Still Using the P-value, Stop It! A Bayesian Analysis of the STOP-IT Trial

Author(s):
James Klugh; Chelsea Guy-Frank; Claudia Pedroza; Robert Sawyer; Jeffrey Claridge; Lillian Kao

Background:

Bayesian statistics is an alternative to traditional frequentist methods and produces a probability (i.e., likelihood) of treatment benefit or harm rather than the commonly misused p-value. The Trial of Short-Course Antimicrobial Therapy for Intraabdominal Infection (STOP-IT) trial demonstrated a four-day course of antibiotics to be similar to liberal antibiotic administration. However, equivalence was not claimed due to inability to reach the full sample size.

Hypothesis:

We hypothesized that a post hoc Bayesian analysis would demonstrate a high probability of no clinically significant difference in outcomes.

Methods:

The primary outcome in the trial was a composite of surgical-site infection (SSI), recurrent intrabdominal infection (rIAI), or death within 30 days after the index source-control procedure. Secondary outcomes included the components of the primary composite and an SSI or rIAI with a resistant pathogen. Bayesian analyses were used to calculate risk ratios (RR) and 95% credible intervals (CrI) under a neutral prior.  Probabilities were calculated for any benefit of a four-day course of antibiotics (RR < 1) and for risk differences (RD) of less than 5% or 10%. Equivalence in the landmark STOP-IT trial was defined as less than a 10% margin of difference.

Results:

The probability of benefit of a four-day antibiotic course was 55% (Table). The probability was 83% for a RD of 5% or less and 99% for a RD of 10% or less. The probability was 80% that a four-day course reduced the risk of SSI and 72% that it reduced the risk of SSI or rIAI with a resistant pathogen. The probabilities that the four-day course decreased mortality or rIAI were 40% and 30% respectively. There was a >87% probability of a RD of 5% or less and >99% of a RD of 10% or less for all secondary outcomes.

Conclusions:

On Bayesian analysis, a four-day treatment of antibiotics after source control had over a 99% probability of being equivalent to liberal antibiotic administration when using the STOP-IT margin of 10%. Bayesian analysis provides useful information for outcomes with low event rates or limited sample sizes and provides more easily interpretable and applicable information to clinicians and patients.