正文

Chapter 6 - More on Interactions of Categorical Predictors

(2014-05-07 09:07:51) 下一個

 

Regression with SAS
Chapter 6 - More on Interactions of Categorical Predictors

Chapter Outline
   6.0 Introduction
   6.1. Analysis with two categorical variables
   6.2. Simple effects
     6.2.1 Analyzing simple effects using PROC GLM
    6.2.2 Analyzing Simple Effects Using  PROC REG
   6.3. Simple comparisons
     6.3.1 Analyzing simple comparisons using PROC REG
     6.3.2 Analyzing simple comparisons using PROC GLM
   6.4. Partial Interaction
     6.4.1 Analyzing partial interactions using PROC GLM
     6.4.2 Analyzing partial interactions using PROC REG
   6.5. Interaction contrasts
     6.5.1 Analyzing interaction contrasts using PROC GLM
     6.5.2 Analyzing interaction contrasts using PROC REG
   6.6. Computing adjusted means
     6.6.1 Computing adjusted means via PROC GLM
     6.6.1 Computing adjusted means via PROC REG
   6.7. More details on meaning of coefficients
   6.8. Simple effects via dummy coding versus effect coding
     6.8.1 Example 1. Simple effects of yr_rnd at levels of mealcat
     6.8.2 Example 2. Simple effects of mealcat at levels of yr_rnd

6.0 Introduction

This chapter will use the elemapi2 data that you have seen in the prior chapters. We assume that you have put the data files in "c:sasreg" directory.

data elemapi2;   set 'c:sasregelemapi2'; run;

For this chapter we will use the elemapi2 data file that we have been using in prior chapters. We will focus on the variables mealcat, and collcat as they relate to the outcome variable api00 (performance on the api in the year 2000. The variable mealcat is the variable meals broken up into three categories, and the variable collcat is the variable some_col broken into 3 categories. We could think of mealcat as being the number of students receiving free meals and broken up into low, middle and high. The variable collcat can be thought of as the number of parents with some college education, and we could think of it as being broken up into low, medium and high. For our analysis, we think that both mealcat and collcat may be related to api00, but it is also possible that the impact of mealcat might depend on the level of collcat. In other words, we think that there might be an interaction of these two categorical variables. In this chapter we will look at how these two categorical variables are related to api performance in the school, and we will look at the interaction of these two categorical variables as well. We will see that there is an interaction of these categorical variables, and will focus on different ways of further exploring the interaction.  Let's have a quick look at these variables. 

proc tabulate data=elemapi2;   class collcat mealcat ;   var api00;   table mealcat='mealcat',          mean=' '*api00='API Index for 2000'*collcat='collcat'*F=10.2          / RTS=13.; run;----------------------------------------------|           |       API Index for 2000       ||           |--------------------------------||           |            collcat             ||           |--------------------------------||           |    1     |    2     |    3     ||-----------+----------+----------+----------||mealcat    |          |          |          ||-----------|          |          |          ||1          |    816.91|    825.65|    782.15||-----------+----------+----------+----------||2          |    589.35|    636.60|    655.64||-----------+----------+----------+----------||3          |    493.92|    508.83|    541.73|----------------------------------------------

 6.1. Analysis with two categorical variables

One traditional way to analyze this would be to perform a 3 by 3 factorial analysis of variance using proc glm, as shown below. The results show a main effect of collcat (F=4.5, p-0.0117), a main effect of mealcat (F=509.04, p=0.0000) and an interaction of collcat by mealcat, (F=6.63, p=0.0000). We also use lsmeans and output statement to output the predicted means for each group and get ourselve ready to graph the cell means.

proc glm data = elemapi2;   class collcat mealcat;   model api00 = collcat | mealcat /ss3;   lsmeans collcat*mealcat;   output out = pred p = pred; run; quit;
The GLM Procedure   Class Level InformationClass         Levels    Valuescollcat            3    1 2 3mealcat            3    1 2 3Number of observations    400The GLM ProcedureDependent Variable: api00   api 2000                                      Sum ofSource                     DF        Squares    Mean Square   F Value   Pr > FModel                       8    6243714.810     780464.351    166.76   <.0001Error                     391    1829957.187       4680.197Corrected Total           399    8073671.998R-Square     Coeff Var      Root MSE    api00 Mean0.773343      10.56356      68.41197      647.6225Source                     DF    Type III SS    Mean Square   F Value   Pr > Fcollcat                     2      42140.566      21070.283      4.50   0.0117mealcat                     2    4764843.563    2382421.781    509.04   <.0001collcat*mealcat             4     124167.809      31041.952      6.63   <.0001Least Squares Meanscollcat    mealcat    api00 LSMEAN1          1            816.9142861          2            589.3500001          3            493.9189192          1            825.6511632          2            636.6046512          3            508.8333333          1            782.1509433          2            655.6376813          3            541.733333

We can now create the graph of cell means of api00 using the dataset pred.

proc sort data = pred;   by mealcat; run;
symbol1 v=circle i=join ci=blue  h= 2; symbol2 v=triangle i=join ci=red h =2; symbol3 v=square i=join ci=black h =2;
proc gplot data = pred;   plot pred*mealcat=collcat ; run; quit;

 

We can do the same analysis using the regression approach via proc reg. We use simple regression coding for both collcat and mealcat. We also create interaction terms for them. The first test statement tests the effect of main effect of collcat, the second the main effect of mealcat and the last one on the effect of overall interaction.

data reg1;   set elemapi2;   s2 = -1/3; s3=-1/3;   if collcat = 2 then s2 = 2/3;   if collcat = 3 then s3 = 2/3;   m2 = -1/3; m3 = -1/3;   if mealcat = 2 then m2 = 2/3;   if mealcat = 3 then m3 = 2/3;   sm22 = s2*m2;   sm23 = s2*m3;   sm32 = s3*m2;   sm33 = s3*m3; run;  proc reg data = reg1;   model api00 = s2 s3 m2 m3 sm22 sm23 sm32 sm33;   Collcat: test s2=s3=0;   Mealcat: test m2=m3=0;   Interaction: test sm22=sm23=sm32=sm33=0;   output out = pred2 p = pred; run; quit;
The REG ProcedureModel: MODEL1Dependent Variable: api00 api 2000                             Analysis of Variance                                    Sum of           MeanSource                   DF        Squares         Square    F Value    Pr > FModel                     8        6243715         780464     166.76    <.0001Error                   391        1829957     4680.19741Corrected Total         399        8073672Root MSE             68.41197    R-Square     0.7733Dependent Mean      647.62250    Adj R-Sq     0.7687Coeff Var            10.56356                            Parameter Estimates                               Parameter      StandardVariable    Label       DF      Estimate         Error   t Value   Pr > |t|Intercept   Intercept    1     650.08826       3.87189    167.90     <.0001s2                       1      23.63531       9.10533      2.60     0.0098s3                       1      26.44625       9.99513      2.65     0.0085m2                       1    -181.04135       9.07713    -19.94     <.0001m3                       1    -293.41027       9.44946    -31.05     <.0001sm22                     1      38.51777      24.19532      1.59     0.1122sm23                     1       6.17754      20.08262      0.31     0.7585sm32                     1     101.05102      22.88808      4.42     <.0001sm33                     1      82.57776      24.43941      3.38     0.0008    Test Collcat Results for Dependent Variable API00                                MeanSource             DF         Square    F Value    Pr > FNumerator           2          21070       4.50    0.0117Denominator       391     4680.19741    Test Mealcat Results for Dependent Variable API00                                MeanSource             DF         Square    F Value    Pr > FNumerator           2        2382422     509.04    <.0001Denominator       391     4680.19741  Test Interaction Results for Dependent Variable API00                                MeanSource             DF         Square    F Value    Pr > FNumerator           4          31042       6.63    <.0001Denominator       391     4680.19741

First, note that the results of the test statements correspond to those from proc glm statement. This is because collcat and mealcat were coded using simple effect coding, a coding scheme where the contrasts sum to 0. If this had been coded using dummy coding,  then the results of the test commands for mealcat and collcat from the proc reg would not have corresponded to the proc glm results. In addition to simple coding, we could have used deviation or helmert coding schemes and the results of the test commands would have matched the result from proc glm, although the meaning of the individual tests would have been different. This point will be explored in more detail later in this chapter.

The graph of the cell means we obtained before illustrates the interaction between collcat and mealcat. The graph shows the 3 levels of collcat as 3 different lines, and the 3 levels of mealcat as the 3 values on the x axis of the graph. We can see that the effect of collcat differs based on the level of mealcat. For example, when mealcat is low, schools where collcat is 3 have the lowest api00 scores, as compared to schools that are medium or high on mealcat, where schools with collcat of 3 have the highest api00 scores.

Let's investigate this interaction further by looking at the simple effects of collcat at each level of mealcat.

   6.2. Simple effects
   6.2.1 Analyzing simple effects using PROC GLM

This analysis looks at the simple effects of collcat at the different levels of mealcat using proc glm. The lsmeans statement with option slice = mealcat gives the test of effects of collcat at each level of mealcat.  

proc glm data= elemapi2;   class collcat mealcat;   model api00 = mealcat|collcat ;   lsmeans mealcat*collcat / slice = mealcat ; run; quit;
The GLM Procedure                                        Sum ofSource                      DF         Squares     Mean Square    F Value    Pr > FModel                        8     6243714.810      780464.351     166.76    <.0001Error                      391     1829957.187        4680.197Corrected Total            399     8073671.998R-Square     Coeff Var      Root MSE    API00 Mean0.773343      10.56356      68.41197      647.6225Source                      DF     Type III SS     Mean Square    F Value    Pr > FMEALCAT                      2     4764843.563     2382421.781     509.04    <.0001COLLCAT                      2       42140.566       21070.283       4.50    0.0117COLLCAT*MEALCAT              4      124167.809       31041.952       6.63    <.0001
COLLCAT    MEALCAT    API00 LSMEAN1          1            816.9142861          2            589.3500001          3            493.9189192          1            825.6511632          2            636.6046512          3            508.8333333          1            782.1509433          2            655.6376813          3            541.733333          COLLCAT*MEALCAT Effect Sliced by MEALCAT for API00                                     Sum ofMEALCAT        DF         Squares     Mean Square    F Value    Pr > F1               2           50909           25455       5.44    0.00472               2           68629           34314       7.33    0.00073               2           29979           14990       3.20    0.0417

6.2.2 Analyzing Simple Effects Using  PROC REG

We have demonstrated how to test the simple effect of collcat at each level of mealcat using PROC GLM in the previous section. That is through the approach of ANOVA. We can also obtain the same analysis through regression approach. After all, Anova is regression. In regression approach, we will create the coding for variable collcat, mealcat and their interaction. The coding scheme is specific for the effect we want to see. For example, in this section, we will do an analysis parallel to the previous section. That is to say that we want to see the simple effect of collcat at each level of mealcat. We will use simple coding for mealcat, though in our case the type of coding for mealcat does not really matter.  The scheme for simple coding is shown chapter 5. The reference group for mealcat is group 1.  We use helmert coding for collcat.  We should note that these terms are not used in the analysis, but are used  for creating the simple effects of collcat at each level of mealcat.

data reg2;   set elemapi2;   mcat1 = 1/3; mcat2 = 1/3;   if mealcat = 3 then mcat1 = -2/3;   if mealcat = 2 then mcat2 = -2/3;   ccat1 = -1/3;    if collcat = 1 then do;        			ccat1 = 2/3; 			ccat2 = 0; 			end;   if collcat = 2 then ccat2 = .5;   if collcat = 3 then ccat2 = -.5;   c1m1 = 0; c2m1 = 0; c1m2 = 0;   c2m2 = 0; c1m3 = 0; c2m3 = 0;
  if ( mealcat = 1)  then do; c1m1 = ccat1;                               c2m1 = ccat2; end;
  if ( mealcat = 2)  then do; c1m2 = ccat1;                               c2m2 = ccat2; end;   if ( mealcat = 3)  then do; c1m3 = ccat1;                               c2m3 = ccat2; end; run;

Now, that we have seen the helmert coding for collcat, we can see how this is used to create the simple effects of collcat at each level of mealcat. First, we look at the two comparisons of collcat at mealcat of 1. Note that the coding is the same as we saw above, but only when mealcat is 1, otherwise these variables are coded 0. Likewise, we look at the terms that form the effects of collcat when mealcat is 2, and we see that the variables are coded the same way when mealcat is 2, and otherwise 0. The same is true for the case when mealcat  is 3. The following matrix is the coding we just used for all the interaction terms.  

collcat mealcat c1m1 c2m1 c1m2 c2m2 c1m3 c2m3
1 1  2/3 0 0 0 0 0
2 1 -1/3  1/2 0 0 0 0
3 1 -1/3 -1/2 0 0 0 0
1 2 0 0  2/3 0 0 0
2 2 0 0 -1/3  1/2 0 0
3 2 0 0 -1/3 -1/2 0 0
1 3 0 0 0 0  2/3 0
2 3 0 0 0 0 -1/3  1/2
3 3 0 0 0 0 -1/3 -1/2

Now we are ready for our regression analysis. The test statements used below are for testing the simple effect of collcat  at each level of mealcat.

proc reg data = reg2;   model api00 = mcat1 mcat2 c1m1 c2m1 c1m2 c2m2 c1m3 c2m3;   mealcat1: test c1m1 = c2m1 = 0;   mealcat2: test c1m2 = c2m2 = 0;   mealcat3: test c1m3 = c2m3 = 0; run; quit;
The REG ProcedureModel: MODEL1Dependent Variable: API00 api 2000
                             Analysis of Variance                                    Sum of           MeanSource                   DF        Squares         Square    F Value    Pr > FModel                     8        6243715         780464     166.76    <.0001Error                   391        1829957     4680.19741Corrected Total         399        8073672Root MSE             68.41197    R-Square     0.7733Dependent Mean      647.62250    Adj R-Sq     0.7687Coeff Var            10.56356                               Parameter Estimates                                  Parameter       StandardVariable     Label        DF       Estimate          Error    t Value    Pr > |t|Intercept    Intercept     1      650.08826        3.87189     167.90      <.0001MCAT1                      1      293.41027        9.44946      31.05      <.0001MCAT2                      1      181.04135        9.07713      19.94      <.0001C1M1                       1       13.01323       13.52800       0.96      0.3367C2M1                       1       43.50022       14.04092       3.10      0.0021C1M2                       1      -56.77117       16.67866      -3.40      0.0007C2M2                       1      -19.03303       13.29175      -1.43      0.1530C1M3                       1      -31.36441       12.86955      -2.44      0.0153C2M3                       1      -32.90000       20.23653      -1.63      0.1048
   Test mealcat1 Results for Dependent Variable API00                                MeanSource             DF         Square    F Value    Pr > FNumerator           2          25455       5.44    0.0047Denominator       391     4680.19741   Test mealcat2 Results for Dependent Variable API00                                MeanSource             DF         Square    F Value    Pr > FNumerator           2          34314       7.33    0.0007Denominator       391     4680.19741   Test mealcat3 Results for Dependent Variable API00                                MeanSource             DF         Square    F Value    Pr > FNumerator           2          14990       3.20    0.0417Denominator       391     4680.19741

6.3 Simple Comparisons

In the analyses above we looked at the simple effect of collcat at each level of mealcat. For example, we looked at the overall effect of collcat when mealcat was 1. This is the simple effect of collcat at mealcat=1. Because collcat has more than 2 levels, we may wish to make further comparisons among the 3 levels of collcat within mealcat=1. Simple comparisons allow us to make such comparisons.

6.3.1 Analyzing Simple Comparisons Using PROC REG

In the previous regression analysis, we used helmert coding for collcat. We choose this coding scheme so we could compare group 1 with groups 2 and 3 and then compare groups 2 and 3 within mealcat = 1. For example, if we wanted to compare collcat 1 vs. 2 and 3, we would want to look at the effect c1m1, and if we wanted to compare collcat groups 2 and 3 when mealcat is 1, then we would look at the effect c2m1. For example, c1m1  is not significant with t-value = 0.96 and p-value = 0.3367. That is to say that the difference between group 1 of collcat with group 2 and group 3 with mealcat = 1 is not significant.

6.3.2 Analyzing Simple Comparisons Using PROC GLM

We can also look at the simple comparisons using PROC GLM. For example, for the comparsion of group 1 vs 2+ of collcat within mealcat = 1, we can do the following. The estimate statement below indicates that the comparison on collcat is between group 1 and all the upper groups and the comparison is restricted to within mealcat = 1.

proc glm data = elemapi2;   class collcat mealcat;   model api00 = collcat mealcat collcat*mealcat/ss3;   estimate 'collcat 1 vs 2+ within mealcat = 1'               collcat 1 -.5 -.5               collcat*mealcat 1  0 0 		             -.5 0 0 		             -.5 0 0; run; quit;
The GLM ProcedureDependent Variable: API00   api 2000                                        Sum ofSource                      DF         Squares     Mean Square    F Value    Pr > FModel                        8     6243714.810      780464.351     166.76    <.0001Error                      391     1829957.187        4680.197Corrected Total            399     8073671.998R-Square     Coeff Var      Root MSE    API00 Mean0.773343      10.56356      68.41197      647.6225Source                      DF     Type III SS     Mean Square    F Value    Pr > FCOLLCAT                      2       42140.566       21070.283       4.50    0.0117MEALCAT                      2     4764843.563     2382421.781     509.04    <.0001COLLCAT*MEALCAT              4      124167.809       31041.952       6.63    <.0001                                                          StandardParameter                                 Estimate           Error    t Value    Pr > |t|collcat 1 vs 2+ within mealcat = 1      13.0132326      13.5279998       0.96      0.3367

6.4 Partial Interaction

A partial interaction allows you to apply contrasts to one of the effects in an interaction term. For example, we can draw the interaction of collcat by mealcat like this below.

  Collcat low Collcat Med Collcat High
Mealcat Low      
Mealcat Med      
Mealcat High      

Say that we wanted to compare, in the context of this interaction, group 1 for collcat vs. groups 2 and 3. The table of this partial interaction would look like this.  The contrast coefficients of -2 1 1 applied to collcat indicate the comparison of  group 1 for collcat vs. groups 2 and 3. 

  -2 1 1
  Collcat low Collcat Med Collcat High
Mealcat Low      
Mealcat Med      
Mealcat High      

Likewise, we also might want to compare groups 2 and 3 of collcat by mealcat, and the table of this interaction would look like this.

  0 -1 1
  Collcat low Collcat Med Collcat High
Mealcat Low      
Mealcat Med      
Mealcat High      

These are called partial interactions because contrast coefficients are applied to one of the terms involved in the interaction.

6.4.1 Analyzing partial interactions using PROC GLM

We wish to compare groups 1 versus 2 on collcat. Similarly, we can also compare  groups 2 and 3 on collcat. For example, we want to test the partial interaction of collcat comparing group 1 vs. 2 and 3 by mealcat, we can do the following contrast statement. Because mealcat has 2 degrees of freedom, the test of partial interaction also has 2 degrees of freedom. The 2 degrees of freedom of factor mealcat can be broken down into 2 comparisons. These two interaction contrasts are separated by a semi-colon, which tells SAS to join these contrasts together into a single test with 2 degrees of freedom. 

proc glm data = elemapi2;   class collcat mealcat;   model api00 = collcat mealcat collcat*mealcat;   contrast 'test of sm11 and sm12' collcat*mealcat    1  -1  0                                                     -.5  .5  0                                                     -.5  .5  0, 				    collcat*mealcat   0   1 -1                                                       0 -.5 .5                                                       0 -.5 .5;   contrast 'test of sm21 and sm22'  collcat*mealcat   0   0  0                                                       1  -1  0                                                      -1   1  0, 				    collcat*mealcat   0   0  0                                                       0   1 -1                                                       0  -1  1; run; quit;
The GLM Procedure<output omitted>Contrast                    DF    Contrast SS    Mean Square   F Value   Pr > Ftest of sm11 and sm12        2    54141.40962    27070.70481      5.78   0.0033test of sm21 and sm22        2    66511.60133    33255.80067      7.11   0.0009

6.4.2 Analyzing partial interactions Using PROC REG

With regression analysis, we can also compare groups 1 vs. 2 and 3 on collcat, or compare groups 2 and 3 on collcat. This implies Helmert coding on collcat, as we did before. 

data reg3;   set elemapi2;   if mealcat = 1 then m1 = 2/3;   if mealcat = 2 then m1 = -1/3;   if mealcat = 3 then m1 = -1/3;   if mealcat = 1 then m2 = 1/3;   if mealcat = 2 then m2 = 1/3;   if mealcat = 3 then m2 = -2/3;     if collcat = 1 then s1 =  2/3;   if collcat = 2 then s1 = -1/3;   if collcat = 3 then s1 = -1/3;   if collcat = 1 then s2 =  0;   if collcat = 2 then s2 =  1/2;   if collcat = 3 then s2 = -1/2;    sm11 = s1*m1;   sm12 = s1*m2;   sm21 = s2*m1;   sm22 = s2*m2; run;  proc reg data = reg3;   model api00 = s1 s2 m1 m2 sm11 sm12 sm21 sm22;   test sm11 = sm12 = 0;   test sm21 = sm22 = 0; run; quit;
The REG ProcedureModel: MODEL1Dependent Variable: api00 api 2000                             Analysis of Variance                                    Sum of           MeanSource                   DF        Squares         Square    F Value    Pr > FModel                     8        6243715         780464     166.76    <.0001Error                   391        1829957     4680.19741Corrected Total         399        8073672Root MSE             68.41197    R-Square     0.7733Dependent Mean      647.62250    Adj R-Sq     0.7687Coeff Var            10.56356                            Parameter Estimates                               Parameter      StandardVariable    Label       DF      Estimate         Error   t Value   Pr > |t|Intercept   Intercept    1     650.08826       3.87189    167.90     <.0001s1                       1     -25.04078       8.34539     -3.00     0.0029s2                       1      -2.81094       9.32938     -0.30     0.7633m1                       1     181.04135       9.07713     19.94     <.0001m2                       1     112.36892       9.90759     11.34     <.0001sm11                     1      69.78440      21.47520      3.25     0.0013sm12                     1     -25.40675      21.06663     -1.21     0.2285sm21                     1      62.53325      19.33438      3.23     0.0013sm22                     1      13.86697      24.21132      0.57     0.5671       Test 1 Results for Dependent Variable api00                                MeanSource             DF         Square    F Value    Pr > FNumerator           2          27071       5.78    0.0033Denominator       391     4680.19741       Test 2 Results for Dependent Variable api00                                MeanSource             DF         Square    F Value    Pr > FNumerator           2          33256       7.11    0.0009Denominator       391     4680.19741

6.5. Interaction Contrasts

Above we saw that a partial interaction allows you to apply contrast coefficients to one of the terms in a 2 way interaction. An interaction contrast allows you to apply contrast coefficients to both of the terms in a two way interaction.

For example, with respect to collcat,  let's say that we wish to compare groups 2 and 3, and with respect to mealcat we wish to compare groups 1 and 2. The table of this looks like this below.

  -1 1 0
Collcat low Collcat Med Collcat High
0 Mealcat Low      
-1 Mealcat Med      
1 Mealcat High      

We also would like to form a second interaction contrast that also compares groups 2 and 3 with respect to collcat, and compares groups 2 and 3 on mealcat. A table of this comparison is shown below.

  0 -1 1
Collcat low Collcat Med Collcat High
0 Mealcat Low      
-1 Mealcat Med      
1 Mealcat High      

If we look at the graph of the predicted values (repeated below) we constructed before, it compares line 2 and 3 (collcat 2 vs. 3) by mealcat 1 vs. 2, and then again by mealcat 2 vs. 3.

6.5.1 Analyzing Interaction Contrasts Using PROG GLM

proc glm data = elemapi2;   class collcat mealcat;   model api00 = collcat mealcat collcat*mealcat;   contrast 'collcat 2v3 with mealcat 1v2' collcat*mealcat 0  0  0                                                           1 -1  0                                                          -1  1  0;   contrast 'somecat 2v3 with mealcat 2v3' collcat*mealcat 0  0  0                                                           0  1 -1 							  0 -1  1; run; quit;
The GLM Procedure <output omitted>Contrast                           DF    Contrast SS    Mean Square   F Valuecollcat 2v3 with mealcat 1v2        1    48958.23687    48958.23687     10.46somceat 2v3 with mealcat 2v3        1     1535.28987     1535.28987      0.33Contrast                       Pr > Fcollcat 2v3 with mealcat 1v2   0.0013somceat 2v3 with mealcat 2v3   0.5671

6.5.2 Analyzing interaction contrasts using PROC REG

In regression analysis, we have seen that difference coding schemes of the variables give us difference contrasts and comparisons. Because we would like to compare groups 1 vs. 2, and then  groups 2 vs. 3 on mealcat,  we will use forward difference coding for mealcat (which will compare 1 vs. 2, then 2 vs. 3). 

data reg4;   set elemapi2;   if mealcat = 1 then m1 = 2/3;   if mealcat = 2 then m1 = -1/3;   if mealcat = 3 then m1 = -1/3;   if mealcat = 1 then m2 = 1/3;   if mealcat = 2 then m2 = 1/3;   if mealcat = 3 then m2 = -2/3;     if collcat = 1 then s1 =  2/3;   if collcat = 2 then s1 = -1/3;   if collcat = 3 then s1 = -1/3;   if collcat = 1 then s2 =  0;   if collcat = 2 then s2 =  1/2;   if collcat = 3 then s2 = -1/2;    sm11 = s1*m1;   sm12 = s1*m2;   sm21 = s2*m1;   sm22 = s2*m2; run;  proc reg data = reg4;   model api00 = s1 s2 m1 m2 sm11 sm12 sm21 sm22; run; quit;
The REG ProcedureModel: MODEL1Dependent Variable: api00 api 2000                             Analysis of Variance                                    Sum of           MeanSource                   DF        Squares         Square    F Value    Pr > FModel                     8        6243715         780464     166.76    <.0001Error                   391        1829957     4680.19741Corrected Total         399        8073672Root MSE             68.41197    R-Square     0.7733Dependent Mean      647.62250    Adj R-Sq     0.7687Coeff Var            10.56356                            Parameter Estimates                               Parameter      StandardVariable    Label       DF      Estimate         Error   t Value   Pr > |t|Intercept   Intercept    1     650.08826       3.87189    167.90     <.0001s1                       1     -25.04078       8.34539     -3.00     0.0029s2                       1      -2.81094       9.32938     -0.30     0.7633m1                       1     181.04135       9.07713     19.94     <.0001m2                       1     112.36892       9.90759     11.34     <.0001sm11                     1      69.78440      21.47520      3.25     0.0013sm12                     1     -25.40675      21.06663     -1.21     0.2285sm21                     1      62.53325      19.33438      3.23     0.0013sm22                     1      13.86697      24.21132      0.57     0.5671

6.6 Computing Adjusted Means

Our model will be almost  the same as before, in addition we include an additional covariate emer. We want to obtain the adjusted means of api00 adjusted for variable emer. These adjusted means compute the mean that would be expected if every school in the sample were at the mean for the variable emer.

6.6.1 Computing Adjusted Means via PROC GLM

The syntax to get the adjusted means using proc glm is as follows. The default is to adjust at the means and it can be changed by using at variable = value option following the lsmeans statement.  

proc glm data = elemapi2;   class collcat mealcat;   model api00 = collcat mealcat collcat*mealcat emer /ss3;   lsmeans collcat*mealcat; run; quit;
The GLM Procedure                                      Sum ofSource                     DF        Squares    Mean Square   F Value   Pr > FModel                       9    6402428.265     711380.918    166.01   <.0001Error                     390    1671243.733       4285.240Corrected Total           399    8073671.998R-Square     Coeff Var      Root MSE    api00 Mean0.793001      10.10801      65.46175      647.6225Source                     DF    Type III SS    Mean Square   F Value   Pr > Fcollcat                     2      34730.090      17365.045      4.05   0.0181mealcat                     2    3017331.845    1508665.923    352.06   <.0001collcat*mealcat             4      96789.116      24197.279      5.65   0.0002emer                        1     158713.455     158713.455     37.04   <.0001
collcat    mealcat    api00 LSMEAN1          1            797.5604281          2            596.9728111          3            509.8722412          1            812.5502482          2            636.4049402          3            523.8846593          1            767.9352413          2            652.9761463          3            550.461628

6.6.2 Computing Adjusted Means via REGRESSION

Now we illustrate how to get the same adjusted means if you were to to the analysis via the proc reg. First, we need to create all the necessary dummy variables for the categorical variables. The choice of coding schemes does not matter for the purpose of obtaining the adjusted means. We choose the same coding scheme we used before for both mealcat and collcat below.  After coding our variables properly, we proceed to proc reg to generate the regression equation used later in the proc score statement to generate predicted valued based on the equation. The proc sql statement below simply generates a new variable meanemer as the mean of emer.

data reg6;   set elemapi2;   if collcat = 1 then s2 = 2/3;   if collcat = 2 then s2 = -1/3;   if collcat = 3 then s2 = -1/3;   if collcat = 1 then s3 = -1/3;   if collcat = 2 then s3 = 2/3;   if collcat = 3 then s3 = -1/3;   if mealcat = 1 then m2 = 2/3;   if mealcat = 2 then m2 = -1/3;   if mealcat = 3 then m2 = -1/3;   if mealcat = 1 then m3 = -1/3;   if mealcat = 2 then m3 = 2/3;   if mealcat = 3 then m3 = -1/3;   sm22 = s2*m2;   sm23 = s2*m3;   sm32 = s3*m2;   sm33 = s3*m3; run;  proc reg data = reg6 outest = pred6 noprint;   yhat: model api00 = s2 s3 m2 m3 sm22 sm23 sm32 sm33 emer; run; quit;  proc sql;   create table xy as   select *, mean(emer) as meanemer   from reg6; quit;

NOTE:  You need to rename meanemer to emer or else the proc score will not work The variables listed on the var statement in the proc score must be the same as the IVs in the regression.  If they are not, you get a cryptic message about not finding a variable , even though you can see the variable in the data set.

data xyz;   set xy;   emer = meanemer; run;    proc score data = xyz score = pred6 out = ep type = parms;   var s2 s3 m2 m3 sm22 sm23 sm32 sm33 emer; run;  proc means data = ep mean;   class collcat mealcat;   var yhat; run; 
The MEANS Procedure            Analysis Variable : yhat             Percentage             free meals                   in 3    N  collcat    categories  Obs      Mean   -------------------------------------------    1             1       35   797.5629402                   2       20   596.9753239                   3       74   509.8747538     2             1       43   812.5527606                   2       43   636.4074521                   3       48   523.8871715     3             1       53   767.9377531                   2       69   652.9786583                   3       15   550.4641407 -------------------------------------------

6.7 More Details on Meaning of the Coefficients

So far we have discussed a variety of techniques that you can use to help interpret interactions of categorical variables in regression, but we have not gone into a great detail about the meaning of the coefficients in these analyses. Let's consider this further. Consider the analysis below using collcat and mealcat, using simple contrasts on both of these variables. The reference group for both variables will be group 1. 

data reg7;   set elemapi2;   if collcat = 1 then s1 = -1/3;   if collcat = 2 then s1 = 2/3;   if collcat = 3 then s1 = -1/3;   if collcat = 1 then s2 = -1/3;   if collcat = 2 then s2 = -1/3;   if collcat = 3 then s2 = 2/3;   if mealcat = 1 then m1 = -1/3;   if mealcat = 2 then m1 = 2/3;   if mealcat = 3 then m1 = -1/3;   if mealcat = 1 then m2 = -1/3;   if mealcat = 2 then m2 = -1/3;   if mealcat = 3 then m2 = 2/3;   sm11 = s1*m1;   sm12 = s1*m2;   sm21 = s2*m1;   sm22 = s2*m2; run;  proc reg data = reg7;   model api00 = s1 s2 m1 m2 sm11 sm12 sm21 sm22;   output out = predreg7 p = yhat; run; quit;
The REG ProcedureModel: MODEL1Dependent Variable: api00 api 2000                             Analysis of Variance                                    Sum of           MeanSource                   DF        Squares         Square    F Value    Pr > FModel                     8        6243715         780464     166.76    <.0001Error                   391        1829957     4680.19741Corrected Total         399        8073672Root MSE             68.41197    R-Square     0.7733Dependent Mean      647.62250    Adj R-Sq     0.7687Coeff Var            10.56356                            Parameter Estimates                               Parameter      StandardVariable    Label       DF      Estimate         Error   t Value   Pr > |t|Intercept   Intercept    1     650.08826       3.87189    167.90     <.0001s1                       1      23.63531       9.10533      2.60     0.0098s2                       1      26.44625       9.99513      2.65     0.0085m1                       1    -181.04135       9.07713    -19.94     <.0001m2                       1    -293.41027       9.44946    -31.05     <.0001sm11                     1      38.51777      24.19532      1.59     0.1122sm12                     1       6.17754      20.08262      0.31     0.7585sm21                     1     101.05102      22.88808      4.42     <.0001sm22                     1      82.57776      24.43941      3.38     0.0008

We can produce the adjusted means as shown below. These will be useful for interpreting the meaning of the coefficients.

proc means data = predreg7 mean;  class collcat mealcat;  var yhat; run;
The MEANS Procedure Analysis Variable : yhat Predicted Value of api00                  Percentage                  free meals                        in 3      N     collcat      categories    Obs            Mean---------------------------------------------------           1               1     35     816.9142857                           2     20     589.3500000                           3     74     493.9189189           2               1     43     825.6511628                           2     43     636.6046512                           3     48     508.8333333           3               1     53     782.1509434                           2     69     655.6376812                           3     15     541.7333333---------------------------------------------------

Let's consider the meaning of the coefficient for s1. The coding for this variable compares group 2 vs. group 1, hence this coefficient corresponds to mean(collcat = 2) - mean(collcat = 1). Note that these are the unweighted means, so we compute the mean for collcat = 2 as the mean of the 3 cells corresponding to collcat = 2, i.e. (825.651+636.605+508.833)/3 . If we compare the result below to the coefficient for s1 we see that they are the same, 

(825.651+636.605+508.833)/3 - (816.914+589.35+493.919)/3 = 23.635333.

Likewise, the coefficient for s2 is mean(collcat = 3) - mean(collcat = 1), computed below. The value below corresponds to the coefficient for s2.

(782.151+655.638+541.733)/3 - (816.914+589.35+493.919)/3 = 26.446333

Likewise, the coefficient for m1 works out to be mean(mealcat = 2) - mean(mealcat = 1), computed below.

(589.35+636.605+655.638)/3 - (816.914+825.651+782.151)/3 = -181.041.

And the coefficient for m2 is mean(mealcat = 3) - mean(mealcat = 1), computed below.

(493.919+508.833+541.733)/3 - (816.914+825.651+782.151)/3 = -293.41033

To get the meaning of the coefficients for the interaction terms, let's write out the regression equation and take a closer look at the coefficients. From the parameter estimates, we have the following linear equation for predicted values:

yhat = 650.090 + 23.635*s1   + 26.446*s2                - 181.042*m1  - 293.412*m2                + 38.518*s1*m1  + 6.178*s1*m2                + 101.051*s2*m1 + 82.578*s2*m2.

Because of the simple coding scheme we use for both variables, we have from the above equation,

yhat(collcat = 2) - yhat(collcat = 1) = 23.635 + 38.518*ms1 + 6.178*ms2.

One way to think about this equation is that for any level of mealcat comparing group 2 vs. group 1 on collcat only involves s1. It then follows that the coefficient for sm11 is to compare the difference of group 2 vs. 1 on collcat when mealcat is 2 with the difference of group 2 vs. 1 on collcat  when mealcat is 1. In other words, sm11 is

[cell(2,2)-cell(1,2)] - [cell(2,1)-cell(1,1)]. 

Plugging all the corresponding cell means to the above formula, we get 

        (636.6047 - 589.3500) - (825.6512 - 816.9143) = 38.5175, 

which is the coefficient for sm11. Using the same argument, we can have the following 

sm11 : [cell(2,2)-cell(1,2)] - [cell(2,1)-cell(1,1)],

sm12 : [cell(2,3)-cell(1,3)] - [cell(2,1)-cell(1,1)],

sm21 : [cell(3,2)-cell(1,2)] - [cell(3,1)-cell(1,1)],

sm22 : [cell(3,3)-cell(1,3)] - [cell(3,1)-cell(1,1)].

We can go through the same process to verify the meaning of the coefficients for the other 3 interaction terms. We verify that sm12 is 6.1775.

(508.8333  - 493.9189)  - (825.6512 - 816.9143) = 6.1775.

We also verify that sm21 is 101.051.

(655.6377 - 589.3500) - (782.1509 - 816.9143) = 101.0511.

Last we verify that sm22 is 82.5778.

(  541.7333 - 493.9189) - ( 782.1509 - 816.9143) = 82.5778.

6.8 Simple Effects via Dummy Coding vs. Effect Coding

We have used in this chapter different types of coding schemes. You may wonder why we have gone to the effort of creating and testing these effects instead of just using dummy coding and what is the difference between different coding schemes and how to choose them. In this section,  let's compare how to get simple effects using the effect coding to how we would get simple effects using dummy coding. We hope to show that it is much easier to use effect coding so that the interpretation of the coefficients is much more intuitive.

6.8.1 Example 1. Simple effects of yr_rnd at levels of mealcat

Let's use an example from Chapter 3 (section 3.5). In that example we looked at and analysis using mealcat and yr_rnd and the interaction of these two variables. First, we look at how to do a simple effects analysis looking at the simple effects of yr_rnd at each level of mealcat using effect coding. To make our results correspond to those from Chapter 3, we will make category 3 of mealcat the reference category.

data reg8;   set elemapi2;  if mealcat = 1 then do; ms1 =2/3;  ms2 = -1/3; end;   if mealcat = 2 then do; ms1 =-1/3; ms2= 2/3; end;  if mealcat = 3 then do; ms1 =-1/3; ms2 = -1/3; end;  if yr_rnd = 0 then yr1 = -1/2;  else yr1 = 1/2;  ym1 = 0;  ym2 = 0;  ym3 = 0;  if mealcat = 1 then  ym1 = yr1;  if mealcat = 2 then  ym2 = yr1;  if mealcat = 3 then  ym3 = yr1; run; proc reg data = reg8;   model api00 = ms1 ms2 ym1 ym2 ym3; run; quit;
The REG ProcedureModel: MODEL1Dependent Variable: API00 api 2000
                             Analysis of Variance                                    Sum of           MeanSource                   DF        Squares         Square    F Value    Pr > FModel                     5        6204728        1240946     261.61    <.0001Error                   394        1868944     4743.51314Corrected Total         399        8073672Root MSE             68.87317    R-Square     0.7685Dependent Mean      647.62250    Adj R-Sq     0.7656Coeff Var            10.63477                               Parameter Estimates                                  Parameter       StandardVariable     Label        DF       Estimate          Error    t Value    Pr > |t|Intercept    Intercept     1      632.23557        5.80048     109.00      <.0001MS1                        1      267.81076       14.61559      18.32      <.0001MS2                        1      114.65715       11.12812      10.30      <.0001ym1                        1      -74.25691       26.75629      -2.78      0.0058ym2                        1      -51.74017       18.88854      -2.74      0.0064ym3                        1      -33.49254       11.77129      -2.85      0.0047
Now we can obtain the simple effect of yr_rnd at mealcat = 1 by inspecting the coefficient for ym1, the simple effect of yr_rnd at mealcat = 2 by inspecting the coefficient for ym2 and the simple effect of yr_rnd at mealcat = 3 by inspecting the coefficient for ym3.

Now let's perform the same analysis using dummy coding. Again, we will explicitly make the 3rd category for mealcat to be the omitted category.

data reg9;   set elemapi2;   if mealcat = 1 then do; md1 = 1; md2 = 0; end;   if mealcat = 2 then do; md1 = 0; md2 = 1; end;   if mealcat = 3 then do; md1 = 0; md2 = 0; end; 	ymd1 = yr_rnd*md1;     ymd2 = yr_rnd*md2; run;
proc reg data = reg9;   model api00 = yr_rnd md1 md2 ymd1 ymd2; run;
The REG ProcedureModel: MODEL1Dependent Variable: API00                        Parameter Estimates                     Parameter       StandardVariable     DF       Estimate          Error    t Value    Pr > |t|Intercept     1      521.49254        8.41420      61.98      <.0001YR_RND        1      -33.49254       11.77129      -2.85      0.0047MD1           1      288.19295       10.44284      27.60      <.0001MD2           1      123.78097       10.55185      11.73      <.0001ymd1          1      -40.76438       29.23118      -1.39      0.1639ymd2          1      -18.24763       22.25624      -0.82      0.4128

In order to form a test of simple main effects we need to make a table like the one shown below that relates the cell means to the coefficients in the regression. Please see Chapter 3, section 3.5 for information on how this table was constructed.

            mealcat=1           mealcat=2         mealcat=3
            -------------------------------------------------
  yr_rnd=0   const               const             const    
             + md1               + md2            
            -------------------------------------------------
  yr_rnd=1  const               const             const    
            + yr_rnd            + yr_rnd          + yr_rnd
            + md1               + md2           
            + ymd1              + ymd2 

Let's start by looking at how to get the simple effect of yr_rnd when mealcat is 3. Looking at the table above, we can see that we would want to compare const with const  +  yr_rnd, , which is the same as testing the coefficient for yr_rnd is zero. This is a single parameter test and is shown in the output above. The t-value is -2.85 and the p-value is .0047.
Note that the coefficient for yr_rnd corresponds to the test of the effect of yr_rnd when all other variables are set to 0 (the reference category), i.e. when mealcat is set to the reference category. You may be tempted to interpret the coefficient for yr_rnd as the overall difference between year round schools and non-year round schools, but in this example we see that it really corresponds to the simple effect of yr_rnd. When using dummy coding people commonly misinterpret the lower order effects to refer to overall effects rather than simple effects.

Now let's look at the simple effect of yr_rnd when mealcat=1. Looking at the table above we see that this involves the comparison of the coefficients for yr_rnd=1 vs. yr_rnd=0 when mealcat=1, i.e. comparing const + yr_rnd +md1 + ymd1 vs. const + md1. Removing the terms that drop out we see that to test the simple effect of yr_rnd when mealcat = 1 is the same to test yr_rnd + ymd1 = 0. We will have to do a test statement here following the previous proc reg

  test yr_rnd + ymd1 = 0; run; quit;       Test 1 Results for Dependent Variable API00                                MeanSource             DF         Square    F Value    Pr > FNumerator           1          36536       7.70    0.0058Denominator       394     4743.51314
These examples illustrate that it is more complicated to form simple effects when using dummy coding, and also that the interpretation of lower order effects when using dummy coding may not have the meaning that you would expect.

 6.8.2 Example 2. Simple effects of mealcat at levels of yr_rnd

Example 1 looked at simple effects for yr_rnd, a variable with only 2 levels and it showed how to use the test statement in SAS for it. In this example, let's consider the simple effects of mealcat at each level of yr_rnd. Because mealcat has more than 2 levels, we will see what is required for doing tests of simple effects for variables with more than 2 levels. We will show both proc glm and proc reg approach here. 

proc glm data = elemapi2;   class yr_rnd mealcat;   model api00 = yr_rnd mealcat yr_rnd*mealcat;   contrast '1' mealcat 1 0 -1                yr_rnd*mealcat 1 0 -1                               0 0  0,                         mealcat 0 1 -1                yr_rnd*mealcat 0 1 -1                               0 0  0;   contrast '2' mealcat 1 0 -1                yr_rnd*mealcat 0 0  0                               1 0 -1,                         mealcat 0 1 -1                yr_rnd*mealcat 0 0  0                               0 1 -1; run; quit;
The GLM Procedure<output omitted>
Contrast       DF    Contrast SS    Mean Square   F Value   Pr > F1              2    3903569.804    1951784.902    411.46   <.00012              2     476157.455     238078.727     50.19   <.0001

Here is how to do it with proc reg. The first test statement below looks at mealcat at yr_rnd = 0 and the second test statement looks at mealcat at yr_rnd = 1.

data reg10;   set elemapi2;   if yr_rnd = 0 then yrrnd = -.5;   if yr_rnd = 1 then yrrnd = .5;   if mealcat = 1 then m1 = 2/3;   if mealcat = 2 then m1 = -1/3;   if mealcat = 3 then m1 = -1/3;   if mealcat = 1 then m2 = -1/3;   if mealcat = 2 then m2 = 2/3;   if mealcat = 3 then m2 = -1/3;   if yr_rnd = 0 then my11 = m1; else my11 = 0;   if yr_rnd = 0 then my21 = m2; else my21 = 0;   if yr_rnd = 1 then my12 = m1; else my12 = 0;   if yr_rnd = 1 then my22 = m2; else my22 = 0; run;
proc reg data = reg10;  model api00 = yrrnd my11 my21 my12 my22;  test my11 = my21 = 0;  test my12 = my22 = 0; run; quit;
The REG ProcedureModel: MODEL1Dependent Variable: api00 api 2000                            Parameter Estimates                               Parameter      StandardVariable    Label       DF      Estimate         Error   t Value   Pr > |t|Intercept   Intercept    1     632.23557       5.80048    109.00     <.0001yrrnd                    1     -53.16321      11.60095     -4.58     <.0001my11                     1     288.19295      10.44284     27.60     <.0001my21                     1     123.78097      10.55185     11.73     <.0001my12                     1     247.42857      27.30218      9.06     <.0001my22                     1     105.53333      19.59588      5.39     <.0001       Test 1 Results for Dependent Variable api00                                MeanSource             DF         Square    F Value    Pr > FNumerator           2        1951785     411.46    <.0001Denominator       394     4743.51314       Test 2 Results for Dependent Variable api00                                MeanSource             DF         Square    F Value    Pr > FNumerator           2         238079      50.19    <.0001Denominator       394     4743.51314

We can also test the simple effects of mealcat at each level of yr_rnd via dummy coding. In SAS, each equal sign in the test statement equals one degree of freedom:  because there are two equals signs in the second test statement, it is a two degree-of-freedom test, which is meant to do. The same logic holds true for the fourth test statement and this test is the simple effect of mealcat when yr_rnd=1.

data reg11;   set elemapi2;   m1 = 0;   if mealcat = 1 then m1 = 1;   m2 = 0;   if mealcat = 2 then m2 = 1;   m1y = m1*yr_rnd;   m2y = m2*yr_rnd; run;
proc reg data = reg11;   model api00 = m1 m2 yr_rnd m1y m2y;   test m1 - m2 = 0;   test m1 = m2 = 0;   test m1 + m1y - m2 - m2y = 0;   test m1 + m1y = m2 + m2y = 0; run; quit;
The REG ProcedureModel: MODEL1Dependent Variable: api00 api 2000       Test 1 Results for Dependent Variable api00                                MeanSource             DF         Square    F Value    Pr > FNumerator           1        1627262     343.05    <.0001Denominator       394     4743.51314       Test 2 Results for Dependent Variable api00                                MeanSource             DF         Square    F Value    Pr > FNumerator           2        1951785     411.46    <.0001Denominator       394     4743.51314       Test 3 Results for Dependent Variable api00                                MeanSource             DF         Square    F Value    Pr > FNumerator           1          96095      20.26    <.0001Denominator       394     4743.51314       Test 4 Results for Dependent Variable api00                                MeanSource             DF         Square    F Value    Pr > FNumerator           2         238079      50.19    <.0001Denominator       394     4743.51314

 

For more information

The content of this web site should not be construed as an endorsement of any particular web site, book, or software product by the University of California.

 
[ 打印 ]
閱讀 ()評論 (0)
評論
目前還沒有任何評論
登錄後才可評論.