Appendix 1

Reported systems for mortality monitoring in general practice.

Criteria	Retrospective examination of Shipman dataset7^,8	Northern Ireland pilot10^,11
Purpose or mortality monitoring	Detection of murder	Primarily, quality improvement; secondarily, detection of murder

Description of monitoring system	Total mortality for individual GPs with correction for over-dispersion was monitored for 1993–2000. Normal log-likelihood ratio CUSUM charts used for detection system. Alarm signals were subsequently investigated in two ways, one based on an open quality-improvement model,9 one more forensic and confidential.4 Neither found cause for concern	Five-year aggregated case-mix-adjusted total mortality for practices was monitored for 1996–2000. Cross-sectional Shewhart charts with three standard deviation control limits used for detection system. Alarm signals were openly investigated using the pyramid model of quality improvement. No cause for concern was found

Coverage — what percentage of practices/GPs/nurses were monitored?	11.8% of practices 37.3% of GPs individually monitored No nurse individually monitored Poor data quality in 1990s meant most practices and GPs excluded	77.5% of practices No individual monitoring of GPs or nurses Nearly a quarter were excluded because of experiencing a major merger or split

Ability to detect murderers	Practice monitoring	Uncertain
	eSDR₇ = 74.7% (k = 1, h = 3)^a
	eSDR₇ = 41.2% (k=1, h = 5)
	GP monitoring
	eSDR₇ = 96.6% (k = 2, h = 3)^a
	eSDR₇ = 82.4% (k = 2, h = 5)

False alarms (percentage of GPs/practices signalling, where later investigation concludes there is no cause for concern)	3.3% of GPs (k = 2, h = 3)	15.8% of practices
	2.3% of GPs (k = 2, h = 5)	Reflects quality of data available for case-mix adjustment

How long must reliable data be available for these conclusions to hold?	7 years	5 years

Comments	Conclusions rely on being able to monitor individual GPs' mortality rates which is no longer even theoretically feasible with the shift to practice-based registration in 2004	Effectiveness of mortality monitoring to detect mass murder is assumed rather than directly examines

↵a eSDR₇ (estimated successful detection rate after 7 years of monitoring) = proportion of truly out of control units successfully detected after 7 years. k and h are chart parameters. k relates to the size of the excess mortality the chart is tuned to detect. For the Shipman dataset, k = 1, ≈ 6 excess deaths for a system monitoring at GP level, and ≈ 13 excess deaths at practice level (for k = 2, ≈12 and 26 respectively). eSDRs shown therefore refer to charts tuned to detect Shipman in his mid-career (approximately 10 murders per year). h is the threshold that the chart statistic has to exceed to trigger an alarm (larger h means deviation from expected mortality has to be greater to signal an alarm). CUSUM = cumulative sum.