Methods of outcomes 1

A core SICP aim was to reduce emergency admissions. Although all mechanisms of integration in the SICP have a potential role in reducing admissions, the MDGs are most clearly focused on providing a rapid reduction in the use of hospital services through intervention with patients at high risk of admission.

Multidisciplinary groups and linked case management interventions have an important place in the NHS as a core mechanism of integration. Since the Evercare pilots,29,30 studies have cast some doubt on the evidence that this model can achieve reductions in hospital admissions.18,19,33,77,78 However, an unpublished survey of CCGs reported that 80% included some variant in their integration plans.79 There are a number of different models of MDGs and some of the ways in which they vary are outlined in Table 2. In line with the realist model (see Chapter 3), there is also an argument that the general case management ‘mechanism’ is effective only in certain contexts, such as a history of previous joint working among staff in an integrated care service.

Methods

The SICP was targeted at all general practices. Therefore, the primary analysis for the effects of the MDGs compared data from practices in Salford with suitable comparators in other parts of England. However, the introduction of MDGs was staged, and we used this to assess any differential impact relating to the staged introduction.

We adopted lagged dependent variable approaches to estimate the effect of the MDGs.80 This approach does not require assumptions of parallel trends between intervention and comparator groups imposed by a difference-in-differences specification. The lagged dependent variable approach uses a fixed vector of lagged values of the outcomes prior to the intervention as explanatory variables. The analysis is conducted only on the time points following the intervention.80

If the parallel trends assumption does not hold, the lagged dependent variable approach is less prone to bias and is more efficient than alternatives such as the creation of synthetic controls.80 The superiority of the lagged dependent variable approach is increased when data are available on more pre-intervention periods, as is the case in this setting.

Data

Data were HES from NHS Digital, stratified by financial quarter and general practice (financial years 2009/10–2015/16), for populations aged ≥ 65 years:

1. the number of accident and emergency (A&E) attendances per person
2. the number of A&E attendances referred by health and social care providers per person
3. the number of self-referred A&E attendances per person
4. the number of emergency admissions per person
5. the number of emergency admissions via A&E per person
6. the number of direct emergency admissions per person
7. the number of ambulatory care-sensitive emergency admissions per person
8. the proportion of patients discharged to usual place of residence.

We also obtained general practice patient registration lists for persons aged ≥ 65 years from two sources: (1) the Personal Demographic Service for the financial years 2009/10 and 2010/11 and (2) NHS Digital for the financial years 2013/14 to 2015/16. For the financial years 2008/9, 2011/12 and 2012/13, we used the closest year of data available.

Intervention sites

Practices in Salford CCG were all considered to be the intervention site, but we also identified distinct subgroups (non-adopters, early adopters and late adopters). Non-adopters (n = 5) were excluded from the analysis, leaving three intervention sites:

1. 9 early adopters, classified as starting the intervention in April 2014
2. 32 later adopters, classified as starting the intervention in April 2015
3. 41 adopters, classified as starting the intervention in April 2015.

Comparator sites

Four comparator sites were used outside Salford CCG (Table 7):

1. all practices in Greater Manchester excluding Salford (‘Greater Manchester’)
2. practices in two CCGs to the west of Greater Manchester (‘West’)
3. practices in nine CCGs to the west of Greater Manchester [‘West (extended)’]
4. all practices in England excluding Salford (‘England’).

Regressions

In total, we estimated 96 models (Table 8). We weighted all analyses by population size and used robust standard errors to allow for heteroscedasticity. We included proportions of the total practice population aged 65–74, 75–84 and ≥ 85 years as additional controls.

Robustness

An additional three models were estimated for the primary outcome (emergency admissions per person). For the first test of robustness, we omitted the data for the first financial year (2009/10). We omitted the first four quarters of data owing to small denominators in the general practice list sizes for the comparators ‘West’, ‘West (extended)’ and ‘England’.

For the second test of robustness, we limited the analysis to the period following the 2011 Census. The 2011 Census resulted in a recalibration of practice populations and may have affected the intervention sites in a different way from the comparators.

For the final robustness analysis, we used a difference-in-differences specification. Our models for difference-in-differences analysis control for Index of Multiple Deprivation,81 quarterly time dummies and proportion of practice list size of certain ages (65–74, 75–84 and ≥ 85 years). Difference-in-differences is not used for the primary analysis as this method relies on the parallel trends assumption. Parallel trends assumption requires that both intervention and comparator sites must have parallel trends pre intervention; violations will result in biased estimated treatment effects.