Mixed Effects Models - Logit Models

Week 9.1: Logit Models
! Introduction to Generalized LMER
! Categorical Outcomes
! Probabilities and Odds
! Logit
! Link Functions
! Implementation in R
! Parameter Interpretation for Logit Models
! Intercept
! Coding the Dependent Variable
! Categorical Variables
! Continuous Variables
! Interactions
! Confidence Intervals

This Week’s Dataset
• New dataset: cuedrecall.csv
• Cued recall task:
• Study phase: See pairs of words
• WOLF--PUPPY
• Test phase: See the first word, have to type in the
second
• WOLF--___?____

cuedrecall.csv
• 120 Subjects, all see the same 36 WordPairs
we arbitrarily created
• Subjects are assigned a Strategy:
• Maintenance rehearsal: Repeat it over & over
• Elaborative rehearsal: Relate the two words
• Subjects choose the StudyTime for each word
• Which independent variables are fixed effects?
• Which independent variables are random?

• With our mixed effect models, we’ve been predicting
the outcome of particular observations or trials
Generalized Linear Mixed Effects Models
= Intercept + +
Study Time
Subject
+ Item
RT +
Strategy

• We sum up the influences on the right hand side as our
model of the DV on the left-hand side
• Works great for normally distributed DVs
yij = β0 + !100x1ij + !200x2ij + ui0 + v0j + eij
Intercept
Study
Time
Strategy Residual
error
Subject
Effect
Item
Effect
Can be any number
β0 + β1X1i + … + ei = -3, 0, 0.13, 1.47, 24…
Recall

• Problem here when we have only 2 possible outcomes:
0 or 1
• This is a binomial (or dichotomous) dependent variable
Intercept
Study
Time
Strategy Residual
error
Subject
Effect
Item
Effect
0 or 1 Can be any number
β0 + β1X1i + … + ei = -3, 0, 0.13, 1.47, 24…
Recall

Binomial Distribution
• Distribution of outcomes when one of
two events (a “hit”) occurs with
probability p
• Examples:
• Word pair recalled or not
• Person diagnosed with depression or not
• High school student decides to attend college or not
• Speaker produces active sentence or passive
sentence

• How can we link the linear model to the two binomial
outcomes?
Intercept
Study
Time
Strategy Residual
error
Subject
Effect
Item
Effect
β0 + β1X1i + … + ei = -3, 0, 0.13, 1.47, 24…
Recall

• What if we modelled the probability (or proportion) of
recall?
• On the right track…
• But, still bounded between 0 and 1
Intercept
Study
Time
Strategy Residual
error
Subject
Effect
Item
Effect
β0 + β1X1i + … + ei = -3, 0, 0.13, 1.47, 24…
Recall

Probabilities, Odds, and Log Odds
• What about the odds of correct recall?
• If the probability of recall is .67, what are the
odds?
• .67/(1-.67) = .67/.33 ≈ 2
• Some other odds:
• Odds of being right-handed: ≈.9/.1 = 9
• Odds of identical twins: 1/375
• Odds are < 1 if the event doesn’t happen more often that
it does happen
p(recall) p(recall)
1-p(recall)
=
p(forgetting)

• If the probability of recall is .67, what are the
odds?
• .67/(1-.67) = .67/.33 ≈ 2
• Some other odds:
• Odds of being right-handed: ≈.9/.1 = 9
• Odds of identical twins: 1/375
• Odds of having five fingers
per hand: ≈500/1
p(recall) p(recall)
1-p(recall)
=
≈ .003
p(forgetting)

• Try converting these probabilities into odds
• Probability of a coin flip being tails: .50
• Probability a random American is a woman: .51
• Probability of maximum shock in Milgram study: .67
• Probability of depression sometime in your life: .17
• Probability of graduating high school in the US: .92
p(recall)
p(forgetting)
p(recall)
1-p(recall)
=

• Try converting these probabilities into odds
• Probability of a coin flip being tails: .50
• = 1.00
• Probability a random American is a woman: .51
• ≈ 1.04
• Probability of maximum shock in Milgram study: .67
• ≈ 2.00
• Probability of depression sometime in your life: .17
• ≈ 0.20
• Probability of graduating high school in the US: .92
• ≈ 11.5
p(recall)
p(forgetting)
p(recall)
1-p(recall)
=

• Creating a model of the odds of correct recall
would be better than a model of the probability
• Odds have no upper bound
• Can have 500:1 odds!
• But, still a lower bound at 0
p(recall)
p(forgetting)
p(recall)
1-p(recall)
=

Logit
• Now, let’s take the logarithm of the odds
• Specifically, the natural log (sometimes written as ln )
• The natural log is what we get by default from log() in R
(and in most other programming languages, too)
• On Google or in Calculator app, need to use ln
• The log odds or logit
p(recall)
1-p(recall)
[ ]
log odds = log

Logit
• The log odds or logit
• If the probability of recall is 0.8, what are the
log odds of recall?
• log(.8/(1-.8))
• log(.8/.2)
• log(4)
• 1.39
p(recall)
1-p(recall)
[ ]
log odds = log

[ ]
Logit
• What are the log odds?
• Probability of a clear day in Pittsburgh: .58
• Probability of precipitation in Pittsburgh: .42
• Probability of dying of a heart attack: .29
• Probability a sq ft. of Earth’s surface is water: .71
• Probability of detecting a gorilla in a crowd: .50
p(recall)
1-p(recall)
log odds = log

[ ]
Logit
• What are the log odds?
• Probability of a clear day in Pittsburgh: .58
• 0.33
• Probability of precipitation in Pittsburgh: .42
• -0.33
• Probability of dying of a heart attack: .29
• -0.90
• Probability a sq ft. of Earth’s surface is water: .71
• 0.90
• Probability of detecting a gorilla in a crowd: .50
• 0
p(recall)
1-p(recall)
log odds = log

Logit
• Probabilities equidistant from .50 have the same
absolute value on the log odds
Probability of
precipitation in
Pittsburgh = .42
Log odds: -0.33
Probability of clear
day in Pittsburgh
= .58
Log odds: 0.33

Logit
• Probabilities equidistant from .50 have the same
absolute value on the log odds
• Magnitude reflects degree to which 1 outcome
dominates
Probability a
square foot of
Earth’s
surface is
water = .71
Log odds: 0.90
Probability a
square foot of
Earth’s surface
is land = .29
Log odds: -0.90

Logit
• When neither outcome is more probable than the
other, log odds of each is 0
Probability of
spotting the
gorilla = .50
Log odds: 0
Probability of
not spotting the
gorilla = .50
Log odds: 0

0.0 0.2 0.4 0.6 0.8 1.0
-4
-2
0
2
4
PROBABILITY of recall
LOG
ODDS
of
recall
As
probability
of hit
approaches
1, log odds
approach
infinity. No
upper
bound.
As
probability
of hit
approaches
0, log odds
approach
negative
infinity. No
lower
bound.
If probability of
hit is .5 (even
odds), log
odds are zero.
Probabilities
equidistant from .5
have log odds with
the same absolute
value (-1.39 and 1.39)
PROBABILITY of recall
LOG
ODDS
of
recall

Generalized LMERs
• To make predictions about a binomial
distribution, we predict the log odds (logit) of a hit
• This can be any number!
• In most other respects, like all linear models
= β0 + !100x1ij + !200x2ij
Intercept Study
Time
Strategy
Can be any number
p(recall)
1-p(recall)
[ ]
log
Can be any number
β0 + β1X1i + β2X2i = -3, 0, 0.13, 1.47, 24…

Generalized LMERs
• Link function that relates the two sides is the logit
• “Generalized linear mixed effect regression” when
we use a link function other than the normal
• Before, our link function was just the identity
= β0 + !100x1ij + !200x2ij
Intercept Study
Time
Strategy
Can be any number
p(recall)
1-p(recall)
[ ]
log
Can be any number
β0 + β1X1i + β2X2i = -3, 0, 0.13, 1.47, 24…

From lmer() to glmer()
• For generalized linear mixed effects models, we
use glmer()
• Part of lme4, so you already have it!
LMER
Linear Mixed Effects
Regression
GLMER
Generalized Linear Mixed
Effects Regression

cuedrecall.csv
• Neither of these strategies is a clear baseline—
how should we code the Strategy variable?
• Effects coding:
• contrasts(cuedrecall$Strategy) <-
c(0.5, -0.5)

cuedrecall.csv
• There’s no such thing as a StudyTime of 0 s …
what should we do this variable?
• Let’s center it around the mean
• cuedrecall %>%
mutate(StudyTime.cen = center(StudyTime))
-> cuedrecall

glmer()
• glmer() syntax identical to lmer() except we
add family=binomial argument to indicate
which distribution we want
• Generic example:
• glmer(DV ~ 1 + Variables +
(1+Variables|RandomEffect),
data=mydataframe, family=binomial)
• For our data:
• glmer(Recalled ~ 1 + StudyTime.cen *
Strategy + (1|Subject) + (1|WordPair),
data=cuedrecall, family=binomial)

Can You Spot the Differences?
Binomial family
with logit link
Fit by Laplace estimation (don’t
need to worry about REML vs ML)
Wald z test: p values automatically given by Laplace estimation.
Don’t need lmerTest for Satterthwaite t test
No residual error
variance. Trial
outcome can only
be “recalled” or
“forgotten,” so each
prediction is either
correct or incorrect.

Interpretation: Intercept
• OK … but what do our results mean?
• Let’s start with the intercept
• Since we centered, this the average log odds of
recall across conditions
• Log odds of recall are 0.31
• One statistically correct way to interpret the model …
but not easy to understand in real-world terms

Logarithm Review
• How “good” are log odds of 0.31?
• log(10) = 2.30 because e2.30 = 10
• “The power to which we raise e (≈ 2.72) to get 10.”
• Natural log (now standard meaning of log)
• Help! Get me out of log world!
• We can undo log() with exp()
• exp(3) means “Raise e to the
exponent of 3”
• exp(log(3))
• Find “the power to which we raise e to get 3” and then
“raise e to that power” (giving us 3)

Interpreting Estimates
• Let’s go from log odds back to regular odds
• Baseline odds of recall are 1.36
• 1.36 correct responses for 1 incorrect response
• About 4 correct responses for every 3 incorrect
• A little better than 1:1 odds (50%)
1.36 0.31
exp()

• This is expressed in terms of the odds of recall
because we coded that as the “hit” (1)
• glmer’s rule:
• If a numerical variable, 0s are considered misses
and 1s are considered hits
• If a two-level categorical variable, the first category
is a miss and the second is a hit
• Could use relevel() to reorder
“Forgotten” listed first, so it’s the “miss”
“Remembered” listed second, so it’s the “hit”

• This is expressed in terms of the odds of recall
because we coded that as the “hit” (1)
• Had we reversed the coding, we’d get the log
odds of forgetting = -0.32
• Same p-value, same magnitude, just different sign
• Remember how logits equally distant from even odds
have the same absolute value?
• Choose the coding that makes sense for your
research question. Do you want to talk about “what
predicts graduation” or “what predicts dropping out”?

Interpretation: Categorical Predictors
• Now, let’s look at a categorical independent
variable
• The study strategy assigned
• Using elaborative rehearsal increases the
chance of recall by 2.29 logits…

• What happens if we exp() this parameter?
• What are…?
• Multiply 2 * 3, then take the log
• Find log(2) and log(3), then add them
• Log World turns multiplication into addition
• Because ea * eb = ea+b
x +
log()
log(6) ≈ 1.79
1.10+0.69 ≈ 1.79

• What happens if we exp() this parameter?
• What are…?
• Multiply 2 * 3, then take the log
• Find log(2) and log(3), then add them
• Find exp(2) and exp(3), then
multiply them
• Add 2 + 3, then use exp()
• Log World turns multiplication into addition
• exp() turns additions back into multiplications
• exp(2+3) = exp(2) * exp(3)
x +
log()
exp()
log(6) ≈ 1.79
1.10+0.69 ≈ 1.79
7.39 * 20 ≈ 148
exp(5) ≈ 148

• Let’s use exp() to turn our effect on log odds
back into an effect on the odds
• Remember that effects that were additive in log
odds become multiplicative in odds
• Elaboration increases odds of recall by 9.87 times
• This can be described as an odds ratio
9.87 + 2.29
exp()
x
Odds of recall with elaborative rehearsal
Odds of recall with maintenance rehearsal
= 9.87

• Let’s use exp() to turn our effect on log odds
back into an effect on the odds
• Remember that effects that were additive in log
odds become multiplicative in odds
• When we study COFFEE-TEA with maintenance rehearsal, our
odds of recall are 2:1. What if we use elaborative rehearsal?
• Initial odds of 2 x 9.87 increase = 19.74 (NOT 11.87!)
9.87 + 2.29
exp()
x

Interpretation: Continuous Predictors
• Next, a continuous predictor variable
• Time (in seconds) spent studying the word pair
• As in all regressions, effect of a 1-unit change
• Each second of study time = +0.40 log odds of recall
1.49 0.40
exp()
x

Interpretation: Continuous Predictors
• Next, a continuous predictor variable
• Time (in seconds) spent studying the word pair
• As in all regressions, effect of a 1-unit change
• Each second of study time = +0.40 log odds of recall
• Each second of study time = increases odds of recall
by 1.49 times

Interpretation: Interactions
• Study time has a + effect on recall
• Elaborative strategy has a + effect on recall
• And, their interaction has a + coefficient
• Interpretation?:
• “Additional study time is more beneficial
when using an elaboration strategy”
• “Elaboration strategy is more helpful if
you devote more time to the item”
(another way of saying the same thing)

Interpretation: Interactions
• We now understand the sign of the interaction
• What about the specific numeric estimate?
• What does 0.28 mean in this context?
• At the mean study time (3.5 s), difference in
log odds between strategies was 2.29 logits
• This difference gets 0.28 logits bigger for each 1
increase in study time
• At 5.5 s: Difference between strategies is 2.85 logits
• Odds of correct recall with elaborative rehearsal are 17
times greater!

Confidence Intervals
• Both our estimates and standard errors are
in terms of log odds
• Thus, so is our
confidence interval
• 95% confidence interval for StudyTime effect in
terms of log odds
• Estimate +/- (1.96 * standard error)
• 2.288 +/- (1.96 * 0.136)
• 2.288 +/- 0.267
• [2.02, 2.56]
• Point estimate is 2.28 change in logits.
• 95% CI around that estimate is [2.02, 2.56]

• Both our estimates and standard errors are
in terms of log odds
• Thus, so is our confidence interval
• For StudyTime effect:
• Point estimate is 2.28 change in logits.
• 95% CI around that estimate is [2.02, 2.56]
• But, log odds hard to understand. Let’s use
exp() to turn the endpoints of the confidence
interval into odds
• 95% CI is exp(c(2.02, 2.56)) =
[7.54, 12.94]
• Need to compute the CI first, then exp()

• For confidence intervals around log odds
• As usual, we care about whether the confidence
interval contains 0
• Adding or subtracting 0 to the log odds doesn’t
change it. It’s the null effect.
• So, we’re interested in whether the estimate of the
effect significantly differs from 0.
• When we transform to the odds
• Now, we care about whether the CI contains 1
• Remember, effects on odds are multiplicative.
Multiplying by 1 is the null effect we test against.
• A CI that contains 0 in log odds will always contain
1 when we transform to odds (and vice versa).

• Strategy effect:
• Point estimate: Elaborative rehearsal increases
odds of recall by 9.78 times
• 95% CI: [7.54, 12.94]
• Our point estimate is 9.78…
• Compare the distance to 7.54 vs. the distance to
12.94
• Confidence intervals are numerically
asymmetric once turned back into odds
9.78
7.54 12.94

-3 -2 -1 0 1 2 3
0
5
10
15
LOG ODDS of recall
ODDS
of
recall
Value of the odds
changes slowly when
logit is small
LOG ODDS of recall
Odds changes quickly
at higher logits
ODDS
of
recall
Asymmetric Confidence Intervals
• We’re more certain about the odds for
smaller/lower logits

Mixed Effects Models - Logit Models

More Related Content

What's hot (8)

Similar to Mixed Effects Models - Logit Models (20)

More from Scott Fraundorf (15)

Recently uploaded (20)

Mixed Effects Models - Logit Models