Cost-effectiveness of a universal strategy of brief dietary intervention for primary prevention in primary care: population-based cohort study and Markov model

A Markov model was designed drawing on previously reported research [20]. Participants entering the model were aged 30 years and over because of the low risk of chronic disease at younger ages. Healthy participants free from chronic disease, referred to as At Risk, could develop one of the disease states of interest including Diabetes mellitus, Coronary Heart Disease, Stroke or Colorectal cancer. Participants were allowed to progress to multiple disease states, representing all potential combinations of the selected conditions, consistent with the frequent development of multiple morbidity in primary care [21]. Participants in each state were allowed to progress to Depression, or to remain Not Depressed [22]. Depression was included in the Model because the prevalence of depression is increased in many chronic diseases and depression is associated with higher health care costs [22]. Depression was considered to be reversible; in each cycle participants transitioned to Depression on a probabilistic basis, based on the prevalence of depression observed in that state. All model states might progress to death. The model included a total of 32 states, representing the potential combinations of included diseases and depression. A schematic diagram of the model was previously reported [19]. The cost analysis was from the health care perspective and only costs of health care utilisation were included. A life time horizon was used for the analysis. The model was stratified by gender and single year of age.

The model was populated with empirical estimates drawn from a large cohort of participants registered with family practices participating in the UK Clinical Practice Research Datalink (CPRD), previously known as the General Practice Research Database (GPRD). Participants were sampled from family practices continuously contributing data to the CPRD between 1 January 2004 and 30 October 2010. Participants comprised a random sample of 299,912 registered patients, aged 30 to 100 years. CPRD data were analysed to estimate, for each state in the model, the incidence of the state, the mortality in each state, the prevalence of depression, and the costs of health care utilisation. Estimates were obtained by 10-year age group and sex. The methodological approach to analysis has been reported previously [22]. Unit costs of health care utilisation were obtained from standard reference sources for 2010 [23]. The Multilex drug code for each prescription record in CPRD was linked to unit costs in the First DataBank Europe database in order to obtain prescription costs [24]. The empirical mean and standard deviation for costs associated with each participant was estimated according to 10-year age group, sex and model state. Utilities for each state were obtained from data published in a compendium of values [25]. Utility values for each state were stratified by single year of age but were the same for men and women.

The Markov model, with probabilistic cohort simulation, was programmed using R software [26]. Outcomes and costs were compared for Intervention and Standard Care over 70 annual cycles, this allowed the entire cohort to progress either to death or to reach age 100 and exit the model. Annual transition probabilities for the model were obtained by sampling from the beta-binomial distribution, using CPRD data as inputs. Utility values were sampled from the beta distribution. The costs of each state were sampled from the gamma distribution with the mean value from CPRD, by10-year age group, sex, condition and depression status, as the empirical input. The model was implemented with a half cycle correction for the estimation of QALYs and costs.

Total costs and quality adjusted life years (QALYs) were obtained by summing across the 70 cycles of the model included in each simulation. There were 2,000 simulations run for each of intervention or standard care scenarios. Results are expressed as rates per 1,000 healthy participants entering the model. Mean costs, and the 95% range, were obtained from the data for 2,000 simulations. Incremental costs and QALYs were obtained as the difference between intervention and standard care scenarios. Costs and QALYs were discounted using a rate of 3.5%, but QALYs were also discounted at a rate of 1.5% as a sensitivity analysis. Not all simulations were associated with positive incremental costs and QALYs, making the incremental cost-effectiveness ratio a less suitable measure for analysis [27]. Net health benefits (NHB), at a threshold value of £30,000 per QALY, were therefore calculated as the difference between the increment in QALYs and the increment in costs divided by the threshold value of cost per QALY [28]. (pages 128-130) A cost-effectiveness acceptability curve was plotted using a range of threshold values.

The intervention was assumed to modify only the incidence of physical disease in healthy participants at risk. Changes in occurrence of multi-disease states, prevalence of depression and mortality were assumed to be secondary to changes in the incidence of diabetes, coronary heart disease, stroke and colorectal cancer. Relative risks of diabetes, coronary heart disease, stroke and colorectal cancer, associated with fruit and vegetable consumption were calculated using two sources of information. The first was a systematic review of randomised trials in primary care [17] that provided an estimated 0.5 (95% confidence interval 0.13 to 0.87) portion per day increment in fruit and vegetable consumption from brief dietary intervention in primary care. This was combined with estimated age-specific relative risks of disease outcomes associated with dietary change derived from the World Health Organization report. (Table 9.28, page 696). The intervention and the intervention effect were modelled to be maintained for five years. In the absence of evidence for the time course of intervention effects, a constant intervention effect was modelled in each of the first five cycles of the model. The cost of the intervention was modelled as a fixed cost per person per year and was assumed initially to be equivalent as the cost of one family practice consultation. Sensitivity analyses was carried out taking costs of intervention equivalent to two family practice consultations per year and 20% of a single family practice consultation per year. In secondary analyses, the effect of intervention was also modelled in selected age groups to allow the comparison of cost-effectiveness between age groups. There were 1,000 simulations performed in each age group because of the greater computational burden.

We used anonymous data approved by CPRD without identifying the patient and written informed consent was not required from the patients for the publication of this paper and any accompanying images.

Our study modelled the health outcomes of a universal strategy for brief intervention to promote dietary change in primary care. We used a very large empirical population registered in UK primary care practices to represent the study population. The intervention effect was derived from a meta-analysis of randomised controlled trials with at least 12 months follow up. The results suggest that even with a very small intervention effect, such as an increase in fruit and vegetable consumption of half a portion per day, when maintained over prolonged periods of time might bring appreciable benefits in terms of number of years lived free from chronic disease. However, even when delivered at very low cost, we estimate that a universal intervention strategy is unlikely to prove cost-effective. We initially considered the possibility that the intervention could be delivered through resource use equivalent to one family practice consultation per year. However, our systematic review revealed that intervention strategies are generally more costly than this, requiring multiple face to face visits, singly or in groups, together with supporting written materials and follow-up communications. An effective intervention may be more costly and even less cost-effective that we have estimated. Our analysis suggests that a targeted strategy focusing on older individuals who may be at higher risk of disease may be a more appropriate strategy for implementation in primary care. However, more research is needed to explore the optimal strategy for targeting intervention. The results also serve to emphasise the importance of developing more effective interventions to promote dietary behaviour change.

Our conclusions are consistent with those of Cobiac et al. [32] who reviewed published literature and modelled the evidence available from range of 23 interventions for promoting fruit and vegetable consumption in adults. Their study concluded that interventions directed at individuals and relying on dietary counselling, telephone contact, worksite promotion and other methods which encourage dietary behaviour change are neither highly effective nor cost effective. Pomerlau et al. [33] also found that dietary change interventions generally had small effects, with larger effect sizes observed in secondary prevention studies. This is also consistent with the suggestion that a targeted intervention strategy may be more appropriate in primary care.

This study was informed by the results of a new systematic review that was tailored to the objective of this research. The model was informed by analysis of a sample of nearly 300,000 participants sampled from the CPRD which is grounded in a general population sample of nearly six million people in the UK. The modelling approach was fully probabilistic which meant that uncertainties in the model inputs were carried through into uncertainties in the model outputs. This was necessary because estimates for incidence and mortality were imprecise for less frequent conditions. We implemented a series of sensitivity analyses that explored the effect of varying key assumptions. We acknowledge that we modelled an intervention duration (five years) that is longer than has been implemented in randomised trials (one year). However, it appears unlikely that a behavioural intervention could be more effective over a longer period than a shorter period, so this does not vitiate our conclusion that a universal intervention strategy in primary care is unlikely to be cost-effective. We focused on the effect of intervention on fruit and vegetable consumption but other dietary changes, such as consumption of fat or fibre, may also be associated with health outcomes. However, self-reported dietary change measures, used in many intervention studies, may have over-estimated the potential effects of intervention. A previous systematic review [18], found a larger intervention effect that but this review included high-risk groups including those with hypertension, hyperlipidaemia or having close relatives with type 2 diabetes or cancer. We also acknowledge that we did not utilise empirical data for the unit costs of intervention. We used a low estimate for the unit cost of intervention and, even with this estimate, the intervention strategy was found not to be cost-effective. More resource intensive interventions are likely to be even less cost-effective, as evidenced by the sensitivity analysis in which a higher intervention cost was used. Our systematic review [MS resubmitted to BMC Public Health] provides information on the resources used in previous intervention studies, and these interventions were generally considerably more resource intensive than the one modelled here. We did not model any social multiplier effects, where the impact of the intervention delivered to one person might motivate the individual’s social contacts to change. It is possible this may have underestimated the benefits of intervention. We did not set an upper age limit for participants as fruit and vegetable intake may be beneficial even in old age [34].We assessed the cost effectiveness of intervention in selected age groups of population which shows that intervention is more cost effective in older age groups. We applied a half cycle correction but this had only a small effect as the perspective of the model, up to 70 years, was long in relation to the length of the annual cycles. However, we acknowledge that there are arguments both for and against the implementation of a half-cycle correction [35, 36]. We included only the health care costs; outcomes might differ if costs and productivity from a wider perspective were to be included. We obtained data from primary care records and we could not follow patient use of secondary care resources beyond referral and admission; this may have underestimated the utilisation of secondary care. We used the average of all costs over all stages of disease which may not have represented the actual resource utilisation costs as health care resource utilisation costs may be higher at the start of illness or at periods closer to death. Utility estimates were drawn from a compendium of values, which may not have provided an entirely accurate assessment for this population. We used a cost-effectiveness threshold of £30,000 per QALY as this is a widely accepted level below which interventions are judged to be cost-effective in the United Kingdom. However, some policy decisions have indicated that higher threshold values might sometimes be accepted in practice [37].

The use of fully anonymised CPRD data was approved by the MHRA Independent Scientific Advisory Committee (Ref. 09-085).