Statistical Analysis of Clustered Binary Response in Oral Health Research
1. Statistical Analysis of Clustered Binary Response in
Oral Health Research
Ronen Ofec*, DMD ; David M. Steinberg, PhD ; Devorah Schwartz-Arad, DND, PhD
* M.Sc. program in Biostatistics, School of Mathematical sciences, Tel-Aviv university
* Praviate dental practice, Tel-Aviv, Israel
The 4th International Meeting on Methodological Issues In Oral Health Research,
Istanbul , Turkey
11. The durability of
Dental implants treatment
Failures (removal of an implant) do occur
Marginal bone loss (MBL) can be an early sign for a failure
MBL: The amount of bone an implant loses during function time
12. Marginal bone loss (MBL)
Fransson et al.(2005)
What are the risk factors for MBL?
Are some patients more prone to MBL?
Are implants within patients correlated to each other?
13. Marginal bone loss (MBL)
Fransson et al.(2005)
What are the risk factors for MBL?
Are some patients more prone to MBL?
Are implants within patients correlated to each other?
14. Marginal bone loss (MBL)
Fransson et al.(2005)
What are the risk factors for MBL?
Are some patients more prone to MBL?
Are implants within patients correlated to each other?
15. Main question of interest and
Objectives of the study
What will be the consequences of a naïve analysis
that doesn't recognize correlation within a patient?
1 To identify risk factors for MBL in a long term follow up study
2 To estimate the Intra patient correlation with regard to MBL
To compare results from a naïve analysis to one that
3 includes intra patient correlation
16. Main question of interest and
Objectives of the study
What will be the consequences of a naïve analysis
that doesn't recognize correlation within a patient?
1 To identify risk factors for MBL in a long term follow up study
2 To estimate the Intra patient correlation with regard to MBL
To compare results from a naïve analysis to one that
3 includes intra patient correlation
17. Main question of interest and
Objectives of the study
What will be the consequences of a naïve analysis
that doesn't recognize correlation within a patient?
1 To identify risk factors for MBL in a long term follow up study
2 To estimate the Intra patient correlation with regard to MBL
To compare results from a naïve analysis to one that
3 includes intra patient correlation
18. Main question of interest and
Objectives of the study
What will be the consequences of a naïve analysis
that doesn't recognize correlation within a patient?
1 To identify risk factors for MBL in a long term follow up study
2 To estimate the Intra patient correlation with regard to MBL
To compare results from a naïve analysis to one that
3 includes intra patient correlation
19. Study design and participants
Historical prospective cohort study design
Schwartz-Arad Surgical center, between January
1996, and July 1998 by a single surgeon (DSA)
Follow-up time was up to 147 months with a mean
of 70 months
20. The exposures,
Binary response and data set
The exposures: Patient-specific and implant-specific
Clustered response: MBL measurement at implant level
Binary response: Acceptable vs. advanced bone loss
Cut point at MBL=0.2 mm/year
The data set: Multilevel data set
195 Patients as the primary sample units (clusters)
721 Implants as the Elementary units
No. of implants per patient [1,16], mode=3
21. The Intra Patient Correlation
1 Way Random Effect ANOVA
Patient effect
Implant effect
Patient and implant effect
are independent
The Intra Class Correlation (ICC)
22. The estimator and the estimate
for ICC
Kappa type estimator proposed by Fleiss and Cuzick (1979)
Confidence Intervals for the estimator formulated
by Zou and Donner (2004)
Simulation results: empirical coverage is close to nominal
with C.I for the kappa type
In our study:
23. The Generalized Estimating
Equations (GEE)
Population average (Marginal) model Liang and Zeger (1986)
1. The mean model:
2. Working variance structure:
3. Working correlation structure:
The empirical/sandwich estimator for the precision of estimates
24. Robustness of the
Sandwich estimator
The estimator is robust to misspecification of the variance and
correlation structures
Our estimates are still valid (consistent) if we use a structure
which is not reflecting reality
Mancl and Leroux (1996): Gain of precision for the “right”
correlation structure
26. Risk factors for MBL
by GEE
0 *** 0.001 ** 0.01 * 0.05
Exposure Function time<3 years Function time≥3 years
Beta S.E PV. Beta S.E PV.
Smoker 1.44 0.41 ***
Coating (HA &TPS) -2.22 0.76 ** 1.34 0.39 ***
Early spontaneous exposure 0.85 0.29 **
Diameter -1.39 0.40 ***
Interaction between function time and risk factors
For smoker:
The odds for MBL for smokers is 4.22 times greater
than for non smokers
The effect of HA & TPS turns from protective to risk
27. Risk factors for MBL
by GEE
0 *** 0.001 ** 0.01 * 0.05
Exposure Function time<3 years Function time≥3 years
Beta S.E PV. Beta S.E PV.
Smoker 1.44 0.41 ***
Coating (HA &TPS) -2.22 0.76 ** 1.34 0.39 ***
Early spontaneous exposure 0.85 0.29 **
Diameter -1.39 0.40 ***
Interaction between function time and risk factors
For smoker:
The odds for MBL for smokers is 4.22 times greater
than for non smokers
The effect of HA & TPS turns from protective to risk
28. Risk factors for MBL
by GEE
0 *** 0.001 ** 0.01 * 0.05
Exposure Function time<3 years Function time≥3 years
Beta S.E PV. Beta S.E PV.
Smoker 1.44 0.41 ***
Coating (HA &TPS) -2.22 0.76 ** 1.34 0.39 ***
Early spontaneous exposure 0.85 0.29 **
Diameter -1.39 0.40 ***
Interaction between function time and risk factors
For smoker:
The odds for MBL for smokers is 4.22 times greater
than for non smokers
The effect of HA & TPS turns from protective to risk
29. The naïve estimation
and GEE
Function time >= 3 years
Naïve GEE
Exposure Beta S.E PV. Beta S.E PV.
Smoker 1.50 0.29 *** 1.44 0.41 ***
Coating (HA &TPS) 1.31 0.27 *** 1.34 0.39 ***
Early exposure 0.85 0.26 ** 0.85 0.29 **
Diameter -1.57 0.35 *** -1.39 0.40 ***
Estimates for exposure effects – it is not bad to be naïve
Correlation doesn’t induce bias to an unbiased estimator
Standard errors of estimates- a naïve analysis leads to bias
Underestimation or overestimation of standard errors
Risk for invalid inference concerning the estimated effect
30. The naïve estimation
and GEE
Function time >= 3 years
Naïve GEE
Exposure Beta S.E PV. Beta S.E PV.
Smoker 1.50 0.29 *** 1.44 0.41 ***
Coating (HA &TPS) 1.31 0.27 *** 1.34 0.39 ***
Early exposure 0.85 0.26 ** 0.85 0.29 **
Diameter -1.57 0.35 *** -1.39 0.40 ***
Estimates for exposure effects – it is not bad to be naïve
Correlation doesn’t induce bias to an unbiased estimator
Standard errors of estimates- a naïve analysis leads to bias
Underestimation or overestimation of standard errors
Risk for invalid inference concerning the estimated effect
31. The naïve estimation
and GEE
Function time < 3 years
Naïve GEE
Exposure Beta S.E PV. Beta S.E PV.
Smoker -0.41 0.41 0.32 -0.43 0.43 0.30
Coating (HA &TPS) -2.2 1.09 0.04 -2.2 0.76 **
Early exposure 0.35 0.42 0.40 0.35 0.39 0.35
Diameter -0.63 0.42 0.13 -0.60 0.47 0.21
Estimates for exposure effects – it is not bad to be naïve
Correlation doesn’t induce bias to an unbiased estimator
Standard errors of estimates- a naïve analysis leads to bias
Underestimation or overestimation of standard errors
Risk for invalid inference concerning the estimated effect
32. The source of exposures
variation
Source of exposure/treatment variation
Between Within
patient/cluster patient/cluster
Patient specific exposure: variation between patient
Similar to treatment effect in Between cluster design
Implant specific exposure: variation within and between patient
Might be similar to treatment effect in Within/Between cluster
design
Depends on the source of variation of Implant specific exposure
33. The Design Effect (Deff)
Between Within
cluster design cluster design
Deff >1 Deff<1
Variance inflation factor (VIF) Variance attenuation factor (VAF)
Therefore, a naïve analysis is Therefore, a naïve analysis is
anti-conservative (underestimate) conservative (overestimate)
34. The Design Effect (Deff)
Between Within
cluster design cluster design
Deff >1 Deff<1
Variance inflation factor (VIF) Variance attenuation factor (VAF)
Therefore, a naïve analysis is Therefore, a naïve analysis is
anti-conservative (underestimate) conservative (overestimate)
35. The Design Effect (Deff)
Between Within
cluster design cluster design
Deff >1 Deff<1
Variance inflation factor (VIF) Variance attenuation factor (VAF)
Therefore, a naïve analysis is Therefore, a naïve analysis is
anti-conservative (underestimate) conservative (overestimate)
36. The answer to the main
question of interest
What will be the consequences of a
naïve analysis that doesn't recognize
correlation within a patient?
No problem with the estimated effect
For a patient specific exposure: underestimation of standard
errors
For an implant specific exposure: underestimation if variance is
from between patients
But, overestimation if variance of exposure is from within patient
Mancl, Leroux, DeRouen (2000) recommended to separate the
effect of a site specific exposure, into within and between effect
37. Conclusions
1 Intra patient correlation for advanced MBL exists
2 The effect of some exposures isn’t constant during function time
Ignoring ICC might bias the precision of estimated effect. Simulation
3 studies should confirm the direction of the bias
38. Conclusions
1 Intra patient correlation for advanced MBL exists
2 The effect of some exposures isn’t constant during function time
Ignoring ICC might bias the precision of estimated effect. Simulation
3 studies should confirm the direction of the bias
39. Conclusions
1 Intra patient correlation for advanced MBL exists
2 The effect of some exposures isn’t constant during function time
Ignoring ICC might bias the precision of estimated effect. Simulation
3 studies should confirm the direction of the bias