Transcript MFP

Willi Sauerbrei
Institut of Medical Biometry and Informatics
University Medical Center Freiburg, Germany
Patrick Royston
MRC Clinical Trials Unit,
London, UK
Multivariable regression models
with continuous covariates
with a practical emphasis on fractional
polynomials and applications in clinical
epidemiology
The problem …
“Quantifying epidemiologic risk factors
using non-parametric regression: model
selection remains the greatest challenge”
Rosenberg PS et al, Statistics in Medicine 2003; 22:3369-3381
Trivial nowadays to fit almost any model
To choose a good model is much harder
2
Overview
• Context and motivation
• Introduction to fractional polynomials for the
univariate smoothing problem
• Extension to multivariable models
• Robustness and stability
• Software sources
• Conclusions
3
Motivation
• Often have continuous risk factors in epidemiology
and clinical studies – how to model them?
• Linear model may describe a dose-response
relationship badly
‘Linear’ = straight line = 0 + 1 X + … throughout talk
• Using cut-points has several problems
• Splines recommended by some – but are not ideal
Lack a well-defined approach to model selection
‘Black box’
Robustness issues
4
Problems of cut-points
• Step-function is a poor approximation to true
relationship
Almost always fits data less well than a suitable
continuous function
• ‘Optimal’ cut-points have several difficulties
Biased effect estimates
Inflated P-values
Not reproducible in other studies
5
Example datasets
1. Epidemiology
• Whitehall 1
17,370 male Civil Servants aged 40-64 years
Measurements include: age, cigarette smoking,
BP, cholesterol, height, weight, job grade
Outcomes of interest: coronary heart disease, allcause mortality  logistic regression
Interested in risk as function of covariates
Several continuous covariates
 Some may have no influence in multivariable context
6
Example datasets
2. Clinical studies
• German breast cancer study group (BMFT-2)
Prognostic factors in primary breast cancer
Age, menopausal status, tumour size, grade, no. of
positive lymph nodes, hormone receptor status
Recurrence-free survival time  Cox regression
686 patients, 299 events
Several continuous covariates
Interested in prognostic model and effect of
individual variables
7
Example:
Systolic blood pressure vs. age
50
100
150
200
250
300
Whitehall 1: BP vs age
40
45
50
55
60
65
Age, years
8
Example: Curve fitting
(Systolic BP and age – not linear)
150
Whitehall 1: BP vs age
125
130
135
140
145
95% CI
Linear function
FP1 function
Running line
40
45
50
55
60
65
Age, years
9
Empirical curve fitting: Aims
• Smoothing
• Visualise relationship of Y with X
• Provide and/or suggest functional form
10
Some approaches
• ‘Non-parametric’ (local-influence) models
Locally weighted (kernel) fits (e.g. lowess)
Regression splines
Smoothing splines (used in generalized additive models)
• Parametric (non-local influence) models
Polynomials
Non-linear curves
Fractional polynomials
 Intermediate between polynomials and non-linear curves
11
Local regression models
• Advantages
Flexible – because local!
May reveal ‘true’ curve shape (?)
• Disadvantages
Unstable – because local!
No concise form for models
 Therefore, hard for others to use – publication,compare results with
those from other models
Curves not necessarily smooth
‘Black box’ approach
Many approaches – which one(s) to use?
12
Polynomial models
• Do not have the disadvantages of local
regression models, but do have others:
• Lack of flexibility (low order)
• Artefacts in fitted curves (high order)
• Cannot have asymptotes
13
Fractional polynomial models
• Describe for one covariate, X
multiple regression later
• Fractional polynomial of degree m for X with powers
p1, … , pm is given by
FPm(X) = 1 X p + … + m X p
1
m
• Powers p1,…, pm are taken from a special set
{2,  1,  0.5, 0, 0.5, 1, 2, 3}
• Usually m = 1 or m = 2 is sufficient for a good fit
14
FP1 and FP2 models
• FP1 models are simple power transformations
• 1/X2, 1/X, 1/X, log X, X, X, X2, X3
8 models
• FP2 models are combinations of these
For example 1(1/X) + 2(X2)
28 models
• Note ‘repeated powers’ models
For example 1(1/X) + 2(1/X)log X
8 models
15
FP1 and FP2 models:
some properties
• Many useful curves
• A variety of features are available:
Monotonic
Can have asymptote
Non-monotonic (single maximum or minimum)
Single turning-point
• Get better fit than with conventional
polynomials, even of higher degree
16
Examples of FP2 curves
- varying powers
(-2, 1)
(-2, 2)
(-2, -2)
(-2, -1)
17
Examples of FP2 curves
- single power, different coefficients
(-2, 2)
4
Y
2
0
-2
-4
10
20
30
x
40
50
18
A philosophy of function selection
• Prefer simple (linear) model
• Use more complex (non-linear) FP1 or FP2
model if indicated by the data
• Contrast to local regression modelling
Already starts with a complex model
19
Estimation and significance testing for
FP models
• Fit model with each combination of powers
FP1: 8 single powers
FP2: 36 combinations of powers
• Choose model with lowest deviance (MLE)
• Comparing FPm with FP(m  1):
compare deviance difference with 2 on 2 d.f.
one d.f. for power, 1 d.f. for regression coefficient
supported by simulations; slightly conservative
20
Selection of FP function
•
•
•
•
•
•
•
Has flavour of a closed test procedure
Use 2 approximations to get P-values
Define nominal P-value for all tests (often 5%)
Fit linear and best FP1 and FP2 models
Test FP2 vs. null – test of any effect of X (2 on 4 df)
Test FP2 vs linear – test of non-linearity (2 on 3 df)
Test FP2 vs FP1 – test of more complex function
against simpler one (2 on 2 df)
21
Example: Systolic BP and age
Model
FP2 v Null
FP2 v Linear
FP2 v FP1
d.f.
4
3
2
Deviance
difference
944.57
29.95
3.29
Pvalue
0.000
0.000
0.2
Reminder:
FP1 had power 3:
1 X3
FP2 had powers (1,1):
1 X + 2 X log X
22
Aside: FP versus spline
• Why care about FPs when splines are more
flexible?
• More flexible  more unstable
More chance of ‘over-fitting’
• In epidemiology, dose-response relationships
are often simple
• Illustrate by small simulation example
23
FP versus spline (continued)
•
•
•
•
•
•
•
•
Logarithmic relationships are common in practice
Simulate regression model y = 0 + 1log(X) + error
Error is normally distributed N(0, 2)
Take 0 = 0, 1 = 1; X has lognormal distribution
Vary  = {1, 0.5, 0.25, 0.125}
Fit FP1, FP2 and spline with 2, 4, 6 d.f.
Compute mean square error
Compare with mean square error for true model
24
FP vs. spline (continued)
2
y
0
-2
-4
-4
-2
y
0
2
4
Sigma = 0.5
4
Sigma = 1
2
4
6
0
2
4
x
Sigma = 0.25
Sigma = 0.125
6
2
y
0
-2
-4
-4
-2
y
0
2
4
x
4
0
0
2
4
x
6
0
2
4
6
x
25
FP vs. spline (continued)
FP1 and spline with 2 df
2
1
0
-1
-2
-2
-1
0
1
2
Solid: FP1; dashed: spline 2 df
4
6
0
2
4
6
0
2
4
6
0
2
4
6
1
0
-1
-2
-2
-1
0
1
2
2
2
0
26
FP vs. spline (continued)
2
1
0
-1
-2
-2
-1
0
1
2
FP2 and spline with 4 df
2
3
4
5
0
1
2
3
4
5
0
1
2
3
4
5
0
1
2
3
4
5
1
0
-1
-2
-2
-1
0
1
2
1
2
0
27
FP vs. spline (continued)
0
.04
.08
.12
FP vs. spline: prediction error
.125
.25
.5
1
sigma
True
Spline 2df
FP1
Spline 4df
FP2
Spline 6df
28
FP vs. spline (continued)
• In this example, spline usually less accurate
than FP
• FP2 less accurate than FP1 (over-fitting)
• FP1 and FP2 more accurate than splines
• Splines often had non-monotonic fitted curves
Could be medically implausible
• Of course, this is a special example
29
Multivariable FP (MFP) models
• Assume have k > 1 continuous covariates and
perhaps some categoric or binary covariates
• Allow dropping of non-significant variables
• Wish to find best multivariable FP model for
all X’s
• Impractical to try all combinations of powers
• Require iterative fitting procedure
30
Fitting multivariable FP models
(MFP algorithm)
• Combine backward elimination of weak
variables with search for best FP functions
• Determine fitting order from linear model
• Apply FP model selection procedure to each X
in turn
fixing functions (but not ’s) for other X’s
• Cycle until FP functions (i.e. powers) and
variables selected do not change
31
Example: Prognostic factors in breast
cancer
• Aim to develop a prognostic index for risk of
tumour recurrence or death
• Have 7 prognostic factors
4 continuous, 3 categorical
• Select variables and functions using 5%
significance level
32
Univariate linear analysis
Variable
X1
X2
X3
X4a
X4b
X5
X6
X7
Name
Age
Menopausal status
Tumour size
Grade 2 or 3
Grade 3
No. of positive lymph nodes
Progesterone receptor status
Oestrogen receptor status
2
0.58
0.28
15.68
19.92
8.19
50.02
34.04
4.70
33
Univariate FP2 analysis
Variable
X1 age
X3 size
X5 nodes
X6 PgR
X7 ER
Powers
(2, 0.5)
(1, 3)
(1, 2)
(0.5, 0)
(2, 1)
2 d.f.
17.61
4
19.81
4
81.36
4
52.73
4
23.07
4
P
0.001
0.001
< 0.001
< 0.001
< 0.001
Gain
17.03
4.13
31.34
18.69
18.37
Gain compares FP2 with linear on 3 d.f.
All factors except for X3 have a non-linear effect
34
Multivariable FP analysis
Variable
X1 age
X3 size
X5 nodes
X6 PgR
X7 ER
X2 mens.
X4a grad 2/3
X4b grad 3
FP etc.
(2, 0.5)
Out
(2, 1)
0.5
Out
Out
In
Out
2
19.33
5.31
74.14
32.70
2.15
0.21
4.59
0.15
d.f.
P
4 0.001
4
0.3
4 <0.001
4 <0.001
4
0.7
1
0.6
1
0.03
1
0.7
35
Comments on analysis
• Conventional backwards elimination at 5%
level selects X4a, X5, X6, and X1 is excluded
• FP analysis picks up same variables as
backward elimination, and additionally X1
• Note considerable non-linearity of X1 and X5
• X1 has no linear influence on risk of
recurrence
• FP model detects more structure in the data
than the linear model
36
Plots of fitted FP functions
Breast cancer: Fitted FP functions
1
Nodes
20
40
-1
-.5
0
.5
Log relative hazard
5
4
3
2
1
0
Log relative hazard
Age
60
80
Age, years
0
10
20
30
40
No. of positive lymph nodes
50
0
-1
-2
-3
Log relative hazard
1
Progesterone receptor
0
500
1000
1500
2000
Progesterone receptor status
2500
37
Survival by risk groups
0.00
0.25
0.50
0.75
1.00
Prognostic classification scheme
0
2
4
Recurrence-free survival, yr
Group = Low risk
Group = High risk
6
8
Group = Medium risk
38
Robustness of FP functions
• Breast cancer example showed non-robust
functions for nodes – not medically sensible
• Situation can be improved by performing
covariate transformation before FP analysis
• Can be done systematically (work in progress)
• Sauerbrei & Royston (1999) used negative
exponential transformation of nodes
exp(–0.12 * number of nodes)
39
0
.5
1
1.5
Making the function for lymph nodes
more robust
-.5
Original
Exponential transformation
0
10
20
30
No. of positive lymph nodes
40
50
40
2nd example: Whitehall 1
MFP analysis
Covariate
Age
Cigarettes
Systolic BP
Total cholesterol
Height
Weight
Job grade
FP etc.
Linear
0.5
-1, -0.5
Linear
Linear
-2, 3
In
No variables were eliminated by the MFP algorithm
Weight is eliminated by linear backward elimination
41
Plots of FP functions
Whitehall 1: multivariable FP analysis
Cigarettes
.5
.4
.3
.2
.1
.08
45
50 55 60
Age at entry
65
0
20
40
Cigarettes/day
5
10
Cholesterol/ mmol/l
100 150 200 250 300
Systolic BP
15
40
60
Height
.08 .09
.1
Probability of death
.2
.1
.12 .14 .16 .18
Probability of death
.14
.12
.1
.08
0
50
Weight
.16
Total cholesterol
60
.11 .12 .13
40
Probability of death
Systolic BP
Probability of death
.1
Probability of death
.15
.1
.05
Probability of death
.2
.12 .14 .16 .18
Age
80 100 120 140
Weight/kgs
140
160
180
Height/cms
200
42
Stability
Models (variables, FP functions) selected by
statistical criteria – cut-off on P-value
Approach has several advantages …
… and also is known to have problems
Omission bias
Selection bias
Unstable – many models may fit equally well
43
Stability
• Instability may be studied by bootstrap resampling
(sampling with replacement)
Take bootstrap sample B times
Select model by chosen procedure
Count how many times each variable is selected
Summarise inclusion frequencies & their dependencies
Study fitted functions for each covariate
• May lead to choosing several possible models, or a
model different from the original one
44
Bootstrap stability analysis of the breast
cancer dataset
• 5000 bootstrap samples taken (!)
• MFP algorithm with Cox model applied to
each sample
• Resulted in 1222 different models (!!)
• Nevertheless, could identify stable subset
consisting of 60% of replications
Judged by similarity of functions selected
45
Bootstrap stability analysis of the breast
cancer dataset
Variable
Model
selected
Age
FP1
FP2
Menopausal status
—
Tumour size
FP1
FP2
Grade 2/3
—
Grade 3
—
Lymph nodes
FP1
Progesterone receptors
FP1
FP2
Oestrogen receptors
FP1
FP2
% bootstraps
model selected
16
76
20
34
6
58
9
100
95
4
13
6
46
Bootstrap analysis: summaries of fitted
curves from stable subset
Log relative hazard
6
1
0
4
-1
2
-2
-3
0
20
30
40
50
60
Age, years
70
80
Log relative hazard
2
0
25
50
75
Tumour size, mm
100
0
250
PgR, fmol/L
500
1
1
0
0
-1
-1
0
10
20
30
Number of positive lymph nodes
47
Presentation of models for continuous
covariates
• The function + 95% CI gives the whole story
• Functions for important covariates should
always be plotted
• In epidemiology, sometimes useful to give a
more conventional table of results in
categories
• This can be done from the fitted function
48
Example: Cigarette smoking and allcause mortality (Whitehall 1)
Cigarettes per day
Number
OR (model based)
Range
Ref. At risk Dying Estimate 95% CI
point
0 (referent) 0
10103 690
1.00
-1-10
5
2254 243
1.69
1.59, 1.80
11-20
15
3448 494
2.25
2.04, 2.49
21-30
25
1117 185
2.60
2.31, 2.91
31-40
35
283
48
2.86
2.52, 3.24
41-50
45
43
8
3.07
2.68, 3.52
51-60
55
12
2
3.25
2.82, 3.75
49
Other issues (1)
• Handling continuous confounders
May use a larger P-value for selection e.g. 0.2
Not so concerned about functional form here
• Binary/continuous covariate interactions
Can be modelled using FPs (Royston & Sauerbrei
2004)
Adjust for other factors using MFP
50
Other issues (2)
• Time-varying effects in survival analysis
Can be modelled using FP functions of time
(Berger; also Sauerbrei & Royston, in progress)
• Checking adequacy of FP functions
May be done by using splines
Fit FP function and see if spline function adds
anything, adjusting for the fitted FP function
51
Software sources
• Most comprehensive implementation is in
Stata
Command mfp is part of Stata 8
• Versions for SAS and R are now available
Contact W Sauerbrei ([email protected])
to request a copy of the SAS macro
R version available on CRAN archive
 mfp package
52
Concluding remarks (1)
• FP method in general
No reason (other than convention) why regression models
should include only positive integer powers of covariates
FP is a simple extension of an existing method
Simple to program and simple to explain
Parametric, so can easily get predicted values
FP usually gives better fit than standard polynomials
Cannot do worse, since standard polynomials are included
53
Concluding remarks (2)
• Multivariable FP modelling
Many applications in general context of multiple
regression modelling
Well-defined procedure based on standard
principles for selecting variables and functions
Aspects of robustness and stability have been
investigated (and methods are available)
Much experience gained so far suggests that
method is very useful in clinical epidemiology
54
Some references
•
•
•
•
•
•
•
Royston P, Altman DG (1994) Regression using fractional polynomials of
continuous covariates: parsimonious parametric modelling. Applied Statistics 43:
429-467
Royston P, Altman DG (1997) Approximating statistical functions by using
fractional polynomial regression. The Statistician 46: 1-12
Sauerbrei W, Royston P (1999) Building multivariable prognostic and diagnostic
models: transformation of the predictors by using fractional polynomials. JRSS(A)
162: 71-94. Corrigendum JRSS(A) 165: 399--400, 2002
Royston P, Ambler G, Sauerbrei W. (1999) The use of fractional polynomials to
model continuous risk variables in epidemiology. International Journal of
Epidemiology, 28: 964-974.
Royston P, Sauerbrei W (2004). A new approach to modelling interactions between
treatment and continuous covariates in clinical trials by using fractional
polynomials. Statistics in Medicine 23: 2509-2525.
Royston P, Sauerbrei W (2003) Stability of multivariable fractional polynomial
models with selection of variables and transformations: a bootstrap investigation.
Statistics in Medicine 22: 639-659.
Armitage P, Berry G, Matthews JNS (2002) Statistical Methods in Medical
Research. Oxford, Blackwell.
55