Please enable JavaScript.
Coggle requires JavaScript to display documents.
General Linear Models - 3 (Interaction Terms (Impact on model (Underweight…
General Linear Models - 3
Prediction
Predicted ex time is different
to mean ex time for each group
Need to use model
to make predictions
Determine the
predicted
ex time
for each BMI group
Estimated
Marginal Means
Present the predicted values for each group,
taking all other parameters (eg HR)
at the sample mean
(av HR across entire sample)
Calculated from model
Exercise time =
SPSS output:
BMI_group
compare to
Descriptive Statistics
ex time predicted using model
vs measured values from descriptive stats
Note that the <21 group has difference
btwn predicted & measured value of ex time
Model is adjusted for HR,
underweight people may have had
incr Rest_HR (decr ex time with incr HR)
Go back and look at descriptives for Rest_HR:
Note that underweight group had
different Rest_HR mean
Benefits of using model:
can control for (eg Rest_HR)
more "accurate" than descriptive stats
Effect Modification
Model structure
restricts each BMI group
to have parallel lines on the graph
But what if the slope of the line is
different for each BMI group?
Fitted line of
Rest_HR vs ex_time
by BMI category:
Gives 3 lines with same slope
but different heights
Different slopes = effect modification
BMI is called an effect modifier
Can be included in model
as interaction terms
Interaction Terms
va's (eg HR & BMI group)
are main effects
interaction is (HR x BMI group)
included in model
represent a multiplicative effect
between 2 (or more) va's
Exercise time =
Used to assess effect modification
Impact on model
Underweight group
Normal group
Overweight group
May also have interactions
btwn 2 cat va's
(eg BMI group & gender)
Underweight group
Normal group
Overweight group
Approach:
If interaction possible:
Include in model
& test for sign.
If not sign. - take out
& focus on main effects
GLMs & Model Building
Often select va's to include in model
by first conducting ANOVA &
multiple linear regression
(sim to filtering)
If you want to consider interactions
in your model - fit a full model first
including interactions
sequentially remove interactions
first if they are not sign.
DO NOT remove main effects
(even if non-sign.) if
interaction
involving
main effect is sign.
User needs to specify the model
No capacity in SPSS
to do stepwise methods
Note:
SPSS will auto create
dummy va's during analysis
Your definition of dummy va's
needs to match that in SPSS analysis
Need to
define dummy va's
when specifying model