F-statistics in the ANOVA output are calculated

    Each F-statistic is a ratio of mean squares. The numerator is the mean square for the term. The denominator is chosen such that the expected value of the numerator mean square differs from the expected value of the denominator mean square only by the effect of interest. The effect for a random term is represented by the variance component of the term. The effect for a fixed term is represented by the sum of squares of the model components associated with that term divided by its degrees of freedom. Therefore, a high F-statistic indicates a significant effect.

    When all the terms in the model are fixed, the denominator for each F-statistic is the mean square of the error (MSE). However, for models that include random terms, the MSE is not always the correct mean square. The expected mean squares (EMS) can be used to determine which is appropriate for the denominator.

    Example

    Suppose you performed an ANOVA with the fixed factor Screen and the random factor Tech, and get the following output for the EMS:

    Source Expected Mean Square for Each Term

    (1) Screen (4) + 2.0000(3) + Q[1]

    (2) Tech (4) + 2.0000(3) + 4.0000(2)

    (3) Screen*Tech (4) + 2.0000(3)

    (4) Error (4)

    A number with parentheses indicates a random effect associated with the term listed beside the source number. (2) represents the random effect of Tech, (3) represents the random effect of the Screen*Tech interaction, and (4) represents the random effect of

  • Error. The EMS for Error is the effect of the error term. In addition, the EMS for Screen*Tech is the effect of the error term plus two times the effect of the Screen*Tech interaction.

    To calculate the F-statistic for Screen*Tech, the mean square for Screen*Tech is divided by the mean square of the error so that the expected value of the numerator (EMS for Screen*Tech = (4) + 2.0000(3)) differs from the expected value of the denominator (EMS for Error = (4)) only by the effect of the interaction (2.0000(3)). Therefore, a high F-statistic indicates a significant Screen*Tech interaction.

    A number with Q[ ] indicates the fixed effect associated with the term listed beside the source number. For example, Q[1] is the fixed effect of Screen. The EMS for Screen is the effect of the error term plus two times the effect of the Screen*Tech interaction plus a constant times the effect of Screen. Q[1] equals (b*n * sum((coefficients for levels of Screen)**2)) divided by (a - 1), where a and b are the number of levels of Screen and Tech, respectively, and n is the number of replicates.

    To calculate the F-statistic for Screen, the mean square for Screen is divided by the mean square for Screen*Tech so that the expected value of the numerator (EMS for Screen = (4) + 2.0000(3) + Q[1] ) differs from the expected value of the denominator (EMS for Screen*Tech = (4) + 2.0000(3) ) only by the effect due to the Screen (Q[1]). Therefore, a high F-statistic indicates a significant Screen effect.

    Return to top

    Why does my ANOVA output include an "x" beside a p-value in the

    ANOVA table and the label "Not an exact F-test"?

    An exact F-test for a term is one in which the expected value of the numerator mean squares differs from the expected value of the denominator mean squares only by the variance component or the fixed factor of interest.

    Sometimes, however, such a mean square cannot be calculated. In this case, Minitab uses a mean square that results in an approximate F-test and displays an "x" beside the p-value to identify that the F-test is not exact.

    For example, suppose you performed an ANOVA with the fixed factor Supplement and the random factor Lake, and the got following output for the expected mean squares (EMS):

    Source Expected Mean Square for Each Term

    (1) Supplement (4) + 1.7500(3) + Q[1]

  • Source Expected Mean Square for Each Term

    (2) Lake (4) + 1.7143(3) + 5.1429(2)

    (3) Supplement*Lake (4) + 1.7500(3)

    (4) Error (4)

    The F-statistic for Supplement is the mean square for Supplement divided by the mean square for the Supplement*Lake interaction. If the effect for Supplement is very small, the expected value of the numerator equals the expected value of the denominator. This is an example of an exact F-test.

    Notice, however, that for a very small Lake effect, there are no mean squares such that the expected value of the numerator equals the expected value of the denominator. Therefore, Minitab uses an approximate F-test. In this example, the mean square for Lake is divided by the mean square for the Supplement*Lake interaction. This results in an expected value of the numerator being approximately equal to that of the denominator if the Lake effect is very small.

    Return to top

    About the "Denominator of F-test is zero or undefined" message

    Minitab will display an error that the denominator of the F-test is zero or undefined for one of the following reasons:

    There is not at least one degree of freedom for error.

    The adjusted MS values are very small, and thus there is not enough precision to display the F and p-values. As a workaround, multiply the response column by 10. Then perform the same regression model, but instead use this new response column for the response.

    Note

    Multiplying the response values by 10 will not affect the F and p-values that Minitab displays the output. However, decimal position will be affected in the remaining output, specifically, the sequential sums of squares, Adj SS, Adj MS, Fit, standard error of the fits, and residual columns.

    What is a variance component?

  • Use variance components to assess sources of variation.

    Variance components assess the amount of variation in the response because of random factors. Random factors have levels that are selected at random; whereas fixed factors have levels that are the only levels of interest. For example, you do a study on the effect of two levels of pressure on output measured by randomly chosen operators. Pressure is fixed (2 levels); and operator is random. The variance components output lists the estimated variance for the operator and error term.

    Interpret a negative variance component

    The following are possible ways to deal with negative estimates of variance components:

    Accept the estimate as evidence of a true value of zero and use zero as the estimate, recognizing that the estimator will no longer be unbiased.

    Retain the negative estimate, recognizing that subsequent calculations using the results might not make much sense.

    Interpret that the negative component estimate indicates an incorrect statistical model.

    Employ a method different from ANOVA for estimating the variance components.

    Collect more data and analyze them separately or in conjunction with the existing data and hope that increased information will yield positive estimates.

    Comparison of data means and fitted means Data means are the raw response variable means for each factor level

    combination whereas fitted means use least squares to predict the mean response values of a balanced design. Therefore, the two types of means are identical for balanced designs but can be different for unbalanced designs.

    Fitted means are useful for assessing response differences due to changes in factor levels rather than differences due to the unbalanced experimental conditions. While you can use raw data with unbalanced designs to obtain a general idea of which main effects may be evident, it is generally good practice to use the fitted means to obtain more precise results.

    Example of data means and fitted means

    For example, you are investigating how time and temperature affect the yield of a chemical reaction. The two factors each have two levels producing four experimental conditions. This is an exaggerated unbalanced experiment to emphasize the difference between the two types of means. All experimental conditions are measured twice except for the time and temperature combination of 50 and 200 which is measured four times. The following tables summarize the designed experiment and results.

    Number of Observations per Experimental Condition

    Temp 150 Temp 200

  • Temp 150 Temp 200

    Time 20 2 2

    Time 50 2 4

    Means by Factor Level

    Data Means Fitted Means

    Time 20 44.01 44.03

    Time 50 47.63 47.02

    Temp 150 44.13 44.14

    Temp 200 47.55 46.90

    The "Time 20" and "Temp 150" data means and fitted means are virtually identical because all experimental conditions involving either one or both of these factor levels are measured exactly twice (top table). However, the combination "Time 50" and "Temp 200" is measured four times which over represents their effects in the raw data means. The fitted means adjust for this and predict what a balanced design would yield.

    What is the overall mean (also called grand mean)? The overall mean is the mean of all observations, as opposed to the mean of

    individual groups. In ANOVA, for example, this is the mean of all observations across factor levels, as opposed to the means of individual levels.

    Here is a main effects plot comparing absenteeism at each of four different local

    schools. Each dot on the plot identifies the mean absenteeism of students from a different school. The line drawn across the plot (called the center line) marks the overall mean: the mean absenteeism of students from all schools.

    What is the variance-covariance matrix?

    A variance-covariance matrix is a square matrix that contains the variances and covariances associated with several variables. The diagonal elements of the matrix contain the variances of the variables and the off-diagonal elements contain the covariances between all possible pairs of variables.

    For example, you create a variance-covariance matrix for three variables X, Y, and Z. In

    the following table, the variances are displayed in bold along the diagonal; the variance

    of X, Y, and Z are 2.0, 3.4, and 0.82 respectively. The covariance between X and Y is -

    0.86.

  • X Y Z

    X 2.0 -0.86 -0.15

    Y -0.86 3.4 0.48

    Z -0.15 0.48 0.82

    The variance-covariance matrix is symmetric because the covariance between X and Y is the same as the covariance between Y and X. Therefore, the covariance for each pair of variables is displayed twice in the matrix: the covariance between the ith and jth variables is displayed at positions (i, j) and (j, i).

    Many statistical applications calculate the variance-covariance matrix for the estimators

    of parameters in a statistical model. It is often used to calculate standard errors of

    estimators or functions of estimators. For example, logistic regression creates this matrix

    for the estimated coefficients, letting you view the variances of coefficients and the

    covariances between all possible pairs of coefficients.

    Note

    For most statistical analyses, if a missing value exists in any column, Minitab ignores the entire row when it calculates the correlation or covariance matrix. However, when you calculate the covariance matrix by itself, Minitab does not ignore entire rows in its calculations when there are missing values. To obtain only the covariance matrix, choose Stat > Basic Statistics > Covariance

    What is the Hotelling's t-squared test?

    Compares two groups in a special case of MANOVA, using one factor that has two levels. The usual T-squared test statistic can be calculated from Minitab's output using the relationship T-squared = (N - 2)U, where N is the total number of observations in the data set and U is the Lawley-Hotelling trace.

    Calculate Levene's test in Minitab

    The modified Levene's test uses the absolute deviation of the observations in each treatment from the treatment median. It then assesses whether or not the mean of these

  • deviations are equal for all treatments. If the mean deviations are equal, the variances of the observations in all treatments will be the same. The test statistic for Levene's test is the ANOVA F-statistic for testing equality of means applied to the absolute deviations.

    You can do this in Minitab by making a new column where each value is the absolute value of the response minus the median of that treatment. Then perform a One-Way ANOVA using the new column as the response. The F-statistic and p-value will be the test statistic and p-value for Levene's test.

    For example, suppose the responses are in C1 and the treatments are in C2, and C3-C6 are empty.

    C1 C2

    Responses Treatments

    10 1

    8 1

    6 1

    4 1

    3 1

    16 2

    14 2

    10 2

    6 2

    2 2

    To perform Levene's test in Minitab for this data set:

    1. Choose Stat > Basic Statistics > 2 Variances.

    2. Click Both samples are in one column.

    3. In Samples, enter C1.

    4. In Sample IDs, enter C2. Click OK.

    Tests

    Test

    Method DF1 DF2 Statistic P-Value

    Bonett 1 2.14 0.143

    Levene 1 8 2.20 0.176

    You can verify these calculations using One-Way ANOVA:

    1. Choose Stat > Basic Statistics > Store Descriptive Statistics.

    2. In Variables, enter C1.

    3. In By variables (optional), enter C2.

  • 4. Click Statistics.

    5. Deselect all fields except Median.

    6. Click OK in each dialog box.

    7. Enter the treatment medians in C5.

    C1 C2 C3 C4 C5

    Responses Treatments ByVar1 Median1 Treatment Medians

    10 1 1 6 6

    8 1 2 10 6

    6 1 6

    4 1 6

    3 1 6

    16 2 10

    14 2 10

    10 2 10

    6 2 10

    2 2 10

    8. Choose Calc > Calculator.

    9. In Store result in variable, enter C6.

    10. In Expression, enter ABSO(C1-C5). Click OK.

    11. Choose Stat > ANOVA > One-Way.

    12. Choose Response data are in one column for all factor levels.

    13. In Response, enter C6.

    14. In Factor, enter C2. Click OK.

    Analysis of Variance

    Source DF Adj SS Adj MS F-Value P-Value

    Treatments 1 12.10 12.100 2.20 0.176

    Error 8 44.00 5.500

    Total 9 56.10

    When you examine the output you see that the F-statistic and p-value in the One-way ANOVA table are identical with the test statistic and p-value for Levenes test.

    Scenario for examples

    Suppose the design has 2 factors (Factor1 and Factor2). Factor1 has two levels (a and b) and Factor2 has three levels (x, y, and z). The data for Factor1 are in C1, Factor2 are

  • in C2, and the responses are in C3. You perform General Linear Model with Factor1, Factor2, and the 2-way interaction Factor1*Factor2 in the model.

    Return to top

    Example of letting Minitab calculate the fitted values and store them in the

    worksheet

    This option allows to you determine the fitted values using the values in the worksheet.

    1. Choose Stat > ANOVA > General Linear Model > Fit General Linear Model.

    2. In Responses, enter C3. In Factors, enter Factor1 Factor2.

    3. Click Model. In the field under Factors and covariates: select both 'Factor 1' and 'Factor 2'. Verify that 2 is selected in the field beside Interactions through order.

    4. Click Add, then click OK.

    5. Click Storage. Check Fits.

    6. Click OK in each dialog box.

    The fitted values are stored in the next available blank column in the worksheet, named FITS1.

    Return to top

    Example of entering coded values into the equation

    Suppose you obtain the following coefficients in the output:

    Term Coef SE Coef T P

    Constant 8.0000 0.5528 14.47 0.000

    Factor1

    a -0.6667 0.5528 -1.21 0.273

    Factor2

    x 5.0000 0.7817 6.40 0.001

    y -2.0000 0.7817 -2.56 0.043

    Factor1*Factor2

    a x -2.8333 0.7817 -3.62 0.011

    a y 1.6667 0.7817 2.13 0.077

    1. Using the coefficients from the table above, you can obtain the following regression equation.

    The equation is:

    Using the default coding that Minitab uses:

    o If Factor1 is a, use a = 1

    o If Factor1 is b, use a = 1

  • o If Factor2 is x, use x = 1 and y = 0

    o If Factor2 is y, use x = 0 and y = 1

    o If Factor2 is z, use x = 1 and y = 1

    2. Put the factor levels in to the equation.

    Suppose the 9th row in the data set has Factor1 = b and Factor2 = z. The fitted value is:

    = 8.0000 + - 0.6667*-1 + 5.0000*-1 - 2.0000*-1 - 2.8333*-1*-1 + 1.6667*-1*-1

    = 8.0000 + 0.6667 - 5.0000 + 2.0000 - 2.8333 + 1.6667

    = 4.5

    If you choose to store the fits as described in Option 1, you will see 4.5 in row 9 (with Factor1 = b and Factor2 = z) of the FITS1 column.

    Return to top

    How to display all the coefficients

    You can have Minitab display the coefficients that are not displayed by default.

    1. Choose Stat > ANOVA > General Linear Model > Fit General Linear Model.

    2. Enter the response columns and factor columns.

    3. Click Results and beside Coefficients select Full set of coefficients.

    4. Click OK in each dialog box.

    To calculate least squares means when you have a single covariate do the following:

    1. Open Least Squares Data.MTW.

    2. Choose Stat > ANOVA > General Linear Model > Fit General Linear Model.

    3. In Responses, enter Size.

    4. In Factors, enter PlasticMix and Subject.

    5. In Covariates, enter Temperature.

    6. Click Options, and beside Means select Main effects.

    7. Click OK in each dialog box.

    You should obtain the following results:

    General Linear Model: Size versus Temperature, PlasticMix, Subject

    Factor Information

  • Factor Type Levels Values

    PlasticMix Fixed 3 1, 2, 3

    Subject Fixed 6 1, 2, 3, 4, 5, 6

    Analysis of Variance

    Source DF Adj SS Adj MS F-Value P-Value

    Temperature 1 158.30 158.304 17.83 0.000

    PlasticMix 2 55.58 27.792 3.13 0.057

    Subject 5 39.99 7.997 0.90 0.492

    Error 34 301.91 8.880

    Lack-of-Fit 32 242.97 7.593 0.26 0.969

    Pure Error 2 58.95 29.473

    Total 42 618.46

    Model Summary

    S R-sq R-sq(adj) R-sq(pred)

    2.97990 51.18% 39.70% 23.15%

    Coefficients

  • Term Coef SE Coef T-Value P-Value VIF

    Constant -8.43 4.61 -1.83 0.076

    Temperature 0.1877 0.0445 4.22 0.000 1.32

    PlasticMix

    1 -1.023 0.731 -1.40 0.171 1.61

    2 2.031 0.813 2.50 0.017 2.31

    Subject

    1 -2.81 1.51 -1.86 0.072 3.75

    2 -0.97 1.31 -0.74 0.466 3.09

    3 2.22 1.43 1.55 0.131 3.34

    4 1.35 1.35 1.00 0.323 3.50

    5 0.207 0.913 0.23 0.822 2.26

    Regression Equation

    PlasticMix Subject

    1 1 Size = -12.27 + 0.1877 Temperature

    1 2 Size = -10.42 + 0.1877 Temperature

    1 3 Size = -7.24 + 0.1877 Temperature

    1 4 Size = -8.10 + 0.1877 Temperature

  • 1 5 Size = -9.24 + 0.1877 Temperature

    1 6 Size = -9.44 + 0.1877 Temperature

    2 1 Size = -9.21 + 0.1877 Temperature

    2 2 Size = -7.37 + 0.1877 Temperature

    2 3 Size = -4.18 + 0.1877 Temperature

    2 4 Size = -5.05 + 0.1877 Temperature

    2 5 Size = -6.19 + 0.1877 Temperature

    2 6 Size = -6.39 + 0.1877 Temperature

    3 1 Size = -12.25 + 0.1877 Temperature

    3 2 Size = -10.41 + 0.1877 Temperature

    3 3 Size = -7.22 + 0.1877 Temperature

    3 4 Size = -8.09 + 0.1877 Temperature

  • 3 5 Size = -9.23 + 0.1877 Temperature

    3 6 Size = -9.43 + 0.1877 Temperature

    Fits and Diagnostics for Unusual Observations

    Obs Size Fit Resid Std Resid

    8 7.10 15.20 -8.10 -3.00 R

    13 9.20 14.36 -5.16 -2.07 R

    31 13.00 7.47 5.53 2.03 R

    R Large residual

    Means

    Fitted

    Term Mean SE Mean

    PlasticMix

    1 9.766 0.972

    2 12.820 0.884

    3 9.780 0.867

    Subject

    1 7.97 1.71

  • 2 9.82 1.47

    3 13.00 1.57

    4 12.14 1.42

    5 10.996 0.893

    6 10.801 0.860

    Covariate Data Mean StDev

    Temperature 102.4 11.9

    8. To get the Least Squares mean corresponding to PlasticMix = 1, you need to calculate 6 fitted values:

    1. PlasticMix = 1, Subject = 1

    Size = -12.27 + .1877 * 102.4 = 6.95048

    2. PlasticMix = 1, Subject = 2

    Size = -10.42 + .1877 * 102.4 = 8.80048

    3. PlasticMix = 1, Subject = 3

    Size = -7.24 + .1877 * 102.4 = 11.98048

    4. PlasticMix = 1, Subject = 4

    Size = -8.10 + .1877 * 102.4 = 11.12048

    5. PlasticMix = 1, Subject = 5

    Size = -9.24 + .1877 * 102.4 = 9.98048

    6. PlasticMix = 1, Subject = 6

    Size = -9.44 + .1877 * 102.4 = 9.78048

    9. Now you need to average the 6 fitted values to obtain 9.766, the Least Squares Mean corresponding to PlasticMix = 1. You can do the same thing for PlasticMix = 2 and PlasticMix = 3.

  • Calculate plotted points on an analysis of means

    interaction effects plot for the normal case

    1. Choose Stat > ANOVA > General Linear Model > Fit General Linear Model.

    2. In Responses, enter the column that contains the response variable.

    3. In Factors, enter the two columns that contain the factors.

    4. Click Storage and check Residuals. Click OK in each dialog box.

    5. Choose Stat > Basic Statistics > Store Descriptive Statistics.

    6. In Variables, enter the residuals column (RESI1 by default).

    7. In By variables (optional), enter the two columns that contain the factors.

    8. Click Statistics and check Mean.

    9. Click OK in each dialog box.

    The values of the plotted points will be stored in the mean column (Mean1 by default).

    What are multiple comparisons?

    Multiple comparisons let you assess the statistical significance of differences between means using a set of confidence intervals, a set of hypothesis tests or both. As usual, the null hypothesis of no difference between means is rejected if and only if zero is not contained in the confidence interval.

    Return to top

    Are the individual error rates and family error rates exact?

    Individual error rates are exact in all cases. Family error rates are exact for equal group sizes. If group sizes are unequal, the true family error rate for Tukey, Fisher, and MCB will be slightly smaller than stated, resulting in conservative confidence intervals. The Dunnett family error rates are exact for unequal sample sizes.

    Return to top

    Perform multiple comparisons using One-Way ANOVA

    1. Choose Stat > ANOVA > One-Way.

    2. Complete the dialog with the appropriate settings for your situation.

    3. Click Comparisons.

    4. Select the comparisons you want Minitab to display.

    5. Click OK in each dialog box.

    Return to top

  • Perform multiple comparisons using a general linear model

    1. Choose Stat > ANOVA > General Linear Model > Fit General Linear Model.

    2. Complete the dialog with the appropriate settings for your situation and click OK.

    3. Choose Stat > ANOVA > General Linear Model > Comparisons.

    4. Under Method, select the comparisons you want Minitab to display.

    5. Under Choose terms for comparisons, double click the factors that you want Minitab to make multiple comparisons for.

    6. Click OK.

    Return to top

    Which multiple comparison method should I use?

    The selection of the appropriate multiple comparison method depends on the inference that you want. It is inefficient to use the Tukey all-pairwise approach when Dunnett or MCB is suitable, because the Tukey confidence intervals will be wider and the hypothesis tests less powerful for a particular family error rate. For the same reasons, MCB is superior to Dunnett if you want to eliminate factor levels that are not the best and to identify those that are best or close to the best. The choice of Tukey versus Fisher's LSD methods depends on which error rate, family or individual, you want to specify.

    The characteristics and advantages of each method are summarized in the following table:

    Method Normal

    Data Strength

    Comparison

    with a Control

    Pairwise

    Comparison

    Tukey Yes Most powerful test when

    doing all pairwise

    comparisons.

    No Yes

    Dunnett Yes Most powerful test when

    comparing to a control.

    Yes No

    Bonferroni's - Robust procedure, but

    produces larger

    Yes Yes

  • Method Normal

    Data Strength

    Comparison

    with a Control

    Pairwise

    Comparison

    inequality confidence intervals.

    Usually conservative.

    Sidk's

    inequality

    Yes Slightly better than

    Bonferronis procedure.

    Usually conservative.

    Yes Yes

    Hsu's MCB

    method

    Yes The most powerful test

    when you compare the

    group with the highest or

    lowest mean to the other

    groups.

    No Yes

    Games-

    Howell

    Yes Used when you do not

    assume equal variances.

    No Yes

    Note

    One-Way ANOVA also offers Fishers LSD method for individual confidence intervals. Fisher's is not a multiple comparison method, but instead contrasts the individual confidence intervals for the pairwise differences between means using an individual error rate. Fisher's LSD method inflates the family error rate, which is displayed in the output.

    Return to top

    Which means should I compare?

    It is important to consider which means to compare when using multiple comparisons; a bad choice can result in confidence levels that are not what you think. Issues that should be considered when making this choice might include:

    1. How deep into the design should you compare means-only within each factor, within each combination of first-level interactions, or across combinations of higher level interactions?

  • 2. Should you compare the means for only those terms with a significant F-test or for those sets of means for which differences seem to be large?

    How deep within the design should you compare means? There is a trade-off: if you compare means at all two-factor combinations and higher orders turn out to be significant, then the means that you compare might be a combination of effects; if you compare means at too deep a level, you lose power because the sample sizes become smaller and the number of comparisons become larger. You might decide to compare means for factor level combinations for which you believe the interactions are meaningful.

    Minitab restricts the terms that you can compare means for to fixed terms or interactions between fixed terms. Nesting is considered to be a form of interaction.

    Usually, you should decide which means you will compare before you collect your data. If you compare only those means with differences that seem to be large, which is called data snooping, then you are increasing the likelihood that the results indicate a real difference where no difference exists. If you condition the application of multiple comparisons on achieving a significant F-test, then you increase the probability that differences exist among the groups but you do not detect them. Because the multiple comparison methods already protect against the detection of a difference that does not exist, you do not need the F-test to guard against this probability.

    However, many people commonly use F-tests to guide the choice of which means to compare. The ANOVA F-tests and multiple comparisons are not entirely separate assessments. For example, if the p-value of an F-test is 0.9, you probably will not discover statistically significant differences between means by multiple comparisons.

    Return to top

    What if the p-value from the ANOVA table conflicts with the multiple

    comparisons output?

    The p-value in the ANOVA table and the multiple comparison results are based on different methodologies and can occasionally produce contradictory results. For example, it is possible that the ANOVA p-value can indicate that there are no differences between the means while the multiple comparisons output indicates that some means that are different. In this case, you can generally trust the multiple comparisons output.

    You do not need to rely on a significant p-value in the ANOVA table to reduce the chance of detecting a difference that doesn't exist. This protection is already incorporated in the Tukey, Dunnett, and MCB tests (and Fisher's test when the means are equal).

    What are individual and family error rates?

  • The type I error rates associated with the multiple comparisons are often used to identify significant differences between specific factor levels in an ANOVA.

    What is the individual error rate?

    The individual error rate is the maximum probability that one or more comparisons will incorrectly conclude that the observed difference is significantly different from the null hypothesis.

    What is the family error rate?

    The family error rate is the maximum probability that a procedure consisting of more than one comparison will incorrectly conclude that at least one of the observed differences is significantly different from the null hypothesis. The family error rate is based on both the individual error rate and the number of comparisons. For a single comparison, the family error rate is equal to the individual error rate which is the alpha value. However, each additional comparison causes the family error rate to increase in a cumulative manner.

    It is important to consider the family error rate when making multiple comparisons because your chances of committing a type I error for a series of comparisons is greater than the error rate for any one comparison alone.

    Example of setting the individual error rate and family error rate

    You do a one-way ANOVA to examine steel strength from five different steel plants using 25 samples from each plant.

    You decide to examine all 10 comparisons between the five plants to determine specifically which means are different. If you assign an alpha of 0.05 to each of the 10 comparisons (the individual error rate), Minitab calculates a family error rate of 0.28 for the set of 10 comparisons. However, if you want the entire set of comparisons to have a family error rate of 0.05, then Minitab automatically assigns each individual comparison an alpha of 0.007.

    Tukey's method, Fisher's least significant difference (LSD), Hsu's multiple comparisons with the best (MCB), and Bonferroni confidence intervals are methods for calculating and controlling the individual and family error rates for multiple comparisons.

    Using confidence levels to identify significant

    differences between factor levels in multiple

    comparisons

  • The confidence levels associated with the confidence intervals often used in multiple comparisons to identify significant differences between specific factor levels in an ANOVA. These confidence levels are analogous to the individual and family error rates but applied to confidence intervals.

    What is an individual confidence level?

    The percentage of confidence intervals that will include the true population parameter or true difference between factor levels if the study were repeated multiple times.

    What is a simultaneous confidence level?

    The percentage of times that a group of confidence intervals will all include the true population parameters or true differences between factor levels if the study were repeated multiple times. The simultaneous confidence level is based on both the individual confidence level and the number of confidence intervals. For a single comparison, the simultaneous confidence level is equal to the individual confidence level. However, each additional confidence interval causes the simultaneous confidence level to decrease in a cumulative way.

    It is important to consider the simultaneous confidence level when you examine multiple confidence intervals because your chances of excluding a parameter or true difference between factor levels for a family of confidence intervals is greater than for any one confidence interval.

    Tukey's method, Fisher's least significant difference (LSD), Hsu's multiple comparisons with the best (MCB), and Bonferroni confidence intervals are methods for calculating and controlling the individual and simultaneous confidence levels.

    Example of individual confidence intervals and simultaneous confidence

    intervals

    For example, you measure the response times for memory chips. You take a sample of 25 chips from five different manufacturers.

    You decide to examine all 10 comparisons between the five plants to determine specifically which means are different. Using Tukey's method, you specify that the entire set of confidence intervals should have a 95% simultaneous confidence level. Minitab calculates that the 10 individual confidence levels need to be 99.35% to obtain the 95% simultaneous confidence level. These wider Tukey confidence intervals provide less precise estimates of the population parameters but limit the probability that one or more of the confidence intervals does not contain the true difference to a maximum of 5%. Understanding this context, you can then examine the confidence intervals to determine whether any do not include zero, which would denote a significant difference.

  • Confidence intervals with 95% individual confidence levels

    Confidence intervals with 99.35% individual confidence levels to obtain a 95% simultaneous confidence level using Tukey's

    Comparison of 95% confidence intervals to the wider 99.35% confidence intervals used by Tukey's in the previous example. The reference line at 0 shows how the wider Tukey confidence intervals can change your conclusions. Confidence intervals that contain zero denote no difference. (Only 5 of the 10 comparisons are shown because of space considerations.)

    What is Tukey's method for multiple comparisons?

    Tukey's method is used in ANOVA to create confidence intervals for all pairwise differences between factor level means while controlling the family error rate to a level you specify. It is important to consider the family error rate when making multiple comparisons because your chances of making a type I error for a series of comparisons is greater than the error rate for any one comparison alone. To counter this higher error rate, Tukey's method adjusts the confidence level for each individual interval so that the resulting simultaneous confidence level is equal to the value you specify.

    Example of Tukey confidence intervals

  • You are measuring the response times for memory chips. You sampled 25 chips from five different manufacturers.

    You decide to examine all 10 comparisons between the five plants to determine specifically which means are different. Using Tukey's method, you specify that the entire set of comparisons should have a family error rate of 0.05 (equivalent to a 95% simultaneous confidence level). Minitab calculates that the 10 individual confidence levels need to be 99.35% in order to obtain the 95% joint confidence level. These wider Tukey confidence intervals provide less precise estimates of the population parameter but limit the probability that one or more of the confidence intervals does not contain the true difference to a maximum of 5%. Understanding this context, you can then examine the confidence intervals to determine whether any do not include zero, indicating a significant difference.

    Confidence intervals with 95% individual confidence levels

    Confidence intervals with 99.35% individual confidence levels to obtain a 95% joint confidence level using Tukey's

    Comparison of 95% confidence intervals to the wider 99.35% confidence intervals used by Tukey's in the previous example. The reference line at 0 shows how the wider Tukey confidence intervals can change your conclusions. Confidence intervals that contain zero indicate no difference. (Only 5 of the 10 comparisons are shown due to space considerations.)

  • What is Dunnett's method for multiple comparisons?

    Dunnett's method is used in ANOVA to create confidence intervals for differences between the mean of each factor level and the mean of a control group. If an interval contains zero, then there is no significant difference between the two means under comparison. You specify a family error rate for all comparisons, and Dunnett's method determines the confidence levels for each individual comparison accordingly.

    Example of Dunnett's method

    You are studying three weight loss pills to determine whether they are significantly different from a placebo. In a double-blind experiment, fifty people receive Pill A, fifty people receive Pill B, fifty people receive Pill C, and fifty people receive a placebo. The placebo group is the control group. You record the average weight loss of each group and perform ANOVA with Dunnett's method to determine whether any of the three pills produce weight loss that is significantly different than the placebo. Dunnett's method produces three confidence intervals: one for the difference in mean weight loss between group A and the placebo group, one for the difference in mean weight loss between group B and the placebo group, and one for the difference in mean weight loss between group C and the placebo group. You set the family error rate for all three comparisons at 0.10, so the confidence level for all the comparisons is 90%.

    The confidence interval for the difference between Pill A and the placebo contains zero; therefore, you conclude that there is no difference between the weight loss in group A and the placebo group. The confidence interval for the difference between Pill B and the placebo contains only negative numbers; therefore, you conclude that subjects in group B lost less weight than subjects in the placebo group. In other words, Pill B prevents weight loss. Finally, the confidence interval for the difference between Pill C and the placebo contains only positive numbers; therefore, you conclude that Pill C produces significantly greater weight loss than a placebo. As a result of this study, you recommend Pill C.

    What is Fisher's least significant difference (LSD)

    method for multiple comparisons?

    Fisher's LSD method is used in ANOVA to create confidence intervals for all pairwise differences between factor level means while controlling the individual error rate to a significance level you specify. Fisher's LSD method then uses the individual error rate and number of comparisons to calculate the simultaneous confidence level for all confidence intervals. This simultaneous confidence level is the probability that all confidence intervals contain the true difference. It is important to consider the family error rate when making multiple comparisons because your chances of committing a

  • type I error for a series of comparisons is greater than the error rate for any one comparison alone.

    Example of Fisher's LSD method

    For example, you are measuring the response times for memory chips. You take a sample of 25 chips from five different manufacturers. The ANOVA resulted in a p-value of 0.01, leading you to conclude that at least one of the manufacturer means is different from the others.

    You decide to examine all 10 comparisons between the five plants to determine specifically which means are different. Using Fisher's LSD method, you specify that each comparison should have an individual error rate of 0.05 (equivalent to a 95% confidence level). Minitab creates these ten 95% confidence intervals and calculates that this set yields a 71.79% simultaneous confidence level. Understanding this context, you can then examine the confidence intervals to determine whether any do not include zero, identifying a significant difference.

    What is Hsu's multiple comparisons with the best

    (MCB)?

    Multiple comparison method that is designed to identify factor levels that are the best, insignificantly different from the best, and those that are significantly different from the best. You can define "best" as either the highest or lowest mean. This procedure is usually used after an ANOVA to more precisely analyze differences between level means.

    Hsu's MCB method creates a confidence interval for the difference between each level mean and the best of the remaining level means. If an interval has zero as an end point, there is a statistically significant difference between the corresponding means. Specifically:

    Highest is best Lowest is best

    Confidence interval contains zero No difference No difference

    Confidence interval entirely above zero Significantly better Significantly worse

  • Highest is best Lowest is best

    Confidence interval entirely below zero Significantly worse Significantly better

    For this method, you specify the family error rate and the individual error rate is adjusted to achieve it. Hsu's MCB method only compares a subset of all possible pairwise comparisons, unlike Tukey's method which does all comparisons. Therefore, Hsu's MCB method will generate tighter confidence intervals and more powerful tests for any specified family error rate.

    Example of Hsu's MCB method

    For example, a memory chip manufacturer randomly samples four production lines to determine which line produces the chips with the fastest response time. The mean response time for each production line is in the following table.

    Production line Mean response time N

    1 4.85 20

    2 10.05 20

    3 7.45 20

    4 1.20 20

    The analyst defines "best" as being the lowest (fastest) mean response time, which is line 4, and uses Hsu's MCB method to identify any production lines that are significantly different from the best. This produces the following confidence intervals [tested level - best of remaining levels].

  • Production line (compared to best) Lower limit Center Upper limit

    1 -1.2 3.65 8.5

    2 0 8.85 13.3

    3 0 6.25 10.2

    4 -8.5 -3.65 1.2

    Based on the confidence intervals, the analyst concludes that lines 2 and 3 are producing chips that are significantly slower (higher mean) than line 4 because their confidence intervals are entirely above zero. However, there is no evidence to indicate a significant difference between lines 1 and 4 because their confidence intervals contain zero (no difference). You might investigate the processes of lines 2 and 3 more carefully.

    Note

    When the tested level is significantly better or worse than the comparison level, Hsu's MCB method does not provide a minimum bound on how much better/worse.

    What is the Bonferroni method?

    Method for controlling the simultaneous confidence level for an entire set of confidence intervals. It is important to consider the simultaneous confidence level when you examine multiple confidence intervals because your chances that at least one of the confidence intervals does not contain the population parameter is greater for a set of intervals than for any single interval. To counter this higher error rate, the Bonferroni method adjusts the confidence level for each individual interval so that the resulting simultaneous confidence level is equal to the value you specify.

    Example of Bonferroni confidence intervals

  • You want to examine the confidence intervals for delivery time in days from five shipping centers. You generate the two sets of five confidence intervals using the same data.

    Unadjusted 95% Confidence Intervals for Delivery Times by Shipping Center

    Bonferroni 95% Confidence Intervals for Delivery Times by Shipping Center (99% Individual Confidence Intervals)

    These graphs compare regular 95% confidence intervals to the Bonferroni 95% confidence intervals. The wider Bonferroni confidence intervals provide less precise estimates of the population parameter but limits the probability that one or more of the confidence intervals does not contain the parameter to a maximum of 5%. In comparison, the family error rate associated with the five regular 95% confidence intervals is 25%.

    This conservative method ensures that the overall confidence level is at least 1 - . To obtain an overall confidence level of 1 - for the joint interval estimates, Minitab constructs each interval with a confidence level of (1 - /g), where g is the number of intervals. In the Bonferroni intervals, Minitab uses 99% confidence intervals (1.00 - 0.05/5 = 0.99) to achieve the 95% simultaneous confidence level.

    Manually calculate Bonferroni confidence intervals for

    the standard deviations (sigmas)

  • Complete the following steps to manually calculate the Bonferroni confidence intervals for the standard deviations (sigmas) of your factor levels instead of using Stat > Basic Statistics > 2 Variances or Stat > ANOVA > Test for Equal Variances.

    Suppose you are using the Minitab sample data set FURNACE.MTW and analyzing the response 'BTU.In' and factor Damper, which has 2 levels. To calculate the confidence interval for level 1 of Damper, using a familywise confidence level of 95% (0.95), do the following:

    1. Calculate K and store the value in a constant called K1.

    1. Open the Minitab sample data set FURNACE.MTW.

    2. Choose Calc > Calculator.

    3. In Store result in variable, enter K1.

    4. In Expression, enter 0.05 / (2 * 2).

    5. Click OK.

    2. Calculate the variance and N for each level of Damper and store the results in the data window.

    1. Choose Stat > Basic Statistics > Store Descriptive Statistics.

    2. In Variables, enter BTU.In.

    3. In By variables (optional), enter Damper.

    4. Click Statistics. Check Variance and N nonmissing.

    5. Click OK in each dialog box.

    For Damper level 1, var = 9.11960 and n = 40.

    3. Calculate U and store the result in worksheet constant named U1.

    1. Choose Calc > Probability Distributions > Chi-Square.

    2. Choose Inverse cumulative probability.

    3. In Degrees of freedom, type 39.

    4. Choose Input constant, and enter K1.

    5. In Optional storage, enter U1.

    6. Click OK.

    4. Calculate the upper bound for the confidence interval.

    1. Choose Calc > Calculator.

    2. In Store result in variable, enter UpperL1.

    3. In Expression, enter ((39 * 9.11960) / U1)**0.5.

    4. Click OK.

    The upper boundary for the 95% Bonferroni confidence interval for level 1 of Damper is 4.02726. The lower bound is calculated the same way, using L instead of U.

  • 5. Calculate L and store the result in worksheet constant named L1.

    1. Choose Calc > Calculator.

    2. In Store result in variable, enter K2.

    3. In Expression, enter 1 - K1.

    4. Click OK.

    5. Choose Calc > Probability Distributions > Chi-Square.

    6. Choose Inverse cumulative probability.

    7. In Degrees of freedom, type 39.

    8. Choose Input constant and enter K2.

    9. In Optional storage, enter L1.

    10. Click OK.

    6. Calculate the lower bound for the confidence interval.

    1. Choose Calc > Calculator.

    2. In Store result in variable, enter LowerL1.

    3. In Expression, enter ((39 * 9.11960) / L1)**0.5.

    4. Click OK.

    The lower bound for the 95% Bonferroni confidence interval for level 1 of Damper is 2.40655.

    Note

    When there is more than one factor, you need to consider each distinct factor level combination as a separate factor level.

    What is the adjusted p-value in multiple comparisons?

    Use for multiple comparisons in General Linear Model ANOVA, the adjusted p-value indicates which factor level comparisons within a family of comparisons (hypothesis tests) are significantly different. If the adjusted p-value is less than alpha, then you reject the null hypothesis. The adjustment limits the family error rate to the alpha level you choose. If you use a regular p-value for multiple comparisons, then the family error rate grows with each additional comparison. The adjusted p-value also represents the smallest family error rate at which a particular null hypothesis will be rejected.

    It is important to consider the family error rate when making multiple comparisons because your chances of committing a type I error for a series of comparisons is greater than the error rate for any one comparison alone.

    Example of adjusted p-values

    Suppose you compare the hardness of 4 different blends of paint. You analyze the data and get the following output:

  • Tukey Simultaneous Tests for Differences of Means

    Difference SE of

    Difference of Levels of Means Difference 95% CI T-Value

    Blend 2 - Blend 1 -6.17 2.28 (-12.55, 0.22) -2.70

    Blend 3 - Blend 1 -1.75 2.28 ( -8.14, 4.64) -0.77

    Blend 4 - Blend 1 3.33 2.28 ( -3.05, 9.72) 1.46

    Blend 3 - Blend 2 4.42 2.28 ( -1.97, 10.80) 1.94

    Blend 4 - Blend 2 9.50 2.28 ( 3.11, 15.89) 4.17

    Blend 4 - Blend 3 5.08 2.28 ( -1.30, 11.47) 2.23

    Adjusted

    Difference of Levels P-Value

    Blend 2 - Blend 1 0.061

    Blend 3 - Blend 1 0.868

    Blend 4 - Blend 1 0.478

    Blend 3 - Blend 2 0.245

    Blend 4 - Blend 2 0.002

    Blend 4 - Blend 3 0.150

    Individual confidence level = 98.89%

    You choose an alpha of 0.05 which, in conjunction with the adjusted p-value, limits the family error rate to 0.05. At this level, the differences between blends 4 and 2 are significant. If you lower the family error rate to 0.01, the differences between blends 4 and 2 are still significant.

    What is an interaction?

    When the effect of one factor depends on the level of the other factor. You can use an interaction plot to visualize possible interactions.

    Parallel lines in an interaction plot indicate no interaction. The greater the difference in slope between the lines, the higher the degree of interaction. However, the interaction plot doesn't alert you if the interaction is statistically significant.

    Example of an interaction plot

    For example, cereal grains must be dry enough before the packaging process. Lab technicians collect moisture data on grains at several oven times and temperatures.

  • This plot indicates an interaction between the oven temperature and oven time. The grain has a lower moisture percentage when baked for a time of 60 minutes as opposed to 30 minutes at 125 and 130 degrees. However, when the temperature is 135 degrees, the grain has a lower moisture percentage when baked for 30 minutes.

    Interaction plots are most often used to visualize interactions during ANOVA or DOE.

    Minitab draws a single interaction plot if you enter two factors, or a matrix of interaction plots if you enter more than two factors.

    Which interaction plots are available in Minitab?

    Minitab provides interaction plots to accompany various analyses. Use the interaction option available through:

    Stat > DOE > Factorial > Factorial Plots to generate interaction plots specifically for factorial designs.

    Stat > DOE > Mixture > Factorial Plots to generate interaction plots specifically for process variables in mixture designs.

    Stat > ANOVA > General Linear Model > Factorial Plots to generate interaction plots for the fitted values from doing an analysis of variance.

    Stat > Regression and then choose either Regression > Factorial Plots, Binary Logistic Regression > Factorial Plots, or Poisson Regression > Factorial Plots to generate interaction plots from a regression model.

    Note

    All of these options let you use fitted means. Usually, plots using fitted values and plots created by the Interaction Plot command using the response will not be the same. They will be the same if the data set is balanced and you fit a full model.

  • Enter factor levels in the worksheet with make

    patterned data

    You can use Calc > Make Patterned Data > Simple Set of Numbers to enter the level numbers of a factor. Here is an easy way to enter the level numbers for a three-way crossed design with a, b, and c levels of factors A, B, C, with n observations per cell:

    1. Enter the levels for factor A.

    1. Choose Calc > Make Patterned Data > Simple Set of Numbers, and press the F3 key to reset defaults.

    2. Enter A in Store patterned data in.

    3. Enter 1 in From first value. Enter the number of levels in A in To last value.

    4. Enter the product of bcn in Number of times to list the sequence.

    5. Click OK.

    2. Enter the levels for factor B.

    1. Choose Calc > Make Patterned Data > Simple Set of Numbers, and press the F3 key to reset defaults.

    2. Enter B in Store patterned data in.

    3. Enter 1 in From first value. Enter the number of levels in B in To last value.

    4. Enter the number of levels in A in Number of times to list each value. Enter the product of cn in Number of times to list the sequence.

    5. Click OK.

    3. Enter the levels for factor C.

    1. Choose Calc > Make Patterned Data > Simple Set of Numbers, and press the F3 key to reset defaults.

    2. Enter C in Store patterned data in.

    3. Enter 1 in From first value. Enter the number of levels in C in To last value.

    4. Enter the product of ab in Number of times to list each value. Enter the sample size n in Number of times to list the sequence.

    5. Click OK.

    What are reduced models and hierarchical models?

    Reduced models

    A reduced model is a model that does not include all the possible terms. For example, suppose you have a three factor design, with factors, A, B, and C. The full model would include:

  • all one factor terms: A, B, C

    all two-factor interactions: A * B, A * C, B * C

    the three-factor interaction: A * B * C

    It becomes a reduced model by omitting terms. You might reduce a model if terms are not significant or if you need additional error degrees of freedom and you can assume that certain terms are zero.

    Hierarchical models

    A hierarchical model is a model where for each term in the model, all lower order terms contained in it must also be in the model. For example, suppose there is a model with four factors: A, B, C, and D. If the term A * B * C is in the model then the terms A, B, C, A*B, A*C, and B*C must also be in the model, though any terms with D do not have to be in the model. The hierarchical structure applies to nesting as well. If B (A) is in the model, then A must be also.

    A model is non-hierarchical if it does not contain all of the lower order terms for each term in the model.

    What are randomized block designs and Latin square

    designs?

    Some designed experiments can effectively provide information when measurements are difficult or expensive to make or can minimize the effect of unwanted variability on treatment inference. The following is a brief discussion of two commonly used designs. To show these designs, two treatment factors (A and B) and their interaction (A*B) are considered. These de