Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects...

124
Digital Public Health Program Modeling Life Science - Dynamical Processes Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de Mathématiques de Bordeaux UMR 5251 Université de Bordeaux

Transcript of Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects...

Page 1: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Digital Public Health Program Modeling Life Science - Dynamical Processes

Nonlinear Mixed-Effects Modeling (population approach)

Fabien Crauste Chargé de Recherches CNRS Institut de Mathématiques de Bordeaux UMR 5251 Université de Bordeaux

Page 2: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

1. Motivation

2. Linear mixed-effects models

3. Nonlinear mixed-effects models

4. Identifiability

Page 3: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Motivation

Page 4: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

MotivationsMain objective: reproducing data using a mathematical model

Math ModelData

Page 5: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

MotivationsMain objective: reproducing data using a mathematical model

Math ModelData

What data?

?

?

Page 6: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

MotivationsMain objective: reproducing data using a mathematical model

Math ModelData

What data?

?

?

What model?

?

?

Page 7: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

MotivationsMain objective: reproducing data using a mathematical model

Math ModelData

What data?

?

?

What model?

?

?

How to compare them?

Page 8: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Collinson-Pautz et al. (2016) Crauste et al. (2015)C

D8+

T c

ell c

ount

s(lo

g sc

ale)

Days post-infection0 5 10 15 20 25 30 35 40 45 50

101

102

103

104

105

106

Time (days)

CD

8 T

cell

counts

Lm

Flu

VV

MotivationsLongitudinal data: mean +/- standard deviation

2 examples: Tumor growth (left) and immune cell counts (right), in various conditions ; mean and standard deviations

Page 9: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

MotivationsLongitudinal data: individual dynamics

CD

8+ T

cel

ls(lo

g sc

ale)

Days after infection

Number of CD8+ T cells in Balbc mice

mean +/- s.d. and individual dynamics

Time (hours) after injection

War

farin

Con

cent

ratio

n (m

g/L)

Concentration of warfarin in 32 healthy patients

(1.5 mg/kg for each patient)

Bazzoli et al. (2014)

Page 10: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

From Lavielle and Bleakley (2015)

1. Viral load of Hepatitis C (4 patients)

MotivationsLongitudinal data: individual dynamics

Page 11: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

From Lavielle and Bleakley (2015)

2. Weight curves (4 rats)

1. Viral load of Hepatitis C (4 patients)

MotivationsLongitudinal data: individual dynamics

Page 12: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

From Lavielle and Bleakley (2015)

2. Weight curves (4 rats)

1. Viral load of Hepatitis C (4 patients)

MotivationsLongitudinal data: individual dynamics

Page 13: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

MotivationsHow to account for inter-individual variability?

Page 14: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Option 1: Use a model to characterize each individual’s behavior-> not enough data to estimate all parameter values -> identifiability issue -> impossibility to characterize individuals’ behaviors

MotivationsHow to account for inter-individual variability?

Page 15: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Option 1: Use a model to characterize each individual’s behavior-> not enough data to estimate all parameter values -> identifiability issue -> impossibility to characterize individuals’ behaviors

Option 2: Use a Population Approach and Mixed-Effects Models-> consider individuals belong to the same population -> use statistical model -> validated in PKPD works

MotivationsHow to account for inter-individual variability?

Page 16: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Option 1: Use a model to characterize each individual’s behavior-> not enough data to estimate all parameter values -> identifiability issue -> impossibility to characterize individuals’ behaviors

Option 2: Use a Population Approach and Mixed-Effects Models-> consider individuals belong to the same population -> use statistical model -> validated in PKPD works

MotivationsHow to account for inter-individual variability?

Page 17: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

(Non)Linear (mixed-effects) models

Page 18: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

(Non)Linear (mixed-effects) modelsLinear models

Assume a linear relationship between observations and explanatory variables

Methodology (simple): - pick up explanatory variables - estimate parameters using least-square method

Page 19: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

(Non)Linear (mixed-effects) modelsLinear models

Assume a linear relationship between observations and explanatory variables

Methodology (simple): - pick up explanatory variables - estimate parameters using least-square method

But: a linear relationship is not always appropriate

Page 20: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

(Non)Linear (mixed-effects) modelsLinear models

Assume a linear relationship between observations and explanatory variables

Methodology (simple): - pick up explanatory variables - estimate parameters using least-square method

But: a linear relationship is not always appropriate

d = 1

d = 3 d = 2

Page 21: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

(Non)Linear (mixed-effects) modelsLinear models

Assume a linear relationship between observations and explanatory variables

Methodology (simple): - pick up explanatory variables - estimate parameters using least-square method

But: a linear relationship is not always appropriate

d = 1

d = 3 d = 2

Question: is weight a polynomial function of time?

Page 22: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Instead of assuming a linear relationship between observations and explanatory variables

(Non)Linear (mixed-effects) modelsNonlinear models

assume a nonlinear relationship between observations and explanatory variables:

where f is a structural model, defined by a set (vector) of parameter values and an explanatory variable .

Page 23: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Instead of assuming a linear relationship between observations and explanatory variables

(Non)Linear (mixed-effects) modelsNonlinear models

assume a nonlinear relationship between observations and explanatory variables:

where f is a structural model, defined by a set (vector) of parameter values and an explanatory variable .

In the case of rats weight, one can for instance assume

Page 24: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Instead of assuming a linear relationship between observations and explanatory variables

(Non)Linear (mixed-effects) modelsNonlinear models

assume a nonlinear relationship between observations and explanatory variables:

where f is a structural model, defined by a set (vector) of parameter values and an explanatory variable .

In the case of rats weight, one can for instance assume

Page 25: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Instead of assuming a linear relationship between observations and explanatory variables

(Non)Linear (mixed-effects) modelsNonlinear models

assume a nonlinear relationship between observations and explanatory variables:

where f is a structural model, defined by a set (vector) of parameter values and an explanatory variable .

In the case of rats weight, one can for instance assume

Page 26: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Assume a linear relationship between observations and explanatory variables for each individual:

(Non)Linear (mixed-effects) modelsLinear mixed-effects models

with

Page 27: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Assume a linear relationship between observations and explanatory variables for each individual:

(Non)Linear (mixed-effects) modelsLinear mixed-effects models

with

fixed effect

Page 28: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Assume a linear relationship between observations and explanatory variables for each individual:

(Non)Linear (mixed-effects) modelsLinear mixed-effects models

with

fixed effect random effect

Page 29: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Assume a linear relationship between observations and explanatory variables for each individual:

(Non)Linear (mixed-effects) modelsLinear mixed-effects models

with

fixed effect random effect

Random effects and residuals are assumed to be random variables, with mean = 0 and a variance to be estimated. Assuming they are normally distributed, then:

Then one needs relevant and appropriate methods to estimate parameter values… (I’ll come back to that)

Page 30: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Assume a linear relationship between observations and explanatory variables for each individual:

(Non)Linear (mixed-effects) modelsLinear mixed-effects models

with

fixed effect random effect

Random effects and residuals are assumed to be random variables, with mean = 0 and a variance to be estimated. Assuming they are normally distributed, then:

Then one needs relevant and appropriate methods to estimate parameter values… (I’ll come back to that)

Page 31: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Nonlinear mixed-effects models

Page 32: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

- observed data of individual i

- explanatory variables

- vector of parameters, with

- the structural model, here a function of explanatory variables and a vector of parameter values

- the error model,

Nonlinear mixed-effects modelsStructural and error models

From the previous considerations, we can then assume

where

[constant]

[proportional]

[combined]

Page 33: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

- observed data of individual i

- explanatory variables

- vector of parameters, with

- the structural model, here a function of explanatory variables and a vector of parameter values

- the error model,

Nonlinear mixed-effects modelsStructural and error models

From the previous considerations, we can then assume

where

[constant]

[proportional]

[combined]

Page 34: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

- observed data of individual i

- explanatory variables

- vector of parameters, with

- the structural model, here a function of explanatory variables and a vector of parameter values

- the error model,

Nonlinear mixed-effects modelsStructural and error models

From the previous considerations, we can then assume

where

[constant]

[proportional]

[combined]

Page 35: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

- observed data of individual i

- explanatory variables

- vector of parameters, with

- the structural model, here a function of explanatory variables and a vector of parameter values

- the error model,

Nonlinear mixed-effects modelsStructural and error models

From the previous considerations, we can then assume

where

[constant]

[proportional]

[combined]

Page 36: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

- observed data of individual i

- explanatory variables

- vector of parameters, with

- the structural model, here a function of explanatory variables and a vector of parameter values

- the error model,

Nonlinear mixed-effects modelsStructural and error models

From the previous considerations, we can then assume

where

[constant]

[proportional]

[combined]

Page 37: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

- observed data of individual i

- explanatory variables

- vector of parameters, with

- the structural model, here a function of explanatory variables and a vector of parameter values

- the error model,

Nonlinear mixed-effects modelsStructural and error models

From the previous considerations, we can then assume

where

[constant]

[proportional]

[combined]

Page 38: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

- observed data of individual i

- explanatory variables

- vector of parameters, with

- the structural model, here a function of explanatory variables and a vector of parameter values

- the error model,

Nonlinear mixed-effects modelsStructural and error models

From the previous considerations, we can then assume

where

[constant]

[proportional]

[combined]

Page 39: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

- observed data of individual i

- explanatory variables

- vector of parameters, with

- the structural model, here a function of explanatory variables and a vector of parameter values

- the error model,

Nonlinear mixed-effects modelsStructural and error models

From the previous considerations, we can then assume

where

[constant]

[proportional]

[combined]

Page 40: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

- observed data of individual i

- explanatory variables

- vector of parameters, with

- the structural model, here a function of explanatory variables and a vector of parameter values

- the error model,

Nonlinear mixed-effects modelsStructural and error models

From the previous considerations, we can then assume

where

[constant]

[proportional]

[combined]

Page 41: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

- observed data of individual i

- explanatory variables

- vector of parameters, with

- the structural model, here a function of explanatory variables and a vector of parameter values

- the error model,

Nonlinear mixed-effects modelsStructural and error models

From the previous considerations, we can then assume

where

[constant]

[proportional]

[combined]

Page 42: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

- observed data of individual i

- explanatory variables

- vector of parameters, with

- the structural model, here a function of explanatory variables and a vector of parameter values

- the error model,

Nonlinear mixed-effects modelsStructural and error models

From the previous considerations, we can then assume

where

[constant]

[proportional]

[combined]

Page 43: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Nonlinear mixed-effects modelsStructural and error models: comments

Comments

Page 44: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Nonlinear mixed-effects modelsStructural and error models: comments

Comments

- observed data of individual i-> the number of observations may vary from one individual to another

Page 45: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Nonlinear mixed-effects modelsStructural and error models: comments

Comments

- observed data of individual i-> the number of observations may vary from one individual to another

- explanatory variables-> time is often an explanatory variable-> solutions of a dynamical system may also be explanatory variables

Page 46: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Nonlinear mixed-effects modelsStructural and error models: comments

Comments

- observed data of individual i-> the number of observations may vary from one individual to another

- explanatory variables-> time is often an explanatory variable-> solutions of a dynamical system may also be explanatory variables

- vector of parameters, with -> the law of random effects has to be chosen-> it may depend on one’s knowledge or on classical laws

Page 47: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Nonlinear mixed-effects modelsStructural and error models: comments

Comments

- observed data of individual i-> the number of observations may vary from one individual to another

- explanatory variables-> time is often an explanatory variable-> solutions of a dynamical system may also be explanatory variables

- vector of parameters, with -> the law of random effects has to be chosen-> it may depend on one’s knowledge or on classical laws

- the error model,

-> often, a constant error is used, -> then

Page 48: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

How does the drug circulate in the body?

Warfarin injections in 32 patients

Nonlinear mixed-effects modelsAn example in PharmacoKinetics-PharmacoDynamics (PK-PD)

Time (hours) after injectionW

arfa

rin C

once

ntra

tion

(mg/

L)

How does it affect the patient?

PCA

(%)

Time (hours) after injectionBazzoli et al. (2014)

Page 49: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Consider a simple model (1 compartment)

Drug Main Compartment

dose F

absorption ka clearance kevolume V

Warfarin injections in 32 patients PK model

Concentration: , where

Nonlinear mixed-effects modelsAn example in PharmacoKinetics-PharmacoDynamics (PK-PD)

Bazzoli et al. (2014)

Page 50: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Consider a classical model: direct response, with inhibitory effect

Effect(Y )PCA synthesis

Rindegradation

kout

function(C(t))

Warfarin injections in 32 patients PD model

Effect of the drug on PCA

Nonlinear mixed-effects modelsAn example in PharmacoKinetics-PharmacoDynamics (PK-PD)

Bazzoli et al. (2014)

Page 51: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Warfarin injections in 32 patients PK-PD model

Effect of the drug on PCA:

Drug concentration:

2 quantities, 9 parameters, and non-identifiability (see for instance F and V)

Nonlinear mixed-effects modelsAn example in PharmacoKinetics-PharmacoDynamics (PK-PD)

Bazzoli et al. (2014)

Page 52: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Warfarin injections in 32 patients PK-PD model

The observation model is

Nonlinear mixed-effects modelsAn example in PharmacoKinetics-PharmacoDynamics (PK-PD)

Bazzoli et al. (2014)

Page 53: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

The observation model is

2 types of data (PK and PD)

Nonlinear mixed-effects modelsAn example in PharmacoKinetics-PharmacoDynamics (PK-PD)

Bazzoli et al. (2014)

Warfarin injections in 32 patients PK-PD model

Page 54: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

The observation model is

structural models

Nonlinear mixed-effects modelsAn example in PharmacoKinetics-PharmacoDynamics (PK-PD)

Bazzoli et al. (2014)

Warfarin injections in 32 patients PK-PD model

Page 55: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

The observation model is

explanatory variable (time)

Nonlinear mixed-effects modelsAn example in PharmacoKinetics-PharmacoDynamics (PK-PD)

Bazzoli et al. (2014)

Warfarin injections in 32 patients PK-PD model

Page 56: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

The observation model is

individual-dependent parameters

Nonlinear mixed-effects modelsAn example in PharmacoKinetics-PharmacoDynamics (PK-PD)

Bazzoli et al. (2014)

Warfarin injections in 32 patients PK-PD model

Page 57: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

The observation model is

One parameter fixed within the population

Nonlinear mixed-effects modelsAn example in PharmacoKinetics-PharmacoDynamics (PK-PD)

Bazzoli et al. (2014)

Warfarin injections in 32 patients PK-PD model

Page 58: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

The observation model is

combined error model

Nonlinear mixed-effects modelsAn example in PharmacoKinetics-PharmacoDynamics (PK-PD)

Bazzoli et al. (2014)

Warfarin injections in 32 patients PK-PD model

Page 59: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

The observation model is

constant error model

Nonlinear mixed-effects modelsAn example in PharmacoKinetics-PharmacoDynamics (PK-PD)

Bazzoli et al. (2014)

Warfarin injections in 32 patients PK-PD model

Page 60: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Nonlinear mixed-effects modelsCovariates

Covariates (age, weight, group…) allow to account for specificities in the population that could explain, at least partially, the data:

Covariates are a priori individual-dependent.

Page 61: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Nonlinear mixed-effects modelsCovariates

Covariates (age, weight, group…) allow to account for specificities in the population that could explain, at least partially, the data:

Covariates are a priori individual-dependent.

Example: weight curves of rats

Page 62: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Nonlinear mixed-effects modelsCovariates

Covariates (age, weight, group…) allow to account for specificities in the population that could explain, at least partially, the data:

Covariates are a priori individual-dependent.

Example: weight curves of rats

males

Page 63: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Nonlinear mixed-effects modelsCovariates

Covariates (age, weight, group…) allow to account for specificities in the population that could explain, at least partially, the data:

Covariates are a priori individual-dependent.

Example: weight curves of rats

males females

Page 64: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Nonlinear mixed-effects modelsCovariates on the PK-PD example

Covariates (age, weight, group…) allow to account for specificities in the population that could explain, at least partially, the data:

Covariates are a priori individual-dependent.

covariate on all parameters?

Bazzoli et al. (2014)

Page 65: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Categorical covariates (age, weight, group…) allow to account for specificities in the population that could explain, at least partially, the data:

Covariates are a priori individual-dependent.

covariate on all parameters?(Wald test)

log(individual weights)

log(

k a)

log(

V)

log(

k e)

gender

Nonlinear mixed-effects modelsCovariates on the PK-PD example

Bazzoli et al. (2014)

Page 66: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Categorical covariates (age, weight, group…) allow to account for specificities in the population that could explain, at least partially, the data:

Covariates are a priori individual-dependent.

covariate on all parameters?(Wald test)

log(individual weights)

log(

k a)

log(

V)

log(

k e)

gender

Nonlinear mixed-effects modelsCovariates on the PK-PD example

Bazzoli et al. (2014)

Page 67: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Categorical covariates (age, weight, group…) allow to account for specificities in the population that could explain, at least partially, the data:

Covariates are a priori individual-dependent.

covariate on all parameters?(Wald test)

log(individual weights)

log(

k a)

log(

V)

log(

k e)

gender

Nonlinear mixed-effects modelsCovariates on the PK-PD example

Bazzoli et al. (2014)

Page 68: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Categorical covariates (age, weight, group…) allow to account for specificities in the population that could explain, at least partially, the data:

Covariates are a priori individual-dependent.

covariate on all parameters?Wald test

Nonlinear mixed-effects modelsCovariates on the PK-PD example

Bazzoli et al. (2014)

Page 69: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Nonlinear mixed-effects modelsEstimating parameter values

This approach aims at modeling the behavior of a population of individuals, hence the objective is to characterize the population behavior. This is done by estimating population parameters, using data of individual dynamics.

Statistical methods used to estimate parameter values are based on specific probability distributions: we denote by the set of population parameters (fixed effects, error parameters…) and by the likelihood of observations y with respect to

Page 70: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Nonlinear mixed-effects modelsEstimating parameter values

This approach aims at modeling the behavior of a population of individuals, hence the objective is to characterize the population behavior. This is done by estimating population parameters, using data of individual dynamics.

Statistical methods used to estimate parameter values are based on specific probability distributions: we denote by the set of population parameters (fixed effects, error parameters…) and by the likelihood of observations y with respect to

Frequentist approach: maximize the likelihood of observations with respect to parameters

Page 71: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Nonlinear mixed-effects modelsEstimating parameter values

This approach aims at modeling the behavior of a population of individuals, hence the objective is to characterize the population behavior. This is done by estimating population parameters, using data of individual dynamics.

Statistical methods used to estimate parameter values are based on specific probability distributions: we denote by the set of population parameters (fixed effects, error parameters…) and by the likelihood of observations y with respect to

Frequentist approach: maximize the likelihood of observations with respect to parameters

Bayesian approach: maximize the conditional probability of parameters with respect to observations

Page 72: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Nonlinear mixed-effects modelsEstimating parameter values

This approach aims at modeling the behavior of a population of individuals, hence the objective is to characterize the population behavior. This is done by estimating population parameters, using data of individual dynamics.

Statistical methods used to estimate parameter values are based on specific probability distributions: we denote by the set of population parameters (fixed effects, error parameters…) and by the likelihood of observations y with respect to

Frequentist approach: maximize the likelihood of observations with respect to parameters

Bayesian approach: maximize the conditional probability of parameters with respect to observations

Page 73: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Nonlinear mixed-effects modelsEstimating parameter values: likelihood maximization

- Model linearization: use an order 1 development of f, compute the integral, and obtain a ‘linearized’ likelihood

- Numerical integration: use a numerical integration of the integral, computationally very expensive

- Stochastic integration: use Monte-Carlo simulations instead of numerical integration of the integral, computationally very expensive

The first question is: how to compute the likelihood?

Page 74: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Nonlinear mixed-effects modelsEstimating parameter values: likelihood maximization

The second question is: how to estimate population parameters?

- Standard minimization algorithms: like the Newton-Ralphson algorithm…

- Specific algorithms: SAEM (stochastic approximation expectation maximization), a modified version of the EM algorithm, is adapted to NMEM, it:

๏ first computes the conditional expectancy of the likelihood, using a stochastic approximation,

๏ then maximizes this quantity with respect to the parameters.

Page 75: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Once population parameters have been estimated, individual parameters are estimated by computing the conditional distribution of

Nonlinear mixed-effects modelsEstimating parameter values: individual parameters

In the frequentist approach, one can focus on the maximum a posteriori (MAP) estimator, given by

and the conditional mean of is the mean of the conditional distribution of

Consequence: the more information the better for the estimation, so individuals with few information may tend to select the population parameter as an individual parameter…

Page 76: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Nonlinear mixed-effects modelsEstimating parameter values: individual parameters & shrinkage

We made the assumption

and when estimating parameter values we obtained values that should follow this law ; hence, we can compute the

Bazzoli et al. (2014)

Page 77: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Nonlinear mixed-effects modelsEstimating parameter values: individual parameters & shrinkage

We made the assumption

and when estimating parameter values we obtained values that should follow this law ; hence, we can compute the

Bazzoli et al. (2014)

PKPD(warfarin)example

75% 10%

Page 78: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Nonlinear mixed-effects modelsEstimating parameter values: individual parameters & shrinkage

We made the assumption

and when estimating parameter values we obtained values that should follow this law ; hence, we can compute the

Are probability laws correct?

Are data informative

enough?

Bazzoli et al. (2014)

PKPD(warfarin)example

75% 10%

Page 79: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Nonlinear mixed-effects modelsValidating model & assumptions

Page 80: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Nonlinear mixed-effects modelsValidating model & assumptions

The whole method is based on a set of assumptions/hypotheses:

Page 81: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Nonlinear mixed-effects modelsValidating model & assumptions

The whole method is based on a set of assumptions/hypotheses: The whole method is based on a set of assumptions/hypotheses:

- Structural model: which formalism? which interactions? how many parameters?

Page 82: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Nonlinear mixed-effects modelsValidating model & assumptions

The whole method is based on a set of assumptions/hypotheses: The whole method is based on a set of assumptions/hypotheses:

- Structural model: which formalism? which interactions? how many parameters?

The whole method is based on a set of assumptions/hypotheses:

- Structural model: which formalism? which interactions? how many parameters?

- Parameters: fixed effects? random effects? which probability law for random effects?

Page 83: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Nonlinear mixed-effects modelsValidating model & assumptions

The whole method is based on a set of assumptions/hypotheses: The whole method is based on a set of assumptions/hypotheses:

- Structural model: which formalism? which interactions? how many parameters?

The whole method is based on a set of assumptions/hypotheses:

- Structural model: which formalism? which interactions? how many parameters?

- Parameters: fixed effects? random effects? which probability law for random effects?

The whole method is based on a set of assumptions/hypotheses:

- Structural model: which formalism? which interactions? how many parameters?

- Parameters: fixed effects? random effects? which probability law for random effects?

- Error model: constant? proportional? combined? something else?

Page 84: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Nonlinear mixed-effects modelsValidating model & assumptions

The whole method is based on a set of assumptions/hypotheses: The whole method is based on a set of assumptions/hypotheses:

- Structural model: which formalism? which interactions? how many parameters?

The whole method is based on a set of assumptions/hypotheses:

- Structural model: which formalism? which interactions? how many parameters?

- Parameters: fixed effects? random effects? which probability law for random effects?

The whole method is based on a set of assumptions/hypotheses:

- Structural model: which formalism? which interactions? how many parameters?

- Parameters: fixed effects? random effects? which probability law for random effects?

- Error model: constant? proportional? combined? something else?

The whole method is based on a set of assumptions/hypotheses:

- Structural model: which formalism? which interactions? how many parameters?

- Parameters: fixed effects? random effects? which probability law for random effects?

- Error model: constant? proportional? combined? something else?

- Covariates: meaningful? meaningless?

Page 85: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Nonlinear mixed-effects modelsValidating model & assumptions

The whole method is based on a set of assumptions/hypotheses: The whole method is based on a set of assumptions/hypotheses:

- Structural model: which formalism? which interactions? how many parameters?

The whole method is based on a set of assumptions/hypotheses:

- Structural model: which formalism? which interactions? how many parameters?

- Parameters: fixed effects? random effects? which probability law for random effects?

The whole method is based on a set of assumptions/hypotheses:

- Structural model: which formalism? which interactions? how many parameters?

- Parameters: fixed effects? random effects? which probability law for random effects?

- Error model: constant? proportional? combined? something else?

The whole method is based on a set of assumptions/hypotheses:

- Structural model: which formalism? which interactions? how many parameters?

- Parameters: fixed effects? random effects? which probability law for random effects?

- Error model: constant? proportional? combined? something else?

- Covariates: meaningful? meaningless?

Need for diagnosis tools

Page 86: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Nonlinear mixed-effects modelsValidating model & assumptions: Residuals

Estimated values versus Theoretical values

- k

- Other ‘shrinkage’ measure:

Individual Weighted Residuals (IWRES) should be normally distributed, with mean 0 and variance 1, so the should be close to 0.

If not, then this may indicate that data are scarce or that too many parameters are used

Page 87: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Nonlinear mixed-effects modelsValidating model & assumptions: Distribution errors

Estimated values versus Theoretical values

Normalized prediction distribution

errors (npde)

[normally distributed]

PK-PD(warfarin)example

npde

npde

time (hours) Predicted population

PK

PD

Bazzoli et al. (2014)

Page 88: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Nonlinear mixed-effects modelsValidating model & assumptions: Visual Predictive Check

Estimated values versus Theoretical values

Visual predictive check (VPC)

[percentiles, based on simulations]

Time (hours) after injection

War

farin

Con

cent

ratio

n (m

g/L)

PK

PCA

(%)

Time (hours) after injection

PD

PK-PD(warfarin)example

Bazzoli et al. (2014)

Page 89: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Nonlinear mixed-effects modelsValidating model & assumptions: Observations vs Predictions

Estimated values versus Theoretical values

Observations versus

Predictions

PD

Population predictions Individual predictions

Obs

erve

d PC

A (%

)

PK

Obs

erve

d C

once

ntra

tion

PK-PD(warfarin)example

Bazzoli et al. (2014)

Page 90: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Nonlinear mixed-effects modelsTo conclude: software

Software Algorithms for estimating parameter values

Monolix SAEM

R nlme, nlmixr, saemix (SAEM), lme (Laplacian approximation)

NONMEM several (Laplacian approximation, SAEM, first-order method…) [but PK modeling only]

SAS NLMIXED (adaptive gaussian quadrature, first-order method)

WinBugs Bayesian algorithm

Page 91: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Identifiability

Page 92: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

IdentifiabiltyWhat is ‘identifiability’?

Main equation:

Initial condition equation:

Observation:

- given inputs - set of parameters (values to be determined) - solution at time t, with the parameters

,

Page 93: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

IdentifiabiltyWhat is ‘identifiability’?

Main equation:

Initial condition equation:

Observation:

- given inputs - set of parameters (values to be determined) - solution at time t, with the parameters

,

Definition. Let be given. The model is identifiable in if for all parameters such that outputs are equal then The model is said to be identifiable if it is identifiable in every parameter

Page 94: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

Methods

IdentifiabiltyTheoretical approaches

Linear models Nonlinear models

Laplace transform Taylor series expansion

Taylor series expansions (of the output) Algebro-Differential elimination

State isomorphism theorem

We will consider the case when the model is either an algebraic equation or a system of ODEs

Page 95: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

IdentifiabiltyTheoretical approaches: an illustrative example

Hemodynamics model (C. Audebert’s PhD thesis)

Model

Page 96: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

IdentifiabiltyTheoretical approaches: an illustrative example

Hemodynamics model (C. Audebert’s PhD thesis)

Model

Time (s)

Obs

erva

tion

(ml/s

)

Observation Qin

Page 97: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

IdentifiabiltyTheoretical approaches: an illustrative example

Hemodynamics model (C. Audebert’s PhD thesis)

Model

Time (s)

Obs

erva

tion

(ml/s

)

Observation QinEstimate

Page 98: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

IdentifiabiltyTheoretical approaches: an illustrative example

Model

, with

Page 99: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

IdentifiabiltyTheoretical approaches: an illustrative example

using Laplace transform,

so

Hence parameters are identifiable if

Page 100: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

IdentifiabiltyTheoretical approaches: an illustrative example

Page 101: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

IdentifiabiltyTheoretical approaches: an illustrative example

Page 102: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

IdentifiabiltyTheoretical approaches: an illustrative example

Page 103: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

IdentifiabiltyTheoretical approaches: an illustrative example

Page 104: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

IdentifiabiltyTheoretical approaches: an illustrative example

Page 105: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

IdentifiabiltyTheoretical approaches: an illustrative example

Page 106: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

IdentifiabiltyTheoretical approaches: an illustrative example

Page 107: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

IdentifiabiltyTheoretical approaches: an illustrative example

Page 108: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

IdentifiabiltyTheoretical approaches

Yet, most of the time it is not possible to prove identifiability

and data are not always ‘ideal’

Page 109: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

IdentifiabiltyStatistical approaches

Raue et al. (2009)

‘Given a model that sufficiently describes the measured data, it is important to infer how well model parameters are

determined by the amount and quality of experimental data’ (Raue et al., 2009)

Page 110: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

IdentifiabiltyStatistical approaches

Raue et al. (2009)

Estimate parameters using likelihood maximizationParameter estimation

Page 111: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

IdentifiabiltyStatistical approaches

Raue et al. (2009)

Estimate parameters using likelihood maximizationParameter estimation

Define confidence intervals of the estimate of a parameter

where is the Hessian matrix, and is the covariance matrix of

the parameter estimates

Confidence intervals

Page 112: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

IdentifiabiltyStatistical approaches

Raue et al. (2009)

One can also define likelihood-based confidence intervals, as a confidence region:

Estimate parameters using likelihood maximizationParameter estimation

Define confidence intervals of the estimate of a parameter

where is the Hessian matrix, and is the covariance matrix of

the parameter estimates

Confidence intervals

Page 113: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

IdentifiabiltyStructural identifiability

Raue et al. (2009)

Identifiability

Structural Identifiability

Page 114: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

IdentifiabiltyStructural identifiability

Raue et al. (2009)

Identifiability

Structural Identifiability

identifiable structurally non-identifiable

Page 115: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

IdentifiabiltyStatistical approaches: an example

Raue et al. (2009)

Page 116: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

IdentifiabiltyStatistical approaches: an example

Raue et al. (2009)

Page 117: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

IdentifiabiltyStatistical approaches: an example

Raue et al. (2009)

Identifiable

Page 118: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

IdentifiabiltyStatistical approaches: an example

Raue et al. (2009)

Page 119: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

IdentifiabiltyStatistical approaches: an example

Raue et al. (2009)

Structurallynon-identifiable

Page 120: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

IdentifiabiltyStatistical approaches: an example

Raue et al. (2009)

Page 121: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

IdentifiabiltyStatistical approaches: an example

Raue et al. (2009)

????

Page 122: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

IdentifiabiltyPractical identifiability

Raue et al. (2009)

identifiable practically non-identifiable

Page 123: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de

IdentifiabiltyConclusions

Identifiability.Some theoretical methods available, but most of them are limited. (There are also methods investigating ‘local’ identifiability)

Structural non-identifiability.The - structural - model is not well fitted to the data. The number of parameters must be reduced (e.g. using correlations).

Practical non-identifiability.Data being ‘non-ideal’, non-identifiability can come either from the model or from the data (amount, quality). Based on profile likelihood.

Page 124: Nonlinear Mixed-Effects Modeling (population …fcrauste/teaching/EUR...Nonlinear Mixed-Effects Modeling (population approach) Fabien Crauste Chargé de Recherches CNRS Institut de