We analysed the data with univariate and multivariate statistical techniques, using SAS (version 6.12). Analyses were all based on intention to treat and were two tailed. We judged a p value of 0·05 as significant. Numerical variables are presented as mean (SD) or median (IQR) and were compared with the t test or Wilcoxon's rank sum test as appropriate for variables displaying non-normal behaviour. Group testing of categorical data are presented as number (%) and were compared with χ 2 test or Mantel-Haenszel χ 2 test for linear association of ordinal variables, with Taylor-series relative risks applied. We analysed differences in knowledge scores between groups at the postnatal assessment for group effects with analysis of variance, controlling for the independent effects of factors that might reasonably be thought to affect knowledge outcomes—ie, age, social class, and baseline knowledge—and also controlling for ethnic origin, social isolation, involvement of the father of the baby, and homelessness, since these factors seemed imbalanced between the two groups. We ascertained the difference between groups with respect to duration of breastfeeding by Kaplan-Meier survival analysis with the log rank test.