davidr Cornell University

Transcription

1 1 NONPARAMETRIC RANDOM EFFECTS MODELS AND LIKELIHOOD RATIO TESTS Oct 11, 2002 David Ruppert Cornell University davidr (These transparencies and preprints available link to Recent Talks and Recent Papers ) Work done jointly with Ciprian Crainiceanu, Cornell University

2 2 OUTLINE Smoothing can be done using standard mixed models software because Splines can be viewed as BLUPs in mixed models This random-effects spline model extends to: Semiparametric models (allows parametric submodels) Longitudinal data nested families of curves

3 3 EXAMPLE Rick Canfield and Chuck Henderson, Jr. at Cornell are working on effects of low-level lead exposure on IQ of children. They have a mixed model but the dose-response curve should be modeled nonparametrically.

4 4 EXAMPLE CONT They asked SAS is a PROC GAMMIXED would be available someday. short answer was no Then, they found Matt Wand s work and then contacted me. Now they know that GAMMIXED GLMMIXED. SAS has GAMMIXED and does not know it!

5 5 TESTING IN THIS FRAMEWORK In principle, likelihood ratio tests (LRTs) could be used to test for effects of interest E.g., hypothesis that a curve is linear or that an effect is zero a variance component (and possibly a fixed effect) is zero allows an elegant, unified theory

6 6 TESTING CONT However, the distribution theory of LRTs is complex: the null hypothesis is on the boundary of the parameter space, so standard theory suggests chi-squared mixtures as the asympototic distribution. but standard asymptotics do not apply because of correlation for the case of one variance component, we now have asymptotics that do apply

7 % " % assume they are iid be a spline controls the amount of shrinkage or smoothing. size of! #" will be treated as random effects letting model UNIVARIATE NONPARAMETRIC REGRESSION 7

8 as a spline: & & & ' ) +* +* as another spline: model the (th subject curve model the population curve & consider the nonparametric model & is th observation on (th subject NONPARAMETRIC MODELS FOR LONGITUDINAL DATA 8

9 effects somewhat different that usual interpretation of random this assumption can be viewed as a Bayesian model %,.- assume they are iid! #" ( /= population ) will be treated as random effects will be treated as fixed effects Recall: POPULATION CURVE 9

10 ) 10 %,21 assume they are iid! #" ( 3= subject ) +* +* will also be treated as random effects this is a typical random effects assumption assume they are iid! #" 0 ) +* +* will be treated as random effects ) +* +* Recall: SUBJECT CURVES

11 4 " " %,21 subject effects are %,21 0 " ) +* +* 11 th degree polynomials no subject effects Recall: NULL HYPOTHESES OF INTEREST

12 12 RELATED WORK Brumback and Rice (1998) Zhang, Lin, Raz, and Sowers (1998) Lin and Zhang (1999) Rice and Wu (2001) See references at end.

13 5 13 < >= " % null hypothesis: ; #" % and & 6 & ( 7 8and ' model: BALANCE 1-WAY ANOVA 7 :9

14 BA C D H < J J J BA C 8 E G? KL under Then the LR test is equivalent to the F-test K. Note: The equivalent fixed effects hypothesis is This is the iid case if we take the subjects as observations (Self and Liang, 1987; Stram and Lee, 1994) FE G? IH H If with9fixed, then 14

15 N N P BA C D B C D BA C D 7 B C D U P P (Crainiceanu and Ruppert, 2002) K L where K L ; _a` VXW Y and KL ; K _a` V W Y G E G ] N K L N K L ^ R S T VXW Y Z \ and E G 8 MON K L N K L QOR S>T VXW Y[Z \ If with8fixed, then 15

16 f / g h C D 16 Probability 1-Way ANOVA: bb c ed ] B FE G ^ Number of levels ML REML 0.9

17 4 H H H i 17 These theoretical results help explain their findings. is better replaced by 7 4 H for 4 kj i. found some empirical evidence that the i mixture simulated the LRT Pinheiro and Bates (2000, p. 87)

18 18 QQ plot for χ 2 1 versus the distributions in equation (9), K=3, 5, Q0.99 = Way ANOVA: asympt. null dist of RLR, given RLRj0

19 19 QQ plot for χ 2 1 versus the distributions in equation (10), K=3, 5, Q0.99 = Way ANOVA: asympt. null dist of LR, given LRj0

20 o p m 20 < q = m alternative hypothesis: < >= m L n L n null hypothesis: l model: PENALIZED SPLINES

21 { L z ~ x z ~ y r d u 21 } with ] m v ^ w r u x r minimize penalized least squares: r s t notation:

22 % " " % 22 (Brumback, Ruppert, and Wand, 1999) w and Cov " same as BLUP in a linear mixed model with

23 n " J J J o " o 23 L or, if j, if new form of null: and % %

24 ed f / g h C D ed f / g h C D 4 o bb c ] B E G ^ i i not.5 and ] B bb c G E G ^ i ƒ not.5 Then, mean) (constant mean versus piecewise constant 20 equally spaced knots s equally spaced Example: (Crainiceanu and Ruppert, 2002) 24

25 f / g h C D 25 Probability P-splines:Bb c ed ] B E G ^ Number of knots ML REML

26 26 ORTHOGONALIZATION one can apply Gram-Schmidt to the design matrix power functions are replaced by orthogonal polynomials Plus functions are replaced by spline basis functions that are orthogonal to polynomials The asymptotics of the LRT are changed by this reparametrization

27 27 Asymptotics are essentially the same as for 1-way ANOVA with 8( # levels) = ( # knots) +1 E.g., 5 levels is like 4 knots

28 Probability Asymptotic null probabilities that log-lr is zero P-splines Orthogonalized = 1-way ANOVA Number of knots Number of levels Probability ML REML ML REML

29 29 Hypotheses: linear trend versus 20-knot linear spline Comparison of finite-sample and asymptotic quantiles Quantiles of the asymptotic distribution (n= ) q 0.66 q 0.95 q 0.99 q Quantiles of distributions n=50 n=100 n= 0.5:0.5 mixture

30 30 Comparison of LRT with other tests Reference: Crainiceanu, Ruppert, Aerts, Claeskens, and Wand (2002, in preparation) Results in next table are for testing constant mean versus general alternative piecewise constant spline, or linear spline

31 31 The comparisons are made with an increasing, concave, and periodic mean function, chosen so that good tests had power approximately 0.8

32 32 R-test is from Cantoni and Hastie (2002) F-test is as in Hastie and Tibshirani (1990) C means alternative is a piecewise constant function L means alternative is a linear spline

33 33 1 means estimate under alternative has DF one greater than under null ML means smoothing parameter under alternative is chosen by ML GCV means smoothing parameter under alternative is chosen by GCV

34 34 Test Average power Maximum power Minimum Power RLRT-C R-GCV-L R-ML-C F-ML-L R-ML-L F-ML-C F-GCV-L LRT-L F-1-C F-1-L R-1-L R-GCV-C

35 35 Conclusions Standard asymptotics are, in general, not suitable Better asymptotics for one variance component are feasible For more than one variance component, one might need to use simulation to get p-values

36 36 References Brumback, B., and Rice, J., (1998), Smoothing spline models for the analsysi of nested and crossed samples of curves, JASA, 93, Brumback, B., Ruppert, D., and Wand, M.P., (1999). Comment on Variable selection and function estimation in additive nonparametric regression using data-based prior by Shively, Kohn, and Wood, JASA, 94, Cantoni, E., and Hastie, T.J., Degrees of freedom tests for smoothing splines, Biometrika, Crainiceanu, C. M., and Ruppert, D., (2002), Asymptotic distribution of likelihood ratio tests in linear mixed models, submitted Crainiceanu, C. M., Ruppert, D., and Vogelsang, T. J., (2002). Probability that the mle of a variance component is zero with applications to likelihood ratio Tests, manuscript

37 37 Crainiceanu, C. M., Ruppert, D., Aerts, M., Claeskens, G., and Wand, M., (2002), Tests of rolynomial regression against a general alternative, in preparation. Hastie, T.J., Tibshirani, R., Generalized Additive Models, London: Chapman and Hall. Lin, X., and Zhang, D. (1999), Inference in generalized additive mixed models by using smoothing splines, JRSS-B, 61, Pinheiro, J., and Bates, D., (2000), Mixed-Effects Models in S and S-PLUS, Springer, New York. Rice, J., and Wu, C., (2001), Nonparametric mixed effects models for unequally sampled noisy curves, Biometrics, 57, Self, S., and Liang, K., (1987). Asymptotic properties of maximum likelihood estimators and likelihood ratio tests under non-standard conditions, JASA, 82,

38 38 Stram, D., Lee, J., (1994). Variance Components Testing in the Longitudinal Mixed Effects Model, Biometrics, 50, Zhang, D., Lin, X., Raz, J., and Sowers, M. (1998), Semi-parametric stochastic mixed models for longitudinal data, JASA, 93,