Can observational data be used to determine drug effectiveness?

Data collected as a standard part of patient care provides an opportunity to provide
evidence on drug effectiveness in a real world (RW) setting. Randomised Controlled Trials
(RCT) typically select a patient population that will maximize the chance of observing a
statistically significant treatment effect and often the patients included are the healthiest
subset of that patient population. As a result evidence on treatment effectiveness can be
lacking for patients who would not have met the inclusion and exclusion criteria of the
RCT such as: those with a more severe or mild form of the disease, patients with
comorbidities, patients with concomitant medications, and pediatric, geriatric, and
pregnant patients.
Patients and doctors must extrapolate RCT results to their patient case which may not be
appropriate. Ideally the doctor and patient would have evidence on the effectiveness of
the drug in patients with a similar profile. Should this project be successful then this would
pave the way for providing such evidence. An additional advantage of RW data is the
availability of long-term outcomes related to the disease which often cannot be included in
RCTs.
Methodology A suitable drug with Phase 3 RCT results available and drug prescription
and outcome measures recorded in the Clinical Practice Research Datalink (CPRD) RW
data will be selected. Data linkage may be employed, linking hospital episode or ONS
mortality data to the CPRD data to increase the likelihood of capturing all available
outcome data. Subpopulations of interest will be identified by comparing the baseline
characteristics of the RCT population to the population who were eventually prescribed
the drug and from current clinical knowledge.
A cohort of patients in the RW data will be selected matching the RCT patients. Various
methods may be trialled to select a suitable RW data cohort: applying the inclusion and
exclusion criteria of the RCT to the RW dataset, selecting a cohort of patients with an
2
overall baseline demographic and medical history distribution matching the trial cohort,
and matching at the patient level on the individual patient characteristics. Though widely
used, propensity score matching (PSM) may increase imbalance and bias (PSM paradox,
King and Nielsen 2016). As part of this project different matching techniques will therefore
be explored to select the best method both in matching from the RCT patients to the RW
cohort and in matching between treatment groups.
Once a suitable RW cohort has been selected the treatment effectiveness will be
analysed. All potential sources of bias will be considered with suitable strategies
implemented to account for them. Sources of confounding will be identified and addressed
with methods such as marginal structural models. Multiple imputation may be used to deal
with missing data in the conduct of sensitivity analyses.
If the analysis of the matched RW cohort provides treatment estimates consistent with the
results of the RCT then this may be considered validation of the use of real world data for
assessing drug effectiveness in that particular drug/disease area. Sensitivity analyses will
be performed to test the robustness of the results; if they support the primary result one
can proceed to looking at the other cohorts of RW patients (those who would not have
been included in the RCT). If the RW results do not show drug effectiveness similar to that
predicted by the RCT then the reasons will be explored - data quality, method of selection
of patients, analysis model used, or finally due to the drug effectiveness truly not being the
same in real life as to the efficacy demonstrated in the trial.
This studentship will meet the MRC skill priority of quantitative skills applied to electronic
health records. Expertise will be gained in pharmacoepidemiology and in medical big data
analysis.