Diarrhoea remains one of the leading causes of mortality in childhood, accounting for 15% of the approximately 8.7 million deaths of under-5 year olds worldwide in 2008 alone [1]. Current best-practice guidance for treatment of diarrhoea is contained within the World Health Organisation's strategy for Integrated Management of Childhood Illness (IMCI), a strategy adopted by over 100 countries worldwide [2]. This strategy is intended to foster correct diagnosis and treatment of common childhood illnesses in outpatient and inpatient settings with guidance for small hospitals published in the Pocket Book of Hospital Care for Children [3]. However reports indicate that success in implementation of diarrhoea case management has been varied [4–8].

Treatment guidelines for diarrhoea emphasise that patients be assessed for presence of blood in stool; non-bloody diarrhoea is to be managed with fluids only (unless co-morbidities are present which require different treatment), while bloody diarrhoea, presumed to be dysentery, should be managed with fluids and antibiotics. This is a pragmatic approach informed by the observation that most non-bloody diarrhoea episodes in the under-5 age group in low-income settings are self-limiting and are caused by pathogens not susceptible (e.g. rotavirus, astrovirus and enteric adenovirus) to antibiotic therapy or for which antibiotics are likely of little value or even deleterious (e.g. salmonellae and campylobacter) [9]. In contrast a significant proportion of episodes of bloody diarrhoea caused by shigella are associated with considerable mortality and are ameliorated by antibiotic therapy [10]. Furthermore correcting dehydration is clinically more important due to its association with adverse outcomes. Despite the fact that the guidelines are unambiguous, inappropriate antibiotic use remains common [11]. Antibiotic misuse may accelerate acquisition of antibiotic resistance [12] necessitating the use of more expensive alternative drugs [13, 14] and is a serious threat to drug therapy of infectious diseases [15, 16]. It may also be associated with missed opportunities to make correct diagnoses and as a result further increase in treatment cost. These undesirable consequences are disproportionately greater in low income countries which bear a bigger burden of infectious diseases [17]. However, research that explores the reasons for unnecessary antibiotic use in low-income settings is unusual; research that explores the effects of interventions on this outcome, particularly amongst inpatient populations, is rare. We conducted a multi-faceted intervention trial in Kenya aimed at improving key inpatient paediatric care practices spanning assessment, diagnosis and treatment of malaria, pneumonia and non-bloody diarrhoea using 14 process-of-care indicators. The effect of this intervention on these practices has been reported previously [18]. Here, we use these data to investigate whether the intervention (described subsequently) reduces inappropriate antibiotic use in the children admitted with non-bloody diarrhoea using a hierarchical modelling approach.

This trial was undertaken between 2006 and 2009 with the main end-point in 2008. It investigated whether an intervention delivered at the hospital level comprising training health workers on the use of evidence-based guidelines for paediatric care, local facilitation, and external supervision and feedback, would improve the quality of care given to children with specific diagnoses admitted to Kenya's district hospitals [18]. In this paper we explore associations between the intervention and evidence for inappropriate antibiotic use in cases of non-bloody diarrhoea. This outcome was not a pre-specified trial endpoint.

Clinical practice guidelines are applied in healthcare settings with the aim of improving the process of care. Studies in a variety of settings and clinical fields generally show that guidelines improve processes and outcomes of care [27]. However most of these studies have been poor quality individually randomised studies or studies in primary healthcare settings that are not comparable to hospitals in a low income country. Guidelines are very important when new evidence makes it necessary to change practice. In cases of non-bloody diarrhoea in children guidelines discourage the use of antibiotics. Focusing on this as a process-of-care measure for diarrhoea case management we sought to determine whether this outcome was influenced by an intervention promoting the uptake of a set of practice guidelines aimed at health workers caring for children admitted to hospital. This outcome is relevant to our setting because childhood diarrhoea and dehydration are leading causes of death, and contemporary evidence favours effective supportive treatments such as rehydration while discouraging antibiotic therapy except in cases of dysentery [9]. Furthermore inappropriate use of antibiotics is a waste of important resources that a low-income country can ill afford. While other interventions are also important in reducing the burden of disease and costs attributable to diarrhoea, such as staff hand-washing that was also promoted within the intervention, our study design did not allow us to evaluate any effect of the intervention on these practices.

We used two approaches to the analysis of the data: an adjusted t-test to assess evidence of a difference in mean effect between the full intervention and control (partial intervention) hospitals-a simple but robust approach [25]-and a hierarchical logistic regression model for these repeated survey data. The latter method allows for the clustering within the data in acknowledgement of the reality that the hospital is the unit of intervention, but can be faulted for underestimating higher-level variances for binary covariates when lower-level units (clusters) are few relative to higher-level units (individuals) [24]. Imbalances may also occur between intervention and control groups with respect to covariates predictive of the outcome when the number of higher-level units is relatively small.

Both procedures provide comparable weak evidence of a possibly substantial intervention effect, with the individual level model indicating that children in the intervention hospitals had just under one-third the odds of receiving antibiotics inappropriately compared to those in the control hospitals by the trial endpoint. This evidence of an intervention effect complements that derived from analyses of the trial's main outcomes (structure indicators of availability of key resources supporting service delivery, process errors in management of pneumonia, malaria and diarrhoea/dehydration, and outcome indicators of adherence to key policy recommendations) [18] and suggests broader benefits than so far reported. A new finding was some evidence that the intervention effectiveness was modified by survey period; this suggests a dose-dependent effect on the outcome. Thus the accelerated improvement in the intervention arm compared with the control arm suggests a true effect of the full intervention. For ethical reasons no comparable data are available from hospitals without any form of intervention although baseline data perhaps indicate typical prevalence of inappropriate antibiotic use. It is therefore not possible to determine if the less marked improvement in hospitals receiving the partial intervention was also an intervention effect or represented just a secular trend. Nevertheless this, along with other results of the trial [18], represents the first evaluation of a complex intervention in our setting and evidence that it promotes good paediatric care practice.

Restricted randomisation was used to create the group allocations. It must be acknowledged that when randomising only 8 units of analysis the potential for residual confounding, even after identifying and adjusting for other associations with the outcome as above, is considerable. Still, the model showed no evidence that the baseline differences in the outcome between the hospitals was more than would have been expected by chance (OR 0.85, p-value 0.843). Furthermore the objective nature of the exposures and outcome further strengthen the analysis. Interestingly between-clinician differences contributed almost two-thirds of overall variability in the data, suggesting that hospitals as organisations do not necessarily promote consistent practice amongst clinicians within them.

We found that children with crackles had increased odds of receiving inappropriate antibiotics. Crackles are commonly associated in health worker training with lower respiratory tract infection. However in the national (and international) guidelines provided and used in training the diagnosis of pneumonia, which requires treatment with antibiotics, does not depend on observing crackles. Rather, it should be made in children with elevated respiratory rates or other signs of respiratory distress. Our results and anecdotal experience however suggest clinicians generally find this diagnostic advice difficult to accept. Nevertheless, and interestingly, in these data the clinicians recording the presence of crackles did not record a diagnosis of pneumonia; in the analysis this therefore was classified as inappropriate antibiotic use. Absence of information on crackles was also associated with inappropriate antibiotic use. An alternative or complementary hypothesis is that clinicians who carefully record clinical signs, and find them to be negative, are more likely to adhere to guidelines. Qualitative research might be particularly useful to examine such hypotheses and facilitate understanding of inappropriate antibiotic use in hospitalised children with diarrhoea.

We are not aware of any other studies evaluating an intervention intending to improve care for multiple, common and life-threatening illnesses in children admitted to hospitals in low-income settings using a randomised controlled design. The intervention provided training on evidence-based paediatric care with facilitation, supervision and feedback delivered at the hospital level. In the analyses presented we show that the intervention substantially reduced the odds of a child being inappropriately treated with antibiotics for non-bloody diarrhoea, an improvement in care that is in addition to the findings of improvement in most of the 14 process of care indicators forming the primary endpoint for the trial [18]. Most previous intervention studies to improve care, in both developed and developing countries, have focused on single diseases and have often demonstrated only modest improvements in practices, The data we bow report indicates even wider benefits of the intervention than suggested in the main study results implying that the intervention effects may cut across all diseases covered by the guideline package approach. Other ongoing analyses of the intervention suggest that it requires modest financial investments for comparatively substantial improvements in quality of care. Enhanced medical education focusing on rational use of antibiotics is necessary to improve clinicians' prescribing habits; we have shown that it is possible to change practices using an integrated approach and that perhaps, at hospital level, this is preferable to the recent focus on disease-specific interventions within low-income settings.

Let Y _ijk be the outcome of patient i who was attended to by clinician j in hospital k. Let x ₁ be a binary variable representing the intervention group, x ₂ a binary variable representing the survey (baseline vs. final post-intervention), and x ₃ an interaction parameter to represent how the intervention is modified by successive surveys. We also define z _l as any other covariate(s) associated with the outcome to be adjusted for in the model.

with clinician-level random-effects u _j ~ N(0, σu _j ²), and, conditional on u _j , hospital-level random-effects u _jk ~ N(0, σu _jk ²).

showing that the relationship between the outcome and the intervention has a slope of β ₁ + β ₃ S and random intercept α + β ₂ S + Σ _l γ _l z _ijkl + u _j + u _jk .

in the intervention group (x ₁ = 1). The difference between these two patients is the intervention effect at baseline, which is β1.

in the intervention group. The intervention effect due to this survey is β ₁ + β ₃ s. To obtain the baseline-adjusted intervention effect we find the difference between effects at these two study points, which is β ₃ s - the interaction between the intervention and survey s.