To identify studies for inclusion in this meta-analysis, two researchers independently conducted searches of MEDLINE®, Cochrane Library documents. The keywords used for these searches were: proton pump inhibitors, PPI, rabeprazole, omeprazole, esomeprazole, lansoprazole, pantoprazole, ranitidine; healing rates, gastric ulcer, stomach ulcer, gastric ulcer disease, peptic ulcer, peptic related disorders; and clinical trials, randomized clinical trials and multicenter studies. Hand searching of medical journals was carried out as well as review of reference lists of other meta-analyses, monographs, pharmacoeconomic studies and reviews.
The search included papers published between January 1990 to July 2001, with no language limitations.
The two researchers independently reviewed the papers and selected all those described as randomized controlled clinical trials, double blind, short-term treatment (less than six months) in patients with endoscopically confirmed diagnosis of gastric ulcer that also included at least one group treated with a PPI. If a paper did not report healing rates, but reported improvement in clinical symptoms, this could be included for the analyses of symptoms.
The decision regarding whether to include a paper was not related to trial results. Papers related to chemical properties of PPI, pharmacokinetic, pharmacodynamic, acidity studies, concomitant duodenal ulcers or gastric ulcers complicated with hemorrhage or perforation, PPI combined with antibiotics or with treatment of Helicobacter pylori were excluded. Papers or abstracts with insufficient data, preliminary results, duplicate reports of the same trials or with more complete reporting in another paper were also excluded. The two researchers reviewed the studies independently and any differences in the data retrieved from the papers were resolved by consensus. The information entered into a database included: study design, population characteristics, diagnosis, severity, PPI and comparator treatment regimens, healing rates and clinical symptoms as well as alcohol consumption, smoking and caffeine use.
Ulcer diagnosis and healing had to be assessed by endoscopy; we accepted the primary study authors' definition for gastric ulcer diagnosis and healing.
Results were retrieved for both intention to treat and per protocol analysis, if the paper provided such data. If the authors reported per protocol results only and mentioned the initial number of subjects in the study, we estimated the healing rates for the intention to treat analysis.
The healing rate ratio (RR) was calculated for each trial by dividing the healing rate of the PPI by the rate with the comparator at the specified time point. Chi-squares and 95% confidence intervals (CI) were calculated using the exact method described by Miettinen [11
The appropriateness of combining results from individual studies was examined using the Wald χ2 for statistical homogeneity. We considered a p value less than 0.05 to be significant for all statistical tests. There were no adjustments for multiple comparisons.
As there was no statistically significant evidence of heterogeneity, the healing rates for each treatment were pooled across trials by time point. Overall healing rates of PPIs at four and eight weeks were estimated by pooling the rates for the recommended daily doses: 30 mg/d for lansoprazole, 20 mg/d for omeprazole, 40 mg/d for pantoprazole. Rabeprazole was pooled using dose of 20 mg/d at three and six weeks. Individual trial results were weighted using the inverse of the variance of the RR estimated for each trial. The overall RR was then estimated by dividing the weighted average healing rates for each treatment with the estimates for placebo, ranitidine or omeprazole. These overall RR were assessed using the Mantel-Haenszel χ2
test and 95% CI were calculated using the exact method [11
To assess the possibility of publication bias, the RR point estimates were plotted against year of publication and sample size. To identify any studies that exerted a disproportionate influence on the summary treatment effect, we deleted individual studies one at a time. To investigate the possible consequences of including abstracts for which we could not obtain full reports, we excluded these data in secondary analyses.