Observation and experiment with the efficacy of drugs: a warning example from a cohort of nonsteroidal anti-inflammatory and ulcer-healing drug users.

Observational data are well suited for many types of medical research, especially when randomized controlled trials are inappropriate. However, some researchers have attempted to justify routine use of observational data in situations in which randomized controlled trials are normally conducted. Literature searches cannot be used to directly compare the results of the two types of research, because invalid observational studies normally are not publishable in the journal literature. The author created a study (1989-1994) to determine the efficacy of one exposure (ulcer-healing drugs) in preventing the serious upper gastrointestinal toxicity associated with another exposure (nonsteroidal anti-inflammatory drugs (NSAIDs)). A cohort of subjects from Tayside, Scotland, receiving both NSAIDs and ulcer-healing drugs appeared to experience a large rise in their risk of gastric bleeding and perforation (e.g., the rate ratio was 10.00 (95% confidence interval: 6.68, 14.97) when this cohort was compared with one receiving NSAIDs alone). This increased risk was due to confounding. Thus, use of a "restricted cohort design" was not able to eliminate uncontrollable bias. It is possible that if many different studies were carried out, then observational research would be found to be only occasionally useful for studying drug efficacy.