Skip to main content

Archived Comments for: The use of segmented regression in analysing interrupted time series studies: an example in pre-hospital ambulance care

Back to article

  1. A tailored mixed methods approach for analysis of multiple interventions at varing time points in different organisations participating in a quality improvement collaborative

    Aloysius Niroshan Siriwardena, University of Lincoln & East Midlands Ambulance Service NHS Trust

    14 July 2014

    Taljaard et al make some interesting suggestions in their commentary on our recent paper [1]. We agree that segmented regression or the broken stick model is useful and we have used it ourselves profitably [2]. In this case though, we wonder whether it is the most suitable or only approach.

    We would point out two features of our study that we considered in our analysis.  

    • The intervention continued over a long time period, up to a year in some ambulance services, and this meant that it was hard to identify a priori the time of the breakpoint for a given ambulance service and even more difficult to do so for the aggregate of all services since they introduced different interventions at different time points. We notice that the re-analysis in the commentary paper did not take into account the notable differences in the various ambulance trusts involved. We found that some services showed considerable reductions in variation and improvements (shifts in control charts) in the care bundles for acute myocardial infarction (AMI) and stroke following the collaborative. For that reason we regard the mixed methods approach we took, using annotated control charts and matching interventions with patterns of change, as most valuable as this enabled us not just to say what happened but why.
    • Fitting any linear model, even to the log odds, assumes that it is realistic for the rise to continue without limit. In this case the performance of ambulance services, especially for stroke, approached 100 per cent; the resultant ceiling meant that a linear model would break down near the end of the period and changes in slope there would be difficult to interpret. We fitted other models to try and account for these features. For example, we fitted a quadratic in time on the basis that if the change was gradual it might look quadratic but this has the disadvantage of also reaching a ceiling and in any event all quadratic models ultimately go off to infinity. We also tried fitting separate slopes to the first and last six months but this has few advantages over the broken stick model.   

    In the absence of a control group we feel that although we presented quantitative summaries these would be inferior to the mixed methods approach for understanding what happened in a Quality Improvement Collaborative (QIC) of this complexity. We welcome further debate on the issue of analysis of QICs involving multiple interventions introduced at different time points in different organisations, including where ceiling effects may be present.    

    References     

    1.   Siriwardena AN, Shaw D, Essam N, Togher FJ, Davy Z, Spaight A, Dewey M: The effect of a national quality improvement collaborative on prehospital care for acute myocardial infarction and stroke in England. Implement Sci 2014, 9:17.   

    2.   Siriwardena A, Fairchild P, Gibson S, Sach T, Dewey M: Investigation of the effect of a countywide protected learning time scheme on prescribing rates of ramipril: interrupted time series study. Fam Pract 2007, 24:26-33.  

     

    A. Niroshan Siriwardena, Michael Dewey, Deborah Shaw, Nadya Essam, Fiona Togher, Zowie Davy and Anne Spaight

    Competing interests

    None. 

Advertisement