Population-Level Estimation

Population-Level Estimation (PLE) enables comparative effectiveness and safety research using causal inference methods applied to observational OMOP CDM data. PLE goes beyond descriptive analyses (characterization, incidence rates) to estimate the causal effect of a treatment on an outcome while accounting for confounding by indication.

Parthenon implements two major causal inference approaches: the New-User Cohort Design (CohortMethod) for between-patient comparisons and Self-Controlled Case Series (SCCS) for within-patient comparisons. Both are available as fully configurable analysis types with design forms, diagnostics, and result visualization.

---

New-User Cohort Design (CohortMethod)

The CohortMethod approach compares outcomes between a target cohort (treatment) and a comparator cohort (alternative treatment or no treatment) of new users, adjusting for baseline confounders.

Study Design Overview

Target Cohort:      New users of Drug A (with washout)
Comparator Cohort:  New users of Drug B (with washout)
Outcome Cohort:     Clinical event of interest (e.g., MI, stroke)

      Index                              Outcome?
        |<---- Time at Risk ----------->|
        |                                |
        Baseline covariates              |
        (pre-index features)             Follow-up end

Creating a CohortMethod Analysis

1. Navigate to Analyses and select the Estimations tab.
2. Click New Analysis and select Estimation.
3. Configure the design:

Cohort Selection:

Setting	Description
Target cohort	Treatment cohort (e.g., new users of Drug A)
Comparator cohort	Alternative treatment cohort (e.g., new users of Drug B)
Outcome cohorts	One or more outcomes to estimate effects for

Model Settings:

Setting	Description	Default
Model type	Cox proportional hazards regression	cox
Time at risk start	Days after index to begin observation	0
Time at risk end	Days after index/cohort end to stop observation	0 (at cohort end)
End anchor	Whether end offset is from cohort start or cohort end	cohort end

Propensity Score Settings:

The propensity score (PS) is the estimated probability of receiving the target treatment given baseline covariates. Parthenon supports three PS-based adjustment methods:

Method	How it works	Configuration
PS Matching	Match each target patient to one or more comparator patients with similar PS	ratio: match ratio (1:1, 1:4); caliper: maximum PS distance for a valid match (default 0.2)
PS Stratification	Divide patients into strata (quintiles) by PS and compare within strata	strata: number of strata (default 5)
IPTW	Weight each patient by inverse of their PS to create a pseudo-population	trimming: truncate extreme PS values (default 0.05)

Setting	Description	Default
Enable PS	Whether to use propensity score adjustment	Enabled
Trimming	Remove patients with extreme PS values (PS < trim or > 1-trim)	0.05

Covariate Settings:

Covariates are the baseline features used to build the propensity score model. Parthenon uses large-scale regularized regression (LASSO) over thousands of covariates automatically extracted from the CDM.

Setting	Description	Default
Demographics	Age, gender, race, ethnicity, year of birth	Enabled
Condition occurrence	All condition diagnoses in the time window	Enabled
Drug exposure	All drug exposures in the time window	Enabled
Procedure occurrence	All procedures in the time window	Disabled
Measurement	Lab values and vitals	Disabled
Observation	Clinical observations	Disabled
Time windows	Lookback period for covariates	[-365, 0]

Negative Control Outcomes:

Negative controls are outcomes with no plausible causal relationship to the exposure. They are used to calibrate effect estimates and detect residual systematic bias.

• Add concept IDs for negative control outcomes (e.g., appendicitis, ingrown toenail for a cardiovascular drug study).
• After execution, the negative control distribution is used to empirically calibrate the true outcome effect estimates.

Negative controls are essential

Including 50--100 negative control outcomes is considered best practice for OHDSI studies. If the negative control effect estimates cluster around HR = 1.0 with appropriate coverage probability, the analysis is well-calibrated. If negative controls show systematic bias (e.g., most HRs < 1.0), confounding adjustment is insufficient.

---

Self-Controlled Case Series (SCCS)

SCCS uses each patient as their own control, comparing event rates during exposed time windows to unexposed time within the same patient. This design inherently controls for all time-invariant confounders (genetics, chronic conditions, socioeconomic status) because each patient serves as their own baseline.

When to Use SCCS

SCCS is particularly suited for:

• Vaccine safety studies: Comparing adverse event rates in the days following vaccination vs. baseline periods
• Acute adverse event detection: Short-term risks after drug initiation
• Drug-event associations: When the exposure is transient and the outcome is acute
• Situations with strong unmeasured confounding: Since patients serve as their own controls, time-invariant confounders are eliminated by design

Creating an SCCS Analysis

1. Navigate to Analyses and select the SCCS tab.
2. Click New Analysis and select SCCS.
3. Configure the design:

Cohort Selection:

Setting	Description
Exposure cohort	The drug or vaccine exposure to study
Outcome cohort	The adverse event or clinical event of interest

Risk Windows:

Risk windows define the time periods during which you hypothesize the exposure increases (or decreases) the event rate. Multiple risk windows can be defined:

Setting	Description	Default
Start	Days after era start (or era end) to begin the risk window	1
End	Days after era start (or era end) to end the risk window	30
Start anchor	era_start or era_end	era_start
End anchor	era_start or era_end	era_start
Label	Descriptive name for this risk window	"Risk window 1"

Example risk windows for a vaccine safety study:

• Window 1: Days 1--7 after vaccination ("acute risk")
• Window 2: Days 8--28 after vaccination ("sub-acute risk")
• Window 3: Days 29--42 after vaccination ("extended risk")

Model Settings:

Setting	Description	Default
Model type	simple (standard SCCS) or with additional adjustments	simple

Study Population:

Setting	Description	Default
Naive period	Days at the start of observation excluded from analysis	180
First outcome only	Whether to count only the first event per patient	Enabled

---

Effect Estimates and Results

CohortMethod Results

After execution, the estimation results page shows:

Metric	Description
Hazard Ratio (HR)	Estimated relative hazard of the outcome in the target vs. comparator cohort
95% Confidence Interval	Precision of the HR estimate
P-value	Statistical significance (null hypothesis: HR = 1.0)
Target events	Number of outcome events in the target cohort
Comparator events	Number of outcome events in the comparator cohort
Target persons	Number of patients in the adjusted target cohort
Comparator persons	Number of patients in the adjusted comparator cohort

Interpretation Guide

HR	Interpretation
HR = 1.0	No difference between target and comparator
HR < 1.0	Target treatment is associated with lower risk
HR > 1.0	Target treatment is associated with higher risk
HR = 0.75 (0.65--0.86)	25% risk reduction, statistically significant
HR = 1.42 (0.98--2.05)	42% risk increase, not statistically significant

SCCS Results

SCCS produces an Incidence Rate Ratio (IRR) comparing event rates during risk windows to unexposed time:

IRR	Interpretation
IRR = 1.0	No change in event rate during risk window
IRR = 3.2 (2.1--4.8)	3.2x higher event rate during exposure, significant
IRR = 0.8 (0.5--1.3)	No significant change in event rate

---

Diagnostics

Proper interpretation of PLE results requires checking diagnostic outputs:

Propensity Score Diagnostics (CohortMethod)

• PS distribution plot: Shows the overlap between target and comparator PS distributions. Good overlap (clinical equipoise) is required for valid estimation.
• Covariate balance: Table of SMDs before and after PS adjustment. All SMDs should be < 0.1 after adjustment.
• Preference score: A rescaled PS that accounts for different cohort sizes.

Negative Control Calibration

• Null distribution: Plot of negative control effect estimates. Should center around 1.0.
• Calibrated estimates: True outcome estimates adjusted for empirical bias observed in negative controls.

Check diagnostics before interpreting results

An HR that appears statistically significant may be entirely driven by residual confounding if diagnostics show poor covariate balance or biased negative controls. Always review diagnostics before drawing causal conclusions.

---

Evidence Synthesis

Parthenon includes an Evidence Synthesis analysis type (accessible from the Evidence Synthesis tab on the Analyses page) for combining PLE results across multiple data sources using meta-analytic methods.

Available Methods

Method	Description
Fixed-effects	Assumes a common true effect across all sources; weights by precision
Random-effects	Allows the true effect to vary across sources; accounts for heterogeneity
Bayesian	Bayesian hierarchical model with configurable chain length and burn-in

Configuration

Setting	Description	Default
Estimates	Select individual PLE results to combine	(required)
Method	Meta-analysis method	bayesian
Chain length	MCMC chain length (Bayesian only)	1,100,000
Burn-in	MCMC burn-in iterations (Bayesian only)	100,000
Sub-sample	MCMC thinning interval (Bayesian only)	1,000

Results include a forest plot showing individual source estimates and the combined meta-analytic estimate with confidence/credible intervals.

---

Best Practices

1. Use active comparators: Comparing Drug A to Drug B (active comparator) is far more robust than comparing Drug A to "no treatment." Active comparator designs reduce confounding by indication.

2. Require new-user designs: Both target and comparator cohorts should be incident (new-user) cohorts with adequate washout periods to avoid prevalent-user bias.

3. Include comprehensive covariates: Enable demographics, conditions, and drugs at minimum. The LASSO regularization handles the dimensionality --- more covariates generally improve confounding adjustment.

4. Always use negative controls: 50--100 negative control outcomes provide the empirical calibration needed to assess and correct for residual systematic bias.

5. Check all diagnostics before interpreting: PS overlap, covariate balance, and negative control calibration are not optional quality checks --- they are required for valid causal inference.

6. Consider SCCS for self-controlled questions: When time-invariant confounding is a major concern and the exposure is transient, SCCS provides stronger causal evidence than CohortMethod.

7. Synthesize across sources: Single-database results may reflect database-specific biases. Meta-analysis across multiple databases provides more robust evidence.

R runtime connection

PLE analyses are designed to leverage the OHDSI HADES R packages (CohortMethod, SelfControlledCaseSeries, EvidenceSynthesis) running in the R runtime container. The configuration UI is fully functional for specifying analysis designs. Execution against the R runtime is being connected. In the interim, designs can be exported as R-ready configuration objects for manual execution.