2. Background
• Bisphosphonates (BP) maintain bone strength
• BP are most commonly prescribed meds for
osteoporosis
• BP treatment has been associated with
Osteonecrosis of the Jaw (ONJ)
• Infection is suggested to play a pivotal role in
the pathogenesis of ONJ
• Reported BP incidence rates; 8799 (Infection)
and 1 (ONJ) per 100000 subject years
3. Objective
• To develop causal relationships between
osteoporotic meds including BP and their risk
factors General Infection and ONJ using
observational claims data
• Need to address; confounding by indication
and informative censoring bias
• Observational data is unblinded, contains
selection bias and time dependent
confounding
5. Data & Study Population
• Marketscan commercial claims database
• Meds, diagnoses, procedures, in/out patient
• Data from 1st Jan 2004 to 30th June 2011
• N=469432 subjects; 1050567 subject years of data
• Postmenopausal women with osteoporosis
• Age > 55 years
• dx osteoporosis, osteo fracture or osteo med
• ≥ 12 months of continuous enrollment
Data:
Study Population Inclusion Criteria (PMO Index date):
6. Baseline Demographic &
Subject Characteristics Table
Covariate BP (Ptyrs=464728) Other OP (Ptyrs=67860) No Treat (Ptyrs=517979)
% % %
Age (55-64 yrs) 51.6 52.5 42.2
(>= 65 yrs) 48.4 47.5 57.8
Diabetes Type II 12.8 13.2 19.1
Fragility fracture 3.7 5.5 15.5
Serious infection 4.2 5.0 5.4
CCI (mean,std) 0.5 (0.8) 0.5 (1.0) 0.6 (1.0)
Corticosteroids 26.3 27.0 22.2
Immunosuppresants 1.1 0.9 0.8
No physician visits
(mean, std)
7.2 (6.1) 7.7 (6.6) 7.5 (6.4)
CCI=Charlson Comorbidity Index
7. Exposure & Endpoint(s)
• Treatment accessed via drug/procedure codes
• 3 Cohorts; BP, Other Osteo Meds, No treat
• Treat duration: Days supplied + 60 days
• Followed from inclusion criteria until
disenrollment, dx/trt malignancy/Paget’s disease,
end of study period, ONJ or general infection
event
Exposure:
Endpoint(s): Time to Event
8. Covariates
• Time fixed (baseline) and time varying covars
• Time fixed: demographics, prior BP use,
healthcare utilization
• Time varying: comorbidities, concomitant
meds, risk factors for ONJ or general infection
• Chronic diseases (diabetes) once identified,
where carried forward throughout the study
period
9. Visit Window (Data Organization)
Unstructured Visit Record Window
X
Time varying covariates: 6 months
Time axis
t t1 t2 t3 t4 t5
Time fixed
covariates:
12 months
Data record at time t is activated by a treatment switch,
ONJ/infection event, or censoring,
time dep var status fragility fracture updated at each new rec,
time at risk defined as days supplied + 60 days (on-treatment)
Time at risk
10. Data Structure: 1 Hypothetical Subject
Id Base-
line
Date
Treat-
ment
Switch
Date
Cohort Time
(t)
Days
Event Base-
line
Covars
Time
Varying
Covars
BP
Prob
Other
Osteo
Prob
No
Osteo
Treat
Prob
1 3Jun10 3Jun05 BP 456 0 x y1 0.42 0.30 0.28
1 3Jun10 1Sep06 Other 123 0 x y2 0.30 0.44 0.26
1 3Jun10 1Jan07 No Trt 702 0 x y3 0.40 0.30 0.30
1 3Jun10 2Dec08 Other 151 1 x y4 0.50 0.25 0.25
x, y1,y2,y3,y4 are vectors of covariates
y1,y2,y3,y4 change over time
LOCF used if data value is missing for a time varying covariate
1=Event, 0=censored
11. Statistical Analysis: MSM model
• IPTW regression models with time dep vars
• Treatment weights: multinominal regression
• Censoring weights: logistic regression
• Wghts inverse cond Prob of obersved treat cat
• Subj with high predicted prob: lower weight
• Subj with low predicted prob: higher weight
• Stabilized weights and truncation introduced
to control extreme weights
Stage 1:
13. Statistical Analysis: MSM Cox model
λ 𝑇 𝑎
(t|V) = λ0 𝑡 exp(β1 𝑎 𝑡 + β2 𝑉)
Where:
λ 𝑇 𝑎
(t|V) is the hazard of ONJ or General infection at time t among subjects
with baseline covariates V in the source population had, contrary to fact,
all subjects followed each treatment cohort history 𝑎 through time t
the scalar β1 and row vector β2 are unknown parameters
λ0 𝑡 is an unspecified baseline hazard
Need to account for within subject correlation: Robust Sandwich Covariance Estimator
Weight and MSM models use different time axes
Stage 2:
14. General Infection Results Table
Treatment Number
of Ptyrs
Number
of Cases
Multivariate
Cox Reg Model I
Multivariate Cox
Reg Model II MSM:Model III
No Osteo
Treatment
330429 78634 1 1 1
BP 335976 82963 1.11 (1.10, 1.13) 1.08 (1.06, 1.09) 0.84 (0.83, 0.85)
Other OP
Meds
47433 12882 1.17 (1.14, 1.19) 1.13 (1.11, 1.15) 0.92 (0.90, 0.93)
Model I: Unweighed Cox model with time fixed covariates
Model II: Unweighted Cox model with time fixed and time varying covariates
Model III: IPTW weighed Cox model with time fixed and time varying covariates
Time fixed covars: demographics, healthcare utilization
Time varying covars: risk factors for general infection (hiv,lupus,diabetes etc.),
comorbidity status, select concomitant medications, malnutrition, obesity,
fragility fracture, etc.
15. ONJ Results Table
Treatment Number
of Ptyrs
Number
of Cases
Multivariate
Cox Reg Model I
Multivariate Cox
Reg Model II MSM
No Osteo
Treatment
515903 108 1 1 1
BP 465060 99 1.04 (0.73, 1.48) 1.03 (0.69, 1.53) 0.94 (0.64, 1.37)
Other OP
Meds
67683 8 0.56 (0.26 1.17) 0.55 (0.26, 1.18) 0.58 (0.27, 1.22)
Model I: Unweighed Cox model with time fixed covariates
Model II: Unweighted Cox model with time fixed and time varying covariates
Model III: IPTW weighed Cox model with time fixed and time varying covariates
Time fixed covars: demographics, healthcare utilization
Time varying covars: risk factors for ONJ (age, gingival bleeding, dental fistula, diabetes etc.),
comorbidity status, select concomitant medications, fragility fracture, etc.
16. MSM Assumptions
• No unmeasured confounders
• Positivity
• Model mis-specification
• Weight Truncation
Claims data does not collect all variables that may impact treatment and outcome,
For instance, bone mineral density (BMD)
All modeled covariates should have a +ve probability for outcome category
The correct model is selected for determining the IPTWs such as using a
multinominal logistic regression model and not an ordinal logistic regression
model when treatments > 2 and are not ordinal
Trade-off between control of confounding and precision of MSM weight estimates
17. Conclusion & Other Approaches
• Unweighted Cox models indicated an increased risk of
general infection for subjects on BP and other OP meds
• Adjusting for time varying confounding covariates such as
fragility fracture using inverse probability of treatment
weights indicated a reduced risk of general infection for
BP and other OP med subjects
• ONJ results were inconclusive due to their low
occurrence rate
• IPTW Kaplan Meier curves are another possible way to
conduct this statistical analysis
18. References
Hernan MA, Brumback B & Robins JM Marginal Structural Models to Estimate the Causal
Effect of Zidovudine on the Survival of HIV-Positive Men. Epidemiology 2000;11(5): 561-570
Westreich D et al. Time Scale and Adjusted Survival Curves for Marginal Structural Cox
Models. Practice of Epidemiology 2010;171(6): 691-700
Wang O, Kilpatrick RD et al. Relationship between Epoetin Alfa Dose and Mortality:
Findings from a Marginal Structural Model. Clin J Am Soc Nephrol. 2010; 5: 182-188
Xue F, Tchetgen Tchetgen E, McMullan T et al. Marginal Structural Model to Estimate the
Effect of Cumulative Osteoporosis Medication on Infection and Potential Osteonecrosis of
the Jaw (ONJ) Using Claims Data (manuscript under progress)
Spreeuwenberg MD et al. The Multiple Propensity Score as Control for Bias in the
Comparison of More Than Two Treatment Arms. Medical Care 2010; 48(2): 166-174
20. Statistical Analysis: Weighed KM
𝑆 𝑥 𝑡 =
𝑡
1 −
𝑑 𝑡𝑥
𝑟𝑡𝑥
𝑑 𝑡𝑥 =
𝑖=1
𝑁
𝑊𝑖𝑡 𝑌𝑖𝑡 (𝑋𝑖𝑡 = 𝑥)
Survivor Function:
where:
𝑑 𝑡𝑥= IPTW weighed number of events for treatment x at week t
𝑊𝑖𝑡=IPTW weight at time t for subject i
𝑌𝑖𝑡 = Event indicator with 1=Event 0=No event
𝑟𝑡𝑥= Subject risk set at time t for treatment t