PDA-OTA

One-shot Transfer Algorithm

Where Distributed
Studies Get Done.

PDA-OTA is the online workspace that powers PDA collaborations. It helps researchers and data partners coordinate studies, exchange aggregated data securely, monitor project status in real time, and keep multi-site analyses moving efficiently. By centralizing workflow instead of centralizing raw data, PDA-OTA makes collaboration more practical at scale.

Launch PDA-OTA
Lorem ipsum dolor sit amet, consectetur adipiscing elit.

How It Works

The OTA Workflow

Local Estimation

Each site fits its own local model using patient-level data that never leaves the institution.

Surrogate Transfer

Sites transmit only lightweight summary statistics — gradients and Hessians — to the coordinating center.

One-Shot Aggregation

The coordinating center aggregates summaries using a single Newton-Raphson update step.

Pooled-Equivalent Result

The resulting estimate is asymptotically equivalent to the pooled analysis — in one round.

PDA-OTA in Action

Linear Regression

Federated OLS with exact pooled equivalence. Supports high-dimensional covariates and heterogeneous site populations.

Logistic Regression

Binary outcome models for comparative effectiveness studies. Validated across PCORnet multi-site deployments.

Cox Proportional Hazards

Survival analysis for time-to-event outcomes. Essential for pharmacovigilance and longitudinal EHR studies.

Poisson Regression

Count data and incidence rate modeling for population-level epidemiological analyses.

SCCS Model

Self-controlled case series for vaccine safety surveillance and drug adverse event detection.

Stratified Cox Regression

Survival models that let baseline hazards and covariate distributions differ by site, so between-site heterogeneity does not bias the hazard ratio (ODACH).

Rare-Event Survival Analysis

One-shot survival estimation tuned for rare outcomes in claims networks, using a Pade approximation of the partial likelihood (Fed-Pade).

Competing Risks Models

Estimation of cause-specific and subdistribution effects when several event types compete, with a heterogeneity-aware variant for differing site populations (ODACoR, ODACoR-H).

Conditional Logistic Regression

Pairwise conditional likelihood that absorbs site-specific baseline risk, staying robust when event rates vary widely across sites (dCLR).

Zero-Inflated and Hurdle Count Models

Two-part models for count outcomes with excess zeros, such as avoidable hospitalizations and adverse event counts (ODAH).

Multi-Categorical Regression

Nominal and ordinal outcome models for multi-level clinical endpoints, with an option to allow site-specific intercepts (ODACAT, ODACAT-H).

Generalized Linear Models

One-shot lossless GLM estimation across the exponential family, with a heterogeneity-aware variant for networks whose populations differ sharply (COLA-GLM, COLA-GLM-H).

Linear Mixed Models

Lossless one-shot fitting of linear mixed models with site-level random effects, for continuous outcomes such as length of stay (DLMM).

Generalized Linear Mixed Models

Random-effects models for binary and other non-normal outcomes, fitted losslessly through distributed penalized quasi-likelihood (dPQL, COLA-GLMM).

Hospital Profiling

Risk-standardized event rates and provider rankings computed from aggregated summaries, so hospitals can be compared without pooling records (dPQL, dGEM).

High-Dimensional Regression

Penalized estimation and variable selection when covariates outnumber events, with adjustment for heterogeneity across sites (ADAP).

Target Trial Emulation

Federated emulation of a hypothetical randomized trial on real-world data, covering cohort construction through effect estimation (LATTE, FLAT-WHITE).

Transfer Learning

Borrowing strength from the full network to sharpen site-specific estimates when one hospital alone has too few cases (hetero-latent-TL).

Latent Class Analysis

Joint subphenotype discovery across institutions, learning shared latent classes with site-specific class proportions (dMLCA).

Finite Mixture Models

Distributed EM for mixture models when the underlying subpopulations differ across sites (dist-EM, one-shot distributed EM).