Mixed Models for Longitudinal Ordinal and Nominal Data

Transcription

1 Mixed Models for Longitudinal Ordinal and Nominal Data Hedeker, D. (2008). Multilevel models for ordinal and nominal variables. In J. de Leeuw & E. Meijer (Eds.), Handbook of Multilevel Analysis. Springer, New York. Hedeker, D. (2005). Generalized linear mixed models. In B. Everitt & D. Howell (Eds.), Encyclopedia of Statistics in Behavioral Science. Wiley. Chapters 10 & 11 in Hedeker, D. & Gibbons, R.D. (2006). Longitudinal Data Analysis. Wiley. 1

2 Why analyze as ordinal? Efficiency: Armstrong & Sloan (1989, Amer Jrn of Epid) report efficiency losses between 89% to 99% comparing an ordinal to continuous outcome, depending on the number of categories and distribution within the ordinal categories. Bias: continuous model can yield correlated residuals and regressors when applied to ordinal outcomes, because the continuous model does not take into account the ceiling and floor effects of the ordinal outcome. This can result in biased estimates of regression coefficients and is most critical when the ordinal variables is highly skewed. Logic: continuous model can yield predicted values outside of the range of the ordinal variable. 2

3 Proportional Odds Model - McCullagh (1980) log P (Y c) = γ c x β 1 P (Y c) c = 1,..., C 1 for the C categories of the ordinal outcome x = vector of explanatory variables (plus the intercept) γ c = thresholds; reflect cumulative odds when x = 0 (for identification: γ 1 = 0 or β 0 = 0) positive association between x and Y is reflected by β > 0 the effect of x is assumed to be the same for each cumulative odds ratio odds that the response is greater than or equal to c (for fixed c) is multiplied by e β for every unit change in x: 1 P (Y c) P (Y c) = e γ c (e β ) x 3

4 Ordinal Model for Dichotomous Response: same as it ever was! log P (Y = 0) 1 P (Y = 0) = 0 x β P (Y = 0) 1 P (Y = 0) = exp(0 x β) 1 P (Y = 0) P (Y = 0) = [exp(0 x β)] 1 1 P (Y = 0) P (Y = 0) = exp(x β) log P (Y = 1) 1 P (Y = 1) = x β 4

5 Ordinal Response and Threshold Concept Continuous y i - unobservable latent variable - related to ordinal response Y i via threshold concept series of threshold values γ 1, γ 2,..., γ C 1 C = number of ordered categories γ 0 = and γ C = Response occurs in category c, Y i = c if γ c 1 < y i < γ c 5

6 The Threshold Concept in Practice How was your day? (what is your level of satisfaction today?) Satisfaction may be continuous, but we sometimes emit an ordinal response: 6

7 Model for Latent Continuous Responses Consider the model with p covariates for the latent response strength y i (i = 1, 2,..., N): y i = x i β + ε i probit: ε i standard normal (mean=0, variance=1) logistic: ε i standard logistic (mean=0, variance=π 2 /3) β estimates from logistic regression are larger (in abs. value) than from probit regression by approximately π 2 /3 = 1.8 Underlying latent variable useful way of thinking of the problem not an essential assumption of the model 7

8 Mixed-effects ordinal logistic regression model (Hedeker & Gibbons, 1994, 1996) i = 1,... N level-2 units (clusters or subjects) j = 1,..., n i level-1 units (subjects or repeated observations) c = 1, 2,..., C response categories Y ij = ordinal response of level-2 unit i and level-1 unit j How was your day? (asked repeatedly each day for a week) 8

9 Mixed-effects ordinal logistic regression model λ ijc = log P ijc = γ c (x (1 P ijc ) ijβ + z ijυ i ) P ijc = Pr (Y ij c υ ; γ c, β, Σ υ ) = 1 1+exp( λ ijc ) p ijc = Pr (Y ij = c υ ; γ c, β, Σ υ ) = P ijc P ijc 1 C 1 strictly increasing model thresholds γ c x ij = p 1 covariate vector z ij = r 1 design vector for random effects β = p 1 fixed regression parameters υ i = r 1 random effects for level-2 unit i N(0, Σ υ ) 9

10 Model for Latent Continuous Responses Model with p covariates for the latent response strength y ij : y ij = x ijβ + υ 0i + ε ij where υ 0i N(0, σ 2 υ), and assuming ε ij standard normal (mean 0 and σ 2 = 1) leads to mixed-effects ordinal probit regression ε ij standard logistic (mean 0 and σ 2 = π 2 /3) leads to mixed-effects ordinal logistic regression 10

11 Underlying latent variable not an essential assumption of the model useful for obtaining intra-class correlation (r) and for design effect (d) r = σ2 υ σ 2 υ + σ 2 d = σ2 υ + σ2 σ 2 = 1/(1 r) ratio of actual variance to the variance that would be obtained by simple random sampling (holding sample size constant) 11

12 Scaling of regression coefficients Fixed-effects model β estimates from logistic regression are larger (in abs. value) than from probit regression by approximately because π 2 /3 1 V (y) = σ 2 = π 2 /3 for logistic V (y) = σ 2 = 1 for probit =

13 Mixed-effects model β estimates from mixed-effects (random intercepts) model are larger (in abs. value) than from fixed-effects model by approximately d = συ 2 + σ 2 σ 2 because V (y) = σ 2 υ + σ2 in mixed-effects (random intercepts) model V (y) = σ 2 in fixed-effects model difference depends on size of random-effects variance σ 2 υ more complex for models with multiple random effects 13

14 Standardization of the random effects With υ = T θ (where T T = Σ υ is the Cholesky) λ ijc = γ c (x ijβ + z ijt θ i ) θ i are distributed multivariate standard normal P ijc = Pr (Y ij c θ ; γ c, β, T ) = 1 1+exp( λ ijc ) p ijc = Pr (Y ij = c θ ; γ c, β, T ) = P ijc P ijc 1 14

15 Conditional probability for response vector y i = n i 1 vector of responses from level-2 unit i Conditional Probability l(y i θ; γ c, β, T ) = l i = n i j=1 C c=1 (p ijc) d ijc where d ijc = 1 if Y ij = c and 0 otherwise Conditional Independence: responses from level-2 unit i are independent given θ 15

16 Maximum (Marginal) Likelihood Estimation Marginal Probability h(y i ) = h i = θ l i g(θ) dθ g(θ) = multivariate standard normal density Maximize the marginal log-likelihood from N level-2 units log L = N log h i with respect to η = [ γ c. β. T ] i Full-likelihood approach found in SAS PROC NLMIXED & GLIMMIX, STATA, SUPERMIX, MIXOR 16

17 Numerical Quadrature for integration over θ method to numerically perform an integration θ f(θ)g(θ)dθ Q q=1 f(b q)a(b q ) where B q (q = 1,..., Q) are the quadrature nodes or points A(B q ) (q = 1,..., Q) are the weights (sum = 1) the more points you use, the more accurate the approximation, but the more time it takes For standard normal distribution, Gauss-Hermite quadrature does yield a likelihood value that can be used for LR tests 17

18 Quadrature: integration over θ Standard normal distribution - tabled values of Q nodes B q and weights A(B q ) logit λ ijcq = γ c (x ij β + z ij T B q) probabilities p ijcq = exp( λ ijcq ) exp( λ ij, c 1, q ) likelihoods l(y i B q ; γ c, β, T ) = l iq = n i j=1 C c=1 (p ijcq) d ijc h(y i ) Q q=1 l iq A(B q ) 18

19 Empirical Bayes estimates (univariate case) ˆθ i = E(θ i y i ) = 1 h i θ θ i l i g(θ) dθ 1 h i Q q=1 B q l iq A(B q ) The variance of this estimator is obtained as: V (ˆθ i y i ) = 1 h i θ (θ i ˆθ i ) 2 l i g(θ) dθ 1 h i Q q=1 (B q ˆθ i ) 2 l iq A(B q ) At convergence, one more round of quadrature and the values of h i = h(y i ) which vary by i units l iq = l(y i B q ; γ c, β, σ υ ) which vary by i units and quad pts 19

20 Subject-specific probabilities estimated logits given θ i ˆλ ijc (θ i ) = ˆγ c x ij ˆβ + z ij ˆT θi estimated probabilities ˆp ijc (θ i ) = Marginal probabilities exp( ˆλ ijc (θ i )) exp( ˆλ ijc 1 (θ i )) ˆp ijc = θ ˆp ijc(θ i ) g(θ) dθ Q q=1 ˆp ijc(b q ) A(B q ) 20

21 Adaptive Quadrature adapt quadrature points and weights for each subject, and at each iteration, using EB estimates of their location ˆθ i and uncertainty s 2 i = V (ˆθ i y i ) requires fewer points to obtain accurate solution especially useful if subject random effects are very spread out (i.e., ICC is high) Adapted quadrature points and weights, from original points B q and weights A q, where φ( ) = normal pdf B iq = ˆθ i + s i B q A iq = 2π s i exp(b 2 q/2) φ(b iq ) A q 21

22 Multiple Random Effects quadrature solution must integrate over each random effect dimension (r = number of random effects) B q = (B q1, B q2,..., B qr ) = r dimension quad pt vector A(B q ) = r h=1 A(B qh) = product of univariate weights curse of dimensionality: Q r total points, where Q is the number of points per dimension (e.g., Q = 10 and r = 3 leads to evaluation at 1000 points) adaptive quadrature especially useful here, since Q can be lower 22

23 Other methods for integration of θ Methods based on first- or second-order Taylor series expansions Marginal quasi-likelihood (MQL) involves expansion around the fixed part of the model Penalized or predictive quasi-likelihood (PQL) also includes the random part in its expansion Both are available in the SAS PROC GLIMMIX and MLwiN fast, but doesn t yield a likelihood for LR tests can yield downwardly biased estimates in certain situations (if N and/or n is small, or ICC is high), especially for MQL 23

24 Laplace approximation - Raudenbush et. al., (2000) a combination of a fully multivariate Taylor series expansion and Laplace approximation fast and computationally accurate yields a likelihood for LR tests available in HLM, though not for all models Other methods Markov Chain Monte Carlo (MCMC) Bayesian approach (in BUGS) Maximum Simulated Likelihood (in some STATA programs) in econometric, transportation, political science literatures 24

25 Treatment-Related Change Across Time Data from the NIMH Schizophrenia collaborative study on treatment related changes in overall severity. IMPS item 79, Severity of Illness, was scored as: 1 = normal or borderline mentally ill 2 = mildly or moderately ill 3 = markedly ill 4 = severely or among the most extremely ill The experimental design and corresponding sample sizes: Sample size at Week Group completers PLC (n=108) % DRUG (n=329) % Drug = Chlorpromazine, Fluphenazine, or Thioridazine Main question of interest: Was there differential improvement for the drug groups relative to the control group? 25

26 26

27 27

28 Within-Subjects / Between-Subjects components Within-subjects model - level 1 (j = 1,..., n i obs) λ ijc = γ c [b 0i + b 1i W eekj ] Between-subjects model - level 2 (i = 1,..., N subjects) b 0i = β 0 + β 2 Grp i + υ 0i b 1i = β 1 + β 3 Grp i υ 0i N ID(0, σ 2 υ) 28

29 0 β 0 = γ1 = week 0 IMPS79 1st logit (1 vs 2-4) γ 2 β 0 = γ2 = week 0 IMPS79 2nd logit (1-2 vs 3-4) γ 3 β 0 = γ3 = week 0 IMPS79 3rd logit (1-3 vs 4) β 1 β 2 β 3 υ 0i = IMPS79 (sqrt) weekly logit change for PLC patients (Grp = 0) = difference in week 0 IMPS79 logit for DRUG patients (Grp = 1) = difference in IMPS79 (sqrt) weekly logit change for DRUG patients (Grp = 1) = individual deviation from group trend 29

30 NIMH Schizophrenia Study: Severity of Illness (N = 437) Ordinal Logistic ML Estimates (se) - random intercept model ML estimates se z p < threshold threshold threshold Drug (0 = plc; 1 = drug) Time (sqrt week) Drug by Time Intercept variance Intra-person correlation = 3.774/( π 2 /3) =.53 2 log L =

31 Within-Subjects / Between-Subjects components Within-subjects model - level 1 (j = 1,..., n i obs) logit cij = γ c [b 0i + b 1i W eekj ] Between-subjects model - level 2 (i = 1,..., N subjects) b 0i = β 0 + β 2 Grp i + υ 0i b 1i = β 1 + β 3 Grp i + υ 1i υ i N ID(0, Σ) 31

32 Ordinal Logistic Random Intercept and Trend Model ML estimates se z p < threshold threshold threshold Drug (0=plc; 1 = drug) Time (sqrt week) Drug by Time Intercept var Int-Time covar (r υ0 υ 1 =.40) Time var log L = , χ 2 2 = 77.3, p <

33 Model Fit of Observed Proportions 33

34 SAS parameterization of ordinal model PROC LOGISTIC, PROC GLIMMIX log P (Y c) = γ c + x β or P (Y c) = Ψ(γ c + x β) 1 P (Y c) where Ψ(z) = 1/[1 + exp( z)] here, negative association between x and Y if β > 0 use DESCENDING option to yield: log P (Y c) = γ c + x β or P (Y c) = Ψ(γ c + x β) 1 P (Y c) SAS labels γ c as Intercepts 34

35 SAS ordinal example without covariates PROC LOGISTIC DESCENDING; MODEL thkso = ; Response Profile Ordered Total Value thkso Frequency Analysis of Maximum Likelihood Estimates Standard Wald Parameter DF Estimate Error Chi-Square Pr > ChiSq Intercept <.0001 Intercept Intercept <

36 Intercept 4 = = logit of response in category 4 Intercept 3 = = logit of response in category 3 or 4 Intercept 2 = = logit of response in category 2, 3 or 4 Probability of response in category 4 = 1/ [1 + exp( ( ))] =.2794 = 447/1600 Probability of response in category 3 or 4 = 1/ [1 + exp( (0.1176))] =.5294 = ( )/1600 Probability of response in category 2, 3 or 4 = 1/ [1 + exp( (1.2548))] =.7781 = ( )/1600 category specific probabilities are obtained by subtraction Probability of response in category 3 = =.25 Intercept 2 > Intercept 3 > Intercept 4 36

37 SAS SYNTAX: SCHZONL.SAS - ordinal fixed- and mixed-effects logistic regression DATA one; INFILE c:\schizx1.dat ; INPUT id imps79 imps79b imps79o int tx week sweek txswk ; /* get rid of observations with missing values */ IF imps79 > -9; PROC FORMAT; VALUE tx 0 = placebo 1 = drug ; /* ordinal fixed-effects logistic regression */ PROC LOGISTIC DATA=one DESCENDING; MODEL imps79o = tx sweek tx*sweek; RUN; /* ordinal random-intercept logistic regression via GLIMMIX */ PROC GLIMMIX DATA=one METHOD=QUAD(QPOINTS=21) NOCLPRINT; CLASS id; MODEL imps79o(desc) = tx sweek tx*sweek / SOLUTION DIST=MULTINOMIAL LINK=CUMLOGIT; RANDOM INTERCEPT / SUBJECT=id; RUN; 37

38 /* ordinal random-intercept logistic regression via NLMIXED */ PROC NLMIXED DATA=one QPOINTS=21; PARMS b tx=0 b sweek=-.54 b txswk=-.75 varu=1 g1=-3.8 g2=-1.8 g3=-.42; z = b tx*tx + b sweek*sweek + b txswk*tx*sweek + u; IF (imps79o=1) THEN p = 1 / (1 + EXP(-(g1-z))); ELSE IF (imps79o=2) THEN p = (1/(1 + EXP(-(g2-z)))) - (1/(1 + EXP(-(g1-z)))); ELSE IF (imps79o=3) THEN p = (1/(1 + EXP(-(g3-z)))) - (1/(1 + EXP(-(g2-z)))); ELSE IF (imps79o=4) THEN p = 1 - (1 / (1 + EXP(-(g3-z)))); loglike = LOG(p); MODEL imps79o GENERAL(loglike); RANDOM u NORMAL(0,varu) SUBJECT=id; ESTIMATE icc varu/((((atan(1)*4)**2)/3)+varu); RUN; 38

39 /* ordinal random trend logistic regression via GLIMMIX */ PROC GLIMMIX DATA=one METHOD=QUAD(QPOINTS=11) NOCLPRINT; CLASS id; MODEL imps79o(desc) = tx sweek tx*sweek / SOLUTION DIST=MULTINOMIAL LINK=CUMLOGIT; RANDOM INTERCEPT sweek / SUBJECT=id TYPE=UN GCORR SOLUTION; ODS LISTING EXCLUDE SOLUTIONR; ODS OUTPUT SOLUTIONR=ebest2; RUN; /* ordinal random trend logistic regression via NLMIXED */ PROC NLMIXED DATA=one QPOINTS=11; PARMS b tx=-.1 b sweek=-.8 b txswk=-1.2 v0=3.8 c01=0 v1=1 g1=-6 g2=-3 g3=-.7; z = b tx*tx + b sweek*sweek + b txswk*tx*sweek + u0 + u1*sweek; IF (imps79o=1) THEN p = 1 / (1 + EXP(-(g1-z))); ELSE IF (imps79o=2) THEN p = (1/(1 + EXP(-(g2-z)))) - (1/(1 + EXP(-(g1-z)))); ELSE IF (imps79o=3) THEN p = (1/(1 + EXP(-(g3-z)))) - (1/(1 + EXP(-(g2-z)))); ELSE IF (imps79o=4) THEN p = 1 - (1 / (1 + EXP(-(g3-z)))); loglike = LOG(p); MODEL imps79o GENERAL(loglike); RANDOM u0 u1 NORMAL([0,0], [v0,c01,v1]) SUBJECT=id out=ebest2b; ESTIMATE re corr c01/sqrt(v0*v1); RUN; 39

40 Model fit of observed marginal proportions 1. ŷ i = X i ˆβ 2. calculate marginalization vector ŝ = 1 σ Diag( ˆV (yi )) 1/2 ˆV (y i ) = Z i ˆΣZ i + σ 2 I i σ = 1 for probit and σ = π/ 3 for logistic Z i = design matrix for random effects for random-intercepts model Z i = 1 i, and so, ŝ = ˆσ 2 υ/σ

41 3. perform element-wise division ẑ i = ŷ i /. ŝ 4. ˆp i = Φ(ẑ i ) for probit and ˆp i = Ψ(ẑ i ) for logistic 5. In practice, for logistic, (15π)/(16 3) works better than π/ 3 as σ (Zeger et al., 1988, Biometrics) 6. Logistic is approximate; relies on cumulative Gaussian approximation to the logistic function 41

42 SAS IML code: SCHZOFIT.SAS - computing marginal probabilities - ordinal models TITLE1 NIMH Schizophrenia Data - Estimated Marginal Probabilities ; PROC IML; /* Results from NLMIXED analysis: random intercept model */; x0 = { , , , }; x1 = { , , , }; varu = {3.774}; beta = {-.058, -.766, }; thresh = {-5.859, , -.709}; 42

43 /* Approximate Marginalization Method */; pi = ATAN(1)*4; nt = 4; ivec = J(nt,1,1); zvec = J(nt,1,1); evec = (15/16)**2 * (pi**2)/3 * ivec; /* nt by nt matrix with evec on the diagonal and zeros elsewhere */; emat = diag(evec); /* variance-covariance matrix of underlying latent variable */; vary = zvec * varu * T(zvec) + emat; sdy = sqrt(vecdiag(vary) / vecdiag(emat)); 43

44 za0 = (thresh[1] - x0*beta) / sdy ; zb0 = (thresh[2] - x0*beta) / sdy; zc0 = (thresh[3] - x0*beta) / sdy; za1 = (thresh[1] - x1*beta) / sdy; zb1 = (thresh[2] - x1*beta) / sdy; zc1 = (thresh[3] - x1*beta) / sdy; grp0a = 1 / ( 1 + EXP(- za0)); grp0b = 1 / ( 1 + EXP(- zb0)); grp0c = 1 / ( 1 + EXP(- zc0)); grp1a = 1 / ( 1 + EXP(- za1)); grp1b = 1 / ( 1 + EXP(- zb1)); grp1c = 1 / ( 1 + EXP(- zc1)); print Random intercept model ; print Approximate Marginalization Method ; print marginal prob for group 0 - catg 1 grp0a [FORMAT=8.4]; print marginal prob for group 0 - catg 2 (grp0b-grp0a) [FORMAT=8.4]; print marginal prob for group 0 - catg 3 (grp0c-grp0b) [FORMAT=8.4]; print marginal prob for group 0 - catg 4 (1-grp0c) [FORMAT=8.4]; print marginal prob for group 1 - catg 1 grp1a [FORMAT=8.4]; print marginal prob for group 1 - catg 2 (grp1b-grp1a) [FORMAT=8.4]; print marginal prob for group 1 - catg 3 (grp1c-grp1b) [FORMAT=8.4]; print marginal prob for group 1 - catg 4 (1-grp1c) [FORMAT=8.4]; 44

45 /* Random Intercept and Trend Model */; varu = { , }; beta = {.057, -.883, }; thresh = {-7.321, , -.813}; zmat = { , , , }; /* variance-covariance matrix of underlying latent variable */; vary = zmat * varu * T(zmat) + emat; sdy = sqrt(vecdiag(vary) / vecdiag(emat)); 45

46 za0 = (thresh[1] - x0*beta) / sdy ; zb0 = (thresh[2] - x0*beta) / sdy; zc0 = (thresh[3] - x0*beta) / sdy; za1 = (thresh[1] - x1*beta) / sdy; zb1 = (thresh[2] - x1*beta) / sdy; zc1 = (thresh[3] - x1*beta) / sdy; grp0a = 1 / ( 1 + EXP(- za0)); grp0b = 1 / ( 1 + EXP(- zb0)); grp0c = 1 / ( 1 + EXP(- zc0)); grp1a = 1 / ( 1 + EXP(- za1)); grp1b = 1 / ( 1 + EXP(- zb1)); grp1c = 1 / ( 1 + EXP(- zc1)); print Random intercept and trend model ; print Approximate Marginalization Method ; print marginal prob for group 0 - catg 1 grp0a [FORMAT=8.4]; print marginal prob for group 0 - catg 2 (grp0b-grp0a) [FORMAT=8.4]; print marginal prob for group 0 - catg 3 (grp0c-grp0b) [FORMAT=8.4]; print marginal prob for group 0 - catg 4 (1-grp0c) [FORMAT=8.4]; print marginal prob for group 1 - catg 1 grp1a [FORMAT=8.4]; print marginal prob for group 1 - catg 2 (grp1b-grp1a) [FORMAT=8.4]; print marginal prob for group 1 - catg 3 (grp1c-grp1b) [FORMAT=8.4]; print marginal prob for group 1 - catg 4 (1-grp1c) [FORMAT=8.4]; 46

47 Proportional and Non-proportional Odds Proportional Odds model log P (Y ij c) = γ c [ x 1 P (Y ij c) ijβ + z ] ijυ i with υ i N(0, T T = Σ) = γ c [ x ij β + z ij T θ ] i with θ i N(0, I) relationship between the explanatory variables and the cumulative logits does not depend on c effects of x variables DO NOT vary across the C 1 cumulative logits 47

48 Hedeker & Mermelstein (1998, Mult Behav Res) extension: log P (Y ij c) 1 P (Y ij c) = γ c(0) [ u ij γ c + x ij β + z ij T θ i u ij = h 1 vector for the set of h covariates for which proportional odds is not assumed effects of u variables DO vary across the C 1 cumulative logits more flexible model for ordinal response relations ] 48

49 Proportional Odds Assumption: covariate effects are the same across all cumulative logits Response group Absent Mild Severe total control cumulative odds = = 2.7 logit -1 1 treatment cumulative odds = = 4.6 logit group difference =.5 for both cumulative logits 49

50 log log P (Y ij = 1) = γ 1 x β 1 P (Y ij = 2 or 3) P (Y ij = 1 or 2) P (Y ij = 3) = γ 2 x β 1 γ 1 = 1, γ 2 = 1, β 1 =.5 (covariate effect is same for both cumulative logits) 50

51 Non-Proportional Odds: covariate effects vary across the cumulative logits Response group Absent Mild Severe total control cumulative odds = = 2.7 logit -1 1 treatment cumulative odds = = 7.3 logit UNEQUAL group difference across cumulative logits 51

52 log log P (Y ij = 1) = γ P (Y ij = 2 or 3) 1 (0) u γ 1 P (Y ij = 1 or 2) P (Y ij = 3) = γ 2 (0) u γ 2 γ 1 (0) = 1, γ 2(0) = 1, γ 1 =.05, γ 2 =.1 (covariate effect varies across the cumulative logits) 52

53 NIMH Schiz Study: Severity of Illness (N = 437) Ordinal LR Estimates (se) - random intercept and trend model Proportional Non-Proportional Odds Model Odds Model 1 vs 2-4 1,2 vs 3,4 1-3 vs 4 Drug (0=plc; 1 = drug) (.391) (1.235) (.510) (.433) Time (sqrt week) (.218) (.598) (.283) (.243) Drug by Time (.253) (.632) (.317) (.297) 2 log L Proportional Odds accepted (χ 2 6 = = 3.74) 53

54 San Diego Homeless Research Project (Hough) 361 mentally ill subjects who were homeless or at very high risk of becoming homeless 2 conditions: HUD Section 8 rental certificates (yes/no) baseline and 6, 12, and 24 month follow-ups Categorical outcome: housing status streets / shelters (Y = 1) community / institutions (Y = 2) independent (Y = 3) Question: Do Section 8 certificates influence housing status across time? 54

55 55

56 56

57 Housing status across time: 1289 observations within 361 subjects Ordinal Mixed Regression Model estimates and standard errors (se) Proportional Odds Non-Proportional Odds Model Non-street 1 Independent 2 term estimate se estimate se estimate se threshold threshold t1 (6 month) t2 (12 month) t3 (24 month) section 8 (y=1) section 8 by t section 8 by t section 8 by t subject variance (ICC.4) 2 log L (χ 2 7 = 52.2) bold indicates p <.05 italic indicates.05 < p <.10 1 = independent + community vs street 2 = independent vs community + street 57

58 58

59 SAS NLMIXED code: SANDONL.SAS - random-intercepts proportional odds model DATA one; INFILE c:\sdhouse.dat ; INPUT Id housing int section8 time1 time2 time3 sect8t1 sect8t2 sect8t3; /* recode missing values */ IF housing = 999 THEN housing =.; PROC NLMIXED DATA=one QPOINTS=21; PARMS btime1=1.4 btime2=1.8 btime3=2.0 bsection8=.4 bsect8t1=.9 bsect8t2=.9 bsect8t3=.4 g1=.2 g2=2.3 varu=1; z = btime1*time1 + btime2*time2 + btime3*time3 + bsection8*section8 + bsect8t1*sect8t1 + bsect8t2*sect8t2 + bsect8t3*sect8t3+ u; IF (Housing=0) THEN p = 1 / (1 + EXP(-(g1-z))); ELSE IF (Housing=1) THEN p = (1/(1 + EXP(-(g2-z)))) - (1/(1 + EXP(-(g1-z)))); ELSE IF (Housing=2) THEN p = 1 - (1 / (1 + EXP(-(g2-z)))); loglike = LOG(p); MODEL housing GENERAL(loglike); RANDOM u NORMAL(0,varu) SUBJECT=Id; ESTIMATE icc varu/((((atan(1)*4)**2)/3) + varu); RUN; 59

60 SAS NLMIXED: SANDONL.SAS - random-intercepts non-proportional odds model using data from the earlier NLMIXED example PROC NLMIXED DATA=one QPOINTS=21; PARMS g1 Time1=1.4 g1 TIme2=1.8 g1 TIme3=2.0 g1 Section8=.4 g1 Sect8T1=.9 g1 Sect8T2=.9 g1 Sect8T3=.4 g2 Time1=1.4 g2 TIme2=1.8 g2 TIme3=2.0 g2 Section8=.4 g2 Sect8T1=.9 g2 Sect8T2=.9 g2 Sect8T3=.4 g1=.2 g2=2.3 varu=1; z1 = g1 Time1*Time1 + g1 Time2*Time2 + g1 Time3*Time3 + g1 Section8*Section8 + g1 Sect8T1*Sect8T1 + g1 Sect8T2*Sect8T2 + g1 Sect8T3*Sect8T3+ u; z2 = g2 Time1*Time1 + g2 Time2*Time2 + g2 Time3*Time3 + g2 Section8*Section8 + g2 Sect8T1*Sect8T1 + g2 Sect8T2*Sect8T2 + g2 Sect8T3*Sect8T3+ u; IF (Housing=0) THEN p = 1 / (1 + EXP(-(g1-z1))); ELSE IF (Housing=1) THEN p = (1/(1 + EXP(-(g2-z2)))) - (1/(1 + EXP(-(g1-z1)))); ELSE IF (Housing=2) THEN p = 1 - (1 / (1 + EXP(-(g2-z2)))); loglike = LOG(p); MODEL Housing GENERAL(loglike); RANDOM u NORMAL(0,varu) SUBJECT=Id; ESTIMATE icc varu/((((atan(1)*4)**2)/3) + varu); RUN; 60

61 SAS IML code: SANDOFIT.SAS - computing marginal probabilities - ordinal models TITLE1 San Diego Homeless Data - Estimated Marginal Probabilities ; PROC IML; /* Results from NLMIXED analysis: proportional odds model */; x0 = { , , , }; x1 = { , , , }; varu = {2.128}; beta = {1.736, 2.315, 2.499, 0.497, 1.408, 1.173,.638}; thresh = {0.220, 2.964}; 61

62 /* number of quadrature points, quadrature nodes & weights */ nq = 10; bq = { , , , , , , , , , }; wq = { , , , , , , , , , }; /* initialize to zero */ grp0a = J(4,1,0); grp0b = J(4,1,0); grp1a = J(4,1,0); grp1b = J(4,1,0); 62

63 DO q = 1 to nq; END; za0 = thresh[1] - (x0*beta + SQRT(varu)*bq[q]); zb0 = thresh[2] - (x0*beta + SQRT(varu)*bq[q]); za1 = thresh[1] - (x1*beta + SQRT(varu)*bq[q]); zb1 = thresh[2] - (x1*beta + SQRT(varu)*bq[q]); grp0a = grp0a + ( 1 / ( 1 + EXP(- za0)))*wq[q]; grp0b = grp0b + ( 1 / ( 1 + EXP(- zb0)))*wq[q]; grp1a = grp1a + ( 1 / ( 1 + EXP(- za1)))*wq[q]; grp1b = grp1b + ( 1 / ( 1 + EXP(- zb1)))*wq[q]; print Proportional odds model ; print Quadrature method - 10 points ; print marginal prob for group 0 - catg 1 grp0a [FORMAT=8.4]; print marginal prob for group 0 - catg 2 (grp0b-grp0a) [FORMAT=8.4]; print marginal prob for group 0 - catg 3 (1-grp0b) [FORMAT=8.4]; print marginal prob for group 1 - catg 1 grp1a [FORMAT=8.4]; print marginal prob for group 1 - catg 2 (grp1b-grp1a) [FORMAT=8.4]; print marginal prob for group 1 - catg 3 (1-grp1b) [FORMAT=8.4]; 63

64 /* Non-Proportional Odds Model */; varu = {2.124}; gam1 = {2.297, 3.346, 2.823,.592,.567, -.959, -.369}; gam2 = {1.080, 1.646, 2.147,.324, 2.022, 2.016, 1.071}; thresh = {.322, 2.700}; /* initialize to zero */ grp0a = J(4,1,0); grp0b = J(4,1,0); grp1a = J(4,1,0); grp1b = J(4,1,0); 64

65 DO q = 1 to nq; END; za0 = thresh[1] - (x0*gam1 + SQRT(varu)*bq[q]); zb0 = thresh[2] - (x0*gam2 + SQRT(varu)*bq[q]); za1 = thresh[1] - (x1*gam1 + SQRT(varu)*bq[q]); zb1 = thresh[2] - (x1*gam2 + SQRT(varu)*bq[q]); grp0a = grp0a + ( 1 / ( 1 + EXP(- za0)))*wq[q]; grp0b = grp0b + ( 1 / ( 1 + EXP(- zb0)))*wq[q]; grp1a = grp1a + ( 1 / ( 1 + EXP(- za1)))*wq[q]; grp1b = grp1b + ( 1 / ( 1 + EXP(- zb1)))*wq[q]; print Non-proportional odds model ; print Quadrature method - 10 points ; print marginal prob for group 0 - catg 1 grp0a [FORMAT=8.4]; print marginal prob for group 0 - catg 2 (grp0b-grp0a) [FORMAT=8.4]; print marginal prob for group 0 - catg 3 (1-grp0b) [FORMAT=8.4]; print marginal prob for group 1 - catg 1 grp1a [FORMAT=8.4]; print marginal prob for group 1 - catg 2 (grp1b-grp1a) [FORMAT=8.4]; print marginal prob for group 1 - catg 3 (1-grp1b) [FORMAT=8.4]; 65

66 Mixed-effects Multinomial Logistic Regression Model for Nominal Responses (Hedeker, 2003) Y ij = nominal response of level-2 unit i and level-1 unit j log p ijc p ij1 = u ijγ c + z ijυ ic C 1 contrasts to reference cell (c = 1) regression effects γ c vary across contrasts (C 1) r random-effects N(0, Σ υ ) c = 2, 3,... C For example, with C = 3 contrast ordinal nominal c1 2 & 3 vs 1 2 vs 1 c2 3 vs 1 & 2 3 vs 1 66

67 Model in terms of the category probabilities p ijc = Pr(Y ij = c υ ic ) = exp(z ijc ) 1 + C h=2 exp(z ijh ) for c = 2, 3,..., C p ij1 = Pr(Y ij = 1 υ ic ) = C h=2 exp(z ijh ) where the multinomial logit z ijc = u ij γ c + z ij υ ic 67

68 68

69 Housing status across time: 1289 observations within 361 subjects Nominal Mixed Regression Model estimates & standard errors (se) Community vs Street Independent vs Street term estimate se estimate se intercept t1 (6 month) t2 (12 month) t3 (24 month) section 8 (y=1) section 8 by t section 8 by t section 8 by t subject variance log L bold indicates p <.05 italic indicates.05 < p <.10 69

70 Housing status across time: 1289 observations within 361 subjects Nominal Mixed Regression Model estimates & standard errors (se) Community vs Street Independent vs Street term estimate se estimate se intercept t1 (6 month) t2 (12 month) t3 (24 month) section 8 (y=1) section 8 by t section 8 by t section 8 by t subject variance covariance r =.73 2 log L χ 2 1 = 33.5 bold indicates p <.05 italic indicates.05 < p <.10 70

71 71

72 72

73 SAS code: SANDNNL.SAS - Nominal models with reference cell contrasts /* Nominal logistic regression via LOGISTIC */ PROC LOGISTIC DATA=one; MODEL Housing(REF=FIRST) = Time1 Time2 Time3 Section8 Sect8T1 Sect8T2 Sect8T3 / LINK=GLOGIT; /* Nominal random intercept (uncorrelated) logistic regression via GLIMMIX */ PROC GLIMMIX DATA=one METHOD=QUAD(QPOINTS=11) NOCLPRINT; CLASS Id Housing; MODEL Housing(REF=FIRST) = Time1 Time2 Time3 Section8 Sect8T1 Sect8T2 Sect8T3 / SOLUTION DIST=MULTINOMIAL LINK=GLOGIT; RANDOM INTERCEPT / SUBJECT=Id GROUP=Housing; /* Nominal random intercept (correlated) logistic regression via NLMIXED */ PROC NLMIXED DATA=one QPOINTS=11; PARMS g1 0=-.49 g1 Time1=1.63 g1 Time2=2.37 g1 Time3=1.8 g1 Section8=.43 g1 Sect8T1=-.46 g1 Sect8T2=-2.12 g1 Sect8T3=-1.1 var1=1 g2 0=-1.7 g2 Time1=1.90 g2 Time2=2.97 g2 Time3=2.9 g2 Section8=.54 g2 Sect8T1=1.13 g2 Sect8T2=-.04 g2 Sect8T3=-.07 var2=1 cov12=0; z1 = g1 0 + g1 Time1*Time1 + g1 Time2*Time2 + g1 Time3*Time3 + g1 Section8*Section8 + g1 Sect8T1*Sect8T1 + g1 Sect8T2*Sect8T2 + g1 Sect8T3*Sect8T3 + u1; z2 = g2 0 + g2 Time1*Time1 + g2 Time2*Time2 + g2 Time3*Time3 + g2 Section8*Section8 + g2 Sect8T1*Sect8T1 + g2 Sect8T2*Sect8T2 + g2 Sect8T3*Sect8T3 + u2; IF (Housing=0) THEN p = 1 / (1 + EXP(z1) + EXP(z2)); ELSE IF (Housing=1) THEN p = EXP(z1) / (1 + EXP(z1) + EXP(z2)); ELSE IF (Housing=2) THEN p = EXP(z2) / (1 + EXP(z1) + EXP(z2)); LogLike = LOG(p); MODEL Housing GENERAL(LogLike); RANDOM u1 u2 NORMAL([0,0],[var1,cov12,var2]) SUBJECT=Id; ESTIMATE RE corr cov12 / SQRT(var1*var2); RUN; 73

74 SAS IML code: SANDNFIT.SAS - computing marginal probabilities - nominal model TITLE1 San Diego Homeless Data - Estimated Marginal Probabilities ; PROC IML; /* Results from GLIMMIX analysis - uncorrelated random effects */; u0 = { , , , }; u1 = { , , , }; varu = {1.099, 3.249}; gam1 = {1.914, 2.686, 1.982,.518, -.639, , }; gam2 = {-2.600, 2.332, 3.677, 3.735,.675, 1.886,.556,.256}; 74

75 /* number of quadrature points, quadrature nodes & weights */ nq = 10; bq = { , , , , , , , , , }; wq = { , , , , , , , , , }; /* initialize to zero */ grp0a = J(4,1,0); grp0b = J(4,1,0); grp0c = J(4,1,0); grp1a = J(4,1,0); grp1b = J(4,1,0); grp1c = J(4,1,0); 75

76 DO q1 = 1 to nq; DO q2 = 1 to nq; z1 0 = u0*gam1 + SQRT(varu[1])*bq[q1]; z2 0 = u0*gam2 + SQRT(varu[2])*bq[q2]; z1 1 = u1*gam1 + SQRT(varu[1])*bq[q1]; z2 1 = u1*gam2 + SQRT(varu[2])*bq[q2]; denom0 = 1 + EXP(z1 0) + EXP(z2 0); denom1 = 1 + EXP(z1 1) + EXP(z2 1); grp0a = grp0a + (1 / denom0)*wq[q1]*wq[q2]; grp0b = grp0b + (EXP(z1 0) / denom0)*wq[q1]*wq[q2]; grp0c = grp0c + (EXP(z2 0) / denom0)*wq[q1]*wq[q2]; grp1a = grp1a + (1 / denom1)*wq[q1]*wq[q2]; grp1b = grp1b + (EXP(z1 1) / denom1)*wq[q1]*wq[q2]; grp1c = grp1c + (EXP(z2 1) / denom1)*wq[q1]*wq[q2]; END;END; print Uncorrelated random effects model fit - Quadrature method - 10 points ; print marginal prob for group 0 - category 1 grp0a [FORMAT=8.4]; print marginal prob for group 0 - category 2 grp0b [FORMAT=8.4]; print marginal prob for group 0 - category 3 grp0c [FORMAT=8.4]; print marginal prob for group 1 - category 1 grp1a [FORMAT=8.4]; print marginal prob for group 1 - category 2 grp1b [FORMAT=8.4]; print marginal prob for group 1 - category 3 grp1c [FORMAT=8.4]; 76

77 /* Results from NLMIXED analysis - correlated random effects */; varu = { , }; gam1 = {-0.622, 2.374, 3.345, 2.589, 0.652, , , }; gam2 = {-2.599, 2.888, 4.394, 4.310, 0.831, 1.969, 0.287, 0.202}; /* get the Cholesky in lower triangular form */ chol = T(ROOT(varu)); /* select the two rows of the Cholesky for the two category contrasts */ chol1 = chol[1,]; chol2 = chol[2,]; /* initialize to zero */ grp0a = J(4,1,0); grp0b = J(4,1,0); grp0c = J(4,1,0); grp1a = J(4,1,0); grp1b = J(4,1,0); grp1c = J(4,1,0); 77

78 DO q1 = 1 to nq; DO q2 = 1 to nq; bqq = bq[q1] // bq[q2]; z1 0 = u0*gam1 + chol1*bqq; z2 0 = u0*gam2 + chol2*bqq; z1 1 = u1*gam1 + chol1*bqq; z1 1 = u1*gam2 + chol2*bqq; denom0 = 1 + EXP(z1 0) + EXP(z2 0); denom1 = 1 + EXP(z1 1) + EXP(z2 1); grp0a = grp0a + (1 / denom0)*wq[q1]*wq[q2]; grp0b = grp0b + (EXP(z1 0) / denom0)*wq[q1]*wq[q2]; grp0c = grp0c + (EXP(z2 0) / denom0)*wq[q1]*wq[q2]; grp1a = grp1a + (1 / denom1)*wq[q1]*wq[q2]; grp1b = grp1b + (EXP(z1 1) / denom1)*wq[q1]*wq[q2]; grp1c = grp1c + (EXP(z2 1) / denom1)*wq[q1]*wq[q2]; END;END; print Correlated random effects model fit - Quadrature method - 10 points ; print marginal prob for group 0 - category 1 grp0a [FORMAT=8.4]; print marginal prob for group 0 - category 2 grp0b [FORMAT=8.4]; print marginal prob for group 0 - category 3 grp0c [FORMAT=8.4]; print marginal prob for group 1 - category 1 grp1a [FORMAT=8.4]; print marginal prob for group 1 - category 2 grp1b [FORMAT=8.4]; print marginal prob for group 1 - category 3 grp1c [FORMAT=8.4]; 78

79 Reference cell coding (3 categories) category z1 z2 1 housing= housing= housing=2 0 1 Helmert contrasts (3 categories) category z1 z2 1 housing=0-2/3 0 2 housing=1 1/3-1/2 3 housing=2 1/3 1/2 contrasts z1 and z2 sum to zero and represent unit differences 79

80 Housing status across time: 1289 observations within 361 subjects Nominal mixed model estimates and std errors (se) - Helmert Contrasts Ind & Comm Independent vs Street vs Community term estimate se estimate se intercept t1 (6 month vs base) t2 (12 month vs base) t3 (24 month vs base) section 8 (yes=1 no=0) section 8 by t section 8 by t section 8 by t subject variance covariance r =.50 2 log L bold indicates p <.05 italic indicates.05 < p <.10 80

81 SAS NLMIXED code: SANDNNL.SAS - random-intercepts nominal model with Helmert contrasts PROC NLMIXED DATA=one QPOINTS=11; PARMS h1 0=-.49 h1 Time1=1.63 h1 Time2=2.37 h1 Time3=1.8 h1 Section8=.43 h1 Sect8T1=-.46 h1 Sect8T2=-2.12 h1 Sect8T3=-1.1 var1=1 h2 0=-1.7 h2 Time1=1.90 h2 Time2=2.97 h2 Time3=2.9 h2 Section8=.54 h2 Sect8T1=1.13 h2 Sect8T2=-.04 h2 Sect8T3=-.07 var2=1 cov12=0; z1 = h1 0 + h1 Time1*Time1 + h1 Time2*Time2 + h1 Time3*Time3 + h1 Section8*Section8 + h1 Sect8T1*Sect8T1 + h1 Sect8T2*Sect8T2 + h1 Sect8T3*Sect8T3 + u1; z2 = h2 0 + h2 Time1*Time1 + h2 Time2*Time2 + h2 Time3*Time3 + h2 Section8*Section8 + h2 Sect8T1*Sect8T1 + h2 Sect8T2*Sect8T2 + h2 Sect8T3*Sect8T3 + u2; IF (Housing=0) THEN p = EXP(-2/3*z1) / (EXP(-2/3*z1) + EXP(1/3*z1-1/2*z2) + EXP(1/3*z1 + 1/2*z2)); ELSE IF (Housing=1) THEN p = EXP(1/3*z1-1/2*z2) / (EXP(-2/3*z1) + EXP(1/3*z1-1/2*z2) + EXP(1/3*z1 + 1/2*z2)); ELSE IF (Housing=2) THEN p = EXP(1/3*z1 + 1/2*z2) / (EXP(-2/3*z1) + EXP(1/3*z1-1/2*z2) + EXP(1/3*z1 + 1/2*z2); LogLike = LOG(p); MODEL Housing GENERAL(LogLike); RANDOM u1 u2 NORMAL([0,0],[var1,cov12,var2]) SUBJECT=Id; ESTIMATE RE corr cov12 / SQRT(var1*var2); RUN; 81

82 Summary Models for longitudinal categorical data as developed as models for continuous data Proportional odds models Non and partial proportional odds models Nominal models (with Reference-cell or Helmert contrasts) Scaling models (Hedeker, Berbaum, & Mermelstein, 2006; Hedeker, Demirtas, & Mermelstein, 2009) Available software includes SAS PROCs GLIMMIX & NLMIXED, STATA, SuperMix, MIXOR, MIXNO,... 82

83 SuperMix Free student and 15-day trial editions Datasets and examples Manual and documentation in PDF form 83