Comparison of statistical methods used to meta-analyse results from interrupted time series studies: an empirical study

Table 1 Characteristics of included meta-analyses and ITS studies

		Meta-analyses (N = 17) n (%) or median (IQR)
Discipline/Topic^a	Public health	15 (88)
Discipline/Topic^a	Crime	2 (12)
Interruption target	Population	10 (59)
	Organisation	5 (30)
	Individual^b	1 (6)
	Combination	1 (1)
Interruption types^a	Policy change	12 (71)
	Practice change	5 (39)
	Communication (campaign)	3 (18)
	Educational method	3 (18)
	Exposure	2 (12)
Outcome types^a	Rate	6 (35)
	Count	5 (29)
	Proportion	2 (12)
	Combination ^c	2 (12)
	Continuous	1 (6)
	Probability	1 (6)
Number of ITS studies	Per meta-analysis	11 (5, 15)
Number of time series datapoints	ITS level (N = 282)	52 (27, 61)
Number of time series datapoints	Meta-analysis level^d	40 (22, 53)
Autocorrelation^e	ITS level (N = 282)	0.22 (0.00, 0.48)
Autocorrelation^e	Meta-analysis level^d	0.17 (0.13, 0.42)
Time interval for time series datapoints	Month	11 (65)
	Year	4 (24)
	4-weeks	1 (6)
	Day	1 (6)

ITS interrupted time series, IQR interquartile range, PW Prais-Winsten, REML restricted maximum likelihood
^aMultiple response options possible therefore percentages sum to greater than 100%
^bInterruptions classified as ‘individual-level interruption’ were the intervention directed at an individual (e.g., delivery of a vaccine), however, the measurements were still aggregated over units of time (e.g., number of vaccinations each year)
^cCombination indicates where reviewers combined multiple data types (e.g., combining studies using proportion and rate outcomes)
^dThe average number of datapoints and autocorrelation were calculated across the series included in each meta-analysis. These averages were then summarised across the meta-analyses using the median and IQR
^eAutocorrelation estimated using REML (or PW where REML failed to converge)

ISSN: 1471-2288