No Title

$next$ $up$ $previous$

STAT 350: Lecture 25

Interaction effects

If a model contains terms then we define the UV interaction term to be , that is, we define a new column of the design matrix which is the product of the U column and the V column.
Something analogous can be done if one of the variables (or both) is categorical.
A factor at K levels will have K-1 columns in the design matrix.
We create an interaction between this factor and a continuous variable X by adding K-1 new columns obtained by multiplying the X column by each of the K-1 columns.
We create an interaction between a factor at levels and another at levels by multiplying each of the columns corresponding to the first factor by each of the columns of the second factor.
Three way interactions are obtained by multiplying columns corresponding to 3 covariates, and so on for higher way interactions.
An interaction term between a continuous covariate and a categorical factor amounts to allowing the slope corresponding to the continuous covariate to depend on the level of the categorical factor.

Examples

Two way analysis of variance

Experiment involving 2 diets, 3 drugs and 2 experimental units per combination - a factorial experiment.
is response for Diet i, Drug j and replicate k.
Additive model is
Impose restriction .
DESIGN MATRIX
First column corresponds to , the grand mean. Second column to and uses . Third and fourth columns for and use .
To put in interaction terms multiply column 2 by 3 and 2 by 4 to get

This design matrix corresponds to a model equation

where the are interaction effects satisfying .

Analysis of Covariance

This is the name given to the analysis of models in which there are categorical factors and continuous covariates. In the car example we had the categorical factor VEHICLE and the continuous covariate MILEAGE. Earlier I gave the design matrix for the model in which there are different intercepts for the two cars but 1 common slope. thus this model is 2 parallel lines. If we use corner point coding and fit a model in which VEHICLE and MILEAGE interact then the design matrix for the small data set above is

displaymath138

where the last column is the product of columns 2 and 3. The design matrix corresponds to a model equation with some slope for the first vehicle and a slope for the second vehicle. That is, the coefficient of the last column of the design matrix is the difference in slopes between the 2 vehicles. Use of the alternative coding based on an average intercept leads now to the design matrix

displaymath144

Again the last column is the product of columns 2 and 3. the coefficient of column 3 is the average slope while the coefficient of the last column is the difference between the slope for vehicle 1 and this average slope.

You saw, in assignment 3, how to test the hypothesis of no interaction in this model.

Analysis of Models with Interaction Terms

Usually: use F-tests to eliminate for higher order interactions (i.e., interactions involving several variables).
Then test for lower order interactions.
To use SAS: put interaction terms in last. Use Type I or sequential SS.
Do not consider models involving a two factor interaction unless individual (also called main) effects included.
Similarly SS for 3 factor interaction normally adjusted for
- all 3 main effects
- all 3 two variable interactions of the 3 variables.

Examples

Two way ANOVA: influence of SCHOOL, REGION on STAY

options pagesize=60 linesize=80;
data scenic;
 infile 'scenic.dat' firstobs=2;
 input Stay  Age Risk Culture Chest Beds 
          School Region Census Nurses Facil;
proc glm  data=scenic;
  class school region ;
  model Stay = School | Region / E 
          SOLUTION SS1 SS2 SS3 SS4 XPX INVERSE;
  output out=scout P=Fitted PRESS=PRESS H=HAT 
   RSTUDENT =EXTST R=RESID DFFITS=DFFITS COOKD=COOKD;
run ;
proc means data=scout;
  var stay;
  class school region;
run;
proc print data=scout;

EDITED SAS OUTPUT (Complete output)

                                 The X'X Matrix
                INTERCEPT     SCHOOL 1     SCHOOL 2     REGION 1     REGION 2
INTERCEPT             113           17           96           28           32
SCHOOL 1               17           17            0            5            7
SCHOOL 2               96            0           96           23           25
REGION 1               28            5           23           28            0
REGION 2               32            7           25            0           32
REGION 3               37            3           34            0            0
REGION 4               16            2           14            0            0
DUMMY001                5            5            0            5            0
DUMMY002                7            7            0            0            7
DUMMY003                3            3            0            0            0
DUMMY004                2            2            0            0            0
DUMMY005               23            0           23           23            0
DUMMY006               25            0           25            0           25
DUMMY007               34            0           34            0            0
DUMMY008               14            0           14            0            0
STAY              1090.26       186.85       903.41       310.49       309.87

                          X'X Generalized Inverse (g2)
                INTERCEPT     SCHOOL 1     SCHOOL 2     REGION 1     REGION 2
INTERCEPT    0.0714285714 -0.071428571            0 -0.071428571 -0.071428571
SCHOOL 1     -0.071428571 0.5714285714            0 0.0714285714 0.0714285714
SCHOOL 2                0            0            0            0            0
REGION 1     -0.071428571 0.0714285714            0 0.1149068323 0.0714285714
REGION 2     -0.071428571 0.0714285714            0 0.0714285714 0.1114285714
REGION 3     -0.071428571 0.0714285714            0 0.0714285714 0.0714285714
REGION 4                0            0            0            0            0
DUMMY001     0.0714285714 -0.571428571            0 -0.114906832 -0.071428571
DUMMY002     0.0714285714 -0.571428571            0 -0.071428571 -0.111428571
DUMMY003     0.0714285714 -0.571428571            0 -0.071428571 -0.071428571
DUMMY004                0            0            0            0            0
DUMMY005                0            0            0            0            0
DUMMY006                0            0            0            0            0
DUMMY007                0            0            0            0            0
DUMMY008                0            0            0            0            0
STAY                 7.89         1.79            0 2.9304347826       1.5372

Dependent Variable: STAY   
                                     Sum of            Mean
Source                  DF          Squares          Square   F Value     Pr > F
Model                    7     132.06558693     18.86651242      7.15     0.0001
Error                  105     277.14479360      2.63947422
Corrected Total        112     409.21038053
                  R-Square             C.V.        Root MSE            STAY Mean
                  0.322733         16.83864       1.6246459            9.6483186
Source                  DF        Type I SS     Mean Square   F Value     Pr > F
SCHOOL                   1      36.08413010     36.08413010     13.67     0.0003
REGION                   3      95.36410217     31.78803406     12.04     0.0001
SCHOOL*REGION            3       0.61735466      0.20578489      0.08     0.9718
Source                  DF       Type II SS     Mean Square   F Value     Pr > F
SCHOOL                   1      27.89404890     27.89404890     10.57     0.0015
REGION                   3      95.36410217     31.78803406     12.04     0.0001
SCHOOL*REGION            3       0.61735466      0.20578489      0.08     0.9718
Source                  DF      Type III SS     Mean Square   F Value     Pr > F
SCHOOL                   1      26.05955792     26.05955792      9.87     0.0022
REGION                   3      47.01938029     15.67312676      5.94     0.0009
SCHOOL*REGION            3       0.61735466      0.20578489      0.08     0.9718
Source                  DF       Type IV SS     Mean Square   F Value     Pr > F
SCHOOL                   1      26.05955792     26.05955792      9.87     0.0022
REGION                   3      47.01938029     15.67312676      5.94     0.0009
SCHOOL*REGION            3       0.61735466      0.20578489      0.08     0.9718
                                        T for H0:    Pr > |T|   Std Error of
Parameter                  Estimate    Parameter=0                Estimate
INTERCEPT               7.890000000 B        18.17     0.0001     0.43420487
SCHOOL        1         1.790000000 B         1.46     0.1480     1.22811685
              2         0.000000000 B          .        .          .        
REGION        1         2.930434783 B         5.32     0.0001     0.55072100
              2         1.537200000 B         2.83     0.0055     0.54232171
              3         1.180588235 B         2.29     0.0241     0.51591227
              4         0.000000000 B          .        .          .        
SCHOOL*REGION 1 1      -0.286434783 B        -0.20     0.8455     1.46660342
              1 2      -0.618628571 B        -0.44     0.6620     1.41099883
              1 3      -0.300588235 B        -0.19     0.8486     1.57026346
              1 4       0.000000000 B          .        .          .        
SCHOOL*REGION 2 1       0.000000000 B          .        .          .        
              2 2       0.000000000 B          .        .          .        
              2 3       0.000000000 B          .        .          .        
              2 4       0.000000000 B          .        .          .        

NOTE: The X'X matrix has been found to be singular and a generalized inverse 
      was used to solve the normal equations.   Estimates followed by the 
      letter 'B' are biased, and are not unique estimators of the parameters.
      SCHOOL        REGION  N Obs    N          Mean       Std Dev       Minimum
--------------------------------------------------------------------------------
           1             1      5    5    12.3240000     3.3527198     9.7800000
                         2      7    7    10.5985714     1.1317454     8.2800000
                         3      3    3    10.5600000     0.7362744    10.1200000
                         4      2    2     9.6800000     0.6788225     9.2000000
           2             1     23   23    10.8204348     2.5061460     8.0300000
                         2     25   25     9.4272000     1.0978635     7.3900000
                         3     34   34     9.0705882     1.1911516     7.0800000
                         4     14   14     7.8900000     0.8332420     6.7000000
--------------------------------------------------------------------------------
  OBS  STAY  AGE RISK CULTURE CHEST BEDS SCHOOL REGION CENSUS NURSES FACIL

   23  9.78 52.3  5.0   17.6   95.9  270    1      1     240    198   57.1
   25  9.20 52.2  4.0   17.5   71.1  298    1      4     244    236   57.1
   26  8.28 49.5  3.9   12.0  113.1  546    1      2     413    436   57.1
   44 10.12 51.7  5.6   14.9   79.1  362    1      3     313    264   54.3
   46 10.16 54.2  4.6    8.4   51.5  831    1      4     581    629   74.3
   47 19.56 59.9  6.5   17.2  113.7  306    2      1     273    172   51.4
   74 10.05 52.0  4.5   36.7   87.5  184    1      1     144    151   68.6
   90 11.41 50.4  5.8   23.8   73.0  424    1      3     359    335   45.7
  100 10.15 51.9  6.2   16.4   59.2  568    1      3     452    371   62.9
  112 17.94 56.2  5.9   26.4   91.8  835    1      1     791    407   62.9

  OBS  FITTED     PRESS      HAT       EXTST      RESID     DFFITS     COOKD

   23 12.3240   -3.18000   0.20000   -1.76835   -2.54400   -0.88418   0.09578
   25  9.6800   -0.96000   0.50000   -0.41618   -0.48000   -0.41618   0.02182
   26 10.5986   -2.70500   0.14286   -1.55177   -2.31857   -0.63351   0.04950
   44 10.5600   -0.66000   0.33333   -0.33029   -0.44000   -0.23355   0.00688
   46  9.6800    0.96000   0.50000    0.41618    0.48000    0.41618   0.02182
   47 10.8204    9.13682   0.04348    6.48789    8.73957    1.38322   0.17189
   74 12.3240   -2.84250   0.20000   -1.57592   -2.27400   -0.78796   0.07653
   90 10.5600    1.27500   0.33333    0.63897    0.85000    0.45182   0.02566
  100 10.5600   -0.61500   0.33333   -0.30774   -0.41000   -0.21761   0.00597
  112 12.3240    7.02000   0.20000    4.15303    5.61600    2.07652   0.46676

Comments on code and results

SS1 SS2 SS3 SS4 request 4 different sums of squares

Type I is sequential. The effect of SCHOOL has not been adjusted for anything, that of REGION has been adjusted for the main effect of SCHOOL and the interaction SS has been adjusted for each main effect.
Type II sums of squares have a complicated definition but here the SS for schools has been adjusted for the main effect of REGION and the SS of REGION has been adjusted for the main effect of SCHOOL. In other words each of these 2 SS is obtained by comparing an additive model with the 2 main effects to a model with just 1 main effect. The SS for SCHOOL*REGION has been adjusted for both main effects; this is definitely the relevant SS for testing for an interaction effect.
Type III sums of squares are generally based on comparing a model in which all the effects are present to one in which a given effect has been deleted. However, when some of the effects are interaction effects this definition depends on how the interaction effects are coded. SAS does its arithmetic on the basis of an overparameterized model. As a result the main effect columns for REGION are linearly dependent on the SCHOOL*REGION interactions columns. As a result if you just adjust for SCHOOL*REGION there are no remaining degrees of freedom for REGION. The package then uses the restriction to define interaction effects and the Type III SS has been adjusted for the corresponding columns of the design matrix.
Almost always Type IV SS are equal to Type II SS; exceptions arise when some combinations of factor levels have no observations.
In general it seems wise to test hypotheses about low order interactions only after concluding that higher order interactions are negligible. This can always be done by an extra sum of squares F-test.
There are only 2 hospitals in region 4 attached to a medical school; this gives the leverages of 0.5 for observations 25 and 46.
Observations 47 and 112 rare gross outliers, looking at the case deleted studentized residuals. Observation 112 exerts a disproportionate influence on the fitted value for case 112 and apparently also on the whole fitted vector. In fact case 112 only influences the fitted vector for the 5 observations in region 1 which are attached to a medical school. The value of STAY is much larger for Observation 112 than for any of the other 4; once again hospital 112 seems unusual.
The options XPX and INVERSE on the model statement request and be printed. These are easier to interpret when you do not have categorical variables; when there are categorical variables is singular and the `inverse' is actually a `generalized inverse'.

Analysis of covariance example

Here I regress STAY on SCHOOL, REGION and FACILITIES. I begin by putting in all the possible interaction effects.

options pagesize=60 linesize=80;
data scenic;
 infile 'scenic.dat' firstobs=2;
 input Stay  Age Risk Culture Chest Beds School Region Census Nurses Facil;
proc glm  data=scenic;
  class school region ;
  model Stay = School | Region | Facil / SS1 SS2 SS3 ;
  output out=scout P=Fitted PRESS=PRESS H=HAT RSTUDENT =EXTST R=RESID DFFITS=DFFITS COOKD=COOKD;
run ;
proc print data=scout;
proc glm  data=scenic;
  class school region ;
  model Stay = School | Region  Facil / SS1 SS2 SS3 ;
run ;

EDITED SAS OUTPUT (Complete output)

Dependent Variable: STAY   
                                     Sum of            Mean
Source                  DF          Squares          Square   F Value     Pr > F
Model                   15     173.90201568     11.59346771      4.78     0.0001
Error                   97     235.30836485      2.42585943
Corrected Total        112     409.21038053
                  R-Square             C.V.        Root MSE            STAY Mean
                  0.424970         16.14289       1.5575171            9.6483186
Source                  DF        Type I SS     Mean Square   F Value     Pr > F
SCHOOL                   1      36.08413010     36.08413010     14.87     0.0002
REGION                   3      95.36410217     31.78803406     13.10     0.0001
SCHOOL*REGION            3       0.61735466      0.20578489      0.08     0.9682
FACIL                    1       9.52496125      9.52496125      3.93     0.0504
FACIL*SCHOOL             1       1.32686372      1.32686372      0.55     0.4613
FACIL*REGION             3      21.28634656      7.09544885      2.92     0.0377
FACIL*SCHOOL*REGION      3       9.69825722      3.23275241      1.33     0.2683
Source                  DF       Type II SS     Mean Square   F Value     Pr > F
SCHOOL                   1       4.73069924      4.73069924      1.95     0.1658
REGION                   3       8.16560072      2.72186691      1.12     0.3441
SCHOOL*REGION            3       7.04260265      2.34753422      0.97     0.4113
FACIL                    1       9.52496125      9.52496125      3.93     0.0504
FACIL*SCHOOL             1       3.76491803      3.76491803      1.55     0.2158
FACIL*REGION             3      21.28634656      7.09544885      2.92     0.0377
FACIL*SCHOOL*REGION      3       9.69825722      3.23275241      1.33     0.2683
Source                  DF      Type III SS     Mean Square   F Value     Pr > F
SCHOOL                   1       2.34679006      2.34679006      0.97     0.3278
REGION                   3       2.46002453      0.82000818      0.34     0.7979
SCHOOL*REGION            3       7.04260265      2.34753422      0.97     0.4113
FACIL                    1       0.70390965      0.70390965      0.29     0.5913
FACIL*SCHOOL             1       1.50831325      1.50831325      0.62     0.4323
FACIL*REGION             3       1.92051520      0.64017173      0.26     0.8513
FACIL*SCHOOL*REGION      3       9.69825722      3.23275241      1.33     0.2683
  OBS  STAY  AGE RISK CULTURE CHEST BEDS SCHOOL REGION CENSUS NURSES FACIL
   25  9.20 52.2  4.0   17.5   71.1  298    1      4     244    236   57.1
   46 10.16 54.2  4.6    8.4   51.5  831    1      4     581    629   74.3
   47 19.56 59.9  6.5   17.2  113.7  306    2      1     273    172   51.4
  OBS  FITTED     PRESS      HAT       EXTST      RESID     DFFITS     COOKD
   25  9.2000     .        1.00000     .        -0.00000     .         .     
   46 10.1600     .        1.00000     .         0.00000     .         .     
   47 11.8970    8.29701   0.07641    5.96177    7.66301    1.71483   0.13553

COMMENTS

The model fits straight line regressions of STAY on FACILITIES with different slopes and intercepts for each combination of SCHOOL and REGION. Observe that the entries for observations 25 and 46 have leverages of 1; the model fits perfectly for these points because there are just 2 hospitals in Region 4 attached to medical schools. You get a slope and an intercept for these 2 schools so the line goes right through the 2 points and the variance of the corresponding residuals is 0.

The slopes and intercepts have been decomposed in the same way that the means in a 2 way layout are decomposed into main effects and interactions. Normally we might begin by looking for any interaction of facility with anything by comparing the full model to a model with no interaction effects. This is what the second proc glm run does. More of the output follows.

Dependent Variable: STAY   
                                     Sum of            Mean
Source                  DF          Squares          Square   F Value     Pr > F
Model                    8     141.59054818     17.69881852      6.88     0.0001
Error                  104     267.61983235      2.57326762
Corrected Total        112     409.21038053
                  R-Square             C.V.        Root MSE            STAY Mean
                  0.346009         16.62612       1.6041408            9.6483186
Source                  DF        Type I SS     Mean Square   F Value     Pr > F
SCHOOL                   1      36.08413010     36.08413010     14.02     0.0003
REGION                   3      95.36410217     31.78803406     12.35     0.0001
SCHOOL*REGION            3       0.61735466      0.20578489      0.08     0.9708
FACIL                    1       9.52496125      9.52496125      3.70     0.0571
Source                  DF       Type II SS     Mean Square   F Value     Pr > F
SCHOOL                   1       8.66242211      8.66242211      3.37     0.0694
REGION                   3      82.48995156     27.49665052     10.69     0.0001
SCHOOL*REGION            3       0.48049197      0.16016399      0.06     0.9796
FACIL                    1       9.52496125      9.52496125      3.70     0.0571
Source                  DF      Type III SS     Mean Square   F Value     Pr > F
SCHOOL                   1       8.45264294      8.45264294      3.28     0.0728
REGION                   3      42.65719728     14.21906576      5.53     0.0015
SCHOOL*REGION            3       0.48049197      0.16016399      0.06     0.9796
FACIL                    1       9.52496125      9.52496125      3.70     0.0571

The error sum of squares has risen from 235.208 to 267.620, an increase of 32.312 on 7 degrees of freedom for a Mean Square of 4.62. The F statistic is 4.62/[235.208/97] which is 1.90 leading to a P-value of 0.077 which is not quite significant. It seems reasonable to drop these interaction terms and just let the intercepts depend on the categorical covariates.
In this new model the SCHOOL*REGION interaction is not significant - look at either the Type II or II SS which have been adjusted for all the other effects.
The main effects of both SCHOOL and FACILITIES are not quite significant and we should look at dropping them from the model but I will leave the analysis at this point.

$next$ $up$ $previous$

Richard Lockhart
Wed Mar 5 22:52:36 PST 1997