Analysis of Registered Clinical Trials in Gastroenterology, 2007-2019

This cross-sectional study investigates the association of gastrointestinal clinical trial characteristics with early discontinuation, results reporting, and methodological rigor from 2007 to 2019.


Statistical Methods
Descriptive statistics were used to summarize trial data and differences between categorical variables, using a 2sided Pearson chi-square test for significance. Missing data were included as a separate category. The 147-month study period was classified at the midpoint into an "early" period (October 1, 2007 to December 31, 2013) and a ''late'' period (January 1, 2014 to December 31, 2019).
Time series analyses summarized all year-to-year changes using absolute average annual growth rates (AAGR). All year-to-year analyses included only years with a full 12-month collection of data (2008 -2019). Significant monotonic trends were evaluated with post hoc Mann-Kendall tests and Ordinary Least-Square (OLS) p-values. To adjust for multiple comparisons, the Bonferroni correction was used for both Mann-Kendall and OLS p-values to control for the familywise error rate. Finally, proportional independence of each group was assessed across ordinal time using the Cochran-Armitage test.
Survival analysis was performed with both primary outcomes of early discontinuation and results reporting. We excluded all trials that reached completion without early discontinuation or that remained ongoing at the cutoff for analysis (September 30th, 2017). Trials lost to follow-up were censored at the date of their last update.
Cross-sectional Cox proportional hazard regression was performed to determine bivariate and multivariate adjusted hazard ratios (HR). The variables that were adjusted for include: phase, number of arms, enrollment number, enrollment type, DMC presence, trial duration, number of facilities, number of regions, low/middle/high income country status, allocation, primary purpose, masking, funding source, study first submitted date, results first submitted date, start date, and primary completion date. Cox regression models were chosen because time to event analyses, such as time to early discontinuation, were important for our research question. Cox regression models are more appropriate for time to event analysis (while logistic regression would be more appropriate for fixed outcomes with less emphasis on time to event). A sensitivity analysis for trial feature selection was conducted by fitting a number of Cox proportional hazards models using (1) all covariates of interest, (2) LASSO regression with a lambda tuning parameter created from 10-fold cross-validation, and (3) two models using stepwise selection (Wald statistics and Akaike Information Criterion, known as AIC) as the criteria of interest. These models also retained variables known to be important by domain knowledge. Models were then assessed using Schoenfeld residuals and Variance Inflation Factors. All analyses were 2-sided with statistical significance set at the α = .01 level and performed using R version 3.5.0 software (R Foundation, Vienna, Austria). Causation cannot be implied.
Missing data were handled using multiple imputation by chained equations (MICE) after a sensitivity analysis revealed that the missing data were not completely random. Recent literature concludes that the number of imputations should be similar to the percentage of incomplete cases, which in our data ranges from 0% to 10.2%. 1-2 A number of 10 imputed datasets was chosen with 20 cycles to reach convergence of the sampling distribution of imputed values. 3 Finally, all analytic variables were included as covariates with continuous, dichotomous, and categorical data modelled using predictive mean-matching, Bayesian logistic regression, and Bayesian polytomous regression, respectively. Rubin's Rules were applied to pool parameter estimates after separate results estimation.

Further Describing the Presence of Rigorous Methodology
As noted in our manuscript, only 12% of phase 3 trials employed a rigorous methodology as defined in our methods. U.S. government-funded trials had the highest proportion of trials meeting this definition (19%), while academicfunded trials had the lowest (5.3%). Table 1 also denotes hazard ratios for early discontinuation and results reporting as analyzed by all of the individual components of our rigorous methodology definition, also as described below. Specifically, we found that U.S. government-funded trials had the lowest risk of early discontinuation of any funding source (HR 0.63, 95% CI 0.48-0.83, P=0.001, industry reference; academic institutions HR 0.99, 95% CI 0.86-1.16, P=0.993). Estimated enrollment had the strongest association with early discontinuation; trials with an estimated enrollment < 50 participants were more likely to be discontinued than larger trials (HR 0.06, 95% CI 0.05-0.07, P<0.001 versus HR 0.01, 95% CI 0.00-0.01, P<0.001). Other factors associated with early discontinuation (without categorizing by funding source) included randomized versus non-randomized trials (HR 1.84, 95% CI: 1.56-2.18, P<0.001). Blinding and DMC presence were not statistically significantly associated with early discontinuation (Table 1). Academic-funded trials also had the lowest odds of reporting results compared to other funders (academic institutions HR 0.39, 95% CI 0.31-0.49, P<0.001; industry reference; U.S. government HR 0.78, 95% CI 0.55-1.11, P=0.166). Other factors associated with results reporting (without categorizing by funding source) included randomized versus non-randomized trials (HR 0.67, 95% CI 0.54-0.83, P<0.001). Estimated enrollment number, blinding, and DMC presence were not statistically significantly associated with results reporting (Table 1).

Clinical Trials Research Team
Each labeler was provided a set of rules and examples to manually categorize clinical trials. Each labeler first achieved >90% agreement in categorizing a training set of trials (set by author Marija Kamceva). A sub-set of each labeler's categorization was reviewed by Marija Kamceva and Nirosha Perera to ensure agreement.