Let’s Interact!

Modeling Interaction Effects in

Linear and Generalized Linear Models using SAS®
Timothy B. Gravelle
Principal Scientist & Director, Insights Lab
PriceMetrix Inc.

© 2000-2012 PriceMetrix Inc. Patents granted and pending.

 Which of these statements are accurate?
The REG Procedure

Dependent Variable: TOTAL_REV_11

Parameter Standard
Variable Estimate Error t Value Pr > |t|

Intercept 739.79385 7.74079 95.57 <.0001

ACCTS_PER_HH 46.24691 22.25271 2.08 0.0379
MEDIAN_HH_ASSETS -0.30548 0.12804 -2.39 0.0172
ACCTS_PER_HH_MED_HH_ASSETS 0.56231 0.17897 3.14 0.0017

1. The coefficient for ACCTS_PER_HH is statistically significant. Therefore, the

hypothesis of a relationship between ACCTS_PER_HH and TOTAL_REV_11 (the
dependent variable) is confirmed.
2. The coefficient for MEDIAN_HH_ASSETS is statistically significant. Therefore, the
hypothesis of a relationship between MEDIAN_HH_ASSETS and TOTAL_REV_11 is
confirmed.
3. The coefficient for the interaction term ACCTS_PER_HH_MEDIAN_HH_ASSETS is
statistically significant. Therefore, the hypothesis relating the combination of
ACCTS_PER_HH and MEDIAN_HH_ASSETS to TOTAL_REV_11 is confirmed.

2
 Topics

• A (brief) review of the theoretical framework for

interaction effects
• Data preparation for testing interactions
• Specifying interactive models in PROC REG and
PROC LOGISTIC
• Graphical displays for interaction effects

3
 Two examples

• Predicting investment advisors’ future productivity

(linear model)
• Canadian attitudes toward Canada-US relations
(ordinal logit model)

4
 Theory

He who loves practice without theory is like the

sailor who boards ship without a rudder and
compass and never knows where he may be cast.
– Leonardo da Vinci
Theorizing and specifying interaction effects

• A “system” comprising 3 variables (Jaccard and Turrisi

2003; Jaccard and Dodge 2004):
- Dependent variable
- “Focal” independent variable
- Moderator variable
• Example: employment income (DV), education (focal IV)
and sex (moderator)

6
Theorizing and specifying interaction effects

• Moderation vs. mediation (Baron and Kenny 1986):

- A moderator variable “affects the direction and/or
strength of the relation between an independent or
predictor variable and a dependent or criterion
variable.”
- A mediator variable “accounts for the relation between
the predictor and the criterion.”

Moderation Mediation

X2
X1 X2 Y
X1 Y
7
Theorizing and specifying interaction effects

• Ŷ = β0 + β1X1 + β2X2 + β12X1X2

predicted value (conditional mean) of the
Ŷ
dependent variable

β0 intercept term

β1 coefficients for the independent variables;

β2 “lower-order” terms

coefficient for the interaction term;

β12
“higher order” term
X1
values taken by the independent variables
X2

8
Theorizing and specifying interaction effects

• When testing an interaction effect, the lower-order terms

(β1 and β2) must still be present in the model. Otherwise,
the model is not “hierarchically well-formulated.”
• Even when included in the model β1 and β2 are not of
primary interest. And they are not interpreted as the
“main effects” of β1 and β2 or the “effects of β1 and β2 in
general.”
• Rather, they indicate the effect of X1 on Ŷ when X2 is 0,
and the effect of X2 on Ŷ when X1 is 0.

9
 Case #1
Predicting Investment Advisor Productivity
Predicting Investment Advisor Productivity

• How can one predict investment advisors’ future

productivity (revenue) given data on advisors’
performance in a baseline period, personal
characteristics and information on the advisors’ books?
• Data are drawn from the proprietary PriceMetrix retail
wealth management database.
- 1,010 investment advisors
- Multiple firms across North America (Canada and US)
- Advisors with 5 to 20 years of industry experience as
of end-of-year 2006

11
 Predicting Investment Advisor Productivity

Advisor Data: Basic Descriptive Statistics

The MEANS Procedure

Variable Mean Minimum Maximum Median Std Dev N

---------------------------------------------------------------------------------------------
TOTAL_REV_11 760.17 14.61 3888.97 630.72 481.83 1010
TRANS_REV_06 327.19 7.40 2184.26 253.86 273.46 1010
TRAILER_REV_06 101.43 0.00 726.07 77.91 86.99 1010
FEE_REV_06 199.22 0.00 2443.41 101.16 257.27 1010
EXPERIENCE_YEARS 11.96 5.05 20.00 11.81 4.27 1010
TEAM 0.06 0.00 1.00 0.00 0.24 1010
CORE_HH_COUNT 74.17 1.00 383.00 65.00 44.75 1010
SMALL_HH_COUNT 173.33 4.00 1269.00 142.00 134.40 1010
RETIREMENT_ACCT_COUNT 127.65 0.00 1141.00 112.00 102.08 1010
ACCTS_PER_HH 2.08 1.07 4.01 2.08 0.40 1010
MEDIAN_HH_ASSETS 150.71 0.10 715.06 139.55 87.65 1010
---------------------------------------------------------------------------------------------

12
Data Preparation
 Data Preparation

• All of the standard assumptions underpinning regression

analysis continue to apply.
- linearity in the predictors
- normality
- constant error variance (homoscedasticity)
- independence of the errors
- absence of high collinearity among the predictors

14
 Mean Centering

• One technique that is especially relevant when modeling

interaction effects is mean-centering.
• Involves subtracting the mean from the original scores,
resulting in new scores with a mean of zero.
• Zero now has the interpretation of a variable’s pre-
transformation mean value.
• Jaccard and Turrisi recommend this strategy as a way to
“force the coefficients to reflect parameters that are of
theoretical interest” (2003: 15).
• Median-centering is also an option.

15
 Mean Centering

• Easy to implement in SAS using PROC STDIZE using

the METHOD=MEAN option (default method creates
standardized (Z) scores).
• Example:
PROC STDIZE DATA=data_1 OUT=data_2 METHOD=MEAN;
VAR TRANS_REV_06 TRAILER_REV_06 FEE_REV_06
EXPERIENCE_YEARS CORE_HH_COUNT
SMALL_HH_COUNT RETIREMENT_ACCT_COUNT
ACCTS_PER_HH MEDIAN_HH_ASSETS;
RUN;

16
 Product Terms

• Once independent variables are mean centered, product

terms for interaction effects can be created in a data
step.
• Unlike PROC GLM, interaction terms cannot be entered
directly into PROC REG.
• Example:
DATA data_2;
SET data_2;
ACCTS_PER_HH_MED_HH_ASSETS=
ACCTS_PER_HH*MEDIAN_HH_ASSETS;
RUN;

17
 Let’s Interact!
Specifying the Model with PROC REG
 Model Specification

PROC REG DATA=data_2 OUTEST=parmest;

MAINEFFECTS: MODEL TOTAL_REV_11 = TRANS_REV_06
TRAILER_REV_06 FEE_REV_06 EXPERIENCE_YEARS TEAM
CORE_HH_COUNT SMALL_HH_COUNT RETIREMENT_ACCT_COUNT
ACCTS_PER_HH MEDIAN_HH_ASSETS
/ADJRSQ CLB STB VIF;
INTERACTION: MODEL TOTAL_REV_11 = TRANS_REV_06
TRAILER_REV_06 FEE_REV_06 EXPERIENCE_YEARS TEAM
CORE_HH_COUNT SMALL_HH_COUNT RETIREMENT_ACCT_COUNT
ACCTS_PER_HH MEDIAN_HH_ASSETS
ACCTS_PER_HH_MED_HH_ASSETS
/ADJRSQ CLB STB VIF;
INT_EFFECT: TEST ACCTS_PER_HH_MED_HH_ASSETS=0;
RUN; QUIT;

19
 Results: Main Effects Model

The REG Procedure

Model: MAINEFFECTS
Dependent Variable: TOTAL_REV_11

Number of Observations Read 1010

Number of Observations Used 1010

Analysis of Variance

Sum of Mean
Source DF Squares Square F Value Pr > F

Model 10 190238759 19023876 431.86 <.0001

Error 999 44007296 44051
Corrected Total 1009 234246055

Root MSE 209.88413 R-Square 0.8121

Dependent Mean 760.17482 Adj R-Sq 0.8103
Coeff Var 27.60998

20
 Results: Main Effects Model
Parameter Estimates

Parameter Standard
Variable Estimate Error t Value Pr > |t|

Intercept 751.26991 6.85512 109.59 <.0001

TRANS_REV_06 0.82323 0.02916 28.24 <.0001
TRAILER_REV_06 0.77207 0.09913 7.79 <.0001
FEE_REV_06 0.97834 0.03272 29.90 <.0001
EXPERIENCE_YEARS -12.56149 1.62893 -7.71 <.0001
TEAM 142.76120 29.46273 4.85 <.0001
CORE_HH_COUNT 1.32491 0.27042 4.90 <.0001
SMALL_HH_COUNT -0.15372 0.10117 -1.52 0.1290
RETIREMENT_ACCT_COUNT 0.47258 0.11506 4.11 <.0001
ACCTS_PER_HH 53.81291 22.22005 2.42 0.0156
MEDIAN_HH_ASSETS -0.11352 0.11302 -1.00 0.3154

21
 Results: Interactive Model
The REG Procedure
Model: INTERACTION
Dependent Variable: TOTAL_REV_11

Number of Observations Read 1010

Number of Observations Used 1010

Analysis of Variance

Sum of Mean
Source DF Squares Square F Value Pr > F

Model 11 190669800 17333618 396.98 <.0001

Error 998 43576255 43664
Corrected Total 1009 234246055

Root MSE 208.95833 R-Square 0.8140

Dependent Mean 760.17482 Adj R-Sq 0.8119
Coeff Var 27.48819

22
 Results: Interactive Model
Parameter Estimates

Parameter Standard
Variable Estimate Error t Value Pr > |t|

Intercept 739.79385 7.74079 95.57 <.0001

TRANS_REV_06 0.80831 0.02941 27.48 <.0001
TRAILER_REV_06 0.76138 0.09875 7.71 <.0001
FEE_REV_06 0.96004 0.03309 29.01 <.0001
EXPERIENCE_YEARS -12.66608 1.62208 -7.81 <.0001
TEAM 140.36661 29.34267 4.78 <.0001
CORE_HH_COUNT 1.64718 0.28810 5.72 <.0001
SMALL_HH_COUNT -0.26999 0.10731 -2.52 0.0120
RETIREMENT_ACCT_COUNT 0.50773 0.11510 4.41 <.0001
ACCTS_PER_HH 46.24691 22.25271 2.08 0.0379
MEDIAN_HH_ASSETS -0.30548 0.12804 -2.39 0.0172
ACCTS_PER_HH_MED_HH_ASSETS 0.56231 0.17897 3.14 0.0017

23
 Results: Interactive Model
Test INT_EFFECT Results for Dependent

Variable TOTAL_REV_11

Mean
Source DF Square F Value Pr > F
Numerator 1 431040 9.87 0.0017
Denominator 998 43664

24
 Results: Summary Dataset

Partial output from the parmest dataset:

ACCTS_PER_
ACCTS_ MEDIAN_ HH_MED_HH_
_MODEL_ _TYPE_ _DEPVAR_ Intercept PER_HH HH_ASSETS ASSETS _RSQ_
MAINEFFECTS PARMS TOTAL_REV_11 751.270 53.813 -0.114 . 0.812
INTERACTION PARMS TOTAL_REV_11 739.794 46.247 -0.305 0.562 0.814

25
A Plot is Worth a Thousand Words
(or Coefficients)
 Graphical Depictions of Interaction Effects

• Two strategies:
- Effect plots (effect displays) depict the strength and
direction of the relationship between the focal
independent variable and dependent variable at
different levels of the moderator variable.
- Coefficient plots display the coefficient (and
confidence interval) for the focal independent variable
with the scores for the moderator variable centered at
different values. This serves to highlight the regions of
significance of the focal independent variable.

27
Effect Plot

28
 Effect Plot

1. Output the means of the variables involved in the

interaction and create macro variables (PROC
UNIVARIATE and CALL SYMPUT).
2. Use nested DO loops in a DATA STEP to generate the
desired values of the variables involved in the
interaction and multiply these values out by the model
coefficients (using the dataset created by the PROC
REG outest option).
3. In a DATA step, restructure the resulting dataset (one
row for each value of the focal IV; multiple columns for
different values of the moderator).
4. Plot the predicted values of the DV (PROC SGPLOT).

29
 Effect Plot

ODS OUTPUT BasicMeasures=means;

PROC UNIVARIATE DATA=data_1;
VAR ACCTS_PER_HH MEDIAN_HH_ASSETS;
RUN;

DATA _NULL_;
SET means;
IF VarName="ACCTS_PER_HH" AND LocMeasure="Mean"
THEN CALL SYMPUT('AVG_ACCTS_PER_HH', LocValue);
IF VarName="MEDIAN_HH_ASSETS" AND LocMeasure="Mean"
THEN CALL SYMPUT('AVG_MEDIAN_HH_ASSETS', LocValue);
RUN;

30
 Effect Plot
DATA plot_1 (DROP=i j _MODEL_);
SET parmest (WHERE=(_MODEL_="INTERACTION") KEEP=_MODEL_ Intercept
MEDIAN_HH_ASSETS ACCTS_PER_HH ACCTS_PER_HH_MED_HH_ASSETS
RENAME=(MEDIAN_HH_ASSETS=b_MED_ASSETS ACCTS_PER_HH=b_ACCTS
ACCTS_PER_HH_MED_HH_ASSETS=b_MED_ASSETS_ACCTS));
DO i=100 TO 600;
DO j=1.5 TO 4 BY 0.5;
MEDIAN_HH_ASSETS=i;
MEDIAN_HH_ASSETS_CTR=i - INPUT(&AVG_MEDIAN_HH_ASSETS, BEST12.);
ACCTS_PER_HH=j;
ACCTS_PER_HH_CTR=j - INPUT(&AVG_ACCTS_PER_HH, BEST12.);
PRED=Intercept + /* Intercept */
(b_MED_ASSETS * MEDIAN_HH_ASSETS_CTR) + /* Median HH Assets */
(b_ACCTS * ACCTS_PER_HH_CTR) + /* Accounts per household */
(b_MED_ASSETS_ACCTS * (MEDIAN_HH_ASSETS_CTR * ACCTS_PER_HH_CTR))
/* Interaction */
;
OUTPUT;
END;
END;
RUN;

31
 Effect Plot

DATA plot_2;
MERGE plot_1 (WHERE=(ACCTS_PER_HH=1.5) RENAME=(PRED=PRED_1_5))
plot_1 (WHERE=(ACCTS_PER_HH=2.0) RENAME=(PRED=PRED_2_0))
plot_1 (WHERE=(ACCTS_PER_HH=2.5) RENAME=(PRED=PRED_2_5))
plot_1 (WHERE=(ACCTS_PER_HH=3.0) RENAME=(PRED=PRED_3_0))
plot_1 (WHERE=(ACCTS_PER_HH=3.5) RENAME=(PRED=PRED_3_5))
plot_1 (WHERE=(ACCTS_PER_HH=4.0) RENAME=(PRED=PRED_4_0));
BY MEDIAN_HH_ASSETS;
RUN;

32
 Effect Plot

ODS GRAPHICS ON /BORDER=OFF HEIGHT=2.5IN WIDTH=4IN;

ODS LISTING IMAGE_DPI=600 STYLE=JOURNAL SGE=OFF;

PROC SGPLOT DATA=plot_2;

TITLE "Interactive Effect of Median HH Size and Accounts per HH";
SERIES Y=PRED_1_5 X=MEDIAN_HH_ASSETS
/LINEATTRS=(THICKNESS=1 PATTERN=SOLID COLOR=CXE1E1E1)
LEGENDLABEL="1.5 Accts per HH";
SERIES Y=PRED_2_0 X=MEDIAN_HH_ASSETS
/LINEATTRS=(THICKNESS=1 PATTERN=SOLID COLOR=CXB4B4B4)
LEGENDLABEL="2.0 Accts per HH";
SERIES Y=PRED_2_5 X=MEDIAN_HH_ASSETS
/LINEATTRS=(THICKNESS=1 PATTERN=SOLID COLOR=CX878787)
LEGENDLABEL="2.5 Accts per HH";
SERIES Y=PRED_3_0 X=MEDIAN_HH_ASSETS
/LINEATTRS=(THICKNESS=1 PATTERN=SOLID COLOR=CX5A5A5A)
LEGENDLABEL="3.0 Accts per HH";

33
 Effect Plot

SERIES Y=PRED_3_5 X=MEDIAN_HH_ASSETS

/LINEATTRS=(THICKNESS=1 PATTERN=SOLID COLOR=CX2D2D2D)
LEGENDLABEL="3.5 Accts per HH";
SERIES Y=PRED_4_0 X=MEDIAN_HH_ASSETS
/LINEATTRS=(THICKNESS=1 PATTERN=SOLID COLOR=CX000000)
LEGENDLABEL="4.0 Accts per HH";
KEYLEGEND /POSITION=RIGHT
LOCATION=OUTSIDE ACROSS=1 DOWN=6 NOBORDER;
YAXIS MIN=400 MAX=1200 VALUES=(400 600 800 1000 1200)
OFFSETMIN=0.02
LABEL="Predicted Revenue (000s), 2011";
XAXIS MIN=100 MAX=600 VALUES=(100 200 300 400 500 600)
OFFSETMIN=0.02
LABEL="Median HH Size (000s)";
RUN;

ODS GRAPHICS OFF;

34
Effect Plot

35
Coefficient Plot

36
 Coefficient Plot

1. Mean-center all of the continuous independent

variables, except for the variables involved in the
interaction (PROC STDIZE)
2. In a DATA step, create an empty dataset (zero
observations) that will contain the model coefficients,
confidence intervals and the centering value of the
moderator.

37
 Coefficient Plot

3. Run the %INTPROBE macro, iteratively performing the

following steps using a macro %DO loop:
1. In a DATA step, increment the value of the
moderator, calculate the new (centred) moderator
and the interaction product term.
2. Run the regression model (PROC REG) and output
the coefficients using the ODS OUTPUT statement.
3. In another DATA step, add the model coefficients to
the coefficients dataset.
4. Delete the datasets created within the iteration of the
macro.

38
 Coefficient Plot

PROC STDIZE DATA=data_1 OUT=data_3 METHOD=MEAN;

VAR TRANS_REV_06 TRAILER_REV_06 FEE_REV_06
EXPERIENCE_YEARS CORE_HH_COUNT SMALL_HH_COUNT
RETIREMENT_ACCT_COUNT;
RUN;
DATA parmsint;

LENGTH Variable $50 ACCTS_PER_HH_CENTER 3.

Estimate LCL UCL 8.;
FORMAT p PVALUE6.4;
RUN;

39
 Coefficient Plot

%MACRO INTPROBE (DataIn=, DataOut= );

%DO ACCT_HH=10 %TO 60;
DATA &DataOut.;
SET &DataIn.;
CENTRE_VALUE=ROUND((&ACCT_HH*0.1), 0.1);
ACCTS_PER_HH=ACCTS_PER_HH - CENTRE_VALUE ;
ACCTS_PER_HH_MED_HH_ASSETS=ACCTS_PER_HH*MEDIAN_HH_ASSETS;
RUN;
PROC REG DATA=&DataOut.;
MODEL TOTAL_REV_11 = TRANS_REV_06 TRAILER_REV_06 FEE_REV_06
EXPERIENCE_YEARS TEAM CORE_HH_COUNT SMALL_HH_COUNT
RETIREMENT_ACCT_COUNT ACCTS_PER_HH MEDIAN_HH_ASSETS
ACCTS_PER_HH_MED_HH_ASSETS
/CLB STB;
ODS OUTPUT ParameterEstimates=parms;
RUN; QUIT;

40
 Coefficient Plot

DATA parmsctr (KEEP=Variable Estimate StandardizedEst

LowerCL UpperCL Probt ACCTS_PER_HH_CENTER
RENAME=(StandardizedEst=StdCoeff LowerCL=LCL UpperCL=UCL
Probt=p));
LENGTH Variable $50 ACCTS_PER_HH_CENTER 3.;
SET parms;
ACCTS_PER_HH_CENTER=ROUND((&ACCT_HH*0.1), 0.1);
FORMAT ACCTS_PER_HH_CENTER 3.1;
RUN;
DATA parmsint;
SET parmsint parmsctr;
IF Variable="" THEN DELETE;
FORMAT ACCTS_PER_HH_CENTER 3.1;
RUN;
PROC DATASETS LIB=work NOLIST;
DELETE &DataOut. parms parmsctr;
RUN; QUIT;
%END;
%MEND INTPROBE;

41
 Coefficient Plot

%INTPROBE(DataIn=data_3, DataOut=data_4);

PROC SORT DATA=parmsint;

BY Variable;
RUN;

42
 Coefficient Plot
Lower 95% Upper 95%
ACCTS_PER_ Parameter CL CL
Variable HH_CENTER Estimate Parameter Parameter Pr > |t|

MEDIAN_HH_ASSETS 1.0 -0.915 -1.462 -0.368 0.0011

MEDIAN_HH_ASSETS 1.1 -0.859 -1.374 -0.344 0.0011
MEDIAN_HH_ASSETS 1.2 -0.803 -1.286 -0.319 0.0012
. . .
MEDIAN_HH_ASSETS 1.9 -0.409 -0.697 -0.121 0.0054
MEDIAN_HH_ASSETS 2.0 -0.353 -0.619 -0.086 0.0096
MEDIAN_HH_ASSETS 2.1 -0.296 -0.545 -0.048 0.0195
MEDIAN_HH_ASSETS 2.2 -0.240 -0.475 -0.006 0.0447
MEDIAN_HH_ASSETS 2.3 -0.184 -0.409 0.041 0.1091
MEDIAN_HH_ASSETS 2.4 -0.128 -0.349 0.093 0.2569
MEDIAN_HH_ASSETS 2.5 -0.072 -0.294 0.151 0.5281
. . .
MEDIAN_HH_ASSETS 3.0 0.210 -0.090 0.509 0.1694
MEDIAN_HH_ASSETS 3.1 0.266 -0.058 0.590 0.1075
MEDIAN_HH_ASSETS 3.2 0.322 -0.028 0.673 0.0716
MEDIAN_HH_ASSETS 3.3 0.378 0.000 0.757 0.0500
MEDIAN_HH_ASSETS 3.4 0.435 0.027 0.842 0.0366
MEDIAN_HH_ASSETS 3.5 0.491 0.054 0.928 0.0279
MEDIAN_HH_ASSETS 3.6 0.547 0.079 1.015 0.0220
. . .

43
 Coefficient Plot

ODS GRAPHICS ON /BORDER=OFF HEIGHT=2.5IN WIDTH=4IN;

ODS LISTING IMAGE_DPI=600 STYLE=JOURNAL SGE=OFF GPATH="C:\NESUG 2012";

PROC SGPLOT DATA=parmsint (WHERE=(Variable="MEDIAN_HH_ASSETS"

AND MOD(ACCTS_PER_HH_CENTER, 0.5)=0));
TITLE "Effect of Median HH Assets on Total Revenue, 2011";
SCATTER X=ACCTS_PER_HH_CENTER Y=Estimate
/YERRORLOWER=LCL YERRORUPPER=UCL MARKERATTRS=(SYMBOL=CIRCLE)
ERRORBARATTRS=(PATTERN=4);
XAXIS TYPE=DISCRETE OFFSETMIN=0.02 LABEL="Accounts per HH";
YAXIS MIN=-2 MAX=4 VALUES=(-2 TO 4 BY 1)
LABEL="b (Median HH Assets)";
REFLINE 0 /AXIS=Y TRANSPARENCY=0.5;
RUN;

ODS GRAPHICS OFF;

44
Coefficient Plot

45
 Case #2
Canadian Attitudes Toward
Canada–US Relations
 Research Questions

• What does the Canadian public think about Canada–U.S.

relations?
- What is the role of political variables (party identification,
ideology) in shaping such attitudes?
- What is the role of proximity to the US?
- How do political variables and proximity interact?

47
Data: Canadian Election Studies (1997–2011)

“Do you think Canada’s ties with the United States should be much
closer, somewhat closer, about the same as now, somewhat more
distant, or much more distant?”

60 55
52 54

50 46
42
38
40 Much/Somewhat Closer
34
30 38
% 30 26 27 25 About the Same as Now

20 24
21 21 Much/Somewhat More Distant
19
17 16
10

0
1997 2000 2004 2006 2008 2011

48
 Let’s Interact Some More!
Specifying the Model with PROC LOGISTIC
 Model Specification
PROC LOGISTIC DATA=data_7;
MODEL CANADA_TIES_US=
POST_PARTY_CONS POST_PARTY_NDP POST_PARTY_BQ
POST_PARTY_OTHER POST_NO_PARTY LEFT_RIGHT
LN_DISTANCE_USA
/LINK=CLOGIT RSQUARE;
WEIGHT WEIGHT;
RUN;

* Control variables included in

the models but not shown.
50
 Model Specification
PROC LOGISTIC DATA=data_7 OUTEST=parmest;
MODEL CANADA_TIES_US=
POST_PARTY_CONS POST_PARTY_NDP POST_PARTY_BQ
POST_PARTY_OTHER POST_NO_PARTY LEFT_RIGHT
LN_DISTANCE_USA
LN_DIST_USA_CONS LN_DIST_USA_NDP LN_DIST_USA_BQ
LN_DIST_USA_OTH_PTY LN_DIST_USA_NO_PTY
LN_DIST_USA_L_R
/LINK=CLOGIT RSQUARE;
INT_EFFECT1: TEST LN_DIST_USA_CONS=LN_DIST_USA_NDP=
LN_DIST_USA_BQ=LN_DIST_USA_OTH_PTY=
LN_DIST_USA_NO_PTY=0;
INT_EFFECT2: TEST LN_DIST_USA_L_R=0;
WEIGHT WEIGHT;
RUN;
* Control variables included in
the models but not shown.
51
 Model Specification
PROC LOGISTIC DATA=data_7 OUTEST=parmest;
MODEL CANADA_TIES_US=
POST_PARTY_CONS POST_PARTY_NDP POST_PARTY_BQ
POST_PARTY_OTHER POST_NO_PARTY LEFT_RIGHT
LN_DISTANCE_USA
LN_DIST_USA_CONS LN_DIST_USA_NDP LN_DIST_USA_BQ
LN_DIST_USA_OTH_PTY LN_DIST_USA_NO_PTY
LN_DIST_USA_L_R
/LINK=CLOGIT RSQUARE;
INT_EFFECT1: TEST LN_DIST_USA_CONS=0,
LN_DIST_USA_NDP=0, LN_DIST_USA_BQ=0,
LN_DIST_USA_OTH_PTY=0, LN_DIST_USA_NO_PTY=0;
INT_EFFECT2: TEST LN_DIST_USA_L_R=0;
WEIGHT WEIGHT;
RUN;
* Control variables included in
the models but not shown.
52
 Results: Main Effects Model
The LOGISTIC Procedure

Model Fit Statistics

Intercept
Intercept and
Criterion Only Covariates

AIC 39777.017 38608.978

SC 39800.129 38878.618
-2 Log L 39771.017 38538.978

R-Square 0.0724 Max-rescaled R-Square 0.0795

Testing Global Null Hypothesis: BETA=0

Test Chi-Square DF Pr > ChiSq

Likelihood Ratio 1232.0390 32 <.0001

Score 1168.6213 32 <.0001
Wald 1205.9667 32 <.0001

53
 Results: Main Effects Model
Analysis of Maximum Likelihood Estimates

Standard Wald
Parameter Estimate Error Chi-Square Pr > ChiSq

Intercept 1 -2.2800 0.0641 1264.3924 <.0001

Intercept 2 -0.9980 0.0611 266.7331 <.0001
Intercept 3 1.4221 0.0618 529.7684 <.0001
POST_PARTY_CONS 0.4089 0.0412 98.4413 <.0001
POST_PARTY_NDP -0.5107 0.0532 92.2785 <.0001
POST_PARTY_BQ -0.4228 0.0659 41.1882 <.0001
POST_PARTY_OTHER -0.5187 0.1050 24.4228 <.0001
POST_NO_PARTY 0.0394 0.0431 0.8325 0.3616
LEFT_RIGHT 0.0594 0.00816 53.0024 <.0001
LN_DISTANCE_USA -0.0253 0.0201 1.5963 0.2064

54
 Results: Interactive Model
The LOGISTIC Procedure

Model Fit Statistics

Intercept
Intercept and
Criterion Only Covariates

AIC 39777.017 38590.735

SC 39800.129 38906.599
-2 Log L 39771.017 38508.735

R-Square 0.0742 Max-rescaled R-Square 0.0813

Testing Global Null Hypothesis: BETA=0

Test Chi-Square DF Pr > ChiSq

Likelihood Ratio 1262.2822 38 <.0001

Score 1195.3992 38 <.0001
Wald 1232.1770 38 <.0001

55
 Results: Interactive Model
Analysis of Maximum Likelihood Estimates

Standard Wald
Parameter Estimate Error Chi-Square Pr > ChiSq

Intercept 1 -2.2824 0.0642 1263.5064 <.0001

Intercept 2 -0.9999 0.0612 266.9235 <.0001
Intercept 3 1.4238 0.0619 529.4537 <.0001
POST_PARTY_CONS 0.4140 0.0413 100.5214 <.0001
POST_PARTY_NDP -0.5161 0.0532 93.9913 <.0001
POST_PARTY_BQ -0.4353 0.0666 42.7814 <.0001
POST_PARTY_OTHER -0.5223 0.1052 24.6627 <.0001
POST_NO_PARTY 0.0365 0.0432 0.7123 0.3987
LEFT_RIGHT 0.0589 0.00817 51.9388 <.0001
LN_DISTANCE_USA 0.0101 0.0311 0.1055 0.7453
LN_DIST_USA_CONS -0.1031 0.0394 6.8508 0.0089
LN_DIST_USA_NDP 0.1041 0.0518 4.0362 0.0445
LN_DIST_USA_BQ -0.0698 0.0803 0.7541 0.3852
LN_DIST_USA_OTH_PTY 0.2790 0.1125 6.1510 0.0131
LN_DIST_USA_NO_PTY -0.0674 0.0408 2.7301 0.0985
LN_DIST_USA_L_R -0.00410 0.00797 0.2647 0.6069

56
 Results: Interactive Model
Linear Hypotheses Testing Results

Wald
Label Chi-Square DF Pr > ChiSq

INT_EFFECT1 26.8584 5 <.0001

INT_EFFECT2 0.2647 1 0.6069

57
 Results: Summary Dataset

Partial output from the parmest dataset:

Intercept_ Intercept_ Intercept_

_LINK_ _TYPE_ _NAME_ 1 2 3

LOGIT PARMS CANADA_TIES_US -2.28241 -0.99994 1.42381

POST_ POST_
PARTY_ PARTY_ POST_ LN_DISTANCE_
CONS NDP PARTY_BQ USA

0.41404 -0.51606 -0.43530 0.010105

LN_DIST_ LN_DIST_ LN_DIST_

USA_CONS USA_NDP USA_BQ

-0.10310 0.10415 -0.069773

58
Effect Plots

59
 Effect Plot
DATA plot_1 (KEEP=DISTANCE_CAN_US_BORDER LN_DISTANCE_USA
LN_DISTANCE_USA_CTR CP_: );
SET parmest (KEEP=Intercept_: POST_PARTY_: LN_DIST:
RENAME=(POST_PARTY_CONS=b_CONS POST_PARTY_NDP=b_NDP
POST_PARTY_BQ=b_BQ LN_DISTANCE_USA=b_LN_DISTANCE_USA
LN_DIST_USA_CONS=b_LN_DIST_USA_CONS
LN_DIST_USA_NDP=b_LN_DIST_USA_NDP
LN_DIST_USA_BQ=b_LN_DIST_USA_BQ));
DO i=0.1, 0.5, 1 TO 2500;
DISTANCE_CAN_US_BORDER=i;
LN_DISTANCE_USA=(LOG(DISTANCE_CAN_US_BORDER));
LN_DISTANCE_USA_CTR=(LOG(DISTANCE_CAN_US_BORDER))
-4.6644800622;

60
 Effect Plot
REG_EQN_1_LIB=Intercept_1 +
(b_LN_DISTANCE_USA*LN_DISTANCE_USA_CTR);
CP_1_LIB=CDF('LOGISTIC',REG_EQN_1_LIB);
REG_EQN_1_CONS=Intercept_1 + (b_CONS*1) +
(b_LN_DISTANCE_USA*LN_DISTANCE_USA_CTR) +
(b_LN_DIST_USA_CONS*(1*LN_DISTANCE_USA_CTR));
CP_1_CONS=CDF('LOGISTIC',REG_EQN_1_CONS);
REG_EQN_1_NDP=Intercept_1 + (b_NDP*1) +
(b_LN_DISTANCE_USA*LN_DISTANCE_USA_CTR) +
(b_LN_DIST_USA_NDP*(1*LN_DISTANCE_USA_CTR));
CP_1_NDP=CDF('LOGISTIC',REG_EQN_1_NDP);
REG_EQN_1_BQ=Intercept_1 + (b_BQ*1) +
(b_LN_DISTANCE_USA*LN_DISTANCE_USA_CTR) +
(b_LN_DIST_USA_BQ*(1*LN_DISTANCE_USA_CTR));
CP_1_BQ=CDF('LOGISTIC',REG_EQN_1_BQ);

61
 Effect Plot
REG_EQN_2_LIB=Intercept_2 +
(b_LN_DISTANCE_USA*LN_DISTANCE_USA_CTR);
CP_2_LIB=CDF('LOGISTIC',REG_EQN_2_LIB);
REG_EQN_2_CONS=Intercept_2 + (b_CONS*1) +
(b_LN_DISTANCE_USA*LN_DISTANCE_USA_CTR) +
(b_LN_DIST_USA_CONS*(1*LN_DISTANCE_USA_CTR));
CP_2_CONS=CDF('LOGISTIC',REG_EQN_2_CONS);
REG_EQN_2_NDP=Intercept_2 + (b_NDP*1) +
(b_LN_DISTANCE_USA*LN_DISTANCE_USA_CTR) +
(b_LN_DIST_USA_NDP*(1*LN_DISTANCE_USA_CTR));
CP_2_NDP=CDF('LOGISTIC',REG_EQN_2_NDP);
REG_EQN_2_BQ=Intercept_2 + (b_BQ*1) +
(b_LN_DISTANCE_USA*LN_DISTANCE_USA_CTR) +
(b_LN_DIST_USA_BQ*(1*LN_DISTANCE_USA_CTR));
CP_2_BQ=CDF('LOGISTIC',REG_EQN_2_BQ);

62
 Effect Plot
REG_EQN_3_LIB=Intercept_3 +
(b_LN_DISTANCE_USA*LN_DISTANCE_USA_CTR);
CP_3_LIB=CDF('LOGISTIC',REG_EQN_3_LIB);
REG_EQN_3_CONS=Intercept_3 + (b_CONS*1) +
(b_LN_DISTANCE_USA*LN_DISTANCE_USA_CTR) +
(b_LN_DIST_USA_CONS*(1*LN_DISTANCE_USA_CTR));
CP_3_CONS=CDF('LOGISTIC',REG_EQN_3_CONS);
REG_EQN_3_NDP=Intercept_3 + (b_NDP*1) +
(b_LN_DISTANCE_USA*LN_DISTANCE_USA_CTR) +
(b_LN_DIST_USA_NDP*(1*LN_DISTANCE_USA_CTR));
CP_3_NDP=CDF('LOGISTIC',REG_EQN_3_NDP);
REG_EQN_3_BQ=Intercept_3 + (b_BQ*1) +
(b_LN_DISTANCE_USA*LN_DISTANCE_USA_CTR) +
(b_LN_DIST_USA_BQ*(1*LN_DISTANCE_USA_CTR));
CP_3_BQ=CDF('LOGISTIC',REG_EQN_3_BQ);
OUTPUT;
END;
RUN;

63
Effect Plots

64
Coefficient Plots

65
 Recap

• Models with interaction effects require a little more care

in their theorizing, specification, testing and interpretation
than strictly main effects models.
• Data preparation is critical. Mean-centering is well-
advised.
• Model coefficients rarely tell the entire story.
• To fully understand an interaction effect, plot it.

66
 References

Aiken, L.S. and S.G. West (1991) Multiple Regression: Testing and Interpreting
Interactions. Thousand Oaks, CA: Sage.
Allison, P.D. (1977) “Testing for Interaction in Multiple Regression.” American
Journal of Sociology 83(1): 144–153.
Baron, R.M. and D.A. Kenny (1986) “The Moderator-Mediator Variable
Distinction in Social Psychological Research.” Journal of Personality and
Social Psychology 51(6): 1173–1182.
Braumoeller, B.F. (2004) “Hypothesis Testing and Multiplicative Interaction
Terms.” International Organization 58(4): 807–820.
Edwards, J.R. (2008) “Seven Deadly Myths of Testing Moderation in
Organizational Research.” Statistical and Methodological Myths and Urban
Legends. Eds. C.E. Lance and R.J. Vandenberg New York: Routledge.
Fox, J. (2008) Applied Regression Analysis and Generalized Linear Models. 2nd
ed. Thousand Oaks, CA: Sage.

67
 References

Friedrich, R.J. (1982) “In Defense of Multiplicative Terms in Multiple Regression

Equations.” American Journal of Political Science 26(4): 797–833.
Hayes, A.F., C.J. Glynn and M.E. Huge (2012) “Cautions Regarding the
Interpretation of Regression Coefficients and Hypothesis Tests in Linear
Models with Interactions.” Communication Methods and Measures 6(1): 1–11.
Jaccard, J. and T. Dodge (2004) “Analyzing Contingent Effects in Regression
Models.” Handbook of Data Analysis. Thousand Oaks, CA: Sage
Jaccard, J. and R. Turrisi (2003) Interaction Effects in Multiple Regression. 2nd
ed. Thousand Oaks, CA: Sage.
Robinson, C. and R.E. Schumacker (2009) “Interaction Effects: Centering,
Variance Inflation Factor, and Interpretation Issues.” Multiple Linear
Regression Viewpoints 35(1): 6–11.
http://mlrv.ua.edu/2009/vol35_1/Robinson_Schumacke_rproof.pdf
Rogosa, D. (1980) “Comparing Nonparallel Regression Lines.” Psychological
Bulletin 88(2): 307–321.

68
 Thank you!

Timothy B. Gravelle
Principal Scientist & Director, Insights Lab
[email protected]