Practice in Growth Curve Modeling - SAS Group... · 2016-03-11 · Practice in Growth Curve...
Transcript of Practice in Growth Curve Modeling - SAS Group... · 2016-03-11 · Practice in Growth Curve...
Practice in Growth Curve Modelingfor Women’s Menstural Cycle
- Fundamental Overview -
Chel Hee Lee1 Angela Baerwald2
1Clinical Research Supporting UnitCollege of Medicine
University of Saskatchewan
2Department of Obstetrics, Gynecology and Reproductive SciencesCollege of Medicine
University of Saskatchewan
Saskatoon SAS Users Group MeetingSeptember 16, 2015
Chel Hee Lee, Angela Baerwald (U of S) Practice in Growth Curve Modeling 2015-09-16 1 / 17
1 IntroductionFacing ProblemsLinear Mixed-effect Model
2 Parameteric Approach to Modeling CurvesUseful ModelsUseful Techniques
3 Trial and Error1st trial: Oh No! ...2nd trial: Making a progress ...3rd trial: It’s not done yet ...4th trial: Much better now ...
4 Questions
Chel Hee Lee, Angela Baerwald (U of S) Practice in Growth Curve Modeling 2015-09-16 2 / 17
Introduction
“Essentially, all models are wrong, but some are useful.”
(Box and Draper, 1986, p. 424)
It is a best guess.
It is an approximation.
It would be a translation of observation into a model.
Chel Hee Lee, Angela Baerwald (U of S) Practice in Growth Curve Modeling 2015-09-16 3 / 17
Introduction Facing Problems
Hormones Involved in Women’s Menstrual Cycle
Subject-specific Trajectory (in Logarithm Scale)
0 10 20 30
Days
0
1
2
3
Fol
licu
lar
Stim
ula
ting
Hor
mon
e
Chel Hee Lee, Angela Baerwald (U of S) Practice in Growth Curve Modeling 2015-09-16 4 / 17
Introduction Linear Mixed-effect Model
Linear Mixed-Effect Models (Singer, 1998)
Level 1 (within person)
Yij =π0j +π1j(TIME)ij + rij, where rij ∼ N(0,σ2)
Level 2 (between-person)
π0j = β00 +u0j,
π1j = β10 +u1j,
where[
u0j
u1j
]∼ N
[(00
),
(τ00 τ01
τ10 τ11
)]
Unconditional Linear Growth Curve
Yij = [β00 +β10TIMEij]︸ ︷︷ ︸FIXED
+ [u0j +u1jTIMEij + rij]︸ ︷︷ ︸RANDOM
Note that a 3-level model can be made if individuals within groups aretracked over time.
Chel Hee Lee, Angela Baerwald (U of S) Practice in Growth Curve Modeling 2015-09-16 5 / 17
Introduction Linear Mixed-effect Model
Using PROC MIXED in SAS
Commands for a linear time trend model
PROC MIXED DATA = d a t a f i l e ;CLASS id time ;MODEL y = time / SOLUTION CHISQ ;REPEATED time / TYPE=UN SUBJECT=id ;
RUN ;
Commands for a subject-specific model
PROC MIXED DATA = d a t a f i l e ;CLASS id ;MODEL y = time / SOLUTION CHISQ ;RANDOM intercept time / TYPE=UN SUBJECT=id ;
RUN ;
Chel Hee Lee, Angela Baerwald (U of S) Practice in Growth Curve Modeling 2015-09-16 6 / 17
Introduction Linear Mixed-effect Model
Concerns Bothering Us
We are working with ...
A small size of participants with many repeated measurements,
An unbalanced data set observed at unequally spaced times, and
An incomplete data due to missing values.
Chel Hee Lee, Angela Baerwald (U of S) Practice in Growth Curve Modeling 2015-09-16 7 / 17
Introduction Linear Mixed-effect Model
Concerns Bothering Us
We are working with ...
A small size of participants with many repeated measurements,
An unbalanced data set observed at unequally spaced times, and
An incomplete data due to missing values.
Chel Hee Lee, Angela Baerwald (U of S) Practice in Growth Curve Modeling 2015-09-16 7 / 17
Introduction Linear Mixed-effect Model
Concerns Bothering Us
We are working with ...
A small size of participants with many repeated measurements,
An unbalanced data set observed at unequally spaced times, and
An incomplete data due to missing values.
Chel Hee Lee, Angela Baerwald (U of S) Practice in Growth Curve Modeling 2015-09-16 7 / 17
Introduction Linear Mixed-effect Model
Concerns Bothering Us
We are faced with difficulties of ...
Determining the level of a model of interest,
Incorporating a person-level covariate,
Addressing a difference between groups,
Dealing with missing data and dropout,
Binning data or using data as it is,
Centralizing, normalizing, orthogonalizing data, and
Applying either or both models,
Choosing an appropriate pattern of covariance structure, etc.
Chel Hee Lee, Angela Baerwald (U of S) Practice in Growth Curve Modeling 2015-09-16 8 / 17
Introduction Linear Mixed-effect Model
Concerns Bothering Us
We are faced with difficulties of ...
Determining the level of a model of interest,
Incorporating a person-level covariate,
Addressing a difference between groups,
Dealing with missing data and dropout,
Binning data or using data as it is,
Centralizing, normalizing, orthogonalizing data, and
Applying either or both models,
Choosing an appropriate pattern of covariance structure, etc.
Chel Hee Lee, Angela Baerwald (U of S) Practice in Growth Curve Modeling 2015-09-16 8 / 17
Introduction Linear Mixed-effect Model
Concerns Bothering Us
We are faced with difficulties of ...
Determining the level of a model of interest,
Incorporating a person-level covariate,
Addressing a difference between groups,
Dealing with missing data and dropout,
Binning data or using data as it is,
Centralizing, normalizing, orthogonalizing data, and
Applying either or both models,
Choosing an appropriate pattern of covariance structure, etc.
Chel Hee Lee, Angela Baerwald (U of S) Practice in Growth Curve Modeling 2015-09-16 8 / 17
Introduction Linear Mixed-effect Model
Concerns Bothering Us
We are faced with difficulties of ...
Determining the level of a model of interest,
Incorporating a person-level covariate,
Addressing a difference between groups,
Dealing with missing data and dropout,
Binning data or using data as it is,
Centralizing, normalizing, orthogonalizing data, and
Applying either or both models,
Choosing an appropriate pattern of covariance structure, etc.
Chel Hee Lee, Angela Baerwald (U of S) Practice in Growth Curve Modeling 2015-09-16 8 / 17
Introduction Linear Mixed-effect Model
Concerns Bothering Us
We are faced with difficulties of ...
Determining the level of a model of interest,
Incorporating a person-level covariate,
Addressing a difference between groups,
Dealing with missing data and dropout,
Binning data or using data as it is,
Centralizing, normalizing, orthogonalizing data, and
Applying either or both models,
Choosing an appropriate pattern of covariance structure, etc.
Chel Hee Lee, Angela Baerwald (U of S) Practice in Growth Curve Modeling 2015-09-16 8 / 17
Introduction Linear Mixed-effect Model
Concerns Bothering Us
We are faced with difficulties of ...
Determining the level of a model of interest,
Incorporating a person-level covariate,
Addressing a difference between groups,
Dealing with missing data and dropout,
Binning data or using data as it is,
Centralizing, normalizing, orthogonalizing data, and
Applying either or both models,
Choosing an appropriate pattern of covariance structure, etc.
Chel Hee Lee, Angela Baerwald (U of S) Practice in Growth Curve Modeling 2015-09-16 8 / 17
Introduction Linear Mixed-effect Model
Concerns Bothering Us
We are faced with difficulties of ...
Determining the level of a model of interest,
Incorporating a person-level covariate,
Addressing a difference between groups,
Dealing with missing data and dropout,
Binning data or using data as it is,
Centralizing, normalizing, orthogonalizing data, and
Applying either or both models,
Choosing an appropriate pattern of covariance structure, etc.
Chel Hee Lee, Angela Baerwald (U of S) Practice in Growth Curve Modeling 2015-09-16 8 / 17
Introduction Linear Mixed-effect Model
Concerns Bothering Us
We are faced with difficulties of ...
Determining the level of a model of interest,
Incorporating a person-level covariate,
Addressing a difference between groups,
Dealing with missing data and dropout,
Binning data or using data as it is,
Centralizing, normalizing, orthogonalizing data, and
Applying either or both models,
Choosing an appropriate pattern of covariance structure, etc.
Chel Hee Lee, Angela Baerwald (U of S) Practice in Growth Curve Modeling 2015-09-16 8 / 17
Parameteric Approach to Modeling Curves Useful Models
Parameteric Approaches to Modeling Curves
Nonlinear Model
yi = u(xi)+εi, i = 1,2, . . . ,n
p-th Degree Polynomial Regession Model (Johnson et al., 2013)
u(xi) =β0 +β1xi +β2x2i +·· ·+βpxp
i
Exponential Growth Curve Model (Zwietering M H et al., 1990)
u(xi) =β0eβ1xi
Trigonometric Model (Cornelissen Germaine, 2014)
u(xi) =µ+Acos(wxi +φ)
Chel Hee Lee, Angela Baerwald (U of S) Practice in Growth Curve Modeling 2015-09-16 9 / 17
Parameteric Approach to Modeling Curves Useful Models
Parameteric Approaches to Modeling Curves
Nonlinear Model
yi = u(xi)+εi, i = 1,2, . . . ,n
p-th Degree Polynomial Regession Model (Johnson et al., 2013)
u(xi) =β0 +β1xi +β2x2i +·· ·+βpxp
i
Exponential Growth Curve Model (Zwietering M H et al., 1990)
u(xi) =β0eβ1xi
Trigonometric Model (Cornelissen Germaine, 2014)
u(xi) =µ+Acos(wxi +φ)
Chel Hee Lee, Angela Baerwald (U of S) Practice in Growth Curve Modeling 2015-09-16 9 / 17
Parameteric Approach to Modeling Curves Useful Models
Parameteric Approaches to Modeling Curves
Nonlinear Model
yi = u(xi)+εi, i = 1,2, . . . ,n
p-th Degree Polynomial Regession Model (Johnson et al., 2013)
u(xi) =β0 +β1xi +β2x2i +·· ·+βpxp
i
Exponential Growth Curve Model (Zwietering M H et al., 1990)
u(xi) =β0eβ1xi
Trigonometric Model (Cornelissen Germaine, 2014)
u(xi) =µ+Acos(wxi +φ)
Chel Hee Lee, Angela Baerwald (U of S) Practice in Growth Curve Modeling 2015-09-16 9 / 17
Parameteric Approach to Modeling Curves Useful Techniques
Modeling Techniques
Piecewise Modeling Approach (Naumova et al., 2001)
u(xi) =β0 +β1xi +β2(xi − c)δ,
where δ= 1 if xi > c and δ= 0 if xi <= c.
Cubic-Splines Modeling Approach (Gurrin et al., 2005)
u(x;θ) =β0 +β1x+β2x2 +β3x3 +m∑
i=1θi(x− ci)
3+
Chel Hee Lee, Angela Baerwald (U of S) Practice in Growth Curve Modeling 2015-09-16 10 / 17
Trial and Error 4th trial: Much better now ...
OH No! This is not what I wanted. (Problems?)
3rd degree polynomial regression
0 5 10 15 20 25
Days
0
1
2
3
Fitt
ed V
alu
es
Chel Hee Lee, Angela Baerwald (U of S) Practice in Growth Curve Modeling 2015-09-16 11 / 17
Trial and Error 4th trial: Much better now ...
OH No! This is not what I wanted. (Problems?)
3rd degree polynomial regression
Predln_fsh
ioi1
ln_
fsh
id = ues112id = twf049
id = tgs105id = smh100id = slb017id = paf120id = ljw126
id = keb118id = jgc015id = jdm097id = img020id = hjr085
id = dps106id = dmp095id = dks091id = cmd109id = cbd086
id = b_j101id = ava119id = arb092id = adl115id = aam076
0 10 200 10 200 10 200 10 200 10 20
0
1
2
3
0
1
2
3
0
1
2
3
0
1
2
3
0
1
2
3
Chel Hee Lee, Angela Baerwald (U of S) Practice in Growth Curve Modeling 2015-09-16 11 / 17
Trial and Error 4th trial: Much better now ...
Oh! It’s better! However, ... Disadvantages???
Piecewise-4th degree polynomial regression
0 5 10 15 20 25
Days
0
1
2
3
Fitt
ed V
alu
es
Chel Hee Lee, Angela Baerwald (U of S) Practice in Growth Curve Modeling 2015-09-16 12 / 17
Trial and Error 4th trial: Much better now ...
Oh! It’s better! However, ... Disadvantages???
Piecewise-4th degree polynomial regression
Predln_fsh
Days
Fol
licle
Stim
ula
ting
Hor
mon
e
id = ues112id = twf049
id = tgs105id = smh100id = slb017id = paf120id = ljw126
id = keb118id = jgc015id = jdm097id = img020id = hjr085
id = dps106id = dmp095id = dks091id = cmd109id = cbd086
id = b_j101id = ava119id = arb092id = adl115id = aam076
0 10 200 10 200 10 200 10 200 10 20
0
1
2
3
0
1
2
3
0
1
2
3
0
1
2
3
0
1
2
3
Chel Hee Lee, Angela Baerwald (U of S) Practice in Growth Curve Modeling 2015-09-16 12 / 17
Trial and Error 4th trial: Much better now ...
Oh! It’s better! However, ... Disadvantages???
Piecewise-4th degree polynomial regression
Conditional Residuals for ln_fsh
BIC 305.98AICC 316.22AIC 315.98Objective 305.98
Fit Statistics
Std Dev 0.3154Maximum 1.2595Mean -2E-15Minimum -0.859Observations 269
Residual Statistics
-3 -2 -1 0 1 2 3
Quantile
-1.0
-0.5
0.0
0.5
1.0
Res
idua
l
-1 -0.6 -0.2 0.2 0.6 1 1.4
Residual
0
5
10
15
20
25
30
Per
cent
0.5 1.0 1.5 2.0 2.5 3.0
Predicted
-1.0
-0.5
0.0
0.5
1.0
Res
idua
l
Chel Hee Lee, Angela Baerwald (U of S) Practice in Growth Curve Modeling 2015-09-16 12 / 17
Trial and Error 4th trial: Much better now ...
Let’s do more work! Still, ... Disadvantages???
3rd degree polynomial regression with trigonometric functions
0 5 10 15 20 25
Days
0
1
2
3
Fitt
ed V
alu
es
10mra
Chel Hee Lee, Angela Baerwald (U of S) Practice in Growth Curve Modeling 2015-09-16 13 / 17
Trial and Error 4th trial: Much better now ...
Let’s do more work! Still, ... Disadvantages???
3rd degree polynomial regression with trigonometric functions
Days
Fol
licle
Stim
ula
ting
Hor
mon
e
id = ues112id = twf049
id = tgs105id = smh100id = slb017id = paf120id = ljw126
id = keb118id = jgc015id = jdm097id = img020id = hjr085
id = dps106id = dmp095id = dks091id = cmd109id = cbd086
id = b_j101id = ava119id = arb092id = adl115id = aam076
0 10 200 10 200 10 200 10 200 10 20
0
1
2
3
0
1
2
3
0
1
2
3
0
1
2
3
0
1
2
3
Chel Hee Lee, Angela Baerwald (U of S) Practice in Growth Curve Modeling 2015-09-16 13 / 17
Trial and Error 4th trial: Much better now ...
Much Better Now! Disadvantages???
Natural Cubic Splines
0 5 10 15 20 25
Days
0
1
2
3
Fitt
ed V
alu
es
Chel Hee Lee, Angela Baerwald (U of S) Practice in Growth Curve Modeling 2015-09-16 14 / 17
Trial and Error 4th trial: Much better now ...
Much Better Now! Disadvantages???
Natural Cubic Splines
Predln_fshBand
Days
Fol
licle
Stim
ula
ting
Hor
mon
e
id = ues112id = twf049
id = tgs105id = smh100id = slb017id = paf120id = ljw126
id = keb118id = jgc015id = jdm097id = img020id = hjr085
id = dps106id = dmp095id = dks091id = cmd109id = cbd086
id = b_j101id = ava119id = arb092id = adl115id = aam076
0 10 200 10 200 10 200 10 200 10 20
0
1
2
3
0
1
2
3
0
1
2
3
0
1
2
3
0
1
2
3
Chel Hee Lee, Angela Baerwald (U of S) Practice in Growth Curve Modeling 2015-09-16 14 / 17
Questions
Questions
QUESTIONS?
Chel Hee Lee, Angela Baerwald (U of S) Practice in Growth Curve Modeling 2015-09-16 15 / 17
Questions
References I
Box, G. E. P. and Draper, N. R. (1986). Empirical Model-building andResponse Surface. John Wiley & Sons, Inc., New York, NY, USA.
Cornelissen Germaine (2014). Cosinor-based rhythmometry. TheoreticalBiology & Medical Modelling, 11:16–16.
Fitzmaurice, G., Laird, N., and Ware, J. (2004). Applied LongitudinalAnalysis. Wiley Series in Probability and Statistics - Applied Probabilityand Statistics Section Series. Wiley.
Gurrin, L. C., Scurrah, K. J., and Hazelton, M. L. (2005). Tutorial inbiostatistics: spline smoothing with linear mixed models. Statistics inMedicine, 24(21):3361–3381.
Johnson, W., Balakrishna, N., and Griffiths, P. L. (2013). Modeling physicalgrowth using mixed effects models. American Journal of PhysicalAnthropology, 150(1):58–67.
Chel Hee Lee, Angela Baerwald (U of S) Practice in Growth Curve Modeling 2015-09-16 16 / 17
Questions
References II
Miles, J. and Shevlin, M. (2001). Applying Regression and Correlation: AGuide for Students and Researchers. SAGE Publications.
Naumova, E. N., Must, A., and Laird, N. M. (2001). Tutorial in Biostatistics:Evaluating the impact of ‘critical periods’ in longitudinal studies ofgrowth using piecewise mixed effects models. International Journal ofEpidemiology, 30(6):1332–1341.
Singer, J. D. (1998). Using SAS PROC MIXED to Fit Multilevel Models,Hierarchical Models, and Individual Growth Models. Journal ofEducational and Behavioral Statistics, 23(4):323–355.
Zwietering M H, Jongenburger I, Rombouts F M, and van ’t Riet, K. (1990).Modeling of the Bacterial Growth Curve. Applied and EnvironmentalMicrobiology, 56(6):1875–1881.
Chel Hee Lee, Angela Baerwald (U of S) Practice in Growth Curve Modeling 2015-09-16 17 / 17