First analysis of ‘pre-registered’ studies shows sharp rise in null findings