Joint Modeling of Longitudinal Item Response Data and Survival

Size: px

Start display at page:

Download "Joint Modeling of Longitudinal Item Response Data and Survival"

Rudolph Francis
5 years ago
Views:

1 Joint Modeling of Longitudinal Item Response Data and Survival Jean-Paul Fox University of Twente Department of Research Methodology, Measurement and Data Analysis Faculty of Behavioural Sciences Enschede, Netherlands

2 Overview Introduction Cross-classified Response Data

3 Overview Introduction Cross-classified Response Data Longitudinal Response Data Mini-Mental State Examination (MMSE)

4 Overview Introduction Cross-classified Response Data Longitudinal Response Data Mini-Mental State Examination (MMSE) Survival Analysis Joint Modeling of Latent Developmental Trajectories and Survival Joint (Response, Survival) Analysis Outcomes

5 Overview Introduction Cross-classified Response Data Longitudinal Response Data Mini-Mental State Examination (MMSE) Survival Analysis Joint Modeling of Latent Developmental Trajectories and Survival Joint (Response, Survival) Analysis Outcomes Discussion

6 Overview Introduction Cross-classified Response Data Longitudinal Response Data Mini-Mental State Examination (MMSE) Survival Analysis Joint Modeling of Latent Developmental Trajectories and Survival Joint (Response, Survival) Analysis Outcomes Discussion

7 Responses to Test Items Collection of responses on tests, i = 1,..., N persons who answer k = 1,..., K items, resulting in N K 0/1 responses Y : Y = Y 1K Y 2K..... Y N Y NK

8 Responses to Test Items Collection of responses on tests, i = 1,..., N persons who answer k = 1,..., K items, resulting in N K 0/1 responses Y : Y = Y 1K Y 2K..... Y N Y NK Develop a model to say something about the structure of this data set

9 Responses to Test Items Collection of responses on tests, i = 1,..., N persons who answer k = 1,..., K items, resulting in N K 0/1 responses Y : Y = Y 1K Y 2K..... Y N Y NK Develop a model to say something about the structure of this data set Structure: person and item effects.

10 Stage 1: Modeling Success Probabilities P (Y ik = 1 θ i, ξ k ) = F (a k θ i b k ) θ i N ( µ θ, σθ 2 )

11 Stage 1: Modeling Success Probabilities P (Y ik = 1 θ i, ξ k ) = F (a k θ i b k ) θ i N ( µ θ, σθ 2 ) Response observations k are nested within persons, random person effects (latent variable)

12 Stage 1: Modeling Success Probabilities P (Y ik = 1 θ i, ξ k ) = F (a k θ i b k ) θ i N ( µ θ, σθ 2 ) Response observations k are nested within persons, random person effects (latent variable) Response observations k are nested within items, fixed/random item effects.

13 Two-Parameter Item Response Model mq qi yik sq ak bk Individual i Item k

14 Likelihood-Model Collection of N K responses, N persons and K items

15 Likelihood-Model Collection of N K responses, N persons and K items P (Y ik = 1 θ i, a k, b k ) = { exp(d(ak θ i b k )) 1+exp(d(a k θ i b k )) Φ(a k θ i b k ) Logistic Model Probit Model

16 Likelihood-Model Collection of N K responses, N persons and K items P (Y ik = 1 θ i, a k, b k ) = { exp(d(ak θ i b k )) 1+exp(d(a k θ i b k )) Φ(a k θ i b k ) Logistic Model Probit Model [ ] p (y θ, a, b) = i where η ik = a k θ i b k k F (η ik ) y ik (1 F (η ik )) 1 y ik

17 Population Model for Item Parameters Stage 2: Prior for Item Parameters (a k, b k ) t N (µ ξ, Σ ξ ) I Ak (a k ), where the set A k = {a k R, a k > 0}

18 Population Model for Item Parameters Stage 2: Prior for Item Parameters (a k, b k ) t N (µ ξ, Σ ξ ) I Ak (a k ), where the set A k = {a k R, a k > 0} Stage 3: Hyper prior Σ ξ IW(ν, Σ 0 ) µ ξ Σ ξ N (µ 0, Σ ξ /K 0 ).

19 Population Model for Person Parameter Stage 2: Prior for Person Parameters θ i N (µ θ, σ 2 θ ). Respondents are sampled independently and identically distributed.

20 Population Model for Person Parameter Stage 2: Prior for Person Parameters θ i N (µ θ, σ 2 θ ). Respondents are sampled independently and identically distributed. Stage 3: Hyper prior σ 2 θ IG(g 1, g 2 ) µ θ σ 2 θ N (µ 0, σ 2 θ /n 0).

21 Longitudinal Item Response Data Discrete response data Y ijk : (subject i, measurement occasion j, item k)

22 Longitudinal Item Response Data Discrete response data Y ijk : (subject i, measurement occasion j, item k) Several measurement occasions j = 1,..., n i, several points in time.

23 Longitudinal Item Response Data Discrete response data Y ijk : (subject i, measurement occasion j, item k) Several measurement occasions j = 1,..., n i, several points in time. Latent Growth Modeling

24 Latent Growth Modeling Model Latent Developmental Trajectories: θ ij = β 0i + β 1i T ime ij + e ij β 0i = γ 00 + u 0i β 1i = γ 10 + u 1i

25 Latent Growth Modeling Model Latent Developmental Trajectories: θ ij = β 0i + β 1i T ime ij + e ij β 0i = γ 00 + u 0i β 1i = γ 10 + u 1i 1. Subjects not measured on the same time points across time (include all data)

26 Latent Growth Modeling Model Latent Developmental Trajectories: θ ij = β 0i + β 1i T ime ij + e ij β 0i = γ 00 + u 0i β 1i = γ 10 + u 1i 1. Subjects not measured on the same time points across time (include all data) 2. Number of observations per subject may vary

27 Latent Growth Modeling Model Latent Developmental Trajectories: θ ij = β 0i + β 1i T ime ij + e ij β 0i = γ 00 + u 0i β 1i = γ 10 + u 1i 1. Subjects not measured on the same time points across time (include all data) 2. Number of observations per subject may vary 3. Follow-up times not uniform across subjects (time a continuous variable, individualized schedule)

28 Latent Growth Modeling Model Latent Developmental Trajectories: θ ij = β 0i + β 1i T ime ij + e ij β 0i = γ 00 + u 0i β 1i = γ 10 + u 1i 1. Subjects not measured on the same time points across time (include all data) 2. Number of observations per subject may vary 3. Follow-up times not uniform across subjects (time a continuous variable, individualized schedule) 4. Handle time-invariant and time-varying covariates

29 Latent Growth Modeling Model Latent Developmental Trajectories: θ ij = β 0i + β 1i T ime ij + e ij β 0i = γ 00 + u 0i β 1i = γ 10 + u 1i 1. Subjects not measured on the same time points across time (include all data) 2. Number of observations per subject may vary 3. Follow-up times not uniform across subjects (time a continuous variable, individualized schedule) 4. Handle time-invariant and time-varying covariates 5. Estimate subject-specific change across time (average change)

30 Latent Growth Modeling Specify curvilinear individual change, e.g., polynomial individual change of any order

31 Latent Growth Modeling Specify curvilinear individual change, e.g., polynomial individual change of any order Model the covariance structure of the level-1 measurement errors explicitly.

32 Latent Growth Modeling Specify curvilinear individual change, e.g., polynomial individual change of any order Model the covariance structure of the level-1 measurement errors explicitly. Model change in several domains simultaneously.

33 Overview Introduction Cross-classified Response Data Longitudinal Response Data Mini-Mental State Examination (MMSE) Survival Analysis Joint Modeling of Latent Developmental Trajectories and Survival Joint (Response, Survival) Analysis Outcomes Discussion

34 Mini-Mental State Examination (MMSE) Data: 4016 measurements of 668 subjects (4-16 measurement occasions)

35 Mini-Mental State Examination (MMSE) Data: 4016 measurements of 668 subjects (4-16 measurement occasions) 26 MMSE (binary) items;

36 Mini-Mental State Examination (MMSE) Data: 4016 measurements of 668 subjects (4-16 measurement occasions) 26 MMSE (binary) items; What day of the week is it? (orientation)

37 Mini-Mental State Examination (MMSE) Data: 4016 measurements of 668 subjects (4-16 measurement occasions) 26 MMSE (binary) items; What day of the week is it? (orientation) pencil - What is this? (language)

38 Mini-Mental State Examination (MMSE) Data: 4016 measurements of 668 subjects (4-16 measurement occasions) 26 MMSE (binary) items; What day of the week is it? (orientation) pencil - What is this? (language) subtract 7 from 100 (attention/concentration)

39 Demographics Table: Demographic information of the study participants. Participants (N = 668) Gender Male 329 Female 339 Age start mean Average sum score <

40 MMSE Sum Scores Follow-up Time (years)

41 Mixture IRT Modeling Modeling of asymmetrical data: Define latent groups g 1 and g 2 p (θ ij Ω) = 2 π ig p (θ ij Ω g ) g=1

42 Mixture IRT Modeling Modeling of asymmetrical data: Define latent groups g 1 and g 2 p (θ ij Ω) = 2 π ig p (θ ij Ω g ) g=1 P (G i = 1 y i, θ i ) = π ni i1 j=1 p (y ij θ ij ) p (θ ij Ω 1 ) g=1,2 π ni ig j=1 p (y ij θ ij ) p (θ ij Ω g )

43 Likelihood-model M 1 Measurement Part M 1 P (Y ijk = 1 θ ij, a k, b k ) = F (a k θ ij b k )

44 Likelihood-model M 1 Measurement Part M 1 P (Y ijk = 1 θ ij, a k, b k ) = F (a k θ ij b k ) Latent Growth Part M 1 p ( θ ij γ 00, τ 2, σ 2) = π i1 φ ( µ ij,1, σ 2) + π i2 φ ( µ ij,2, σ 2) µ ij,1 = γ 00,1 + u i0,1 µ ij,2 = γ 00,2 + u i0,2 ( γ (2) < γ (1))

45 Table: MMSE: Parameter Estimates of Model M 1. Mixture MLIRT M 1 Decline Stable Mean SD Mean SD Fixed Effects γ 00 Intercept Random Effects Within-individual σθ 2 Residual variance Between-individual τ00 2 Intercept

46 Measurement Part M 2 Likelihood-model M 2 P (Y ijk = 1 θ ij, a k, b k ) = F (a k θ ij b k )

47 Likelihood-model M 2 Measurement Part M 2 P (Y ijk = 1 θ ij, a k, b k ) = F (a k θ ij b k ) Latent Growth Part M 2 p ( θ ij γ, T, σ 2) = π i1 φ ( µ ij,1, σ 2) + π i2 φ ( µ ij,2, σ 2) µ ij,g = β i0,g + β i1,g T ime ij β i0,g = γ 00,g + u i0,g β i1,g = γ 10,g + u i1,g, and u i,g N (0, T g ) with T g a diagonal matrix with elements τ00,g 2 and τ 11,g 2 for g = 1, 2.

48 Table: MMSE: Parameter estimates of Model M 2 Mixture MLIRT M 2 Decline Stable Mean SD Mean SD Fixed Effects γ 00 Intercept Time Variables γ 10 Follow-up time Random Effects Within-individual σθ 2 Residual variance Between-individual τ00 2 Intercept Follow-up time τ 2 11

49 Estimated random effects of cognitive impairment 0.2 Patients Controls Not Classified 0.0 Individual Trend Initial Score

50 Survival Time Data Time to certain (non-repeatable) events (e.g., death, response, failure time)

51 Survival Time Data Time to certain (non-repeatable) events (e.g., death, response, failure time) Persons were followed to death or (right-)censored in a study { ti t survival time v i = i c i observed (uncensored) c i t i > c i not observed (censored)

52 Distribution of Survival Times Survivor function: probability survives longer than t S(t) = P (T > t) = 1 F (t) Probability density function f(t) 0, t 0 Hazard function: conditional failure rate h(t) = f(t) 1 F (t) = f(t) S(t)

53 Right-Censored Observations Joint probability of observing data v: f(v η) = r n f(t i, η) S(c i, η) i=1 i=r+1

54 Problems in Survival Modeling Censoring: data missingness, subject does not undergo the event

55 Problems in Survival Modeling Censoring: data missingness, subject does not undergo the event Unobserved between-individual variation in the probability to experience the event

56 Problems in Survival Modeling Censoring: data missingness, subject does not undergo the event Unobserved between-individual variation in the probability to experience the event Presence of time-varying covariates (e.g., prognostic factors)

57 Identify Prognostic Factors Prognosis, course, outcome of a disease

58 Identify Prognostic Factors Prognosis, course, outcome of a disease Model probability of surviving given (possible) prognostic factors (risk factors, individual characteristics)

59 Identify Prognostic Factors Prognosis, course, outcome of a disease Model probability of surviving given (possible) prognostic factors (risk factors, individual characteristics) Popular model: Cox Proportional Hazards Model: h(t x 1 ) h(t x 2 ) = constant h(t x) = h 0 (t)g(x) = h 0 (t) exp(η t x)

60 Identify Prognostic Factors Prognosis, course, outcome of a disease Model probability of surviving given (possible) prognostic factors (risk factors, individual characteristics) Popular model: Cox Proportional Hazards Model: h(t x 1 ) h(t x 2 ) = constant h(t x) = h 0 (t)g(x) = h 0 (t) exp(η t x) Violate proportional hazards assumption when using time-varying covariates

61 Overview Introduction Cross-classified Response Data Longitudinal Response Data Mini-Mental State Examination (MMSE) Survival Analysis Joint Modeling of Latent Developmental Trajectories and Survival Joint (Response, Survival) Analysis Outcomes Discussion

62 Joint Modeling Approach Patients and controls with different (latent, time-continuous) backgrounds may have different survival prognosis

63 Joint Modeling Approach Patients and controls with different (latent, time-continuous) backgrounds may have different survival prognosis Longitudinal factor/covariates are measured infrequently and with measurement error

64 Joint Modeling Approach Patients and controls with different (latent, time-continuous) backgrounds may have different survival prognosis Longitudinal factor/covariates are measured infrequently and with measurement error Subjects enter the study at different time-points, measured at different times, different number of measurements

65 Joint Modeling: Two-stage procedure Survival information not used in modeling the covariate process.

66 Joint Modeling: Two-stage procedure Survival information not used in modeling the covariate process. New growth curves are estimated at each new event (interpretability)

67 Joint Modeling: Two-stage procedure Survival information not used in modeling the covariate process. New growth curves are estimated at each new event (interpretability) Handle measurement error in time-dependent (latent) covariate(s) survival model

68 Joint Modeling Joint distribution (survival data, response data): p (t, y x) = = = p (t, y η, x) p (η x) dη p (t η, x) p (y η) p (η Ω, x) p(ω)dηdω p (t η, Ω, x) p (y η) p (η Ω, x) p(ω)dηdω

69 Time Density Function Define v i = min(t i, c i ), D i = { 1 Event observed 0 Censored observation

70 Define v i = min(t i, c i ), Time Density Function D i = { 1 Event observed 0 Censored observation Density Function: f (v i, d i η, Ω) = h (v i η, Ω) d i S (v i η, Ω) [ vi ] = h (v i η, Ω) d i exp h (u η(u), Ω) du 0

71 Define v i = min(t i, c i ), Time Density Function D i = { 1 Event observed 0 Censored observation Density Function: f (v i, d i η, Ω) = h (v i η, Ω) d i S (v i η, Ω) [ vi ] = h (v i η, Ω) d i exp h (u η(u), Ω) du Define subject-specific time-intervals, t il t i(l+1) ; f il ( di, t i(l+1), t il η l ) = S ( t i(l+1) η l ) 1 di f ( t i(l+1) η l ) di 0 S (t l η l )

72 Time Density Function For subject i f i (d i, t i η) = = l=l i 1 l=0 l=l i 1 l=0 f il ( di, t il, t i(l+1) η l ) S ( t i(l+1) η l ) 1 di f ( t i(l+1) η l ) di S (t l η l )

73 Overview Introduction Cross-classified Response Data Longitudinal Response Data Mini-Mental State Examination (MMSE) Survival Analysis Joint Modeling of Latent Developmental Trajectories and Survival Joint (Response, Survival) Analysis Outcomes Discussion

74 Estimated Parameters Mixture Model Decline Stable EAP SD EAP SD Fixed Effects γ 00 Intercept γ 01 Time Slope Variance Components τ 2 Between Individual σ 2 Residual Mixture Proportion π

75 Results: Comparing Models Model Density Covariates Groups DIC M 1 Exponential M 2 Weibull M 3 Lognormal M 4 Exponential 1, θ M 5 Weibull 1, θ M 6 Lognormal 1, θ M 7 Weibull 1, θ, Male, Age M 8 Lognormal 1, θ, Male, Age M 9 Weibull 1, θ, Male, Age M 1 0 Lognormal 1, θ, Male, Age

76 Stratified Lognormal Survival Model M 10 Decline (g=2) Stable (g=1) EAP SD EAP SD Fixed Effects β 0,g Intercept β 1,g Male β 2,g Age (standardized) Λ g Cognitive Function Variance Components σs 2 Residual

77 1.0 Stable Decline P(Survival) Time (years)

78 P(Survival) year 2 years 3 years 4 years 5 years 0.6 Stable Fifty-Fifty Decline Cognitive Function

79 MMSE: Cognitive Function Sum score = 15 Latent variable estimate P(survival) Sum score Sum score = Time in years Sum score = 25 P(survival) P(survival) Time in years Time in years

80 Discussion Mental health (individual trajectory of cognitive impairment) serves as a time-varying covariate

81 Discussion Mental health (individual trajectory of cognitive impairment) serves as a time-varying covariate Making inferences at the level of individuals (patients) and their disease trajectories

82 Discussion Mental health (individual trajectory of cognitive impairment) serves as a time-varying covariate Making inferences at the level of individuals (patients) and their disease trajectories Combine cognitive test outcomes with other indicators to serve as a diagnostic tool

Factor Analytic Models of Clustered Multivariate Data with Informative Censoring (refer to Dunson and Perreault, 2001, Biometrics 57, )

Factor Analytic Models of Clustered Multivariate Data with Informative Censoring (refer to Dunson and Perreault, 2001, Biometrics 57, 302-308) Consider data in which multiple outcomes are collected for