On the Behavior of Marginal and Conditional Akaike Information Criteria in Linear Mixed Models

Transcription

1 On the Behavior of Marginal and Conditional Akaike Information Criteria in Linear Mixed Models Thomas Kneib Department of Mathematics Carl von Ossietzky University Oldenburg Sonja Greven Department of Biostatistics Johns Hopkins University Lübeck,

2 Outline Outline Akaike Information Criterion Linear Mixed Models Marginal Akaike Information Criterion Conditional Akaike Information Criterion Application: Childhood Malnutrition in Nigeria On the Behavior of Marginal and Conditional Akaike Information Criteria in Linear Mixed Models 1

3 Akaike Information Criterion Akaike Information Criterion Most commonly used model choice criterion for comparing parametric models. Definition: AIC = 2l( ˆψ) + 2k. where l( ˆψ) is the log-likelihood evaluated at the maximum likelihood estimate ˆψ for the unknown parameter vector ψ and k = dim(ψ) is the number of parameters. Properties: Compromise between model fit and model complexity. Allows to compare non-nested models. Selects rather too many than too few variables in variable selection problems. On the Behavior of Marginal and Conditional Akaike Information Criteria in Linear Mixed Models 2

4 Akaike Information Criterion Data y generated from a true underlying model described in terms of density g( ). Approximate the true model by a parametric class of models f ψ ( ) = f( ; ψ). Measure the discrepancy between a model f ψ ( ) and the truth g( ) by the Kullback- Leibler distance K(f ψ, g) = [log(g(z)) log(f ψ (z))] g(z)dz = E z [log(g(z)) log(f ψ (z))]. where z is an independent replicate following the same distribution as y. Decision rule: Out of a sequence of models, choose the one that minimises K(f ψ, g). On the Behavior of Marginal and Conditional Akaike Information Criteria in Linear Mixed Models 3

5 Akaike Information Criterion In practice, the parameter ψ will have to be estimated as ˆψ(y) for the different models. To focus on average properties not depending on a specific data realisation, minimise the expected Kullback-Leibler distance E y [K(f ˆψ(y), g)] = E y[e z [ log(g(z)) log(f ˆψ(y) (z)) ]] Since g( ) does not depend on the data, this is equivalent to minimising 2 E y [E z [log(f ˆψ(y) (z))]] (1) (the expected relative Kullback-Leibler distance). On the Behavior of Marginal and Conditional Akaike Information Criteria in Linear Mixed Models 4

6 Akaike Information Criterion The best available estimate for (1) is given by 2 log(f ˆψ(y) (y)). While (1) is a predictive quantity depending on both the data y and an independent replication z, the density and the parameter estimate are evaluated for the same data. Introduce a correction term. On the Behavior of Marginal and Conditional Akaike Information Criteria in Linear Mixed Models 5

7 Akaike Information Criterion Let ψ denote the parameter vector minimising the Kullback-Leibler distance. Then AIC = 2 log(f ˆψ(y) (y)) + 2 E y[log(f ˆψ(y) (y)) log(f ψ (y))] + 2 E y [E z [log(f ψ (z)) log(f ˆψ(y) (z))]] is unbiased for (1). On the Behavior of Marginal and Conditional Akaike Information Criteria in Linear Mixed Models 6

8 Consider the regularity conditions Akaike Information Criterion ψ is a k-dimensional parameter with parameter space Ψ = R k (possibly achieved by a change of coordinates). y consists of independent and identically distributed replications y 1,..., y n. In this case, the AIC simplifies since ] 2 [log(f ˆψ(y) (y)) log(f ψ (y)) [ ] 2 E z log(f ψ (z)) log(f ˆψ(y) (z)) a χ 2 k, a χ 2 k and therefore an (asymptotically) unbiased estimate for (1) is given by AIC = 2 log(f ˆψ(y) (y)) + 2k. On the Behavior of Marginal and Conditional Akaike Information Criteria in Linear Mixed Models 7

9 Linear Mixed Models Linear Mixed Models Mixed models form a very useful class of regression models with general form y = Xβ + Zb + ε where β are usual regression coefficients while b are random effects with distributional assumption [ ] ([ [ ]) ε 0 σ N, b 0] 2 I 0. 0 D In the following, we will concentrate on mixed models with only one variance component where b N(0, τ 2 I) or b N(0, τ 2 Σ) with Σ known. On the Behavior of Marginal and Conditional Akaike Information Criteria in Linear Mixed Models 8

10 Special case I: Random intercept model for longitudinal data Linear Mixed Models y ij = x ijβ + b i + ε ij, j = 1,..., J i, i = 1,..., I, where i indexes individuals while j indexes repeated observations on the same individual. The random intercept b i accounts for shifts in the individual level of response trajectories and therefore also for intra-subject correlations. On the Behavior of Marginal and Conditional Akaike Information Criteria in Linear Mixed Models 9

11 Linear Mixed Models Special case II: Penalised spline smoothing for nonparametric function estimation y i = m(x i ) + ε i, i = 1,..., n, where m(x) is a smooth, unspecified function. Approximating m(x) in terms of a spline basis of degree d leads (for example) to the truncated power series representation m(x) = d β j x j + j=0 K b j (x κ j ) d + j=1 where κ 1,..., κ K denotes a sequence of knots. Assume random effects distribution b N(0, τ 2 I) for the basis coefficients of truncated polynomials to enforce smoothness. Works also for other basis choices (e.g. B-splines) and other types of flexible modelling components (varying coefficients, surfaces, spatial effects, etc.). On the Behavior of Marginal and Conditional Akaike Information Criteria in Linear Mixed Models 10

12 Linear Mixed Models Additive mixed models consist of a combination of random effects and flexible modelling components such as penalised splines. Example: Childhood malnutrition in Zambia. Determine the nutritional status of a child in terms of a Z-score. We consider chronic malnutrition measured in terms of insufficient height for age (stunting), i.e. zscore i = cheight i med, s where med and s are the median and standard deviation of (age-stratified) height in a reference population. On the Behavior of Marginal and Conditional Akaike Information Criteria in Linear Mixed Models 11

13 Additive mixed model for stunting: Linear Mixed Models zscore i = x iβ + m 1 (cage i ) + m 2 (cfeed i ) + m 3 (mage i ) + m 4 (mbmi i ) with covariates +m 5 (mheight i ) + b si + ε i, csex gender of the child (1 = male, 0 = female) cfeed duration of breastfeeding (in months) cage age of the child (in months) mage age of the mother (at birth, in years) mheight height of the mother (in cm) mbmi body mass index of the mother medu education of the mother (1 = no education, 2 = primary school, 3 = elementary school, 4 = higher) mwork employment status of the mother (1 = employed, 0 = unemployed) s residential district (54 districts in total) The random effect b si captures spatial variability induced by unobserved spatially varying covariates. On the Behavior of Marginal and Conditional Akaike Information Criteria in Linear Mixed Models 12

14 Marginal perspective on a mixed model: Linear Mixed Models y N(Xβ, V ) where V = σ 2 I + ZDZ Interpretation: The random effects induce a correlation structure and therefore enable a proper statistical analysis of correlated data. Conditional perspective on a mixed model: y b N(Xβ + Zb, σ 2 I). Interpretation: Random effects are additional regression coefficients (for example subject-specific effects in longitudinal data) that are estimated subject to a regularisation penalty. On the Behavior of Marginal and Conditional Akaike Information Criteria in Linear Mixed Models 13

15 Interest in the following is on the selection of random effects: Compare Linear Mixed Models M 1 : y = Xβ + Zb + ε, b N(0, τ 2 Σ) and M 2 : y = Xβ + ε. Equivalent: Compare model with random effects (τ 2 > 0) and without random effects (τ 2 = 0). Random Intercept: τ 2 > 0 versus τ 2 = 0 corresponds to the inclusion and exclusion of the random intercept and therefore to the presence or absence of intra-individual correlations. Penalised splines: τ 2 > 0 versus τ 2 = 0 differentiates between a spline model and a simple polynomial model. In particular, we can compare linear versus nonlinear models. On the Behavior of Marginal and Conditional Akaike Information Criteria in Linear Mixed Models 14

16 Akaike Information Criteria in Linear Mixed Models Akaike Information Criteria in Linear Mixed Models In linear mixed models, two variants of AIC are conceivable based on either the marginal or the conditional distribution. The marginal AIC relies on the marginal model and is defined as y N(Xβ, V ) where the marginal likelihood is given by maic = 2l(y ˆβ, ˆτ 2, ˆσ 2 ) + 2(p + 2), l(y ˆβ, ˆτ 2, ˆσ 2 ) = 1 2 log( ˆV ) 1 2 (y X ˆβ) ˆV 1 (y X ˆβ) and p = dim(β). On the Behavior of Marginal and Conditional Akaike Information Criteria in Linear Mixed Models 15

17 The conditional AIC relies on the conditional model Akaike Information Criteria in Linear Mixed Models y b N(Xβ + Zb, σ 2 I) and is defined as where caic = 2l(y ˆβ, ˆb, ˆτ 2, σ 2 ) + 2(ρ + 1), l(y ˆβ, ˆb, ˆτ 2, σ 2 ) = n 2 log(ˆσ2 ) 1 2ˆσ 2(y X ˆβ Zˆb) (y X ˆβ Zˆb) is the conditional likelihood and (( X ρ = tr X X ) 1 ( )) Z X X X Z Z X Z Z + σ 2 /τ 2 Σ Z X Z Z are the effective degrees of freedom (trace of the hat matrix). On the Behavior of Marginal and Conditional Akaike Information Criteria in Linear Mixed Models 16

18 Akaike Information Criteria in Linear Mixed Models The conditional AIC seems to be recommended when the model shall be used for predictions with the same set of random effects (for example in penalised spline smoothing). The marginal AIC is more plausible when observations with new random effects shall be predicted (e.g. new individuals in longitudinal studies). On the Behavior of Marginal and Conditional Akaike Information Criteria in Linear Mixed Models 17

19 Marginal AIC Marginal AIC Model M 1 (τ 2 > 0) is preferred over M 2 (τ 2 = 0) when maic 1 < maic 2 2l(y ˆβ 1, ˆτ 2, ˆσ 2 1) + 2(p + 2) < 2l(y ˆβ 2, 0, ˆσ 2 2) + 2(p + 1) 2l(y ˆβ 1, ˆτ 2, ˆσ 2 1) 2l(y ˆβ 2, 0, ˆσ 2 2) > 2. The left hand side is simply the test statistic for a likelihood ratio test on τ 2 = 0 versus τ 2 > 0. Under standard asymptotics, we would have 2l(y ˆβ 1, ˆτ 2, ˆσ 1) 2 2l(y ˆβ 2, 0, ˆσ 2) 2 a,h 0 χ 2 1 and the marginal AIC would have a type 1 error of P (χ 2 1 > 2) Common interpretation: AIC selects rather too many than too few effects. On the Behavior of Marginal and Conditional Akaike Information Criteria in Linear Mixed Models 18

20 Marginal AIC In contrast to the regularity conditions for likelihood ratio tests, τ 2 is on the boundary of the parameter space for model M 2. The likelihood ratio test statistic is no longer χ 2 -distributed but (approximately) follows a mixture of a point mass in zero and a scaled χ 2 1 variable. The point mass in zero corresponds to the probability that is typically larger than 50%. P (ˆτ 2 = 0) Similar difficulties appear in more complex models with several variance components when deciding on zero variances. On the Behavior of Marginal and Conditional Akaike Information Criteria in Linear Mixed Models 19

21 The classical assumptions underlying the derivation of AIC are also not fulfilled. Marginal AIC The high probability of estimating a zero variance yields a bias towards simpler models: The marginal AIC is positively biased for twice the expected relative Kullback- Leibler-Distance. The bias is dependent on the true unknown parameters in the random effects covariance matrix and this dependence does not vanish asymptotically. Compared to an unbiased criterion, the marginal AIC favors smaller models excluding random effects. This contradicts the usual intuition that the AIC picks rather too many than too few effects. On the Behavior of Marginal and Conditional Akaike Information Criteria in Linear Mixed Models 20

22 Marginal AIC Simulated example: y i = m(x) + ε where m(x) = 1 + x + 2d(0.3 x) 2. The parameter d determines the amount of nonlinearity. m(x) x On the Behavior of Marginal and Conditional Akaike Information Criteria in Linear Mixed Models 21

23 Marginal AIC ML REML selection frequency of the larger model selection frequency of the larger model nonlinearity parameter d nonlinearity parameter d On the Behavior of Marginal and Conditional Akaike Information Criteria in Linear Mixed Models 22

24 Conditional AIC Conditional AIC Vaida & Blanchard (2005) have shown that the conditional AIC from above is asymptotically unbiased for the expected relative Kullback Leibler distance for given random effects covariance matrix. Intuition: Result should carry over when using a consistent estimate. Simulation results indicate that this is not the case. On the Behavior of Marginal and Conditional Akaike Information Criteria in Linear Mixed Models 23

25 Conditional AIC ML REML selection frequency of the larger model maic caic selection frequency of the larger model maic caic nonlinearity parameter d nonlinearity parameter d On the Behavior of Marginal and Conditional Akaike Information Criteria in Linear Mixed Models 24

26 Conditional AIC Surprising result of the simulation study: The complex model including the random effect is chosen whenever ˆτ 2 > 0. If ˆτ 2 = 0, the conditional AICs of the simple and the complex model coincide (despite the additional parameters included in the complex model). The observed phenomenon can be shown to be a general property of the conditional AIC: ˆτ 2 > 0 caic(ˆτ 2 ) < caic(0) ˆτ 2 = 0 caic(ˆτ 2 ) = caic(0). Principal difficulty: The degrees of freedom in the caic are estimated from the same data as the model parameters. On the Behavior of Marginal and Conditional Akaike Information Criteria in Linear Mixed Models 25

27 Conditional AIC Liang et al. (2008) propose a corrected conditional AIC, where the degrees of freedom ρ are replaced by n ( ) ŷ i ŷ Φ 0 = = tr y i y if σ 2 is known. i=1 For unknown σ 2, they propose to replace ρ + 1 by Φ 1 = σ2 ˆσ 2 tr ( ŷ y ) + σ 2 (ŷ y) ˆσ 2 y + 1 ( 2ˆσ 2 ) 2 σ4 tr y y, where σ 2 is an estimate for the true error variance. On the Behavior of Marginal and Conditional Akaike Information Criteria in Linear Mixed Models 26

28 Conditional AIC ML REML selection frequency of the larger model maic caic corrected caic with df Φ 0 corrected caic with df Φ 1 selection frequency of the larger model maic caic corrected caic with df Φ 0 corrected caic with df Φ nonlinearity parameter d nonlinearity parameter d On the Behavior of Marginal and Conditional Akaike Information Criteria in Linear Mixed Models 27

29 The corrected conditional AIC shows satisfactory theoretical properties. Conditional AIC However, it is computationally cumbersome: Liang et al. suggested to approximate the derivatives numerically (by adding small perturbations to the data). Numerical approximations require n and 2n model fits. In our application, computing the corrected conditional AICs would take about 110 days. In addition, the numerical derivatives were found to be instable in some situations (for example the random intercept model with small cluster sizes). We have developed a closed form representation of Φ 0 that is available almost instantaneously. On the Behavior of Marginal and Conditional Akaike Information Criteria in Linear Mixed Models 28

30 Application: Childhood Malnutrition in Zambia Application: Childhood Malnutrition in Zambia Model equation: zscore i = x iβ + m 1 (cage i ) + m 2 (cfeed i ) + m 3 (mage i ) + m 4 (mbmi i ) +m 5 (mheight i ) + b si + ε i. Parametric effects in x β are not subject to model selection. 2 6 = 64 models to consider in the model comparison. On the Behavior of Marginal and Conditional Akaike Information Criteria in Linear Mixed Models 29

31 Application: Childhood Malnutrition in Zambia The eight best fitting models: ML REML cfeed cage mage mheight mbmi s caic maic caic maic Linear effects are selected for age, height and body mass index of the mother. Some nonlinearity is detected for age of the child and duration of breastfeeding. On the Behavior of Marginal and Conditional Akaike Information Criteria in Linear Mixed Models 30

32 Application: Childhood Malnutrition in Zambia REML ML linear age of the child duration of breastfeeding age of the mother height of the mother On the Behavior of Marginal and Conditional Akaike Information Criteria in Linear Mixed Models 31

33 Application: Childhood Malnutrition in Zambia Inclusion of the region-specific random effect is required to capture spatial variation On the Behavior of Marginal and Conditional Akaike Information Criteria in Linear Mixed Models 32

34 Summary Summary The marginal AIC suffers from the same theoretical difficulties as likelihood ratio tests on the boundary of the parameter space. The marginal AIC is biased towards simpler models excluding random effects. The conventional conditional AIC tends to select too many variables. Whenever a random effects variance is estimated to be positive, the corresponding effect will be included. The corrected conditional AIC rectifies this difficulty and is now available in closed form. On the Behavior of Marginal and Conditional Akaike Information Criteria in Linear Mixed Models 33

35 References: Summary Greven, S. & Kneib, T. (2009): On the Behavior of Marginal and Conditional Akaike Information Criteria in Linear Mixed Models. Technical Report. Liang, H., Wu, H. & Zou, G. (2008): A note on conditional AIC for linear mixed-effects models. Biometrika 95, Vaida, F. & Blanchard, S. (2005): Conditional Akaike information for mixed-effects models. Biometrika 92, A place called home: On the Behavior of Marginal and Conditional Akaike Information Criteria in Linear Mixed Models 34