Background
Methods
Repository of ITS studies
Methods to obtain time series data
Interrupted time series model
Interrupted time series analysis methods
-
ordinary least squares regression (OLS), which provides no adjustment for autocorrelation, and in the presence of positive autocorrelation will yield standard errors that are too small [16];
-
OLS with Newey-West standard errors (NW), which yield OLS estimates of the model regression parameters, but with standard errors that are adjusted for autocorrelation [17];
-
restricted maximum likelihood (REML) (with and without the small sample Satterthwaite approximation (Satt)), which addresses bias in maximum likelihood estimators of variance components by separating the log-likelihood into two terms (one of which is only dependent on variance parameters) and using the appropriate number of degrees of freedom (d.f.) [20, 21]; and,
-
autoregressive integrated moving average (ARIMA), which explicitly models the influence of previous time points by including regression coefficients from lagged values of the dependent variable and errors [22].
Statistical method | Autocorrelation adjustment | Abbreviation |
---|---|---|
Ordinary least squares | None | OLS |
Newey-West standard error adjustment with lag-1 autocorrelation | NW | |
Generalised least squares | Prais-Winsten | PW |
Restricted maximum likelihood | Lag-1 autocorrelation model | REML |
Lag-1 autocorrelation model with small sample Satterthwaite approximation | REML-Satt | |
Autoregressive integrated moving average | Lag-1 autocorrelation model (i.e. ARIMA(1,0,0)) | ARIMA |
Analysis of the ITS datasets
-
outcome variable;
-
time variable t, beginning at 1 and incrementing by 1 up to time point N;
-
an interruption time indicator \({D}_{t}\); coded 0 pre-interruption and 1 post-interruption; and,
-
a slope change variable \(\left[t-{T}_{I}\right]{D}_{t}\) , equal to zero at the time of the interruption (\({T}_{I}\)) and incrementing by 1 up to time point N.
Comparison of results from the different ITS analysis methods
Estimates of level and slope changes, and their standard errors
Confidence Intervals
p-values
Autocorrelation coefficient estimates
Results
Time series dataset acquisition
Characteristics of the included ITS
Study level characteristics | All ITS studies (n = 200) | ITS studies with available data (n = 166) | ||
---|---|---|---|---|
n
| % |
n
| % | |
Type of interruption | ||||
Exposurea
| 12 | 6 | 10 | 6 |
Intervention | 188 | 94 | 156 | 94 |
Intervention type | ||||
Policy change | 104 | 52 | 82 | 49 |
Practice change | 40 | 20 | 36 | 22 |
Communication | 29 | 15 | 24 | 14 |
Organisation of care | 13 | 7 | 12 | 7 |
Clinical intervention | 2 | 1 | 2 | 1 |
Time interval type | ||||
Daily | 3 | 2 | 2 | 1 |
Weekly | 9 | 5 | 6 | 4 |
Two weekly | 1 | 1 | 1 | 1 |
Monthly | 120 | 60 | 96 | 58 |
Quarterly | 31 | 16 | 28 | 17 |
Six monthly | 3 | 2 | 3 | 2 |
Annually | 20 | 10 | 17 | 10 |
Other | 12 | 6 | 12 | 7 |
Can't determine | 1 | 1 | 1 | 1 |
Series level characteristics | ITS (n = 230) | ITS with available data (n = 190) | ||
median | IQR | median | IQR | |
Number of time points per series | 48 | (30, 100) | 41 | (25, 71) |
Number of time points in the segments used to calculate estimates for the first interruption | 18 | (10, 34) | 16 | (10, 28) |
Comparison of results from the different ITS analysis methods
Estimates of level and slope changes, and their standard errors
N
|
Absolute value of effect estimate
| ||
---|---|---|---|
Level change
Median (IQR)
|
Slope change
Median (IQR)
| ||
ARIMA | 189 | 1.40 (0.63,2.90) | 0.13 (0.05,0.26) |
OLS (NW)a
| 190 | 1.49 (0.60,3.03) | 0.13 (0.06,0.27) |
PW | 189 | 1.33 (0.57,2.81) | 0.13 (0.05,0.26) |
REML (REML-Satt)a
| 181 | 1.22 (0.47,2.56) | 0.13 (0.05,0.25) |
Number of comparisons | ARIMA | OLS | NW | PW | REML | REML-Satt |
---|---|---|---|---|---|---|
ARIMA | 189 | 189 | 188 | 185 | 175 | 175 |
OLS | 190 | 189 | 186 | 175 | 175 | |
NW | 189 | 186 | 174 | 174 | ||
PW | 186 | 171 | 171 | |||
REML | 175 | 175 | ||||
REML-Satt | 175 |
Confidence Intervals
p-values
Autocorrelation coefficient estimates
Statistical method
|
Autocorrelation coefficient (ρ) estimate
| |||||
---|---|---|---|---|---|---|
All available datasets
|
Series with ≥ 24 points
|
Series with ≥ 100 points
| ||||
N
|
median (IQR)
|
N
|
median (IQR)
|
N
|
median (IQR)
| |
ARIMA | 189 | 0.04 (-0.15,0.30) | 154 | 0.07 (-0.10,0.36) | 31 | 0.19 (0.04,0.54) |
PW | 186 | 0.05 (-0.14,0.33) | 155 | 0.07 (-0.10,0.38) | 31 | 0.19 (0.04,0.54) |
REML | 175 | 0.20 (-0.01,0.54) | 147 | 0.20 (-0.01,0.53) | 31 | 0.23 (0.08,0.57) |
Restricted to datasets where all methods can be compared
| ||||||
ARIMA | 171 | 0.05 (-0.14,0.30) | 147 | 0.06 (-0.11,0.35) | 31 | 0.19 (0.04,0.54) |
PW | 171 | 0.05 (-0.14,0.31) | 147 | 0.07 (-0.11,0.35) | 31 | 0.19 (0.04,0.54) |
REML | 171 | 0.20 (-0.01,0.54) | 147 | 0.20 (-0.01,0.53) | 31 | 0.23 (0.08,0.57) |