Download as pdf or txt
Download as pdf or txt
You are on page 1of 45

REGRESSION

WITH
INTERACTION
(MODERATION)
IN SPSS

September 2020
Ineke Nagel
STEP 1: CAUSAL MODEL OF INTERACTION (MODERATION)

Causal model 1: effect X -> Y

X Y

Causal model 2 : does the effect X -> Y vary according to M (moderator)?

X Y

M
STEP 1: EXAMPLE
Causal model 1: effect X -> Y

CULTURAL PARTICIPATION CULTURAL PARTICIPATION


MOTHER CHILD

Causal model 1 : does the effect X -> Y vary according to M (moderator)?

CULTURAL PARTICIPATION CULTURAL PARTICIPATION


MOTHER CHILD

GENDER
STEP 2: PREPARATION 1: NULCATEGORY X-VARIABLES

• For interpretation purposes it is important to have a well-defined 0-category that is in the data.

• For gender (GIRL) 0 = boy

4
STEP 2: PREPARATION 1: NULCATEGORY X-VARIABLES
MCUL mother s cultural participation (scale 0-1)
Cumulative
Frequency Percent Valid Percent Percent
Valid ,00 1 ,0 ,0 ,0
,00 1 ,0 ,0 ,1
,10 539 17,5 17,5 17,5
,20 3 ,1 ,1 17,6
,21 3 ,1 ,1 17,7
,21 1 ,0 ,0 17,8
,21 1 ,0 ,0 17,8
,21 23 ,7 ,7 18,5
,22 1 ,0 ,0 18,6
,22 17 ,6 ,6 19,1

1,00 1 ,0 ,0 99,9
1,00 1 ,0 ,0 99,9
1,00 1 ,0 ,0 99,9
1,00 1 ,0 ,0 100,0
1,00 1 ,0 ,0 100,0
Total 3084 100,0 100,0

• For mother’s cultural participation (MCUL) 0 = lowest value on the continuous scale (0-1) = the least

cultural participating mother

5
STEP 2: PREPARATION 1: NULCATEGORY X-VARIABLES

If there is no 0-category, you have to transform the variables in order to create one:

• By recoding the variable (variables that have a restricted number of values)

• By standardizing the variable (continuous variables) = creating z-scores that have on


average 0; the units are then standarddeviations, zmcul 0 = average mcul

6
STEP 2: PREPARATION 2: COMPUTE THE INTERACTION-VARIABLE

Calculate a new variable as the product of the two variables: transform -> compute.
In the syntax:
Compute MCUL_GENDER = MCUL * GENDER.
EXECUTE. 7
STAP 3: REGRESSION ANALYSIS

• In 2 steps, via ‘NEXT’:


• Model 1 (no moderation): effects of the X variables MCUL and GIRL
• Model 2 (or moderation): adding the interaction term
RESULTS MODEL 1: NO MODERATION MODELLED

How does the effect of mothers' cultural participation vary between boys and girls?

CULPAR = 25.272 + 5.516 GIRL + 46.139 MCUL


There is no moderation :
The effect of MCUL
For boys (fill in GIRL = 0):
does not vary between
CULPAR = 25.272 + (5.516 * 0) + 46.139 MCUL
boys and girls
= 25.272 + 46.139 MCUL
For girls (fill in GIRL =1):
CULPAR = 25.272 + (5.516 * 1) + 46.139 MCUL
= 25.272 + 5.516 + 46.139 MCUL
= 30.778 + 46.139 MCUL
RESULTS MODEL 1: NO MODERATION MODELLED

• Equal slopes: effects of MCUL are equal for boys and girls

• Different intercepts: girls participate more in culture than boys - regardless of their
mother's cultural participation.
https://psychology.victoria.ac.nz/modgraph/modgraph.php 10
There is moderation :
RESULTS MODEL 2: MODERATION The effect of MCUL
varies significantly
between
boys and girls

How does the effect of mothers' cultural participation vary between boys and girls?
The complete model is:
CULPAR = 27.903 + 41.009 MCUL + .216 GIRL + 10.218 MCUL_GIRL
= 27.903 + 41.009 MCUL + .216 GIRL + 10.218 MCUL * GIRL
For boys (fill in GIRL = 0):
CULPAR = 27.903 + 41.009 MCUL + (.216 * 0) + 10.218 (MCUL * 0)
= 27.903 + 41.009 MCUL
For girls (fill in GIRL = 1):
CULPAR = 27.903 + 41.009 MCUL + (.216 * 1) + 10.218 (MCUL * 1)
= 27.903 + 41.009 MCUL + .216 + 10.218 MCUL
= 28.119 + 51.227 MCUL
RESULTS MODEL 2: MODERATION

• Unequal slopes: effect of MCUL differs between boys and girls

• The effect of MCUL is stronger for girls; or, the gender differences increases with MCUL.

• For MCUL = Low, boys and girls are quite similar, but that changes when MCUL is bigger.
12
RESULTS MODEL 2: MODERATION

For boys : CULPAR = 27.903 + 41.009 MCUL


For girls : CULPAR = 28.119 + 51.227 MCUL

Interpretation in the regression table:


The main effect of MCUL = the effect of MCUL for boys (GIRL = 0)
The main effect of MCUL + 1* interaction = the effect of MCUL for girls
The main effect of GIRL = the effect of GIRL when MCUL = 0
The main effect of GIRL + 1* interaction = how the effect of GIRL varies when MCUL changes by 1 unit

13
QUANTITATIVE
RESEARCH
METHODOLOGY
Lecture 4:
Regression analysis – with moderation

Ineke Nagel

27 september 2021
EXAMPLE: READING BEHAVIOR AMONG YOUNG ADULTS

• It is known that education is rather strongly related to


reading books: higher educated on average read more
book than lower educated.

• We estimated the effect of education, one effect of


education on the sample, holding for everyone.

• NOW: is the effect the same for everyone, or does the


effect differ between groups of respondents?
2
EXAMPLE: READING BEHAVIOR AMONG YOUNG ADULTS

X Y
Education Reading
(secondary school frequency as a
level) young adult

M
gender

• Does the effect of education hold for boys and girls equally?

• Or: does the effect of education interact with gender, doe sgender
3
moderate the effect of education?
MODERATION BY GENDER: DIFFERENT EFFECTS FOR BOYS AND GIRLS?

Boys

Girls

4
FREQUENCIES Y-VARIABLE:
READING BEHAVIOR AMONG YOUNG ADULTS

read Fractional Rank Percent of xread


Valid Cumulative
Frequency Percent Percent Percent
Valid 7,09 71 14.7 14.7 14.7
14,25 2 0.4 0.4 15.1
15,30 10 2.1 2.1 17.2
16,26 1 0.2 0.2 17.4
16,59 1 0.2 0.2 17.6
17,39 6 1.2 1.2 18.9
18,36 2 0.4 0.4 19.3

88,41 2 0.4 0.4 88.2


89,13 3 0.6 0.6 88.8
91,38 18 3.7 3.7 92.5
93,24 1 0.2 0.2 92.7
96,70 35 7.3 7.3 100.0
Total 482 100.0 100.0

• READING: reading frequency scale (1-100) 5


PREPARATION 1 A: CHECK X-VARIABLES ON 0 -CATEGORY

6
PREPARATION 1 B: RECODING

7
PREPARATION 2: COMPUTE INTERACTION VARIABLE

8
REGRESSION: FIRST MODEL WITH NO MODERATION

9
REGRESSION: NEXT MODEL WITH MODERATION

10
REGRESSION ANALYSIS: NO MODERATION YET

• We are interested in the effect SEDUC -> READ

• Let’s find out what the effects of SEDUC are for boys and girls.
REGRESSION ANALYSIS: NO MODERATION YET

• READ = 29.680 + 6.389 SEDUC + 14.735 FEMALE

• For boys: fill in 0 for the variable FEMALE

READ = 29.680 + 6.389 SEDUC + 14.735 * 0 = 29.680 + 6.389 SEDUC

• For girls: fill in 1 for the variable FEMALE

READ = 29.680 + 6.389 SEDUC + 14.735 * 1

= 29.680 + 6.389 SEDUC + 14.735 = 44.415 + 6.389 SEDUC


NO MODERATION: PARALLEL REGRESSION LINES

13
REGRESSION ANALYSIS: WITH MODERATION

• READ = 24.386 + 9.041 SEDUC + 26.817 FEMALE + -5.933 SEDUC_FEMALE

• READ = 24.386 + 9.041 SEDUC + 26.817 FEMALE + -5.933 SEDUC * FEMALE

• For boys: fill in 0 for the variable FEMALE

READ = 24.386 + 9.041 SEDUC + 26.817 * 0 + -5.933 SEDUC * 0 = 24.386 + 9.041 SEDUC

• For girls: fill in 1 for the variable FEMALE

READ = 24.386 + 9.041 SEDUC + 26.817 * 1 + -5.933 SEDUC * 1

= 24.386 + 9.041 SEDUC + 26.817 -5.933 SEDUC

= 51.203 + 3.108 SEDUC


MODERATION BY GENDER: DIFFERENT EFFECTS FOR BOYS AND GIRLS

X Y
9.041
Education Reading
(secondary school frequency as a
level) young adult

X Y
Education
3.108 Reading
(secondary school frequency as a
level) young adult
15
MODERATION: NO PARALLEL REGRESSION LINES

16
REGRESSION ANALYSIS: WITH MODERATION

• Interpretation ‘main effects’ in a moderation / interaction model:


– The effect of SEDUC is the effect of SEDUC for FEMALE = 0

– The effect of FEMALE is the effect of FEMALE for SEDUC = 0

• Interpretation ‘interaction effect’ in a moderation / interaction model:


– The degree to which the effect of SEDUC varies when FEMALE increases by 1 unit

– The degree to which the effect of FEMALE varies when SEDUC increases by 1 unit
REGRESSION ANALYSIS: WITH MODERATION

• Of course we only can expect effects for the population if:


– The t-value of the interaction-effect > |1.96|, and p < .05

– And the confidence interval does not contain 0


REGRESSION AND CAUSATION

• We did not mention causation so far. We studied variations in the effect


of SEDUC, but we did not control for confounding variables.

• To do that we should control PARENTS’ EDUCATION, and also variations in


that effect with gender.

• However that is a bit complicated to do in this course.


EXAMPLE OF AN EXAM QUESTION

Based on these results, what statements can we make about the population?

1: For boys, one educational level higher leads to a higher reading frequency than
for girls.

2: The differences in reading frequency between boys and girls are greater at lower
education levels than at higher education levels.

A. Only statement 1

B. Only statement 2

C. Both statements

D. Neither statement
TO DO
WEEK 4 REGRESSION ANALYSIS WITH MODERATING VARIABLES

Scheduled meetings Learning materials What to do? Why?


and self-study
Monday 8:30- TEST VISION: MIDTERM ASSESSMENT WEEK 3 – Learning objectives:
10:45 • calculation
• interpretation analysis SPSS
• statements about population
• interpretation in terms of causal models
11:00-
13:00 Introduction lecture regression analysis with mediating variables
Tuesday Self- Field: Video’s: SPSS: Watch the videos Preparation
study Ch 11: 11.3.1- • SPSS – regression • Home and study the practical
11.3.5 (482-489) with interaction practical literature. Then do
the home-practice
assignments, which
test if you have
understood the
material.
Wednesday Self- Read this Preparation
study introduction on the practical
group assignment
13:30- Do the group Learning
17:30 assignment objective 1:
-interpretation
analysis SPSS
-statements
about population

21
Thursday Self- Study the literature Prepare working
study and watch the group
videos.
Download the
answer file of and
the SPSS output.
15:30- Working group Do the group Learning
17:00 Group assignment based on SPSS - output assignment objective 2:
-interpretation in
terms of causal
models
Friday Self- Practice questions in Test Vision Test if you have Preparation
study understood the midterm
material. assessment

22
Practicum IV – Wednesday, September 29 2021

Download the file ESSWAVE9NL alongside this assignment. The assignment is worth 10 points.

This dataset is from the ninth wave of the European Social Survey, a large scale survey that takes
place in several European countries. The subsample in this subsection of the dataset is wholly made
up by Dutch people without a migration background. In this assignment we look at a relationship
that is often studied in literature on migration studies, that between political orientation and
attitudes towards migrants. The literature shows that in general, the more right wing a respondent
is, the more negative their attitude toward migrants will be. We will also be investigating whether
education level (higher or lower education) moderates this relationship, i.e. whether the effect of
right wing (compared to left wing) political orientation differs between higher and lower educated
respondents.

In this assignment you estimate a moderation model in which you use the variable attitude toward
immigrants (imwbcnt) as a dependent variable (Y), right wing orientation (lrscale) as an independent
variable (X) and the variable denoting whether a respondent has higher education (hbo or university)
or lower education (all other) (highedu) as a moderator (M).

1 – Which one of the models visually represents this assumed relationship, where X is right wing
political orientation, Y is attitude toward migrants, and M is education level? 0.5 points

The correct answer is B.

2 – Check the categories of the variables. Is it necessary to recode or standardize the predictor and
moderator variables? Why/why not? 0.5 points

You can check using Frequencies and Descriptives. It is not necessary to recode or
standardize the variables because both the predictor and the moderator contain the value 0
in their answer categories, and this is a meaningful value. It is necessary for 0 to be a
meaningful value for all X-variables involved in a moderation.

1
Compute a new variable that denotes the interaction between right wing orientation and education
(highedu). Conduct a hierarchical regression analysis, where you include the two main predictors in
block 1, and the interaction term in block two.

3 – What does the constant in block 1 tell us? 1 point

The constant in block 1 tells us that if people score 0 on the left-right scale (therefore are
very left wing) and if they score 0 on highedu, therefore have lower education, the average
score on attitudes toward migrants is 6.256.

4 – Look at the coefficients in block 1. Which of the following statements about right wing
orientation is most applicable?. 5 points

A - We can accept H0 because p > .05 and the confidence interval does contains the value 0.

B – We can reject H0 because p > .05 and the confidence interval contains the value 0.

C – We can reject H0 because p < .05 and the confidence interval does not contain the value 0.

The correct answer is C.

Right wing orientation has an effect of -.152. P is smaller than .001, and therefore the
effect is significant. The confidence interval ranges between -.200 and -.104. Both values
are negative and 0 is not part of these values. Taken together with the significance level,
this indicates that we can reject H0. Right wing political orientation has a significant effect
on attitude toward migrants – with every one unit change from more left wing to more
right wing, attitudes toward migrants become more negative by -.152 units.

2
5 – Look at the coefficients in block 1. Which of the following statements about education level is
most applicable for the population?. 5 points

A – Higher educated people display more negative attitudes toward migrants than lower educated
people.

B – Higher educated people display more positive attitudes toward migrants that lower educated
people.

C – There is insufficient evidence that higher and lower educated people differ in their attitudes
towards migrants.

The correct answer is B.

Education level has an effect of .367. P is smaller than .001, and therefore significant. The
confidence interval ranges between .173 and .561. Both values are positive and 0 is not
part of these values. Taken together with the significance level, this indicates that we can
reject H0. Higher education has a significant effect on attitude toward migrants – higher
educated respondents score .367 higher on positive attitude toward migrants than lower
educated responders (on the 0-10 attitude scale).

6 - Write down and interpret the coefficients of both predictors in block 2 (the main effects). Include
significance and confidence intervals. Why do the values of the coefficients change? 2 points

3
Right wing orientation has an effect of -.113, it applies to those with highedu = 0. P is
smaller than .001, and therefore the effect is significant. The confidence interval ranges
between -.173 and -.053. Both values are negative and 0 is not part of these values. Taken
together with the significance level, this indicates that we can reject H0. We can expect that
right wing political orientation has an effect on attitude toward migrants in the population
for lower educated (=0): for lower educated holds that with every one unit change from
more left wing to more right wing, attitudes toward migrants become more negative by -
.113 units.

Education level has an effect of .901, it applies to those with lrscale = 0. P is smaller than
.001, and therefore significant. The confidence interval ranges between .362 and 1.440.
Both values are positive and 0 is not part of these values. Taken together with the
significance level, this indicates that we can reject H0. Higher education has a significant
effect on attitude toward migrants – higher educated respondents score .367 higher on
positive attitude toward migrants than lower educated responders - for those with a right
wing orientation of 0 (so very much left wing).

The coefficients change because of the inclusion of the interaction term. Without the
interaction, the coefficients represent the unique effect of the predictors on the dependent
variable while the other predictor is kept constant. With the interactions, the effects of the
predictors on the dependent variable only hold for those who have a 0 on the other
predictor that is involved in the interaction.

7 - Interpret the interaction effect lrscale*highedu, indicating how the effect of right wing
orientation changes when education level changes. 2 points

The interaction between political orientation and education level has an effect of -.104. P is
.038, smaller than .05, and therefore significant. The confidence interval ranges between -
.203 and -.006. Both values are negative and 0 is not part of these values. Taken together
with the significance level, this indicates that we can reject H0. There is a significant
interaction effect between right wing political orientation and level of education (i.e. the
effect of right wing political orientation on attitude toward migrants is significantly
different for higher and lower educated people). To interpret this effect, we need to
investigate what it means for the different education levels.

We do this by looking at effect of right wing orientation that is added for higher educated
people (higher educated people score a 1, therefore a higher value category) on top of the
effect of right wing orientation on attitudes toward migrants for lower educated people.
For lower educated people (value 0) the effect was -.113. For higher educated people it is
1*-.104 (the interaction effect). The effect of right wing orientation for higher educated
people is = -.113 + (-.104) = -.217. Therefore, we can see that for higher educated people,
the effect of right wing orientation on attitudes toward migrants is more strongly negative
than for lower educated people.

4
8 – Interpret the interaction based on the plot below. 1.5 points

We can see that the regression line for lower educated people is flatter than the regression
line for higher educated people. Therefore, we can see visually what we concluded in
exercise 5: for higher educated people, the effect of right wing orientation on attitudes
toward migrants is stronger than for lower educated people.

9 – Explain the results in your own words and draw substantive conclusions on this model. 1.5 points

5
In general, we can expect for the population that the more right wing people are, the more
negative they are toward migrants in the society (the worse they think migrants make
society). At the same time, higher educated people are more positive toward migrants in
society than lower educated people.

There is a moderation effect. The relationship between right wing political orientation and
attitudes toward migrants in society is moderated by the level of education. This effect is
significant for the lowest level of education and even more negative for the highest level of
education (therefore, for both lower and higher educated people, the more right wing they
are, the more negative they are toward migrants). However this effect is stronger for
higher educated people – the more right wing they are, the more negative they are toward
migrants as compared to their lower educated counterparts.

6
WORKING GROUP ASSIGNMENT KWOM/QNRM 2021-2022 WEEK 4

The data are from the research project Youth and Culture (Ganzeboom and Nagel, 1998-2002) and concern
more than 3000 students in classes 3, 4 and 5 (ages 14-17) who were questioned about their cultural
participation. They were questioned just before and just after the year in which an arts course (CKV) was
introduced. Furthermore, there is information about their social background and the cultural participation of
their parents.

Now we investigate whether participation in youth culture (pop concerts, cinema, dj/vj events and youth
events) depends on school level and age, and in particular whether age differences are stronger at the lower
(vmbo) or the higher school levels (havo/vwo).

1.Study both age variables. What is the relation between the variable age15 and age (see output)? In other
words: how is age15 calculated in spss?

For the variable age15 15 years is subtracted from age:

compute age15 = age - 15.


exec.

2. Why is the variable age so edited?

To get a well interpretable 0-category that appears in the data.


The 0 in age15 now means 'age = 15' and it appears in the data.

3.In the REGRESSION 1, the main effects of age and school level are examined.

What is the interpretation of the constant in this analysis?

A The constant 49.315 indicates the lowest value on participation in youth culture in the data, with age
and school level kept constant
B The constant 49.315 indicates the average value on participation in youth culture, with age and
school level kept constant
C The constant 49.315 indicates the score on participation in youth culture for an lower level school
student 12 years old.
D The constant 49.315 indicates the score on participation in youth culture for an lower level school
student 15-years old.

D - the constant indicates the predicted score at Y, for all X = 0, so school level =0 and age15=0.

4. How much more, according to model 1 does a 16-year-old higher school student participate in youth
culture than a 15-year-old student of the same school level?

A 2.563
B 4.368
C 53.683
D 51.878

B – this is about the effect of 1 year (unit) age, school level kept constant, that is the unstandardized
regression coefficient b of age15.
5. If this sample is one of 95% that contains the population value then we can conclude from the confidence
interval of age15 in model 1 that ....

A The population value cannot be 0, and that therefore H0 is unlikely


B The population value cannot be 0, so that H0 is possible
C The estimated b 4.368 is part of the confidence interval, so there is likely to be a positive effect in
the population, and H0 therefore unlikely
D The R-square .022 is not part of the confidence interval, so that H0 is likely

A - only if this sample is one of 95% which contain the population value; the sample can of course also
belong to the other 5%

6.Formulate a (H1) hypothesis on school level as moderator for the effect of age on youth culture
participation.

A H1: The effect of age on youth culture participation is stronger than that of school level.
B H1: The effect of age on youth culture participation is positive when school level is kept constant
C H1: Age differences in youth culture participation vary between both school levels
D H1: School level provides an explanation for age differences in youth culture participation

C - moderation involves the effect of age on youth culture participation that is different / varies, depending
on the value of a third variable, in this case school level

7.In REGRESSION 2 the interaction between age and school level is added to the model.

What is the interpretation of the regression coefficient b of age15?

A The b of age 15 .219 is the effect of age on the lower school level.
B The b of age 15 .219 is the effect of age on the higher school level
C The b of age 15 .219 is the effect of age where school level is kept constant.
D The b of age 15 .219 indicates the differences between school levels for age = 15 years

A - in an interaction model the main effects indicate the effect for the value at which the other variable is 0 -
in this case the effect of age for school level =0

8.What is the interpretation of the regression coefficient b of school level? State the coefficient in your
answer.

B = -2,624 The difference in youth culture participation between lower and higher school level students for
15-year-old students (i.e. where the other variable is 0). That is -2.624: at the age of 15, pupils of the higher
school level participate less in youth culture (on the 0-100 scale) than students of the lower school level.
9. Write out the full regression equation for model 2.

Fill in: Y = intercept + b1 * x1 + b2 * x2 + b3 * x1 *x2

JCULPAR =

JCULPAR = 49.636 + -2.624 SCHOOLLEVEL + .219 *AGE15 + 4.979 SCHOOLLEVEL_AGE15

10.What is the predicted score on youth culture participation of a 17-year-old student at the higher school
level?

Fill in (everything because it is a predicted score):

JCULPAR = 49.636 + -2.624 SCHOOLLEVEL + .219 *AGE15 + 4.979 * SCHOOLLEVEL_AGE15

JCULPAR = (49.636) + (-2.624 * 1) + (.219 *2) + (4.979 *1 * 2) = 57.408

11. Write the regression equation of JCULPAR on age15 separately for the lower and higher school level: ‘fill
in’ lower and higher school level, while leaving age15 unchanged as a variable (see video slides):

For the lower school level (vmbo):

JCULPAR =

JCULPAR = 49.636 + -2.624 * 0 + .219 *AGE15 + 4.979 * 0 * AGE15 = 49.636 + .219 AGE15

For the higher school level (havo-vwo):

JCULPAR =

JCULPAR = 49.636 + -2.624 * 1 + .219 *AGE15 + 4.979 * 1 * AGE15 = 47.012 + 5.198 AGE15

12.Study the plot: how do you see that the effect of age15 varies according to school level?

The plot shows that in the sample the age effect is positive for the higher school level (5.198) and (very)
slightly positive for the lower school level (.219).

13.Study the plot: how do you see that - conversely - the effect of school level depends on age15?

For low values of age, pupils of higher school level participate less often in youth culture than students of
lower school levels. For high values of age, the situation is reversed: students of lower school level
participate less often than students of higher school levels. The effect of school level is negative (meaning
higher school level students participate less) at the age of 15 (-2.624), but increases by 4,979 with each year,
so that at the age of 16 it is 2.355 and at the age of 17 it is 7.334.
14.What is the result of the hypothesis test, i.e. the hypothesis that the effect of age depends on school
level?

The interaction effect between school level and age is significant (t= 3.571, p < .001) and also lower than the
significance level 0.05 that we (implicitly) use here. The H0 that the effect of age does not vary between the
two school levels is therefore very unlikely. If this sample is one of 95% that contains the population value,
the interaction effect is between 2.245 and 7.713, so in all cases it is above 0. Based on this, we expect a
positive interaction between age and school level in the population as well.

REFERENTIES:

Ganzeboom, H., Nagel, I. 1998-2002 [p.i.]. Youth and Culture. A Multi-actor Panel Study. [machine readable
datafiles] Utrecht: ICS [producer]. DANS, p1724 [distributor] .

You might also like