Pandemic response COVID-19 Research Collaboration Platform

Lead Investigator: Dan Hanley, Johns Hopkins University
Title of Proposal Research: Pandemic response COVID-19 Research Collaboration Platform
Vivli Data Request: 6765
Funding Source: Funding is via an administrative supplement mechanism of an existing NCATS award
Potential Conflicts of Interest: None

Summary of the Proposed Research:

Project background: Over 500,000 Americans have died of COVID-19 and many more COVID-19 deaths are expected. Lessons must be learned from successes and failures in response to the pandemic. Therefore, we are conducting a pooled analysis of prior studies and its efficacy for hydroxychloroquine (and chloroquine) in hospitalized patients with COVID-19 – to provide closure for these drugs in this pandemic – we summarize relevant key studies.

Furthermore, vaccines which are being deployed, and therapeutics which have been difficult to operationalize with evidence basis, are needed. Herein, after publication of cognizance of the negative saga might help prevent repetition with other therapeutics and vaccines in the future pandemic response.

Vivli will be a repository for the 7 studies that will be analyzed in response to the pooling effort of the COVID-19 research collaboration platform

Statistical Analysis Plan:

Statistical analyses

Design-stage checks

Before the outcome data are shared with the analysis team, we will examine the distributions of baseline covariates within and between trials. If there are individuals with extreme values of baseline covariates, we will check their data with trial investigators, and potentially exclude them from the outcome analysis.

If there are substantial covariate imbalances (with respect to covariates not included in our model) between treated and control groups, either within a larger trial or overall, we will consider including that covariate in our outcome model. We will also examine baseline covariates for missing data.

Primary outcome – model

We will fit a Bayesian proportional odds ordinal regression model for NCOSS measured at day 28-30. In addition to treatment whether the treatment included azithromycin, and study, the model will include the following individual-level covariates: sex, age, Charlson score, BMI, and baseline COVID outcome scale.

The coding and reference levels for these variables are as follows:

Covariate Coding Reference level Binned version
Sex female = +1/2, male = -1/2 0 (midpoint) n/a
Age (age in years – 60)/10 60 years old 18-30, 30-50, 50-70, 70-80, 80+
azithromycin arm

1 = in an azithromycin arm,

0 = in a non-azithromycin arm

0 = in a nonazithromycin arm n/a
Charlson score

0, 1, 2, 3, 4, ≥5

(the last category will be coded as a numeric 5 in the fixed effect part of the model)

0 0, 1-4, ≥5
BMI (BMI – 25)/5 BMI of 25 ≤20, 20-25, 25-30, 30-35, ≥35
baseline COVID ordinal scale (NCOSS) indicators for levels 2-5 of the NCOSS 5 = hospitalized, not requiring supplemental oxygen (the lowest possible value for inpatients) n/a

Model description. Let individual patients be indexed by i. Each patient has a vector of baseline covariates ci and is assigned to treatment ti, either hydroxychloroquine or chloroquine (ti = 1) or standard of care/placebo (ti = 0). Let the primary outcome for individual i be denoted yi with levels indexed by l. Finally, let yi(0) and yi(1) be the potential outcomes for individual i under control t = 0 and treatment t = 1 respectively, regardless of whether in fact ti = 0 or ti = 1.

For each individual i, the model fits a vector of predicted probabilities that individual i has outcome yi, based on their covariates ci and treatment status ti. Changing the value of an individual’s ti in the model generates a prediction for the counterfactual outcome. In a cumulative ordinal model, the probabilities are linked to a linear predictor ηi via

P( yil given ci , ti )=logit−1(θlηi )

for l = 1, …, 6. The θl are cut points (also called intercepts) that are common to all individuals.

The linear predictor takes the form

 ηi = xiT  β+αstudy+γCharlson score+δbaseline NCOSS

+ti×[ τtreat+xi T βtreat+αstudy treat +γCharlson score treat +δbaseline treat NCOSS ]


  • xi (a function of ci) contains the predictors sex; age and (age)2; BMI and (BMI)2; the

numeric Charlson score, with all scores 5 and above coded as 5; the numeric baseline NCOSS status; and an indicator for azithromycin arm;

  • the coefficients β and βtreat are given uniform priors;
  • the coefficients αstudy, γCharlson score, δbaseline NCOSS, and the corresponding treatment terms, are all modelled as independent mean-zero Normal random effects, each with their own standard deviation parameter distributed as a half Student-t distribution with 3 degrees of freedom and scale parameter 10.

The model will be fit using R, and the library ‘brms’.

To assess the sensitivity of our conclusions to modeling choices, we will

  • repeat the core analysis with weakly informative N(0, 52) priors on the fixed effect coefficients, and more conservative half Student-t (df = 3, scale = 5) priors on the group level standard deviations,
  • repeat the core analysis with a stopping ratio (i.e., hazard) model allowing for hazardspecific effects,
  • repeat the core analysis with study-level covariates, for example age and/or Charlson score.

Primary outcome – effects of interest

For all of our effects of interest—the overall effect of treatment; the between-study heterogeneity; the individual-level treatment-covariate interactions—we will produce two kinds of effect estimates:

  • Conditional effect estimates. These correspond to coefficients and fitted curves from the regression model. They admit conditional interpretations, with the other included covariates set to their reference levels (see above). When reporting conditional estimates, we will average over the superpopulation distribution of between-study effects.

For the conditional effect estimates, our scale will be the relative risk of “mechanical ventilation or death” (levels 1 and 2 of the NCOSS outcome scale).

  • Standardized (post-stratified) effect estimates. These represent the effect of the treatment in a given population, averaged over that population’s distribution of individual-level covariates.

For examining treatment heterogeneity within categorical covariates and between studies, we will standardize to the empirical covariate distribution. For treatment interactions with continuous covariates, we will divide the study population into covariate-based bins (see table above) and standardize to the empirical covariate distribution within each bin.

We will produce standardized estimates by drawing from the posterior predictive distribution of unobserved, counterfactual outcomes, combining these with the (fixed) observed outcomes, and regressing these potential outcomes against the corresponding treatment indicator.

The advantage of standardized over conditional effect estimates is that they are directly comparable to unadjusted “plug-in” maximum likelihood effect estimates. We will present both of these in our results, with the maximum likelihood confidence intervals uncorrected for multiple testing.

For the standardized effect estimates, our effect measure will be the “proportional odds ratio” corresponding to the estimate from a proportional odds model fit to the data by maximum likelihood. We will present this estimate as our primary assessment of the effect of treatment.

We will also consider presenting selected absolute effects: probabilities of the different outcome categories under the treatment and control conditions.

Secondary outcomes and safety outcomes (excluding death and lowest NCOSS score)

Because both secondary outcomes are potentially compromised by truncation-by-death, we will not pre-register a model-based analysis; our analysis of these outcomes will be primarily exploratory.

Model checking and robustness

For the primary outcome model, we will conduct and report posterior predictive checks and other model checking diagnostics, and comment on the extent to which they cast doubt on the main analysis. We may conduct follow-up analyses if model fit is poor.

For each treatment-covariate interaction in the model, we will examine, using exploratory plots, the association between the estimated study-specific treatment effect and the study-specific mean or median level of the covariate.

We may also examine in an exploratory manner the effect of treatment duration and dose.

Requested Studies:

Treating COVID-19 With Hydroxychloroquine: A Multicenter Randomized, Double-blind, Placebo-controlled Clinical Trial in Hospitalized Adults
Sponsor: NYU Langone Health
Study ID: NCT04369742
Sponsor ID: 20-00463

Hydroxychloroquine vs. Azithromycin for Hospitalized Patients With Suspected or Confirmed COVID-19 (HAHPS): A Prospective Pragmatic Trial
Sponsor: Intermountain Health Care, Inc.
Study ID: NCT04329832
Sponsor ID: 1051355

WU 352: Open-label, Randomized Controlled Trial of Hydroxychloroquine Alone or Hydroxychloroquine Plus Azithromycin or Chloroquine Alone or Chloroquine Plus Azithromycin in the Treatment of SARS CoV-2 Infection
Sponsor: Washington University School of Medicine
Study ID: NCT04341727
Sponsor ID: 202003188

COVIDMED-Comparison Of Therapeutics for Hospitalized Patients Infected With SARS-CoV
Sponsor: Bassett Research Institute
Study ID: NCT04328012
Sponsor ID: NCT04328012

A Randomized, Controlled Clinical Trial of the Safety and Efficacy of Hydroxychloroquine for the Treatment of COVID-19 in Hospitalized Patients
Sponsor: Queen’s Medical Centre
Study ID: NCT04345692
Sponsor ID: RA-2020-018

Pragmatic Factorial Trial of Hydroxychloroquine, Azithromycin, or Both for Treatment of Severe SARS-CoV-2 Infection
Sponsor: Duke University
Study ID: NCT04335552
Sponsor ID: PRO00105339

Treatment in Patients With Suspected or Confirmed COVID-19 With Early Moderate or Severe Disease: A Randomized Clinical Trial
Sponsor: LCMC Health
Study ID: NCT04344444
Sponsor ID: COVID 2020-001

Public Disclosures:

Di Stefano L, Ogburn EL, Ram M, Scharfstein DO, Li T, Khanal P, Baksh SN, McBee N, Gruber J, Gildea MR, Clark MR, Goldenberg NA, Bennani Y, Brown SM, Buckel WR, Clement ME, Mulligan MJ, O’Halloran JA, Rauseo AM, Self WH, Semler MW, Seto T, Stout JE, Ulrich RJ, Victory J, Bierer BE, Hanley DF, Freilich D. Pandemic Response COVID-19 Research Collaboration Platform for HCQ/CQ Pooled Analyses. Hydroxychloroquine/chloroquine for the treatment of hospitalized patients with COVID-19: An individual participant data meta-analysis. PLoS One. 2022 Sep 29;17(9):e0273526. doi: 10.1371/journal.pone.0273526