Exposure and Outcome Ascertainment

Practical Skills in Protocol Writing and
Computer Programming: Exposure and
Outcome Ascertainment
Jessica Jalbert, PhD
ICPE, Taipei, Taiwan
October 23rd, 2014
Disclosures
• None
Exposure Ascertainment
Types of Exposures
• Medications
• Devices
• Procedures
Identifying Exposures in Claims
• Wish list of what you might need:
– Date medication was dispensed
– Molecule (e.g. atorvastatin, citalopram)
– Number of units (e.g. 30 pills, 90 pills)
– Dose
– Intended duration of treatment (e.g. days supply )
– Formulation (e.g. oral, injectable)
Exposure Identification (in claims)
• What you are not going to get:
– Data on over-the-counter medications
– Medications given in-hospital
• What is available?
– Depends on dataset and drug coding system
• Examples of drug coding systems
– National Drug Codes (NDCs)
– Anatomical Therapeutic Chemical (ATCs) Classification
System
– Others
What are NDCs?
• Unique identifier for prescription and non-prescription
drugs in the US
– 10-digit NDC used on drug packaging (and by FDA)
– 11-digit NDC usually used in claims
• NDC allows for identification of manufacturer, dose,
formulation, package form, and size
Examples of NDCs:
• 00002-7597-01: Zyprexa® 10 mg vial
• 50242-0040-62: Xolair® 150 mg vial
• Need crosswalks to make NDCs useable for research!
– ~100,000 NCD codes
What are ATCs?
• Created by WHO in the 1980s
• Hierarchical drug classification system:
– Anatomical main groups (Level 1); Therapeutic subgroup
(Level 2) ; Pharmacological subgroup (Level 3); Chemical
subgroup (Level 4); Chemical Substance (Level 5)
Examples
• sdf
http://www.whocc.no/atc_ddd_index/
Example Dataset #1 (USA)
Example Dataset #2 (USA)
Example Dataset #3 (from Europe)
Estimating Exposure Duration
• Ways to calculate duration of exposure:
– Date of dispensation + days supply
– Date of dispensation + [quantity dispensed/(number of
tablets/day)]
– Date of dispensation + [(quantity dispensed*dose)/DDD]
• Defined daily dose (DDD): assumed average maintenance dose/day for a
drug used for its main indication in adults
• Grace periods between dispensations may account for:
– Underestimations of duration of exposure
– Scheduling difficulties (weekends/holidays)
– Even if patient stopped taking the drug, they could still have
therapeutic concentrations of the drug in their systems
Exposure Considerations
• Healthy-User Effect
– Initiators of preventative therapies might be more likely to have other healthy behaviors
contributing to better overall health and outcomes
– Observational studies: Hormone replacement therapy (HRT) associated with 35-60%
reduction in coronary heart disease (Petitti et al. JAMA 1998)
– RCTs: No effect (Herrington et al. NEJM 2000) or increased risk (Rossouw et al. JAMA
2002)
• Healthy-Adherer Effect
– Patients adhering to treatment might be more likely to have other healthy behaviors
– Patients adherent to statins less likely to have car accidents and more likely to use
screening services (Dormuth et al. Circulation, 2009)
– Patients adherent to placebos less likely to die in a meta-analysis of RCTs (Simpson, BMJ
2005)
• Sick-Stopper Bias
– Patients may stop their medications at the end of life (e.g. hospice)
• Protopathic bias
– Undiagnosed disease leads to treatment of early symptoms which exposure seems to
cause
– Preclinical pancreatic cancer can cause diabetes mellitus and treatment for diabetes (or
diabetes itself) can be associated with pancreatic cancer (Gullo et al. NEJM 1994)
Incident vs. Prevalent Users
• Compare new users (initiators) of one drug to new
users of another drug (preferably) or non-users (Ray
W, AJE 2003)
– Avoid under ascertainment of early events
• Benzodiazepines and falls
– Avoid controlling for factors affected by the drug
– Minimize potential bias from healthy adherer effect and
depletion of susceptibles
• Identifying new users:
– Identify the first dispensation of a given drug (drug-naïve)
– Require a washout period for the drug
Outcome Ascertainment
Types of Outcomes
• Death
• Disease event
• Surgeries or procedures
• Treatment regimen changes
(i.e. discontinuations, switches, therapy
intensification, etc.)
Identifying Outcomes in Claims
• What files you are likely to get:
– Drug files (already described)
– Date of death
– Inpatient file with dates of hospitalization and
diagnostic/procedure codes
– Outpatient file with dates of encounter and
diagnostic/procedure codes
Sources of Information About
Diseases, Procedures, and Surgeries
Primary diagnosis
Secondary diagnoses (ordered, up to a maximum)
Admitting diagnosis
Procedure codes (primary vs. others or not
ordered)
• Diagnosis-Related Groups (DRGs)
• Others
•
•
•
•
Institutional Claim: UB-04 Part I
Institutional Claim: UB-04 Part II
Considerations in Identifying Outcomes in Claims
• Serious events more likely captured in claims data
– White spots on nails vs. hemorrhagic stroke
• Rule-out diagnoses
– Conditions that could explain symptoms
– “Chest pain” rule-out “MI”
• Possibility of “upcoding”
– TIA vs. Ischemic Stroke
Considerations in Selecting the Most Valid
Outcome Algorithm
• Primary position dxs may represent most serious
or resource-intensive condition or the reason for
the encounter
– More valid than diagnoses listed in other positions
• Inpatient diagnoses considered to be more valid
than outpatient diagnoses
• Confirmatory diagnosis or medication
– 2 separate outpatient diagnoses of diabetes or 1
outpatient diagnosis of diabetes + dispensation for
antidiabetic drug
N.B. All increase specificity at the expense of sensitivity!
Algorithm Validation
• Validate your algorithm using a gold standard
– Limited by time/cost/feasibility
• Use a previously-validated algorithm
– Applicability?
• What is more important: sensitivity, specificity,
PPV, or NPV?
– Often PPV is only measure available
– Depends on study objectives
• 100% specificity, RR are unbiased
– High PPV with low sensitivity may hurt power
Thank you for your attention