Current Practices in Missing Data Handling for Interrupted Time Series Studies Performed on Individual-Level Data: A Scoping Review in Health Research

Figures & data

Figure 1 PRISMA diagram for the scoping review.

Notes: PRISMA figure adapted from Liberati A, Altman D, Tetzlaff J, et al. The PRISMA statement for reporting systematic reviews and meta-analyses of studies that evaluate health care interventions: explanation and elaboration. Journal of Clinical Epidemiology. 2009;62(10)e1-e34. Creative Commons.²⁸

Table 1 Characteristics of the Included Interrupted Time Series Studies (N=60)

	n	(%)
Country of Study
Australia	1	(1.7)
Bangladesh	1	(1.7)
Brazil	2	(3.3)
Cambodia	1	(1.7)
Canada	4	(6.7)
China	2	(3.3)
France	2	(3.3)
Germany	1	(1.7)
Israel	1	(1.7)
Italy	1	(1.7)
Japan	1	(1.7)
Malawi	1	(1.7)
Netherlands	1	(1.7)
Rwanda	1	(1.7)
Saudi Arabia	1	(1.7)
South Korea	1	(1.7)
Spain	2	(3.3)
Switzerland	1	(1.7)
UK	7	(11.7)
USA	28	(46.7)
Study Design
CITS	10	(16.7)
ITS	48	(80)
SR	2	(3.3)
Participants
Children	3	(5)
Firefighters	1	(1.7)
General population	4	(6.7)
Health personnel	6	(10)
Health personnel and patients	6	(10)
Insured women	1	(1.7)
Medications	1	(1.7)
Patients	38	(63.3)
Type of Intervention
Guideline/protocol/sound publication or evidence	9	(15)
Focused intervention	15	(25)
Policy	16	(26.7)
Programme	14	(23.3)
Relevant or historical event	3	(5)
Treatment	3	(5)
Level of Intervention
Cities, group of	1	(1.7)
City/district	3	(5)
Country	17	(28.3)
Hospital	18	(30)
Hospitals, group of	8	(13.3)
Individual level	2	(3.3)
State/province/county	10	(16.7)
Fire departments	1	(1.7)
Most Granulated Cluster Available
GP	3	(5)
District	2	(3.3)
Fire department	1	(1.7)
Group of patients (by diagnosis)	1	(1.7)
Health facility	3	(5)
Hospital	13	(21.7)
Hospital unit	3	(5)
Household	1	(1.7)
Individual level	32	(53.3)
Medications	1	(1.7)
Longitudinal Follow-up
Prospective cohort (individuals)	8	(13.3)
Prospective panel (cluster)	6	(10)
Retrospective cohort (individuals)	25	(41.7)
Retrospective panel (cluster)	21	(35)

Abbreviations: CITS, controlled interrupted time series; ITS, interrupted time series; SR, segmented regression; GP, general practice.

Table 2 Data and Statistical Analyses of the Included Interrupted Time Series Studies (N=60)

	n	(%)
Data Source
Collected for the study (prospective)	14	(23.3)
Routinely collected (retrospective)	46	(76.7)
Linked Data
No	50	(83.3)
Yes	10	(16.7)
Outcome Type
Continuous	10	(16.7)
Count	11	(18.3)
Proportion	39	(65)
Time-Points, Number
Median (IQR)	38	(55)
Minimum	6
Maximum	1217
Time-Points, Unit
Day	3	(5)
Half-year	1	(1.7)
Month	36	(60)
Quarter-year	8	(13.3)
Two-month	1	(1.7)
Week	5	(8.3)
Year	6	(10)
Averaging Step
No	11	(18.3)
Yes	47	(78.4)
Unclear	2	(3.3)
Statistical Model
ARIMA	7	(11.7)
Joint-point (exploratory method)	1	(1.7)
SR-GEE	7	(11.6)
SR-GLM	15	(25)
SR-GLS	1	(1.7)
SR-OLS	23	(38.3)
Mixed effects (random intercept only)	4	(6.7)
Mixed effects (random intercept and slopes)	2	(3.3)
Confounder Reported
No	19	(31.7)
Yes	41	(68.3)
Confounder Adjusted
No	27	(45)
Yes	33	(55)

Abbreviations: IQR, interquartile range; ARIMA, autoregressive integrated moving average; SR, segmented regression; GEE, generalised estimating equation; GLS, generalised least squares; OLS, ordinary least squares.

Table 3 Reporting and Handling of Methodological Issues in the Included Interrupted Time Series Studies (N=60)

	n	(%)
Autocorrelation – Considered (n=60)
No	19	(31.7)
Yes	41	(68.3)
Autocorrelation – Tested With (n=41)
Breusch–Godfrey	2	(4.9)
Cumby–Huizinga	1	(2.4)
Durbin–Watson	8	(19.5)
Within-individual correlation by design	11	(26.8)
Autocorrelation function	3	(7.3)
Autocorrelation probability	2	(4.9)
Not specified	13	(31.7)
Residuals examination	1	(2.4)
Autocorrelation – Concluded by Test (n=41)
No	5	(12.2)
Yes	36	(87.8)
Autocorrelation – Controlled By (n=36)
Cochrane–Orcutt	1	(2.8)
GEE models	6	(16.7)
Newey–West standard errors	7	(19.4)
Prais–Winsten	2	(5.6)
Autoregressive error term	8	(22.2)
Mixed models	5	(13.9)
Not specified	7	(19.4)
Seasonality – Considered (n=60)
No	41	(68.3)
Yes	19	(31.7)
Seasonality – Tested With (n=19)
Dickey–Fuller	1	(5.3)
Autocorrelation/partial autocorrelation function	2	(10.5)
No formal test	14	(73.7)
Not possible (short period)	1	(5.3)
Regression diagnosis test	1	(5.3)
Seasonality – Concluded by Test (n=19)
No	1	(5.3)
Yes	18	(94.7)
Seasonality – Controlled By (n=18)
ARIMA parameter	1	(5.6)
Covariate in the model	12	(66.7)
Decomposition	1	(5.6)
Not handled (reported as limitation)	2	(11.1)
Seasonal ARIMA	2	(11.1)
Time-Dependent Variable – Considered (n=60)
No	11	(18.3)
Yes	49	(81.7)
Time-Dependent Variable – Handled By (n=49)
Control group	10	(20.4)
Control outcome	1	(2)
Covariate (exploration)	1	(2)
Covariate in the model	3	(6.1)
Reported as a limitation, not handled	34	(69.4)
Other Issues – Considered (n=60)
No	35	(58.3)
Yes	25	(41.7)
Other Issues – Handled By (n=25)
Bonferroni adjustment (p values)	1	(4)
Adjusted for survey design	1	(4)
Aggregate ecological design (reported as a limitation)	1	(4)
Confounders not controlled (reported as a limitation)	1	(4)
Minimising immortal time bias	1	(4)
Non-stationary (ARIMA controlled)	2	(8)
Overdispersion evaluation (Poisson models)	2	(8)
Secular trends (reported as a limitation)	2	(8)
Sensitivity analysis (extracting patients)	3	(12)
Sensitivity analysis (impact model)	3	(12)
Sensitivity analysis (various)	6	(24)
Subgroup analysis	2	(8)

Abbreviations: GEE, generalised estimating equation; ARIMA, autoregressive integrated moving average.

Table 4 Reporting and Handling of Missing Data Issues in the Included Interrupted Time Series Studies (N=60)

	n	(%)
Missing Data – Considered (n=60)
No	47	(78.3)
Yes	13	(21.7)
Missing Data – % Reported (n=13)
% Not reported, but declared as an issue to be solved	2	(15.4)
Covariates <30%/outcome <50%	1	(7.7)
Covariates at baseline (<1% each, not combined)	1	(7.7)
Covariates at baseline (<10% each, not combined)	2	(15.4)
Covariates at baseline (<2%, flow chart)	1	(7.7)
Covariates at baseline (<25% each, not combined)	1	(7.7)
Covariates at baseline (<25%, flowchart)	1	(7.7)
Covariates at baseline (<30% each, not combined)	1	(7.7)
Covariates at baseline (<5%, flowchart)	1	(7.7)
Outcome <60%	1	(7.7)
Smoking (one case), outcome irregularly recorded	1	(7.7)
Missing Data Mechanism – Considered (n=13)
No	11	(84.6)
Yes	2	(15.4)
Missing Data Mechanism – Reported (n=2)
MAR	1	(50)
MNAR	1	(50)
Method for Handling Missing Data – Considered (n=13)
No	0	(0)
Yes	13	(100)
Method for Handling Missing Data – Reported (n=13)
CCA	11	(84.6)
Mixed intercept model for handling missing outcomes	1	(7.7)
Mixed intercept and slope model for handling missing outcomes	1	(7.7)
Sensitivity Analysis for Missing Data Mechanism – Considered (n=13)
No	11	(84.6)
Yes	2	(15.4)
Sensitivity Analysis for Missing Data Mechanism – Reported (n=2)
Comparing results from MICE versus CCA	1	(50)
Comparing results from using a “missing data category” versus CCA	1	(50)

Abbreviations: MAR, missing at random; MNAR, missing not at random; CCA, complete case analysis; MICE, multiple imputation by chained equations.

Liberati A, Altman D, Tetzlaff J, et al. The PRISMA statement for reporting systematic reviews and meta-analyses of studies that evaluate health care interventions: explanation and elaboration. Journal of Clinical Epidemiology. 2009;62(10)e1–e34.19631507

 PubMedGoogle Scholar