mixed models: misconceptions, pitfalls, and many opportunities · • misconceptions • problems...

55
Mixed models: Misconceptions, pitfalls, and many opportunities 4th Channel Network Conference: St-Andrews, July 3-5, 2013 Geert Verbeke I-Biostat: International Institute for Biostatistics and statistical Bioinformatics Katholieke Universiteit Leuven & Universiteit Hasselt, Belgium http://perswww.kuleuven.be/geert verbeke

Upload: others

Post on 24-Aug-2020

2 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

Mixed models:

Misconceptions, pitfalls, and many opportunities

4th Channel Network Conference: St-Andrews, July 3-5, 2013

Geert Verbeke

I-Biostat: International Institute for Biostatistics and statistical Bioinformatics

Katholieke Universiteit Leuven & Universiteit Hasselt, Belgium

http://perswww.kuleuven.be/geert verbeke

Page 2: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

Outline

• Mixed models in action

• Random-effects assumptions

• Mixed models in large data sets

4th Channel Network Conference: St-Andrews, July 3-5, 2013 1

Page 3: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

I will focus on . . .

• Model formulation

• Parameter interpretation

• Misconceptions

• Problems often encountered in practice

• Random-effects distributional assumptions

• Issues with large data sets

4th Channel Network Conference: St-Andrews, July 3-5, 2013 2

Page 4: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

I will NOT talk about . . .

• Estimation methods

• Inferential procedures

• Asymptotics

• Algorithms

• Model selection

• . . .

4th Channel Network Conference: St-Andrews, July 3-5, 2013 3

Page 5: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

Mixed models in action

4th Channel Network Conference: St-Andrews, July 3-5, 2013 4

Page 6: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

The Diabetes Project Leuven

(Borgermans et al. 2009)

• The impact of offering GP’s assistance of a diabetes care team,consisting of a nurse educator, a dietician, an ophthalmologist and aninternal medicine doctor, for the treatment of their diabetes patients

• GP’s randomized to one of two programs:

. LIP: Low Intervention Program (group A)

. HIP: High Intervention Program (group R)

• We consider the HIP group only

. 61 GP’s with 1577 patients

. # patients per GP between 5 and 138, with median of 47

4th Channel Network Conference: St-Andrews, July 3-5, 2013 5

Page 7: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

• Patients were measured twice:

. When the program was initiated (time T0)

. After one year (time T1)

• HbA1c: glycosylated hemoglobin:

. Molecule in red blood cells that attaches to glucose (blood sugar)

. High values reflect more glucose in blood

. Gives a good estimate of how well diabetes has been managed overlast 2 or 3 months

. Non-diabetics have values between 4% and 6%

. HbA1c above 7% means diabetes is poorly controlled, implyinghigher risk for long-term complications.

4th Channel Network Conference: St-Andrews, July 3-5, 2013 6

Page 8: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

A logistic mixed model

• Dichotomized version of HbA1c:

Y =

1 if HbA1c < 7%

0 if HbA1c ≥ 7%

• A three-level logistic mixed model:

Yijk ∼ Bernoulli(πijk)

logit(πijk) = log(

πijk

1−πijk

)= β0 + β1tk + ai + bj(i),

ai ∼ N (0, σ2GP ), bj(i) ∼ N (0, σ2

PAT )

4th Channel Network Conference: St-Andrews, July 3-5, 2013 7

Page 9: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

Fixed effects

Effect Estimate (se) p-value

Intercept β0: 0.1662 (0.0796) 0.0410

Time β1: 0.6240 (0.0812) < .0001

“Fixed effects model systematic trends”

6=“Fixed effects model average trends”

4th Channel Network Conference: St-Andrews, July 3-5, 2013 8

Page 10: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

Logistic random-intercepts model

E[Yijk|ai, bj(i)

]= πijk =

exp[β0 + β1tk + ai + bj(i)

]

1 + exp[β0 + β1tk + ai + bj(i)

]

4th Channel Network Conference: St-Andrews, July 3-5, 2013 9

Page 11: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

Average subject treated by average GP

E[Yijk|ai = 0, bj(i) = 0

]=

exp [β0 + β1tk + 0 + 0]

1 + exp [β0 + β1tk + 0 + 0]

4th Channel Network Conference: St-Andrews, July 3-5, 2013 10

Page 12: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

Average evolution

E[Yijk

]= E

exp[β0 + β1tk + ai + bj(i)

]

1 + exp[β0 + β1tk + ai + bj(i)

]

4th Channel Network Conference: St-Andrews, July 3-5, 2013 11

Page 13: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

Conclusion

Average evolution 6= Evolution average subject

• Parameters in the mixed model have a subject-specific interpretation,not a population-averaged one.

• Calculation of the marginal average population requires computation of∫∫

exp[β0 + β1tk + ai + bj(i)

]

1 + exp[β0 + β1tk + ai + bj(i)

]

f (ai)f (bj(i)) daidbj(i)

4th Channel Network Conference: St-Andrews, July 3-5, 2013 12

Page 14: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

Variance components

Effect Estimate (se) p-value

Between GP variance σ2GP : 0.1399 (0.0528) ?

Between patient variance σ2PAT : 1.1154 (0.1308) ?

!!! Tests for variance components not standard !!!

4th Channel Network Conference: St-Andrews, July 3-5, 2013 13

Page 15: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

Random effects predictions

4th Channel Network Conference: St-Andrews, July 3-5, 2013 14

Page 16: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

Scatterplot of random effects predictions

4th Channel Network Conference: St-Andrews, July 3-5, 2013 15

Page 17: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

• For each GP, we observe at most 7 different patient predictions.

• These correspond to the 7 possible response profiles:0 −→ 0, 0 −→ 1, 1 −→ 1, 0 −→ ·, 1 −→ ·, · −→ 0, and · −→ 1.

• The negative trends are also a side effect of the discrete nature of theoutcomes.

• Two patients, j1 and j2, treated by different GP’s, i1 and i2, with thesame response profile should get identical predicted probabilities

⇒ ai1 + bj1(i1)= ai2 + bj2(i2)

⇒ ai + bj(i) is constant

⇒ Observed non-normality not necessarily problematic

4th Channel Network Conference: St-Andrews, July 3-5, 2013 16

Page 18: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

The reverse ?

• Simulation of 1000 subjects with 5 measurements each

• Histogram of true random intercepts:

4th Channel Network Conference: St-Andrews, July 3-5, 2013 17

Page 19: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

• Histogram of predictions assuming normality:

• The normal “prior” forces the predictions to satisfy normality

4th Channel Network Conference: St-Andrews, July 3-5, 2013 18

Page 20: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

Conclusion

(Verbeke & Lesaffre 1996)

The normality assumption for random effectscannot be tested using their predictions

Other techniques needed

4th Channel Network Conference: St-Andrews, July 3-5, 2013 19

Page 21: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

Random-effects assumptions

4th Channel Network Conference: St-Andrews, July 3-5, 2013 20

Page 22: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

The general mixed model

• Conditional model for vector yi of repeated measurements:

yi|bi ∼ Fi(bi), with density fi(yi|bi),

possibly depending on unknown parameters.

• Marginal model, assuming mixing distribution G for bi:

yi with density fi(yi|G) =

∫fi(yi|b)dG(b)

• Inference based on marginal log-likelihood:

`(G) =

N∑

i=1

ln[fi(yi|G)]

4th Channel Network Conference: St-Andrews, July 3-5, 2013 21

Page 23: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

The importance of G: Literature

• Neuhaus, Hauck, and Kalbfleisch (1992)• Butler and Louis (1992)• Magder and Zeger (1996)• Verbeke and Lesaffre (1996, 1997)• Heagerty and Kurland (2001)• Zhang and Davidian (2001)• Chen, Zhang, and Davidian (2002)• Agresti, Caffo, and Ohman-Strickland (2004)• Ghidey, Lesaffre, and Eilers (2004)• Ritz (2004)• Pan and Lin (2005)• Tchetgen and Coull (2006)• Litiere, Alonso, and Molenberghs (2007, 2008)• Huang (2008)• Muthen and Asparouhov (2009)• Tsonaka, Verbeke, and Lesaffre (2009)• . . .

4th Channel Network Conference: St-Andrews, July 3-5, 2013 22

Page 24: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

The importance of G: Conclusions

• Mixed feelings: Sometimes important, sometimes not

• Some aspects of inference sensitive, some not

• Some models sensitive, some not

• Proposed solutions:

. ‘Omnibus’ goodness-of-fit testsProblem: No specific alternative

. Compare model to extended versionsProblem: Lack of software

4th Channel Network Conference: St-Andrews, July 3-5, 2013 23

Page 25: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

Our aim

To develop a simple, widely applicable, diagnostic tool tocheck the random-effects assumption in any mixed model

To develop a simple, widely applicable, diagnostictool which indicates how the random-effects

model can be improved

Focus on models with (multivariate) normal random effects,but equally well applicable to other latent variable models

4th Channel Network Conference: St-Andrews, July 3-5, 2013 24

Page 26: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

General idea

• Suppose a mixed model with specific parametric assumption for G hasbeen fitted, maximizing the marginal log-likelihood `(G)

• Question:

Can `(G) be increased considerably by replacing G byanother mixing distribution H ?

• Ideal situation:

G maximizes `(G) over all possible mixing distributions G

4th Channel Network Conference: St-Andrews, July 3-5, 2013 25

Page 27: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

The gradient function and its properties

∆(G, b) =1

N

i

fi(yi|b)

fi(yi|G)

G maximizes `(G) over all G

if and only if

∆(G, b) does not exceed 1.

Moreover, ∆(G, b) equals 1 in support points of G

4th Channel Network Conference: St-Andrews, July 3-5, 2013 26

Page 28: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

Implications for checking normality

(Verbeke & Molenberghs, 2013)

• Graphical check for goodness-of-fit:

(1) Fit mixed model under normality for G

(2) Check whether ∆(G, b) = 1N

∑i

fi(yi|b)

fi(yi|G)≡ 1

• Attention can be restricted to region I of unique modes of all fi(yi|b)

• Pointwise confidence band around ∆(G, b) possible using CLT

4th Channel Network Conference: St-Andrews, July 3-5, 2013 27

Page 29: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

Application: The onychomycosis trial

• De Backer et al (Brit. J. Dermatology, 1996), Verbeke and Molenberghs(Springer, 2000), Molenberghs and Verbeke (Springer, 2005).

• A randomized, double-blind, parallel group, multicenter longitudinalstudy for the comparison of two oral treatments (A and B) for toenaildermatophyte onychomycosis (TDO)

• 146 and 148 subjects, measured at time-points:

. During treatment: Month 0, 1, 2, 3

. After treatment: Month 6, 9, 12

• Outcome: Severity of infection (0: not severe, 1: severe)

4th Channel Network Conference: St-Andrews, July 3-5, 2013 28

Page 30: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

Observed proportions

4th Channel Network Conference: St-Andrews, July 3-5, 2013 29

Page 31: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

Logistic mixed model

yij ∼ Bernoulli(πij), logit(πij) = β0 + bi + β1Ti + β2tj + β3Titj

Normal model

Linear predictor: β0 -1.6306 (0.4345)

β1 -0.1146 (0.5852)

β2 -0.4041 (0.0460)

β3 -0.1613 (0.0718)

Mixing distribution: σ 4.0133 (0.3763)

−2` 1247.8

4th Channel Network Conference: St-Andrews, July 3-5, 2013 30

Page 32: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

Gradient function

4th Channel Network Conference: St-Andrews, July 3-5, 2013 31

Page 33: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

Logistic mixture mixed model

• Clear evidence for non-normality

• A more flexible model:

bi ∼ π1N (µ1, σ2) + π2N (µ2, σ

2) + π3N (µ3, σ2)

with π1 + π2 + π3 = 1 and π1µ1 + π2µ2 + π3µ3 = 0

• Multimodal as well as unimodal, symmetric as well as skewed

4th Channel Network Conference: St-Andrews, July 3-5, 2013 32

Page 34: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

Gradient function

4th Channel Network Conference: St-Andrews, July 3-5, 2013 33

Page 35: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

Results

Normal model Mixture model

Linear predictor: β0 -1.6306 (0.4345) -1.5160 (0.4854)

β1 -0.1146 (0.5852) 0.4479 (0.4306)

β2 -0.4041 (0.0460) -0.3992 (0.0466)

β3 -0.1613 (0.0718) -0.1562 (0.0758)

Mixing distribution: σ 4.0133 (0.3763) 0.8561 (0.1889)

µ1 -2.5617 (0.4831)

µ2 2.7744 (0.3146)

µ3 9.5282 (1.2788)

π1 0.5770 (0.0422)

π2 0.3779 (0.0426)

π3 0.0451 (0.0129)

−2` 1247.8 1219.5

4th Channel Network Conference: St-Andrews, July 3-5, 2013 34

Page 36: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

Fitted mixing distributions G

4th Channel Network Conference: St-Andrews, July 3-5, 2013 35

Page 37: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

Components in mixing distribution

• Component 1: µ1 = −2.5617, π1 = 57.70%

Patients with no severe infections at all, 163/294 = 55.44%

• Component 2: µ2 = 2.7744, π2 = 37.79%

Patients with non-constant profiles, 115/294 = 39.12%

• Component 3: µ3 = 9.5282, π3 = 4.51%

Patients with only severe infections, 16/294 = 5.44%

4th Channel Network Conference: St-Andrews, July 3-5, 2013 36

Page 38: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

Gradient function to check G

• Applicable to any type of mixed model

• Applicable to check any G (mixtures, latent class, . . . )

• No computations needed additional to the fitting of the mixed model

• Straightforward multivariate extension

• Ongoing:

. Construction of a formal test based on the gradient

. Studying operating characteristics

4th Channel Network Conference: St-Andrews, July 3-5, 2013 37

Page 39: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

Mixed models in large data sets

4th Channel Network Conference: St-Andrews, July 3-5, 2013 38

Page 40: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

Large data sets

Measurements → 1 2 3 4 n

↓ Subjects#1 • • • • • • • • • • • • • • • • • • •#2 • • • • • • • • • • • • • • • • • • •#3 • • • • • • • • • • • • • • • • • • •#4 • • • • • • • • • • • • • • • • • • •

• • • • • • • • • • • • • • • • • • •• • • • • • • • • • • • • • • • • • •• • • • • • • • • • • • • • • • • • •• • • • • • • • • • • • • • • • • • •• • • • • • • • • • • • • • • • • • •

#N • • • • • • • • • • • • • • • • • • •

N large, or n large, or both ?

4th Channel Network Conference: St-Andrews, July 3-5, 2013 39

Page 41: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

Situations leading to large data sets

• N large: Observational longitudinal data

• n large: Statistical genetics / functional data analysis

• N and n large: Large multivariate longitudinal data

4th Channel Network Conference: St-Andrews, July 3-5, 2013 40

Page 42: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

Large N

Measurements → 1 2 3 4 n

↓ Subjects#1 • • • • • • • • • • • • • • • • • • •#2 • • • • • • • • • • • • • • • • • • •#3 • • • • • • • • • • • • • • • • • • •#4 • • • • • • • • • • • • • • • • • • •

• • • • • • • • • • • • • • • • • • •• • • • • • • • • • • • • • • • • • •• • • • • • • • • • • • • • • • • • •• • • • • • • • • • • • • • • • • • •• • • • • • • • • • • • • • • • • • •

#N • • • • • • • • • • • • • • • • • • •

{

{{

=⇒ Independent sub-samples

4th Channel Network Conference: St-Andrews, July 3-5, 2013 41

Page 43: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

Large n

Measurements → 1 2 3 4 n

↓ Subjects#1 • • • • • • • • • • • • • • • • • • •#2 • • • • • • • • • • • • • • • • • • •#3 • • • • • • • • • • • • • • • • • • •#4 • • • • • • • • • • • • • • • • • • •

• • • • • • • • • • • • • • • • • • •• • • • • • • • • • • • • • • • • • •• • • • • • • • • • • • • • • • • • •• • • • • • • • • • • • • • • • • • •• • • • • • • • • • • • • • • • • • •

#N • • • • • • • • • • • • • • • • • • •︸ ︷︷ ︸ ︸ ︷︷ ︸ ︸ ︷︷ ︸

=⇒ Dependent sub-samples

4th Channel Network Conference: St-Andrews, July 3-5, 2013 42

Page 44: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

The general split sample idea

(Molenberghs, Verbeke, & Iddi 2011)

• Split sample in M sub-samples

• Analyse each sub-sample separately

• Combine results in appropriate way

• Inference follows from pseudo likelihood ideas

4th Channel Network Conference: St-Andrews, July 3-5, 2013 43

Page 45: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

Pseudo likelihood

(Arnold & Strauss 1991)

• (Log-)Likelihood:

`(Θ) =∑

i

`(yi|Θ), Θd→ N (Θ, I−1

0 )

• Pseudo (log-)likelihood:

p`(Θ) =∑

i

s

δs `(yi(s)|Θ), Θ

d→ N (Θ, I−1

0 I1I−10 )

4th Channel Network Conference: St-Andrews, July 3-5, 2013 44

Page 46: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

Example: Multivariate longitudinal data

• Threshold sound pressure levels (dB), on both ears,11 frequencies: 125 → 8000 Hz

• Observations from 603 males, with up to 15 obs./subject.

× 603

4th Channel Network Conference: St-Andrews, July 3-5, 2013 45

Page 47: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

Linear mixed models for hearing data

• Linear mixed model for one outcome:

Yi(t) = (β1 + β2 Fagei + β3 Fage2i + ai)

+ (β4 + β5 Fagei + bi) t + β6 visit1(t) + εi(t)

• Joint model:

Y1i(t) = µ1(t) + a1i + b1it + ε1i(t)

Y2i(t) = µ2(t) + a2i + b2it + ε2i(t)

...

Y22i(t) = µ22(t) + a22i + b22it + ε22i(t)

4th Channel Network Conference: St-Andrews, July 3-5, 2013 46

Page 48: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

Joint model

• Distributional assumptions:

(a1i, a2i, . . . , a22i, b1i, b2i, . . . , b22i)′ ∼ N (0, D44×44)

(ε1i(t), ε1i(t), . . . , ε1i(t))′ ∼ N (0, Σ22×22) , for all t

• Full multivariate joint model

. 44 × 44 covariance matrix for random effects

. 22 × 22 covariance matrix for error components

. 990 + 253 = 1243 covariance parameters

=⇒ Computational problems!

4th Channel Network Conference: St-Andrews, July 3-5, 2013 47

Page 49: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

Pairwise approach

(Fieuws & Verbeke 2006)

• Fit all 231 bivariate models using (RE)ML (SAS PROC MIXED):

(Y1, Y2), (Y1, Y3), . . . , (Y1, Y22), (Y2, Y3), . . . , (Y2, Y22), . . . , (Y21, Y22)

• Equivalent to maximizing pseudo (log-)likelihood:

p`(Θ) = `(Y1, Y2|Θ1,2) + `(Y1, Y3|Θ1,3) + . . . + `(Y21, Y22|Θ21,22)

• Inferences follow from pseudo likelihood theory

4th Channel Network Conference: St-Andrews, July 3-5, 2013 48

Page 50: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

Overlapping sub-samples

Measurements → 1 2 3 n

↓ Subjects#1 • • • • • • • • • • • • • • • • • • •#2 • • • • • • • • • • • • • • • • • • •#3 • • • • • • • • • • • • • • • • • • •#4 • • • • • • • • • • • • • • • • • • •

• • • • • • • • • • • • • • • • • • •• • • • • • • • • • • • • • • • • • •• • • • • • • • • • • • • • • • • • •• • • • • • • • • • • • • • • • • • •• • • • • • • • • • • • • • • • • • •

#N • • • • • • • • • • • • • • • • • • •︸ ︷︷ ︸ ︸ ︷︷ ︸ ︸ ︷︷ ︸

︸ ︷︷ ︸ ︸ ︷︷ ︸

4th Channel Network Conference: St-Andrews, July 3-5, 2013 49

Page 51: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

Hearing data: Joint tests for fixed effects

• Example: Interaction between the linear time effect and age.

• Estimates and standard errors:

χ210 = 90.4, p < 0.0001 χ2

10 = 110.9, p < 0.0001

4th Channel Network Conference: St-Andrews, July 3-5, 2013 50

Page 52: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

Hearing data: Association of evolutions

• Association between underlying random effects: D44×44 of interest

• PCA on correlation matrix of random slopes, left side:

4th Channel Network Conference: St-Andrews, July 3-5, 2013 51

Page 53: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

Overall conclusions

4th Channel Network Conference: St-Andrews, July 3-5, 2013 52

Page 54: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

• Mixed models provide flexible tools for hierarchical data:

. Unbalanced data

. Multiple levels

. Natural way to incorporate association by modeling variability

. Natural extension of ‘standard models’

. Large data sets can be handled (pseudo-likelihood)

• However:

. Parameter interpretation needs careful reflection

. Inference not always standard

. Computational issues, especially in large data sets

. Model assessment more involved

4th Channel Network Conference: St-Andrews, July 3-5, 2013 53

Page 55: Mixed models: Misconceptions, pitfalls, and many opportunities · • Misconceptions • Problems often encountered in practice ... • Some aspects of inference sensitive, some not

Thanks !4th Channel Network Conference: St-Andrews, July 3-5, 2013 54