Lesson 11: Interactions Continued

Nicky Wakim

2024-02-26

Learning Objective

Interpret the interaction component of a model with two continuous covariates, and how the main variable’s effect changes.

When there are only two covariates in the model, test whether one is a confounder or effect modifier.

Do we think food supply is an effect modifier for female literacy rate?

We can start by visualizing the relationship between life expectancy and female literacy rate by food supply
Questions of interest: Does the effect of female literacy rate on life expectancy differ depending on food supply?
- This is the same as: Is food supply is an effect modifier for female literacy rate? Is food supply an effect modifier of the association between life expectancy and female literacy rate?
Let’s run an interaction model to see!

Model with interaction between two continuous variables

Model we are fitting:

$L E = β_{0} + β_{1} F L R^{c} + β_{2} F S^{c} + β_{3} F L R^{c} \cdot F S^{c} + ϵ$

$L E$ as life expectancy
$F L R^{c}$ as the centered around the mean female literacy rate (continuous variable)
$F S^{c}$ as the centered around the mean food supply (continuous variable)

Code to center FLR and FS

gapm_sub = gapm_sub %>% 
  mutate(FLR_c = FemaleLiteracyRate - mean(FemaleLiteracyRate), 
         FS_c = FoodSupplykcPPD - mean(FoodSupplykcPPD))
mean_FS = mean(gapm_sub$FoodSupplykcPPD) %>% round(digits = 0)
mean_FLR = mean(gapm_sub$FemaleLiteracyRate) %>% round(digits = 2)

In R:

m_int_fs = lm(LifeExpectancyYrs ~ FLR_c + FS_c + FLR_c*FS_c, data = gapm_sub)

m_int_fs = lm(LifeExpectancyYrs ~ FLR_c*FS_c, data = gapm_sub)

Displaying the regression table and writing fitted regression equation

tidy_m_fs = tidy(m_int_fs, conf.int=T) 
tidy_m_fs %>% gt() %>% tab_options(table.font.size = 35) %>% fmt_number(decimals = 5)

term	estimate	std.error	statistic	p.value	conf.low	conf.high
(Intercept)	70.32060	0.72393	97.13721	0.00000	68.87601	71.76518
FLR_c	0.15532	0.03808	4.07905	0.00012	0.07934	0.23130
FS_c	0.00849	0.00182	4.67908	0.00001	0.00487	0.01212
FLR_c:FS_c	−0.00001	0.00008	−0.06908	0.94513	−0.00016	0.00015

$\begin{aligned} \hat{L E} = & {\hat{β}}_{0} + {\hat{β}}_{1} F L R^{c} + {\hat{β}}_{2} F S^{c} + {\hat{β}}_{3} F L R^{c} \cdot F S^{c} \\ \hat{L E} = & 70.32 + 0.16 \cdot F L R^{c} + 0.01 \cdot F S^{c} - 0.00001 \cdot F L R^{c} \cdot F S^{c} \end{aligned}$

Comparing fitted regression lines for various food supply values

To identify different lines, we need to pick example values of Food Supply:

Food Supply of 1812 kcal PPD

$\begin{aligned} \hat{L E} = & {\hat{β}}_{0} + {\hat{β}}_{1} F L R^{c} + \\ {\hat{β}}_{2} \cdot (- 1000) + \\ {\hat{β}}_{3} F L R^{c} \cdot (- 1000) \\ \hat{L E} = & ({\hat{β}}_{0} - 1000 {\hat{β}}_{2}) + \\ ({\hat{β}}_{1} - 1000 {\hat{β}}_{3}) F L R^{c} \end{aligned}$

Food Supply of 2812 kcal PPD

$\begin{aligned} \hat{L E} = & {\hat{β}}_{0} + {\hat{β}}_{1} F L R^{c} + \\ {\hat{β}}_{2} \cdot 0 + \\ {\hat{β}}_{3} F L R^{c} \cdot 0 \\ \hat{L E} = & ({\hat{β}}_{0}) + \\ ({\hat{β}}_{1}) F L R^{c} \end{aligned}$

Food Supply of 3812 kcal PPD

$\begin{aligned} \hat{L E} = & {\hat{β}}_{0} + {\hat{β}}_{1} F L R^{c} + \\ {\hat{β}}_{2} \cdot 1000 + \\ {\hat{β}}_{3} F L R^{c} \cdot 1000 \\ \hat{L E} = & ({\hat{β}}_{0} + 1000 {\hat{β}}_{2}) + \\ ({\hat{β}}_{1} + 1000 {\hat{β}}_{3}) F L R^{c} \end{aligned}$

Interpretation for interaction between two continuous variables

$\begin{aligned} \hat{L E} = & {\hat{β}}_{0} + {\hat{β}}_{1} F L R^{c} + {\hat{β}}_{2} F S^{c} + {\hat{β}}_{3} F L R^{c} \cdot F S^{c} \\ \hat{L E} = & [{\hat{β}}_{0} + {\hat{β}}_{2} \cdot F S^{c}] + \underset{FLR’s effect}{\underset{⏟}{[{\hat{β}}_{1} + {\hat{β}}_{3} \cdot F S^{c}]}} F L R \end{aligned}$

Interpretation:
- $β_{3}$ = mean change in female literacy rate’s effect, for every one kcal PPD increase in food supply
In summary, the interaction term can be interpreted as “difference in adjusted female literacy rate effect for every 1 kcal PPD increase in food supply”
It will be helpful to test the interaction to round out this interpretation!!

term	df.residual	rss	df	sumsq	statistic	p.value
LifeExpectancyYrs ~ FLR_c + FS_c	69.000	2,005.556	NA	NA	NA	NA
LifeExpectancyYrs ~ FLR_c + FS_c + FLR_c * FS_c	68.000	2,005.415	1.000	0.141	0.005	0.945

Learning Objective

Interpret the interaction component of a model with two continuous covariates, and how the main variable’s effect changes.

When there are only two covariates in the model, test whether one is a confounder or effect modifier.

Reminder from Lesson 9: General steps for F-test

Met underlying LINE assumptions

State the null hypothesis

\begin{aligned} H_{0} & : β_{1} = β_{2} = \dots = β_{k} = 0 \\ vs. H_{A} & : At least one β_{j} \neq 0, for j = 1, 2, \dots, k \end{aligned}

Specify the significance level.

Often we use $α = 0.05$

Specify the test statistic and its distribution under the null

The test statistic is $F$ , and follows an F-distribution with numerator $d f = k$ and denominator $d f = n - k - 1$ . ( $n$ = # obversation, $k$ = # covariates)

Compute the value of the test statistic

The calculated test statistic is

$F^{=} \frac{\frac{S S E (R) - S S E (F)}{d f_{R} - d f_{F}}}{\frac{S S E (F)}{d f_{F}}} = \frac{M S R_{f u l l}}{M S E_{f u l l}}$

Calculate the p-value

We are generally calculating: $P (F_{k, n - k - 1} > F)$

Write conclusion for hypothesis test

We (reject/fail to reject) the null hypothesis at the $100 α %$ significance level.

Step 1: Testing the interaction

We test with $α = 0.10$
Follow the F-test procedure in Lesson 9 (MLR: Inference/F-test)
- This means we need to follow the 7 steps of the general F-test in previous slide (taken from Lesson 9)
Use the hypothesis tests for the specific variable combo:

Binary & continuous variable (Lesson 11, LOB 2)

Testing a single coefficient for the interaction term using F-test comparing full model to reduced model

Multi-level & continuous variables (Lesson 11, LOB 3)

Testing group of coefficients for the interaction terms using F-test comparing full to reduced model

Binary & multi-level variable (Lesson 11, LOB 4)

Testing group of coefficients for the interaction terms using F-test comparing full to reduced model

Two continuous variables (Lesson 11, LOB 5)

Testing a single coefficient for the interaction term using F-test comparing full to reduced model

Testing for percent change ( $Δ %$ ) in a coefficient

Let’s say we have $X_{1}$ and $X_{2}$ , and we specifically want to see if $X_{2}$ is a confounder for $X_{1}$ (the explanatory variable or variable of interest)
If we are only considering $X_{1}$ and $X_{2}$ , then we need to run the following two models:
- Fitted model 1 / reduced model (mod1): $\hat{Y} = {\hat{β}}_{0} + {\hat{β}}_{1} X_{1}$
  - We call the above ${\hat{β}}_{1}$ the reduced model coefficient: ${\hat{β}}_{1, mod1}$ or ${\hat{β}}_{1, red}$
- Fitted model 2 / Full model (mod2): $\hat{Y} = {\hat{β}}_{0} + {\hat{β}}_{1} X_{1} + {\hat{β}}_{2} X_{2}$
  - We call this ${\hat{β}}_{1}$ the full model coefficient: ${\hat{β}}_{1, mod2}$ or ${\hat{β}}_{1, full}$

Calculation for % change in coefficient

$Δ % = 100 % \cdot \frac{{\hat{β}}_{1, mod1} - {\hat{β}}_{1, mod2}}{{\hat{β}}_{1, mod2}} = 100 % \cdot \frac{{\hat{β}}_{1, red} - {\hat{β}}_{1, full}}{{\hat{β}}_{1, full}}$

Is food supply a confounder for female literacy rate? (1/3)

Run models with and without food supply:

Model 1 (reduced): $L E = β_{0} + β_{1} F L R^{c} + ϵ$

mod1_red = lm(LifeExpectancyYrs ~ FLR_c, data = gapm_sub)

Model 2 (full): $L E = β_{0} + β_{1} F L R^{c} + β_{2} F S^{c} + ϵ$

mod2_full = lm(LifeExpectancyYrs ~ FLR_c + FS_c, data = gapm_sub)

Note that the full model when testing for confounding was the reduced model for testing an interaction
Full and reduced are always relative qualifiers of the models that we are testing

Is food supply a confounder for female literacy rate? (2/3)

Record the coefficient estimate for centered female literacy rate in both models:

Model 1 (reduced):

term	estimate	std.error	statistic	p.value	conf.low	conf.high
(Intercept)	70.29722	0.72578	96.85709	0.00000	68.84969	71.74475
FLR_c	0.22990	0.03219	7.14139	0.00000	0.16570	0.29411

Model 2 (full):

term	estimate	std.error	statistic	p.value	conf.low	conf.high
(Intercept)	70.29722	0.63537	110.63985	0.00000	69.02969	71.56475
FLR_c	0.15670	0.03216	4.87271	0.00001	0.09254	0.22085
FS_c	0.00848	0.00179	4.72646	0.00001	0.00490	0.01206

Calculate the percent change:

$Δ % = 100 % \cdot \frac{{\hat{β}}_{1, mod1} - {\hat{β}}_{1, mod2}}{{\hat{β}}_{1, mod2}} = 100 % \cdot \frac{0.22990 - 0.15670}{0.15670} = 46.71 %$

Determining if income level is an effect modifier, confounder, or neither

Step 1: Testing the interaction/effect modifier
- Compare model with and without interaction using F-test to see if interaction is significant
- Models
  - Model 1 (red): $\begin{aligned} L E = & β_{0} + β_{1} I (Americas) + β_{2} I (Asia) + β_{3} I (Europe) + β_{4} I (high income) + ϵ \end{aligned}$
  - Model 2 (full): $\begin{aligned} L E = & β_{0} + β_{1} I (Americas) + β_{2} I (Asia) + β_{3} I (Europe) + β_{4} I (high income) + \\ β_{5} \cdot I (high income) \cdot I (Americas) + β_{6} \cdot I (high income) \cdot I (Asia) + \\ β_{7} \cdot I (high income) \cdot I (Europe) + ϵ \end{aligned}$
Step 2: Testing a confounder (only if not an effect modifier)
- Compare model with and without main effect for additional variable (income level) using F-test to see if additional variable (income level) is a confounder
- Models
  - Model 1 (reduced): $\begin{aligned} L E = & β_{0} + β_{1} I (Americas) + β_{2} I (Asia) + β_{3} I (Europe) + ϵ \end{aligned}$
  - Model 2 (full): $\begin{aligned} L E = & β_{0} + β_{1} I (Americas) + β_{2} I (Asia) + β_{3} I (Europe) + β_{4} I (high income) + ϵ \end{aligned}$

term	df.residual	rss	df	sumsq	statistic	p.value
LifeExpectancyYrs ~ income_levels2 + four_regions	67.000	1,693.242	NA	NA	NA	NA
LifeExpectancyYrs ~ income_levels2 + four_regions + income_levels2 * four_regions	64.000	1,681.304	3.000	11.938	0.151	0.928

Step 2: See if income is a confounder

Fit the reduced and full model for testing the confounder

Model 1 (reduced): $\begin{aligned} L E = & β_{0} + β_{1} I (Americas) + β_{2} I (Asia) + β_{3} I (Europe) + ϵ \end{aligned}$

mod1_wr_inc_red = lm(LifeExpectancyYrs ~ four_regions, 
                   data = gapm_sub)

Model 2 (full): $\begin{aligned} L E = & β_{0} + β_{1} I (Americas) + β_{2} I (Asia) + β_{3} I (Europe) + β_{4} I (high income) + ϵ \end{aligned}$

mod1_wr_inc_full = lm(LifeExpectancyYrs ~ four_regions + income_levels2, 
                   data = gapm_sub)

Step 2: See if income is a confounder

Record the coefficient estimate for centered female literacy rate in both models:
Model 1 (reduced): $\begin{aligned} \hat{L E} = & {\hat{β}}_{0} + {\hat{β}}_{1} I (Americas) + {\hat{β}}_{2} I (Asia) + {\hat{β}}_{3} I (Europe) \end{aligned}$

term	estimate	std.error	statistic	p.value	conf.low	conf.high
(Intercept)	61.27000	1.16508	52.58870	0.00000	58.94512	63.59488
four_regionsAmericas	14.33000	1.90257	7.53193	0.00000	10.53349	18.12651
four_regionsAsia	8.11824	1.71883	4.72313	0.00001	4.68837	11.54810
four_regionsEurope	14.78217	1.59304	9.27924	0.00000	11.60332	17.96103

Model 2 (full): $\begin{aligned} \hat{L E} = & {\hat{β}}_{0} + {\hat{β}}_{1} I (Americas) + {\hat{β}}_{2} I (Asia) + {\hat{β}}_{3} I (Europe) + {\hat{β}}_{4} I (high income) \end{aligned}$

term	estimate	std.error	statistic	p.value	conf.low	conf.high
(Intercept)	60.54716	1.16190	52.11048	0.00000	58.22800	62.86632
four_regionsAmericas	12.04102	2.05816	5.85038	0.00000	7.93292	16.14912
four_regionsAsia	7.77808	1.66414	4.67394	0.00001	4.45645	11.09971
four_regionsEurope	12.51938	1.79139	6.98864	0.00000	8.94375	16.09501
income_levels2Higher income	3.61419	1.46967	2.45917	0.01651	0.68070	6.54767

Step 2: See if income is a confounder

Calculate the percent change for ${\hat{β}}_{1}$ : $Δ % = 100 % \cdot \frac{{\hat{β}}_{1, mod1} - {\hat{β}}_{1, mod2}}{{\hat{β}}_{1, mod2}} = 100 % \cdot \frac{14.33000 - 12.04102}{12.04102} = 19.01$
Calculate the percent change for ${\hat{β}}_{2}$ : $Δ % = 100 % \cdot \frac{{\hat{β}}_{2, mod1} - {\hat{β}}_{2, mod2}}{{\hat{β}}_{2, mod2}} = 100 % \cdot \frac{8.11824 - 7.77808}{7.77808} = 4.37$
Calculate the percent change for ${\hat{β}}_{3}$ : $Δ % = 100 % \cdot \frac{{\hat{β}}_{3, mod1} - {\hat{β}}_{3, mod2}}{{\hat{β}}_{3, mod2}} = 100 % \cdot \frac{14.78217 - 12.51938}{12.51938} = 18.07$
Note that two of these % changes are greater than 10%, and one is less than 10%…

General interpretation of the interaction term (reference)

$E [Y ∣ X_{1}, X_{2}] = β_{0} + \underset{X_{1} ’s effect}{\underset{⏟}{(β_{1} + β_{3} X_{2})}} X_{1} + \underset{X_{2} held constant}{\underset{⏟}{β_{2} X_{2}}}$
$E [Y ∣ X_{1}, X_{2}] = β_{0} + \underset{X_{2} ’s effect}{\underset{⏟}{(β_{2} + β_{3} X_{1})}} X_{2} + \underset{X_{1} held constant}{\underset{⏟}{β_{1} X_{1}}}$

Interpretation:
- $β_{3}$ = mean change in $X_{1}$ ’s effect, per unit increase in $X_{2}$ ;
- $β_{3}$ = mean change in $X_{2}$ ’s effect, per unit increase in $X_{1}$ ;
- where the “ $X_{1}$ effect” equals the change in $E [Y]$ per unit increase in $X_{1}$ with $X_{2}$ held constant, i.e. “adjusted $X_{1}$ effect”
In summary, the interaction term can be interpreted as “difference in adjusted $X_{1}$ (or $X_{2}$ ) effect per unit increase in $X_{2}$ (or $X_{1}$ )”

A glimpse at how interactions might be incorporated into model selection

Identify outcome (Y) and primary explanatory (X) variables
Decide which other variables might be important and could be potential confounders. Add these to the model.
- This is often done by indentifying variables that previous research deemed important, or researchers believe could be important
- From a statistical perspective, we often include variables that are significantly associated with the outcome (in their respective SLR)
(Optional step) Test 3 way interactions
- This makes our model incredibly hard to interpret. Our class will not cover this!!
- We will skip to testing 2 way interactions
Test 2 way interactions
- When testing a 2 way interaction, make sure the full and reduced models contain the main effects
- First test all the 2 way interactions together using a partial F-test (with $a l p h a = 0.10$ )
  - If this test not significant, do not test 2-way interactions individually
  - If partial F-test is significant, then test each of the 2-way interactions
Remaining main effects - to include of not to include?
- For variables that are included in any interactions, they will be automatically included as main effects and thus not checked for confounding
- For variables that are not included in any interactions:
  - Check to see if they are confounders by seeing whether exclusion of the variable(s) changes any of the coefficient of the primary explanatory variable (including interactions) X by more than 10%
    - If any of X’s coefficients change when removing the potential confounder, then keep it in the model

1 / 31

Lesson 11: Interactions Continued Nicky Wakim 2024-02-26

Lesson 11: Interactions Continued
Learning Objective
Do we think food supply is an effect modifier for female literacy rate?
Model with interaction between two continuous variables
Displaying the regression table and writing fitted regression equation
Comparing fitted regression lines for various food supply values
Poll Everywhere Question??
Interpretation for interaction between two continuous variables
Test interaction between two continuous variables
Test interaction between two continuous variables
Learning Objective
Deciding between confounder and effect modifier
Reminder from Lesson 9: General steps for F-test
Step 1: Testing the interaction
Poll Everywhere Questions 2-4
Step 2: Testing a confounder
Testing for percent change ( $Δ %$ ) in a coefficient
Is food supply a confounder for female literacy rate? (1/3)
Is food supply a confounder for female literacy rate? (2/3)
Is food supply a confounder for female literacy rate? (3/3)
Let’s try this out on one of our potential effect modifiers or confounders
Determining if income level is an effect modifier, confounder, or neither
Step 1: Results from Lesson 11 LOB 4
Step 2: See if income is a confounder
Step 2: See if income is a confounder
Step 2: See if income is a confounder
Step 2: See if income is a confounder
If you want extra practice
Extra Reference Material
General interpretation of the interaction term (reference)
A glimpse at how interactions might be incorporated into model selection