Substitution of physicians by nurses in primary care: a systematic review and meta-analysis

We included peer reviewed randomised controlled trials (RCTs) from any country published in English in which nurses (in any type of role) substituted physicians by acting as the main figure of care with autonomous or delegated clinical responsibility for tasks that would have formerly been performed by physicians alone: where nurse-led care was compared to physician-led care (family physicians, paediatricians, and geriatricians); the intervention had taken place in general practices, community or ambulatory care settings; in patients of all ages seeking care for all conditions including mental health and addiction restricted to primary care; and which reported on patient satisfaction, quality of life (QoL), hospital admission, mortality and cost of health services. Following the framework published in a Cochrane review [3], we excluded studies in which nurses firstly, provided services which supplemented or extended the care provided by physicians or tasks that are not part of the usual care of physicians and secondly, where nurses collaborated with other clinicians in a team and thus the effect of nurse-led care, as the main intervention, could not be distinguished.

We searched OVID Medline, Embase, CINAHL and The Cochrane Library of Systematic Reviews which includes the Cochrane Effective Practice and Organisation of Care Group, from all available dates until August 2012. The searches, not age-, date- or country-specific included ‘primary care’, ‘skill-mix’, ‘physicians’-‘nurse’ substitution’ (Additional file 1: Table S2). We also manually searched the reference lists of included studies and relevant reviews.

We assessed the risk of bias of all trials without the calculation of a composite score following available guidelines [6‐8]. We considered bias due to attrition of more than 20% to be of significant concern; and adequate intention-to-treat (ITT) if trial authors analysed participants based on their original group allocation regardless of protocol violations or non-compliance [9].

Two authors independently screened titles and abstracts, assessed both the full-text of eligible publications and the risk of bias of included studies, and extracted data. Differences were resolved through consensus.

We calculated the individual and pooled unadjusted relative risks (RR) and the standardised mean differences (SMD); and performed meta-analyses when at least three trials reported appropriate data, using the inverse variance fixed-effects (FE) method and repeated the analyses using a random-effects (RE) model in Cochrane RevMan (Version 5.1) [10]. We report the summary statistics, their 95% confidence intervals (CI) and consider p < 0.05 statistically significant. When scales pointed in opposite directions, we subtracted the mean from the maximum possible value of the scale and estimated the standard deviations (SD) using well-established techniques [11]. We analysed dichotomous and continuous data together by converting ORs to an effect size expressed in SMD using available methods [12]. We decided to use a FE model in keeping with: 1) having no basis to assume that the effects had a normal distribution, 2) the small number of studies in at least two of the analyses, 3) the accuracy in estimates and CIs that FE provides even in a small number of studies and the more weight assigned to larger studies; RE gives similar weight to small and larger studies. We quantified heterogeneity using the I² statistic [13] and explored the effects of nurse-led care and potential sources of I² by pooling data into pre-specified subgroup analyses by clinical characteristics: nurses’ roles (based on reported details: nurse practitioner with higher degree courses/specialisation (NP+) versus nurse practitioner (NP) versus registered/licensed nurse (RN/LN), type of care (single contact versus on-going care; urgent versus non-urgent) and length of follow-up (months: <6 versus ≥6; <12 versus ≥12). We explored the effect of potential sources of bias by study size (small, N < 200 versus large, N ≥ 200), allocation concealment (adequate versus inadequate/unclear) and attrition (<20% versus ≥20%), and inspected publication bias using funnel plots where there were at least 10 trials [14]. We performed sensitivity analyses by excluding trials with potentially contaminated samples (i.e. patient crossover between groups), quasi and cluster design and in which nurses had full clinical autonomy (to perform tasks) and/or where this information was not reported. For data not combined in meta-analyses, individual trial estimates were compared.

A total of 4,133 original records were identified. We excluded 12 of 44 relevant publications for the reasons provided in Table S3 (Additional file 1). Twenty-six studies reported in 32 publications met the inclusion criteria and comprised a total of 38,974 randomised participants (Figure 1) [15‐46]. Twenty-four of the studies were RCTs and the other two were economic evaluations based on three of the appraised RCTs [17, 26, 38]. Table 1 and Table S4 (Additional file 1) report the summary and detailed characteristics of participants, interventions and outcomes of the trials included in review.

There were twenty RCTs of parallel design, three cluster-RCTs, one quasi-RCT and two studies [31, 38] with cost data from three of the included RCTs [17, 26, 39]. The trials were conducted in the UK (n = 9), the Netherlands (n = 6), the USA (n = 6), Russia (n = 1) and South Africa (n = 2). Median follow-up was 14.8 (range: 0.5 to 122.4) months with at least 12 months in fourteen trials, less than 6 months in seven and 6 to 12 months in the other three. The median number of participants was 1,624 (range: 50 to 12,894) with less than 200 in eleven trials and more than 200 in the other thirteen. Mean age was reported in twenty trials and ranged from 10 to 83 years. Twenty-two trials reported on gender and 38.3% of the participants were male; one included women only.

A summary of settings, interventions and nurses’ roles are reported in Table 1 and Table S5 (Additional file 1). Nurses worked as physician substitutes in a range of care settings. The interventions were carried-out in general practices [17, 18, 26, 29, 30, 33, 39, 44], nurse clinics [23, 36, 40, 42] and in hospital-based, health care centres, specialised practices, community or university clinics [15, 16, 22, 25, 27, 28, 32, 34, 35, 37, 46]. In the controlled intervention, nurses were the main figure of care with autonomous or delegated responsibility in various clinical domains including a whole range of possible (undifferentiated/minor acute/common) or specific conditions (e.g. hypertension, heart failure, diabetes, HIV, etc.). In one trial, the clinical domain was assumed to represent undifferentiated care [35]. Nurses’ specific qualifications and training were not reported in sufficient detail but using the information provided by study authors we grouped nurses’ roles. Nurses’ roles were described in some detail in sixteen trials (two reported in one publication) [15‐18, 22, 23, 27‐29, 32, 36, 37, 39, 42, 46]. Seven trials employed NP+ only [16, 25, 28, 32, 39, 40, 42], six employed NPs [18, 22, 26, 30, 34, 44], eight employed RN and/or LN [15, 23, 27, 29, 33, 35‐37], one employed NP and NP+ [17], and one employed NP and LN [46]. Nurses’ interventions were guideline- or protocol-based in eighteen trials, while six had no report of having followed specific guidelines. Nurses’ clinical autonomy was obtainable from twenty-two trials. In three, nurses had full clinical autonomy to manage patients with diabetes type II [30] or undifferentiated conditions [34, 35]. In the other nineteen, nurses made independent decisions to perform several tasks (e.g. adopting, initiating and prescribing treatment, ordering tests or referrals) but they still required minor support or contact with the physicians (e.g. to sign prescriptions, referrals and tests, to discuss patients’ records or to develop action plans). Although the interventions in the control group were not clearly described in at least a few trials, these were assumed to represent physicians-usual-led care. Ten trials addressed single contact care [15, 39], single contact and on-going care [27, 35, 37], single contact and urgent care [17, 18, 26, 34] and single, on-going and urgent care [22]. The other fourteen included patients in on-going care for complex conditions (e.g. HIV, Asthma, hypertension, heart failure, etc.).

The overall quality of the studies varied substantially when assessed against current reporting standards [6, 11] (Table 2). Only 54.3% of the trials measured the success of the intervention by defining a primary outcome. Random sequence generation was adequate in 54.0%, allocation concealment in 42.0%, blinding of patients and providers in 4% and blinding of outcome assessors in 21.0%. Patient or clinician crossover between groups was reported in 12.5% of the trials. At baseline, groups were comparable for all tested factors in 70.8% of the trials. Both inclusion and exclusion criteria were reported in 70.8%. Sample size calculation based on power (80.0% to 90.0%) was performed in 70.8%, but only ten trials held the least target sample size to achieve power in at least one outcome. Rates of missing data varied widely (range: 5.0% to 65.5%). While three did not report any attrition, more than half (13/24) of the trials had an attrition rate of at least 20%: nine had more than 20% in both arms in at least one outcome (range: 10.0 to 65.5%), three had at least 45.0% per arm, and four had more than 20% (range: 8.0 to 30.0%) in one arm with a differential rate of 9.0% to 15.0% across the treatment and control groups. Only 29.2% of the trials reported the use of intention to treat (ITT) techniques (type not always reported) to deal with missing data.

We identified several significant limitations in the current evidence which should be considered in future research. The trials included were highly heterogeneous in terms of tasks, settings, collection and reporting of outcome measurements. There is a considerable amount of data that are reported in descriptive accounts only, limiting both their pooled validity and the interpretation of their results. Additionally, many studies failed to report some important statistical information (e.g. sample sizes, mean scores, SDs) required to calculate trial estimates and to integrate them in a meta-analysis.

No study fulfilled the set of methodological quality criteria assessed, despite widely available guidelines for RCTs. Trials of lower methodological quality (small study, at least 20% attrition and lack/unclear allocation concealment) tended to inflate the results and only less than 50% of the trials maintained the least target sample required to achieve power, which makes results less trustworthy. The most probable small study bias affecting the effect sizes are the results of small negative studies which are generally less likely to be published than small studies with positive results (i.e. publication bias). Blinding (clinicians, patients and outcome assessors) was reported in only a few trials and we don’t rule out the possibility that patient satisfaction, a subjective outcome, may have been especially positively affected by this. The trials consisted of follow-up episodes of variable length (0.5 to 122.4 months) which may have limited the true effect of care especially in multi-morbid or serious illnesses. Our analyses partly explained the reasons for heterogeneity where this was present but several other variables, which we could not account for, may have also caused this. Patients’ perception and evaluation of satisfaction may be inherently subjective due to socio-demographic differences, experiences from previous care, the physical environment, and patient-care provider interactions. Therefore, measurements of outcome using validated tools are preferred. Of the trials appraised, less than 50% used validated questionnaires for patient satisfaction.

We also identified a lack of trials of cluster randomisation. Although these may be more complex in design, if accounted for all key factors including clustering effect, appropriate sampling and analyses, cluster RCTs could add important value to the current evidence.

Surprisingly, there is a dearth of economic data. The little evidence available on the cost of physician-nurse substitution relies on results which are mainly based on direct costs and use variable approaches. The more recent literature reports more economic data, but it seems difficult to integrate these results especially because cost evaluations differ across countries and thus in cost measurements. We found only two publications [31, 38] providing economic data related to three of the included trials. Despite continued claims of substituting physicians by nurses based on healthcare costs, the evidence can only suggest that substitution is cost neutral. Therefore, as suggested in a recent systematic review of economic evaluations [49], to meaningfully place the costs and consequences of substitution in the context of healthcare, studies should address all types of costs. Relevant and appropriate data should be generated by means of a systematic collection of economic measures, and specific rules for cost data estimations should also be defined and followed.

More intensive implementation could enhance the outcomes of nursing care, but most studies do not provide the necessary information. In the evaluated studies, the assumption is that nurses possess the competence required for substituting physicians, but the level of substitution does not seem equal among studies. While the level of training may be a critical factor for an effective outcome, the studies report incomplete descriptions of nurses’ roles and competencies. The level of clinical autonomy in nurses does not seem consistent with the level of training and the tasks performed. Also, nurses still require support or communication with the physician for various tasks. It seems then that the level of qualification and training required to carry out substitution requires yet a better definition of practice boundaries including a classification of tasks. Better criteria conceptualised to define nurses’ roles and responsibilities are needed. In addition, the various differences between countries’ definitions and their organisation of nurse care should be taken into account. Lastly, more than half of the evidence reviewed (62.5%) has been conducted in Europe, mainly the UK and the Netherlands.

It is apparent that there is much room for primary studies that include larger numbers of patients, methodologically more rigorous in terms of quality, comprehensive in terms of data and statistical methods and with longer follow-up episodes. Furthermore, in order to gain a better understanding of substitution, future research should map a wider range of nurses, the various levels of training and clinicians’ characteristics, which are provided in many countries. As suggested previously [50], each method of skill-mix may have its own strengths and weaknesses. The implementation of methodologies aiming at the standardisation of skill-mix studies could support a sound assessment such that health sector reform may also benefit from the publication of evidence.

Our review updates and extends earlier systematic reviews [3, 4] and benefits from a thorough assessment of RCTs, in which the nurse acted as the main figure of care. It also presents (where available) the results by nurses’ roles. Having used the fixed effect model, we can only make inferences about the studies included in the meta-analyses performed here. We only included RCTs because these are at a lower risk of bias and allow for the identification of causal relationships. Although non-randomised trials may overestimate the benefits of nurse-led care it would be recommended to scrutinise the current evidence with such designs. These may not only provide an opportunity for an update but also allow for the collection of data from long term (more than 12 months) follow-up designs which may consist of larger sample sizes. We only included publications in English. We did however screen the reference lists of relevant reviews (some in foreign languages) and searched the reference lists of all included studies. We did not contact authors for further information nor did we search for grey literature. A further limitation is that it was often difficult to understand in detail what role and responsibilities nurses had, when substituting physicians. In many cases, they remain embedded in patient care teams that also involved physicians.