I have read a lot of studies using an interrupted time series design, e.g., studies using the ‘unexpected event during survey design‘. I have noticed that the better studies using an interrupted time series design all provide great visual presentations of the results.
There are relatively easy ways to improve the visual presentation of findings in interrupted time series studies. In the paper Creating effective interrupted time series graphs: Review and recommendations you will find specific recommendations to improve such figures.
The recommendations are related to the data points, the interruption, trend lines, the counterfactual, additional lines and general graph components. In addition, the authors emphasise that some of the recommendations are core whereas other are additional.
For data points, it is important to 1) plot each data point, 2) show the same points as used in the analysis, 3) line up the data points with x-axis tick marks, and 4) not join data points with lines. Accordingly, in good papers, it is easy to get a sense of the data simply by looking at the visualisation.
For the interruption in the visualisation, it is important to 1) clearly show the interruption (e.g., with a vertical line), 2) show any potential transition in the period (if it is not a clear interruption), and 3) label the interruption line. In other words, a good interrupted time series graph provides a lot of details on the actual interruption being studied.
For the trend lines, it is important to 1) plot both the fitted pre- and post-interruption trends, 2) use bold and solid lines for fitted trends, and 3) match the colours for the trend line and data points. Without these trend lines, it is next to impossible to evaluate whether there is an interruption in the outcome of interest. Here, it is also relevant to add a counterfactual trend line (i.e., the estimated trend in the absence of the event of interest), and in particular to use a different line pattern.
For additional lines, the most important thing to consider is reporting uncertainty. This is something that the paper is not devoting significant attention to, but I believe this is paramount for a good visualisation. Specifically, make sure to show 95% confidence intervals around the trend lines (or something to that effect).
Last, there is a series of aspects to consider in relation to general graph components. These core aspects include: 1) showing axis tick marks, 2) label axes, 3) align axis labels with axis tick marks, 4) include axis titles with information on name and unit of measurement. In addition, it is worth considering the use of grid lines, the scale, the visual impact of additional text, horizontal text, and colourblind-friendly colours.
There are some good examples in the paper as well. For example, consider Figure 6 where panel A is a figure from a published paper and panel B is the revised figure taking the recommendations into account:
In the original figure, vertical bars are used to show the monthly proportions. Such bar charts are not ideal as it is difficult to get a sense of the variation in the data over time, and it needs to include 0 on the y-axis in order not to mislead. Here is how Turner et al. (2021) describe the improved version:
The data points have been plotted, which allows the spread of the data to be more easily seen, allows the data to be extracted and reduces the visual clutter. The interruption has been represented by a vertical line which is labeled. The counter-factual and trend lines have been plotted, allowing the reader to more easily see the impact of the intervention. The x-axis has been adjusted so that the points are more clearly aligned with the tick marks to facilitate data extraction. The range of the y-axis has been decreased to allow the data to fill the available space. Using additional text, the level and slope changes are given, along with 95% confidence intervals.
Turner et al. (2021) further looked at 217 interrupted time series graphs related to public health interventions to examine what recommendations were followed in the literature. 73% of the graphs had a line representing the time of the interruption, but only 17% of the graphs had a line for the counterfactual trend.
In sum, interrupted time series graphs are great and there are easy ways to make such graphs effective.