Modelling the Cost Effectiveness of Treatments for Parkinson’s Disease
The objective of this review was to assess models of cost effectiveness for Parkinson’s disease (PD) published after July 2002 and to derive recommendations for future modelling.
A systematic literature search was performed in the databases PubMed, Current Contents, EMBASE, EconLit, the Cochrane Database of Systematic Reviews, and DARE (Database of Abstracts of Reviews of Effectiveness), NHS EED (Economic Evaluation Database) and HTA (Health Technology Assessment) of the UK NHS Centre for Review and Dissemination (July 2002 to March 2010). Only fully published studies using decision trees, Markov models, individual simulation models or sets of mathematical equations were included.
Most of the 11 studies identified used Markov models (n = 9) and two employed were based on decision trees. Based on the Hoehn & Yahr (HY) scale, authors evaluated the cost effectiveness of drug treatments (n = 6), surgical approaches such as deep brain stimulation (n = 1) or striatal cell grafting (n = 1), and diagnostic procedures such as single photon emission computed tomography (SPECT) testing (n = 3) over a time horizon of 1 year to lifetime. Costs were adapted to address a societal and/or healthcare provider/ third-party payer perspective. All but one of the interventions investigated were considered cost effective or cost saving.
Cost-effectiveness modelling in PD between 2003 and 2010 showed only minor improvement when compared with our earlier review of models published from 1998 up to 2003. Cost-effectiveness modelling recommendations were complied with to only a limited extent, leaving room for quality improvement. More advanced modelling approaches may, so far, be underrepresented, but may be used in the future, driven by the research question. Adverse events of treatment, co-morbidities or disease complications are not yet sufficiently included in the models to adequately represent clinical reality.
In recent years, evaluation of cost effectiveness has become an increasingly important issue in the context of reimbursement decisions by healthcare providers. The benefit of new treatment options for society or healthcare providers must be proven through cost-effectiveness analyses, particularly in chronic diseases with high prevalence and high lifetime costs.
Parkinson’s disease (PD) has a prevalence of 66–258 per 100 000 and, as such, is among the most common neurodegenerative disorders.[1, 2, 3, 4, 5] The disease is characterized by tremor, bradykinesia, rigidity and postural instability; however, non-motor complications such as behavioural and psychological symptoms (e.g. depression, dementia and hallucinations), sleep disorders and gastrointestinal problems are common and measurably affect quality of life (QOL). In addition to levodopa, a large variety of drug treatment options have been developed in the last 20 years, including dopamine agonists, monoamine oxidase (MAO-B)/catechol-O-methyl transferase (COMT) inhibitors, anticholinergic drugs and N-methyl d-aspartate receptor (NMDA) antagonists (amantadine). These, alone or in combination, are used in the treatment of motor symptoms.[7, 8, 9] Moreover, surgical approaches such as deep brain stimulation (DBS) have become an increasingly important and reliable cornerstone of treatment for severely incapacitated but selected PD patients.[10,11]
Comparing these treatments from a health-economic perspective is a difficult task for several reasons: (i) different adverse events may occur; (ii) different costs and utilities may result; (iii) the sociodemographic characteristics of PD patients, such as age, affect the course of the disease and therefore influence treatment; and (iv) only short longitudinal studies are available that may not be sufficient to capture the effects of a lifelong therapy. Nevertheless, several modelling approaches have been presented for evaluating the cost effectiveness of treatment options for PD.[12, 13, 14] In this review, we provide an overview of mathematical modelling approaches in PD and derive recommendations for future modelling. Common modelling techniques employed to date include decision trees, Markov models, discrete-event simulation (DES) models and sets of mathematical equations. We restricted this review to papers published after July 2002. For models published before this date, we refer to a previous work by our group. Modelling studies were included if they consisted of a health economic evaluation of two or more treatment options.
Key points for decision makers
Markov models are still the predominant methodology used to assess the cost effectiveness of treatments for Parkinson’s disease
Decision makers should ensure models have a sufficiently long time horizon
There is room for improvement in the quality of models. Clinically relevant aspects such as treatment of adverse effects, co-morbidities and disease complications are not yet considered to an appropriate extent
Transparent reporting and critical evaluation of models is crucial. Limitations in the transferability of data and results should be outlined in detail
1. Literature Review and Assessment
1.1 Search Methods
Consistent with our previous publication, a literature search was conducted in the databases PubMed, Current Contents, EMBASE, EconLit, Cochrane Database of Systematic Reviews, DARE (Database of Abstracts of Reviews of Effectiveness), NHS EED (Economic Evaluation Database) and HTA (Health Technology Assessment) of the UK NHS Centre for Review and Dissemination (from July 2002 to March 2010). The following keywords were employed: ‘decision analysis’, ‘decision-analytic’, ‘decision model’, ‘health care model’, ‘health care evaluation model’, ‘decision tree’, ‘Markov model’, ‘discrete event simulation’, ‘cost-effectiveness’, ‘cost-utility’, ‘cost-benefit’, ‘cost-minimiz(s)ation’, ‘QALY’ and ‘Parkinson’.
In addition, we examined reference lists of studies identified and reviewed articles from the archives of the authors. The search was restricted to the following six languages: English, German, Spanish, Portuguese, Italian and Polish.
We only included published full papers that employed a decision-analytic model or another type of mathematical healthcare model (including decision trees, Markov models, DES and sets of mathematical equations), and that evaluated therapeutic interventions or diagnostic procedures for PD, following the definition of a “model” as suggested by Weinstein et al.
Additionally, at least two treatment options had to be compared and the primary aim of the studies had to be a health economic evaluation of these treatments. We excluded studies using models only as an illustration of methodological aspects or those published as abstracts.[16, 17, 18, 19, 20]
1.2 Data Extraction and Model Assessment
study reference (bibliographical notes, setting and year)
decision-analytic framework (target population, study question/objective, study type, alternatives compared, time horizon, outcome measures, model structure, statistical analysis, perspective, annual discount rate)
data sources (epidemiology/natural history, efficacy, health-related QOL [HR-QOL]/utilities, costs)
model validation (key structural and parametric assumptions)
results (for costs and effectiveness and their relationship; sensitivity analysis)
limitations (reported by the author)
conclusions (reported by the author) and
We used a checklist by the German Scientific Working Group to characterize the quality and transparency of the health economic evaluations. This checklist comprises 56 questions regarding subjects such as problem, frame of evaluation, methods of analysis and modelling, health outcomes, costs, discounting, presentation of results, uncertainty, discussion and conclusion. Each single criterion could be valued by points as follows: 1 = criterion fulfilled, ½ = criterion partly fulfilled, 0 = criterion not fulfilled.
1.3 Search Results
The literature search resulted in 90 hits in PubMed, 96 hits in Current Contents, 263 hits in EMBASE, 50 hits in the Cochrane Library, ten hits in EconLit and five hits in the NHS database. After screening the references, a total of 11 relevant studies were included and assessed in greater detail.
2. Systematic Description and Assessment of Studies
2.1 Evaluation of Drug Treatment
Lindgren et al. evaluated the cost effectiveness of early treatment of PD with the dopamine agonist cabergoline1 (with possible later addition of levodopa) compared with standard levodopa therapy with respect to onset of motor complications. A Markov model was developed with seven states, including a state without signs of disease, states for Hoehn and Yahr (HY) scale stages I–IV, a state for patients with motor complications (regardless of HY stage) and a death state. The authors performed a deterministic cohort simulation over a 5-year period, with 6-month cycles, to determine the incremental costs per year with motor complications avoided. A randomized controlled trial (RCT) over 5 years provided data for the estimation of transition probabilities. Using a regression equation, transitions between HY stages were modelled as a function of age of disease onset, actual HY ‘on’ stage2 and treatment status. The risk of motor complications was estimated by a Weibull regression model. Mortality was assumed to be the same as for the general population. Given the third-party payer perspective of Sweden, only direct costs without co-payments were included. The costs associated with PD (including treatment costs) were taken from a Swedish study of 127 patients. Costs for patients with motor complications were estimated to be twice as high as those for patients without motor complications. Data were from published European studies,[27,28] assuming that 50% of patients had motor complications. Costs and effects were discounted by 3% annually, and all costs were adjusted to year 2000 values.
The study estimated that cabergoline resulted in a gain of 0.31 years without motor complications, at additional costs of about €4300 over a period of 5 years (€13 900 per year of motor complications avoided). Extensive sensitivity analyses led to incremental cost-effectiveness ratios (ICERs) of €6124–14 384 per year of motor complications avoided and demonstrated the robustness of the model. Multiway (probabilistic) sensitivity analysis or validation attempts were not reported.
The model presented by Lindgren et al. was based on a broadly accepted functional clinical scale (HY scale) and presence of motor complications as Markov states, together with an elegant approach to model transition probabilities. The authors assumed mean, HY-stage weighted costs for patients with motor complications, which may represent an over-simplification. QALYs were not calculated, and calculation may require adaption of the model structure. A time horizon of 5 years (10 years in sensitivity analysis) may not be sufficient, given the chronic progressive nature of the disease. In summary, they presented a transparently reported, conclusive study with a robust model. A more detailed description of the model structure and transition probabilities would have been desirable. Furthermore, some aspects may have been over-simplified. Validation results were lacking.
Smala et al. evaluated the cost effectiveness of cabergoline versus levodopa in a Markov model for patients with early PD. Markov states were based on HY ‘on’ stages, the analyses were performed stratified for age group (< vs ≥ 60 years). Cost effectiveness was measured in a cohort simulation by costs per Unified Parkinson’s Disease Rating Scale (UPDRS) total score point decrease and per motor complication omitted. Data on treatment efficacy were taken from a published multicentre RCT over 5 years; data on the natural course of the disease were from an observational study and cost data were from a longitudinal German cost-of-illness study. The target population of the analysis was chosen to be similar to the RCT study population (i.e. HY stages I–III). The time horizon was 10 years, with a cycle length of 1 year. Costs were evaluated from a societal perspective in the German healthcare system, and indirect costs were included in the analysis. All costs were adjusted to year 2002 values and discounted by 5% annually. One-way sensitivity analyses were performed over several input variables.
Incremental costs per patient without additional motor complications were estimated as €104 400 for those aged <60 years and €57 900 for those aged ≥60 years. For patients aged ≥60 years, a cost effectiveness of €1031 per decreased (i.e. improved) UPDRS score point was calculated. For the group aged <60 years, levodopa strongly dominated the cabergoline strategy. From the sensitivity analyses performed, the authors concluded robust results, although cost-effectiveness measures were not shown.
Smala et al. presented a Markov model based on the progression of PD. However, a death state was obviously omitted, and QALYs were not calculated, although it would have been feasible to do so. No age-specific death rates were applied, the risk of motor complications was assumed to be constant over all HY stages modelled and costs were calculated independent of age. The costs used to value the various HY stages or motor complications were not reported. The authors did not adjust 5-year cumulative incidence to 1-year probabilities by first transforming them into rates as would have been the correct approach. Median instead of mean values were used for costs, and costs — but not effects — were discounted. The inclusion of indirect costs may be questionable because of the relatively high age of the PD patients. A 10-year horizon may not be sufficient to capture all relevant effects given the chronic progressive nature of the disease. No sensitivity analyses were performed on efficacy parameters. Multi-way (probabilistic) sensitivity analyses were not performed, and validation efforts were not reported.
Hudry et al. evaluated the cost utility of the second-generation MAO-B inhibitor rasagiline and the COMT inhibitor entacapone as adjunctive therapies to levodopa and compared this with standard levodopa care in patients with motor fluctuations. They used a Markov model that was based on the previously published model of Nuijten et al. and Palmer et al. The Markov model consisted of three states consistent with the UPDRS item 39. These included ‘off’ stages ≤25% and >25% of the waking time, and a state for death. No adverse events were taken into account. A 2-year time horizon was used, with 4-month cycles. The model was evaluated by Monte Carlo simulation. As an effectiveness outcome, Hudry et al. used QALYs and the number of months with ≤25% ‘off’ time per day. Probabilistic sensitivity analysis was performed for several scenarios and demonstrated differences in efficacy and drug prices.
All drug-specific transition probabilities were derived from the LARGO (Lasting effect in Adjunct therapy with Rasagiline Given Once daily) RCT. The authors assumed that patients in the levodopa-only arm could improve their daily ‘off’ time only in the first cycle, which is regarded as a conservative assumption in the evaluation of new combination treatments. For patient demographics and the calculation of QALYs, they used results of a recent publication. Resource utilization and costs were derived from a Finnish cost-of-illness study, except for drug costs, which were estimated using a retail price per defined daily dose of levodopa and entacapone. Mortality rates were from a Norwegian community-based cohort study and utilities were taken from a US study. Costs for patients with >25% ‘off’ time per day were assumed to be twice as high as for those with less ‘off’ time per day. Costs and effects were discounted by 5% annually (no undiscounted results were reported), and costs were adjusted to year 2004 values. The perspective was societal (including indirect costs) and from a third-party payer perspective (with direct costs only) within the Finnish healthcare system.
From a third-party payer perspective, rasagiline adjunctive therapy resulted in an ICER of €17 800 per QALY over 2 years; entacapone adjunctive therapy resulted in an ICER of €18 600 per QALY. If indirect costs were included in the model, both therapies were cost saving. The model was sensitive to prices for entacapone and rasagiline and the cost proportions between the health states. No validation of the model was reported.
Hudry et al. kept the model structure simple by determining the course of disease through ‘off’ time per day. No adverse events, further co-morbidities or complications were included. This may not entirely reflect the disease and, instead, a more complex scale, such as the HY scale, may map the course of PD in more detail. The authors described all input data and assumptions transparently and in appropriate detail. In the sensitivity analysis, drug costs had a major impact (a 20% decrease resulted in cost savings of 61%). Utilities were taken from a US population, which may not reflect the Finnish cultural background. In addition, no sensitivity analyses were performed concerning utilities. A more detailed sensitivity analysis of the relationships between the health states would have been of interest, given the nature of this assumption (only drug costs were considered in the sensitivity analysis). The presentation of more extensive multi-way and especially one-way sensitivity analyses would be desirable, as these confer different information from probabilistic sensitivity analysis alone. Unfortunately, the authors only used a 2-year perspective, which does not reflect the long-term progression of PD, although they argued that the model was not constructed to reflect the course of the disease or treatment beyond a short time interval. A more detailed discussion on input data and influence on the results, as well as a discussion and comparison with other (independent) health economic studies would have completed this modelling approach.
Haycox et al. developed a Markov model to compare the MAO-B inhibitor rasagiline with the dopamine agonist pramipexole in early idiopathic PD patients. Patients who were initially treated with rasagiline could switch to levodopa or pramipexole monotherapy and patients who initially received pramipexole could switch to levodopa until the emergence of treatment-related dyskinesias, which were regarded as an endpoint (and an absorbing state). A death state was not included in the model. The outcomes to be estimated from the model were time to levodopa treatment, time to dyskinesia onset, QALYs and direct costs. The authors took a third-payer perspective in the context of the UK NHS. The time horizon was 5 years, with 6-month cycles.
Data for clinical effectiveness were taken from three RCTs: one for rasagiline[37,38] (6 years of follow-up) and one for pramipexole[39,70] (4 years of follow-up) and a third, from which the rate of dyskinesias within the levodopa treatment arm was estimated. Disease duration was significantly shorter for rasagiline patients, whereas the UPDRS score was higher for pramipexole patients in those RCTs. The model included direct costs as reported by a UK study, together with drug costs based on the NHS wholesale purchase price (WPP). All costs were adjusted to £, year 2007 values using the UK ‘consumption index’. Utilities were taken from Palmer et al. using the EQ-5D in a sample of 63 US patients. Discounting was performed according to the UK National Institute for Health and Clinical Excellence (NICE) guidelines (at the time), with 1.5% for health effects and 6% annually for costs.
Rasagiline proved to be the dominant strategy. Costs were reduced by £3931 (−18%) and QALYs increased by 0.19 (5%) per patient over 5 years in the rasagiline treatment arm. Rasagiline was associated with a 10% longer time span until the onset of dyskinesias and a 25% longer time until levodopa initiation than pramipexole. Sensitivity analysis showed that the model was robust.
This model considered strategies with the possibilities of switching treatments in patients with PD. However, the authors did not explicitly highlight that the results did not represent a direct comparison of pramipexole and rasagiline treatment, but rather a strategy combining both drugs. The model was transparently reported, although it remained unclear where Monte Carlo simulation was applied and (partly) how transition probabilities were derived from the sources. The time horizon of 5 years was quite short for a chronic disease such as PD. However, for a longer time horizon, a death state would have been necessary, and this was omitted. Although the authors justified this decision, a more thorough discussion was warranted. The authors discussed differences in the study populations of the underlying RCTs and concluded that these acted against the rasagiline strategy. Cost data were from 1998 and could be outdated. A societal perspective (as suggested by Gold et al.) was not presented. Extensive sensitivity analyses were lacking, and no undiscounted costs and effects were presented. Furthermore, a comparison of the results with independent guidelines or the influence of input data and model structure on results was missing. Utilities were taken from a different cultural context, which may not be fully appropriate, but were subjected to univariate sensitivity analysis.
Findley et al. presented a Markov model originally developed by Linna et al. for determining the cost effectiveness of a combination of levodopa, carbidopa and entacapone (LCE) over a 10-year period. Patients entering the model were aged 66 years, with an HY distribution derived from a Finish naturalistic study. The Markov model consisted of eight states: seven represented by the modified HY scale, and a death state. Cycle length was 6 months. Progression probabilities were taken from levodopa- and placebo-controlled trials of entacapone.[40,45] Costs were modelled from a societal and a third-party payer perspective (the UK NHS), comprising only direct costs (NHS costs, social service costs and private PD-related expenditures) and medication costs using official tariff lists from 2005.[35,42,72] All costs were adjusted to year 2003 values using the healthcare-specific and the overall consumer price index of the UK. Utilities were taken from a study reporting EQ-5D data from a sample of 124 patients in the UK. Mortality rates were taken from UK registries for the general population, assuming a 2.3-fold higher PD-specific mortality. Costs and effects were discounted by 3.5%.
From a societal perspective, LCE provided a gain of 1.04 QALYs and a cost reduction of about £10 200 for the UK over the 10-year time horizon and was therefore dominant. Taking the perspective of the NHS, incremental costs of about £3200 per QALY resulted in an ICER of £3100 per QALY.
Univariate sensitivity analysis on discount rates and with a time horizon of 5 years showed robust results. Second-order Monte Carlo simulation over costs, utilities, transition probabilities and initial HY distribution still supported the dominance of LCE for the societal perspective and cost effectiveness at a threshold of £30 000 per QALY from the NHS perspective.
Findley et al. transparently reported a Markov model representing the natural course of the disease based on HY stages. No further considerations were made for adverse events such as motor or non-motor complications. Nevertheless, a more detailed description of the course of disease and transition probabilities would be desirable. National utility values were used and PD-specific mortality was modelled, which enhanced plausibility and acceptance of the model. However, reporting of more extensive one-way sensitivity analyses would have further increased transparency. Methods used to extract HR-QOL data were not reported in detail. The extensive probabilistic sensitivity analyses demonstrated the robustness of the model and were a strength of this study. The use of Normal distributions for utilities in the probabilistic sensitivity analysis could have led to numerical problems and may not be appropriate. The time horizon of 10 years may not be sufficient for a lifetime disease, but the presentation of a sensitivity analysis on the time horizon proved very helpful.
Kristiansen et al. developed a decision tree with a 2-year time horizon to evaluate the cost effectiveness of duodenal levodopa infusion (DLI) versus oral levodopa for the treatment of advanced PD. Patients were offered either conventional polypharmacotherapy or DLI. If treatment with DLI was not tolerated, the patient could switch to conventional therapy after 5 days and 6, 12, 18 and 24 months. The model also considered possible death, although only a short time horizon was chosen. One-way and probabilistic sensitivity analyses were performed to evaluate the impact of all model parameters.
Input data were taken from an open, randomized, controlled, crossover study over 3 + 3 weeks (DIREQT [Duodopa Infusion-Randomized Efficacy and Quality of Life Trial]) and from an observational study. The costs included those for drug acquisition, drug administration, initiation and monitoring of DLI, and treatment of complications associated with DLI. Costs were from the perspective of the Swedish healthcare system and were reported in Swedish kronor (SEK) in year 2004 values. HR-QOL was evaluated within the DIREQT study using the 15D, a 15-dimensional generic utility instrument shown to be sensitive in PD. Assessments took place at baseline and after the two 3-week trial periods, and then every 3 weeks during the subsequent 6 months within the DIREQT study. Both costs and HR-QOL were discounted by 3% per annum; results were also reported undiscounted.
The expected per-patient 2-year costs of the DLI strategy was SEK562 000, while it was SEK172 000 for conventional therapy (undiscounted: SEK569 000 and SEK175 000). Mean QALYs were 1.48 and 1.42 (undiscounted: 1.5 and 1.44), resulting in a ICER of SEK6.1 million per QALY. Using other assumptions in sensitivity analyses to account for indirect costs, avoided nursing costs and use of (expensive) apomorphine therapy in all patients in the conventional therapy strategy, the incremental costs per QALY could be as low as SEK456 000. The ICER was most sensitive to daily costs of DLI and costs of the conventional therapy, the change in HR-QOL, the discount factor and the costs of neurologist visits. Thus, Kristiansen et al. suggested that DLI may be associated with significant health benefits for patients with advanced PD depending on how the health benefits were measured. They also raised questions as to how these benefits should be valued by society, as DLI has been granted orphan drug status.
Kristiansen et al. developed a simple model structure with few and conservative assumptions alongside a clinical trial to evaluate the cost effectiveness of DLI. No clinical scale was used in the model. Costs were modelled from a third-party payer perspective only, and not from a societal perspective, as suggested by Gold et al. The main input data for clinical efficacy and HR-QOL were limited to a single study including only 24 patients. All data used were extensively and transparently reported. The discussion of costs was kept short. Although the authors argued that a lack of long-term data precluded a longer time horizon, estimates for the effects within a longer timeframe would have been of interest. This could have been conducted in two stages: the first including the trial data and the second representing the extrapolation. However, this probably would have required changes in the model structure. No validation attempts were reported. The performance of extensive sensitivity analyses adds considerably to the quality of the study. Kristiansen et al. modelled the impact of treating advanced PD with DLI, which had not been previously evaluated. This was a carefully conducted and very transparently reported study with a detailed discussion on input data and their impact on results.
2.2 Evaluation of Surgical Options
Goulionis and Vozikis presented a Markov decision model that evaluated the cost effectiveness of DBS compared with best medical treatment (before DBS). Mild, moderate and severe symptoms (referred to as ‘adverse events’) were chosen as Markov states. Instead of directly determining deterioration of the patient’s health, which could not be ascertained using clinical data, patients’ health was linked to possible observations such as activities of daily living, induced tremor and motor functions. Because patient progression was unobservable, Goulionis and Vozikis chose a partially observable Markov model as their modelling approach.
Costs were calculated dependent on observed symptoms (e.g., motor functions) and treatment; clinical (and treatment) effects were approximated by using the UPDRS and Sickness Impact Profile data. Data were from a cohort of 150 patients from the Greece General Hospital (year 2009).
Instead of considering cost effectiveness using ICERs, Goulionis and Vozikis performed their calculations with the aim of optimizing the total lifetime costs. The calculated thresholds at which medical treatment was more cost effective than DBS were linked to the inital probability of having mild, moderate or severe adverse events. If patients did not develop severe adverse events, a 40% chance of experiencing moderate adverse events was needed for DBS to be cost effective. Furthermore, a ≥75% chance of experiencing severe adverse events implied that DBS was cost effective. Cost effectiveness was preserved if the chance of developing moderate adverse events was <4% and was balanced by a higher probability of the occurrence of severe adverse events. Further interpretation of these results was difficult, especially with respect to clinical implications. The model was presented in a very technical and abstract way, which obviously limits its acceptability. Data were presented in a very disaggregated way, if at all. Data were not from published sources, which severely compromises transparency. Results represented the probabilities of achieving a certain state (mild, moderate or severe adverse events), based on a cost-minimization approach. Together with the very rough selection of only three Markov states, further interpretation is considerably limited.
Hjelmgren et al. developed a model for the early evaluation of hypothetical interventions capable of modifying progression of PD and applied it in assessing the cost effectiveness of a future dopamine cell replacement therapy as compared with standard drug therapy. The model also served to estimate a possible price premium for the compensation of drug development costs. Based on a Markov model, six health states were defined, including ‘off’ time HY stages I–V and a death state. The authors used ‘off’-phase HY stages because the model aimed at interventions that did not relieve symptoms but rather altered the underlying disease progression and pathology. The cycle length was not reported.
The natural history of disease was based on data from an original study of 79 patients in a routine clinical setting. A multivariate regression model was developed that calculated disease progression as a function of age at diagnosis, HY stage and time. Utilities stratified by HY stage estimated using the EQ-5D came from two published studies.[44,52] Cost data by off-phase HY stage were taken from an earlier Swedish PD cost-of-illness study. The model included PD-specific mortality, estimated from Swedish registries and a cohort study. For the application of the model to the intervention of an intrastriatal graft of dopamine-rich human embryonic ventral mesencephalic tissue, efficacy data came from an earlier published clinical study of 14 patients at the researchers’ institution. This non-randomized, non-controlled study provided data over up to 10 years of follow-up. Based on this study, the authors assumed a progressive improvement over 2 years, followed by a stationary period of 5 years and natural disease progression after this point. Efficacy assumptions were extensively evaluated in sensitivity analyses. Furthermore, this clinical study also provided intervention costs, including intervention-related care. The costs of complications were estimated with data from a literature review of clinical transplantation trials and conservative assumptions about the costs incurred by the most frequent adverse events, haematoma and dyskinesias. Assumptions about the incidence of adverse events were also varied in a wide range of sensitivity analyses. All cost data were adjusted to year 2002 values and costs and effects were discounted by 3% annually. A third-party payer perspective was adopted.
The study results indicated long-term effectiveness and cost savings for the neurological graft procedure if direct non-medical costs (transportation, cost of care) were included. If only direct medical costs were included, the new intervention would still be regarded as cost effective over a considerable range of price premiums.
This study was performed carefully in order to evaluate a hypothetical intervention. A major strength of the model was that it modelled the course of the disease using correlations of the underlying pathological process. The model was partially based on primary data and on carefully researched, published data. Where no data were available, conservative assumptions were used and explicitly described. All parameters that were based on assumptions were evaluated in sensitivity analyses, which showed mostly robust results. However, as utility values were from the UK, more extensive sensitivity analysis would have been desirable. In addition, the sensitivity analysis with regard to treatment efficacy could have tested even more conservative assumptions (e.g. fall-back assumption). As this parameter is the most crucial of the whole model, even more in-depth analyses could have provided more insight into cost effectiveness. Costs were modelled from a third-party payer perspective, but not from a societal perspective, as suggested by Gold et al. The ICER results were presented as a function of the price premium of the intervention, which represents a rarely used modelling application. The authors were cautious in the interpretation of their results, recognizing the weak evidence for the efficacy data used. They described the results of external validation of the disease progression of the model, although not in detail. In addition, the authors stated that they were not willing to “report either internal or external validity tests.” The model can be helpful in decision making regarding future interventions that are able to change the pathology and course of the disease.
2.3 Diagnostic Evaluations
Antonini et al. performed a cost-effectiveness study of 123I-FP-CIT (Ioflupane 123I) single photon emission computed tomography (SPECT) for the differential diagnosis of patients with movement disorders, either Parkinson syndrome (PS) [including idiopathic PD, progressive supranuclear palsy and multiple system atrophy] or essential tremor (ET). The diagnostic SPECT technique was compared with current diagnostic practice over a 5-year time horizon. The outcomes of interest were projected years on ‘potentially beneficial therapy’ (PBTY), for either ET or PS, and costs. The authors developed a Markov model with eight states. In addition to ‘dead’ and ‘no therapy’, there were states for correct ET and correct first- and second-line PS therapies and wrong treatments (ET therapy for PS, first- and second-line PS therapy for ET). Falsely diagnosed patients were considered as not being on PBTY. False-positive and false-negative diagnoses (with inadequate treatment) were considered equal (except for the costs). Non-response rates and adverse event rates were not stratified by true or false diagnoses. Cycle lengths of 6 months were used. Results obtained through a deterministic cohort simulation were supplemented through one-way sensitivity analyses.
The diagnostic accuracy of SPECT and the clinical exam, and probabilities for withdrawal from therapy because of adverse advents, were taken from published clinical studies.[56,57] The advantage of the new technology was its high sensitivity (97%) and specificity (100%) compared with conventional diagnosis (sensitivity 97%, specificity 82%). The underlying prevalence of PS and ET, probability of non-response to PD therapy and resource use were based on expert surveys (Delphi panel of 12 neurologists). Mortality estimates were based on national statistics, and unit costs were gathered from public databases, the SPECT manufacturer and official tariff lists or DRGs. Antonini et al. evaluated the cost effectiveness from the perspective of the Italian healthcare system, and all costs were adjusted to year 2005 values. Costs and effects were discounted by 5% annually; however, undiscounted results were also reported.
The study concluded that the SPECT-based diagnosis led to an increase of 1.68 years of PBTY per patient over a time horizon of 5 years while being cost saving (-€341). After the time horizon of 5 years, the distributions of proportions of patients untreated or on PBTY were similar between the two strategies. Univariate sensitivity analyses showed robust results; the most influential parameter was the prevalence of PS in the respective setting. It is important to note that the study results changed from cost saving (i.e. more effective and less costly) to ‘cost effective’ (i.e. more effective and more costly) within the range of prevalence of PS in movement disorders clinics reported in the published literature.
The model was clearly described. Data sources and assumptions were explicitly stated. The rationale for most assumptions was given and most of the assumptions were considered to be conservative by the modellers. Data for diagnostic accuracy of clinical examination were used from a single publication without justification. In addition, the parameter for specificity was at the lower bound of published estimates, which acted in favour of the SPECT strategy. The model was based largely on expert opinion. A societal perspective was not presented, as suggested by Gold et al. Sensitivity analyses have been performed on several important parameters but not on the critical parameters such as diagnostic accuracy of clinical assessment, with the values used differing considerably from those seen in other evaluations. Probabilistic sensitivity analysis would have led to additional insights. Cross-validation was attempted, but reasons for differences of the results were not discussed in detail.
The authors admitted that clear patient-relevant health outcomes were not presented, which would only have been possible with additional assumptions. They asserted that the calculation of QALYs was not possible due to missing data and, moreover, considered it inappropriate because the QALY concept would not capture intangible costs (“less tangible costs”) assumed by the authors to be present in a relevant extent in this specific context (e.g. “provide patients and their families with a better understanding of the likely course of the disease”). However, the study would have been considerably more relevant if the inclusion of patient-relevant health outcomes had been attempted. This may also be of importance because, as the authors reported, there was no difference in treatment status between the two diagnostic strategies after 5 years. As long as the SPECT strategy is cost saving, this may be of reduced importance, but becomes crucial if SPECT becomes ‘cost effective’, leading to the question, ‘what value is society or the NHS willing to pay for what kind of improved outcomes?’.
No conflict of interest was explicitly declared; however, the model was developed for the use of the manufacturer in seeking reimbursement for the tracer, and some of the authors were employed by, or had received (and declared) grants from, the manufacturer.
Van Laere et al. presented a 5-year Markov model to evaluate the cost effectiveness of using 123I-FP-CIT SPECT in the diagnosis of neurodegenerative Parkinsonism (NDP) in contrast to ET in a patient population of ‘difficult’ differential diagnosis between NDP and ET in an early stage of the disease. The model consisted of eight Markov states: (i) NDP with first-line NDP therapy (e.g. levodopa, dopamine agonists); (ii) ET with first-line NDP therapy (i.e. inappropriate therapy); (iii) NDP with second-line NDP therapy (e.g. selegiline, anticholinergics, amantadine); (iv) ET with second-line NDP therapy (inappropriate therapy); (v) ET with ET therapy (e.g. β-adrenoceptor antagonist therapy); (vi) NDP with ET therapy (inappropriate therapy); (vii) no therapy (inappropriate); and (viii) death. The model aimed to estimate the time on appropriate therapy (adequately treated years [ATY]) and the proportion of changes in treatment due to results of the SPECT procedure. A cycle length of 6 months was used.
Estimates for epidemiological and clinical data (prevalence of ET/NDP, proportion of patients eligible for SPECT and uncertain cases, SPECT availability, incidence of disease complications, sensitivity and specificity) for model input were taken from published studies or (predominantly) estimated from a Delphi panel consisting of 13 Belgian neurologists and nuclear medicine specialists. Estimates were collected with two rounds of postal questionnaires. Members of the panel were not stated. Additional data from a multicentre registry with 1701 consecutive patients undergoing SPECT imaging were used to validate input estimates into, and projections of, the Markov model.[57,59, 60, 61] Costs were from the perspective of the Belgium healthcare payers and included direct medical costs for diagnosis, treatment and adverse events. Cost data were estimated from the Delphi panel and were reported in an aggregated manner. The costs’ reference year was not stated. Costs and health effects were discounted by 5% per annum.
Incremental discounted costs for the SPECT strategy were €429 at an NDP prevalence of 60% (as estimated from registry data) and a clinical efficacy of 1.198 ATY, leading to an ICER of €358 per ATY. No undiscounted results were presented. The model results were reported to be robust with respect to rates of adverse effects, but were sensitive to NDP prevalence. When validating the model results with data from the registry, the model prediction that therapy would not be changed in 51.5% of patients after SPECT (changes in 48.5%), corresponded well to the respective proportion of 40% from the observed registry data (changes: 49%; therapy data for 11% of patients were missing from the registry).
This study simulated cost effectiveness for diagnostic strategies in the work-up of unclear early PD in the Belgian healthcare system and reproduced the results by Dodel et al. Data sources relied primarily on estimates from a Delphi panel, and cost data in particular were presented in an aggregated form. The respective costs for the Markov states used were not reported. It appears that a review of the existing literature for data sources was not performed, and some input data could be out of date (e.g. sensitivity and specificity of SPECT). The target population of the model was not clearly described. It is unclear how old the cost data were. A societal perspective was not taken. Given the large uncertainty in input parameters, the extent of sensitivity analyses seems inappropriate, and multi-way sensitivity analyses were lacking. The most critical parameter, the prevalence of NDP, was calculated from registry data using SPECT performance data. No attempts were made to correlate the outcome measure of ATY to a patient-relevant outcome. On the other hand, the validation efforts (external and cross-model validation), were a major strength of the study even though they were not discussed in detail.
Dodel et al. published a decision tree for evaluating the clinical and economic impact of different diagnostic strategies for PD involving 123I-FP-CIT SPECT. The hypothetical study cohort consisted of patients with PD symptoms (HY stages I or II) presenting to a specialized movement disorder clinic. Four strategies were considered for diagnosis (with subsequent treatment): (i) diagnosis based only on clinical examination; (ii) diagnosis based only on SPECT results in all patients; (iii) SPECT only if clinical examination was negative; PD is assumed (and treated) if either clinical examination or SPECT is positive; or (iv) SPECT only if clinical examination was positive; PD is assumed (and treated) if both clinical examination and SPECT were positive. The outcome was measured in adequate treatment months. The months of adequate treatment were weighted by a consequence ratio for false test results (i.e. wrongly treating non-PD patients to withholding treatment in PD patients, adequate treatment month equivalents [ATME]).
Data were extracted through a systematic literature search. Prevalence data were taken from Meara et al. and test accuracy data were from several published studies.[56,59,61,63, 64, 65] Treatment costs came from a German cost-of-illness study and diagnostic costs were determined by evaluating resource consumption in small samples of patients undergoing diagnosis in a German university hospital and by valuing these with prices from official price and tariff lists. All cost data were in year 2002 values. Given the time horizon of 12 months, no discounting was performed. All costs were from a healthcare provider (hospital) perspective in Germany.
Results indicated that strategies with only SPECT, or SPECT following a negative clinical examination, had lower effectiveness at higher costs (33 ATME, €2003 and €1731, respectively) than the strategies with clinical examination only (52.85 ATME, €946) and SPECT following positive exam (53.40 ATME, €1352). This strategy had an ICER of €733 per ATME compared with the exam-only strategy. SPECT should therefore be used as a confirmatory test in patients with a positive clinical examination before treatment initiation. Extensive one- and two-way sensitivity analyses showed robust results.
This study reported assumptions, data sources and the modelling approach in a very detailed and explicit manner. The rationale for the assumptions was given by the authors and the assumptions were appropriate or probably biased against SPECT-based strategies. A societal perspective was not taken, as suggested by Gold et al. The structure of the model was validated by clinical experts as were parameter estimates. Because data for the calculation of QALYs were missing, the authors tried to estimate the relevance for patients, taking into account the fact that harming non-PD patients with adverse effects from a treatment based on a false-positive diagnosis is less acceptable than withholding treatment from PD patients due to a false-negative diagnosis.
This review assessed models for PD published after July 2002 to supplement a previous review and evaluation. We presented in detail the structures and practical aspects of the respective modelling approaches (tables I–III). The studies reflected the healthcare systems of Belgium, Finland, France, Germany, Greece, Italy, Sweden and the UK. Patients with early and advanced PD stages were evaluated, with particular consideration given to motor complications. Different medical treatment options, including surgical approaches such as DBS or cell-replacement therapy and drug treatments with cabergoline, rasagiline, pramipexole, LCE or DLI[24,29,32,36,43,47] were compared. Three studies evaluated 123I-FP-CIT SPECT as a diagnostic test for early PD.[55,58,62]
Among the 11 studies identified, six were cost-utility analyses[32,36,43,47,49,50] and five were cost-effectiveness studies.[24,29,55,58,62] Outcomes included QALYs,[32,36,43,47,50] life expectancy, decreases in UPDRS score, time on PBTYs, ATYs and ATMEs. Data for utilities, costs and transition probabilities were taken from clinical trials or national databases.[3,25, 26, 27, 28,30,31,33, 34, 35,37,38,40,42,44,45,52,54,56,57,59,61, 62, 63, 64,74] Expert opinion was frequently used when data were missing. All studies except one demonstrated that the strategy of interest was cost saving or cost effective. The analyses had time horizons between 1 year and (in only one study) lifetime and were performed from the societal or (in the majority) the healthcare provider or payer perspective.
There were two primary, unique types of models: decision trees and Markov models. Only two studies used a decision tree for a time-period of 1 year from the healthcare payer perspective.[47,62] All remaining models were based on Markov assumptions.[24,29,32,43,49,50,55,58] So far, no microsimulation approach or DES model has been published in the field of PD.
Two main aims were distinguishable among the studies: to determine the cost effectiveness of treatment options, or to evaluate new diagnostic test procedures or test strategies. In the latter, the impact of the diagnostic test 123I-FP-CIT SPECT was evaluated.[55,58,62] The decision as to whether clinical examination or 123I-FP-CIT SPECT adequately identified PD was measured in PBTY, ATY or ATMEs for the particular healthcare provider. Two studies indicated that 123I-FP-CIT SPECT testing was likely to be cost effective compared with clinical examination,[55,58] whereas the third showed that 123I-FP-CIT SPECT testing is only effective in combination with clinical examination.
Drug treatments were addressed in most of the models included in this review: cabergoline versus levodopa,[24,29] LCE versus standard treatment, DLI versus oral levodopa, rasagiline versus pramipexole and rasagiline versus entacapone. One model identified the impact of DBS in a partially observable Markov model based on HY scale and considered the improvement in motor complications. Another working group determined the impact of DLI in advanced PD, allowing patients to receive levodopa if DLI was not tolerated. In contrast to determining disease progression (by the UPDRS or HY),[17,29] a number of evaluations were based on QALYs[19,32,36,43,49] or motor complications avoided[16,19,24,32] as an effectiveness outcome. Only two studies considered other complications or the time to levodopa treatment and time to levodopa-induced dyskinesias. Lindgren et al. and Smala et al. studied the cost effectiveness of cabergoline. Both concluded that cabergoline dominated levodopa; however, Lindgren et al. demonstrated a greater effectiveness in younger patients and Smala et al. in patients aged >60 years. Two other studies showed that, compared with standard treatment or pramipexole, rasagiline was cost effective.[32,36] Furthermore, entacapone alone was preferable to standard treatment, as well as within an LCE combination.
3.1 Model Structure
Different model structures were applied. In order to determine the cost effectiveness of a diagnostic test, comparisons between the clinical examination and the test were considered.[55,58,62] When evaluating the costs and effectiveness of drug treatment or surgery, progression was measured by ‘off’ times per day[16,17,19,32] or the HY scale,[24,29,43,50] and adverse events such as motor complications were also considered.[24,29,36] In order to assess disease progression, different outcomes such as the ‘off’ time per day or the HY scale were chosen. For the former, patients in different stages of PD were clustered into one group, so that varying health outcomes could not be adequately distinguished. This could be avoided by using the HY scale; however, a differentiation into ‘on’ and ‘off’ stages would be desirable. One model distinguished between patients aged above and below 60 years by including the long-term adverse events of levodopa while also considering patients in higher HY stages.
Although costs can (and have been, in the models described here) be evaluated from different perspectives (healthcare provider/third-party payer/societal), it has been proposed that, depending on the research question, to improve comparability of results, costs should be evaluated and reported from the societal as well as the chosen perspective. However, only a minor fraction of the modelling studies reported cost-effectiveness results from the societal perspective.[29,32,43]
3.3 Adverse Events and Co-Morbidities
Motor complications such as dyskinesias and motor fluctuations particularly affect HR-QOL and PD-related treatment costs. Most of the studies included motor complications to different degrees. All studies focused primarily on motor symptoms and motor disability. To our knowledge, only one study (not included in this review, because no treatment options were compared) has considered the impact of non-motor symptoms, such as dementia, on cost effectiveness. Behavioural and psychological symptoms, and other co-morbidities such as gastrointestinal symptoms or sleep disorders, which have been shown to considerably affect HR-QOL and costs, were not included in the existing models.
In a model, simulations can be performed at the individual (also called microsimulation) or cohort level, in which individuals or groups of patients, respectively, pass through the model. Cohort models can be designed in a deterministic or probabilistic fashion, the latter applying Monte Carlo simulation. Most of the PD models used deterministic cohort simulations;[24,29,36,43,47,49,50,55,58,62] only one study used a probabilistic approach with second-order Monte Carlo simulation. Although microsimulation approaches are complex and time consuming, it is surprising that no microsimulation studies have been published in the field of PD so far.
It is the very nature of a decision-analytic model to draw conclusions for a decision problem, which arises due to insufficient or inappropriate primary data. Evidence from various sources is synthesized within the model. Therefore, uncertainty is inherent in such models. Main sources of uncertainty arise from assumptions made by the modeller in designing the modelling approach (structural uncertainty) and within the selection of data sources or the data themselves (parameter uncertainty). No study reported having considered structural uncertainty (e.g. excluding states by setting transition probabilities to zero), although one study referred the structure of their model to an expert panel for review. However, Kristiansen et al. reported on sensitivity of alternative cost groups, which could be considered a structural sensitivity analysis. Parameter uncertainty arising from the selection of data sources could be reduced by performing a thorough literature search and including all relevant data of sufficient quality. This was best demonstrated, for example, by two studies[55,62] dealing with SPECT imaging for the diagnostic process in PD. The usual way of handling parameter uncertainty is to use one- and multi-way sensitivity analyses, which were performed regularly, but still to a varying extent.[24,29,36,47,50,55,58,62] In settings with many parameters of relevance, probabilistic sensitivity analysis is the appropriate approach for evaluating uncertainty. Probabilistic sensitivity analysis was applied by only a few authors.[24,32,43,47] However, while we would not advocate the general use of probabilistic sensitivity analysis, there are situations in which these would be considered necessary (e.g. Antonini et al.). However, it should be noted that probabilistic sensitivity analysis should supplement rather than replace one- or two-way deterministic sensitivity analysis.
Because financial dependency may introduce a bias,[76,77] we looked closely at study sponsors. Several studies received no funding from industry (e.g. but did receive federal funding).[49,50,62] Seven studies performed or supported by the pharmaceutical industry are included in this review (table I).[24,32,36,43,47,55,58] No detailed information on funding was provided by Smala et al. (cabergoline vs levodopa), but the study was funded in part by Pharmacia, similar to the study by Lindgren et al. All authors funded by the pharmaceutical industry may have a conflict of interest, as the treatments evaluated were produced by the supporting companies.
3.7 Model Validation
In the model validation process, modellers assess the concordance of the model results with input or calibration data (internal validation), comparing them with independent outcome data (external validation) or with other models (between-model validation). This provides quality assessment and enhances credibility of both the results and the chosen model structure. In most PD modelling studies, no validation was performed,[32,36,43,49,55] despite explicit recommendation since 2003.
First attempts for validation were undertaken by Hjelmgren et al. (external validation with respect to disease progression), Dodel et al., Van Laere et al. (model structure and parameter plausibility [face validity]) and by Antonini et al. (between-model validity). However, detailed results were not reported or discussed in any of the studies.
3.8 Economic Evaluation
Tables I–III show that most models were well developed and reported in detail (table III shows scores from a study quality assessment instrument). Almost all authors clearly described the objectives of their studies and the medical context, addressed the target population and justified the chosen diagnostic or treatment options. For diagnostic evaluations, the time horizon differed between 1 and 5 years, which was considered long enough to capture crucial effects.[55,58,62] For the evaluation of drug treatment, a duration of <5 years was often chosen, which may be too short.[24,32,36,47] Appropriate time horizons were applied by Findley et al., Hjelmgren et al. and Smala et al.
Only two models[49,50] did not specify the chosen model type. Model structure, transition probabilities and assumptions were well described by eight studies.[29,32,36,43,47,55,58,62] No graphical representations of the model structure were found in the articles by Findley et al., Goulionis and Vozikis and Lindgren et al. As evaluated by the checklist of the German Scientific Working Group (see table III), quality and quantity of input data varied considerably. Data for health outcomes used by Findley et al., Goulionis and Vozikis and Dodel et al. could have been presented more clearly. Similarly, details of cost data and/or discounting presented by Van Laere et al. and Goulionis and Vozikis were sketchy. Good descriptions of input data were found in Antonini et al. and Hjelmgren et al.
Three studies[43,47,49] could have illustrated their results in a better fashion. Results were often presented as either individual or incremental costs and utilities,[43,47,49] but not both. Finally, some authors did not disaggregate their results in tables or graphics.[47,49] The papers by Lindgren et al. and Smala et al. are examples of good presentation of results.
Van Laere et al. and Goulionis and Vozikis gave no (detailed) description of sensitivity analysis. Multi-way sensitivity analysis was found in five studies.[32,43,47,50,62] Sensitivity analysis was essentially incomplete in the articles by Haycox et al. and Smala et al. with respect to testing relevant assumptions or potentially influential model parameters. Only two authors[43,62] did discuss the results of the sensitivity analysis (e.g. direction of influence of parameter values and model structure). Furthermore, only a few authors related their results to independent studies[43,62] and presented the main limitations.[29,36,47,55,58] Only Haycox et al. discussed initial population assumptions.
Except for Hjelmgren et al., study conclusions were presented according to standard criteria. Kristiansen et al. could have drawn conclusions from their results (that even in sensitivity analysis DLI might not be cost effective compared with oral levodopa) in a more conservative way.
3.9 Excluded Literature
We excluded five articles from the search results over the period from July 2002 to March 2010 because the authors did not compare treatment options or because the comparison of treatments was only used as an example.[16, 17, 18, 19] In four communications, Nuijten and colleagues[16, 17, 18, 19] used PD as an example and demonstrated primarily methodological aspects of the impact of confounding variables on cost-effectiveness analysis as well as proposing the combination of budget-impact and cost-effectiveness analysis.
Van den Hout et al. presented primarily technical work for combining Markov models with mixed effect models and focused on technical approaches and the validation of model structures. He evaluated in detail the impact of dementia in PD patients. Transition probabilities were dependent on sex, duration and mortality and were built into the Markov model through a statistical mixed-effects model. Furthermore, van den Hout et al. calculated life expectancy using microsimulation.
3.10 Recommendations for Future Modelling
Today, three main types of model approaches are available: decision trees, Markov models and microsimulation/DES. Decision trees consist of different paths that map the course of the disease. They create a difficult model structure for complex diseases, such as PD, considering patients along a long time horizon. However, for modelling diagnostic testing, one could map the problem using a simple approach. Conversely, to precisely image the course of a certain disease for a number of years or even a lifetime, decision trees can be tedious and should be replaced by Markov models. The methodological framework of Markov models guarantees that patients cannot be in two states at the same time. This implies that the course of disease must be represented by discrete, mutually exclusive and complete states. Therefore, continuous scales (e.g. UPDRS) are less appropriate (or even inappropriate) than such discrete scales as the HY scale. To overcome constraints of Markov models, which may limit the modelling architecture, especially when continuous scales have to be implemented, micromodelling approaches or DES are an alternative methodological framework. Patients and the disease itself are, in the case of DES, modelled by so-called entities, which could experience events (e.g. a change in health, deriving a special medication, needing care or even visiting the doctor) and carry properties (‘attributes’, e.g. UPDRS scores). However, the flexibility is hampered by the need for a lot of input data that are often not generated directly in clinical trials. If no data are available, estimation of such data (e.g. using expert opinion or extrapolation from other data) may be a major source of bias. Therefore, as in the evaluation of modelling in PD, models presented in the health economic literature often use Markov models, as this represents a systematic and (theoretically) very transparent extrapolation method that reveals potential bias.
It has been recommended that the model should closely represent the natural course of the disease under investigation. The use of very few health states (e.g. two that only represent the occurrence of ≥25% or <25% ‘off’ times per day) seem a crude and probably inappropriate approach in the context of modelling complex diseases such as PD. Therefore, basing the architecture of the model on scales that closely mirror the natural progression of the disease would be ideal. The HY scale has been used; however, more elaborate scales should also be investigated for modelling PD progression. The widely used and accepted UPDRS scale may be problematic, as no definite states are classified, which is a requisite for Markov models. To approach reality more closely, future modelling should also consider complications and co-morbidities beyond motor complications (e.g. dyskinesias), such as behavioural and psychological symptoms (e.g. dementia, depression, hallucinations), gastrointestinal symptoms or sleep disorders. However, care should be taken to avoid double counting the effects of these conditions on (clinical or QOL) effectiveness outcomes and costs.
We reviewed cost-effectiveness models from 2003 onwards. In line with earlier models summarized by Siebert et al., the majority used Markov models to evaluate drug therapy. Most of the models used HY stages as Markov states to represent the course of the disease; however, most of them used HY ‘on’ stages that do not mirror exactly the underlying disease progression, because disease progression is mixed up with treatment effects in HY ‘on’ states (i.e. HY state under treatment). Recommendations for good modelling practice and cost effectiveness were published in 1996 and 2003, but have, on the whole, not been followed, as evidenced by the lack of use of the societal perspective (as an additional perspective for a reference case) for costs or, at best, sparse attempts at model validation. Apart from motor complications such as fluctuations or dyskinesias, adverse (drug) events, co-morbidities or complications were almost never included in the models, which is not in line with our current understanding of the disease. The use of probabilistic approaches is growing, but probably still under-represented. Individual simulation methods (micromodelling, DES) are recent developments to address more specific research questions not assessed in PD so far (such as the relevance of specific patient characteristics or for sufficient treatment possibilities, e.g, in DBS). We predict that patient-level simulation techniques will be used more frequently in the future, especially in the evaluation of chronic neurodegenerative disorders.
Cabergoline, along with other ergoline dopamine agonists, was removed for the treatment of PD from the market in 2007 due to cardiac adverse effects.
The disease is considered to be in an ‘on’ stage when the symptoms respond well to treatment. The disease is considered to be in an ‘off’ stage when it is either not or not adequately responding to treatment and the patient is experiencing significant symptoms.
The authors have no conflicts of interest that are directly relevant to the content of this study. Full financial disclosure for Professor Dr R. Dodel is available as SDC 2, http://links.adisonline.com/PCZ/A122.
- 7.Oertel WH, Quinn NP. Parkinsonism. In: Brandt T, Diener HC, Caplan LR, et al., editors. Neurological disorders: course and treatment. San Diego (CA): Academic Press, 2003Google Scholar
- 21.Drummond MF, O’Brien B, Stoddard GL, et al. Methods for the economic evaluation of health care programs. New York: Oxford University Press, 1997Google Scholar
- 22.Gold MR, Siegel JE, Russell LB, et al. Cost-effectiveness in health and medicine. New York: Oxford University Press, 1996Google Scholar
- 23.Siebert U, Behrend C, Mühlberger N, et al. Entwicklung eines Kriterienkataloges zur Beschreibung und Bewertung ökonomischer Evaluationsstudien in Deutschland. In: Leidel R, Graf von der Schulenburg JM, Wasem J, editors. Approaches and methods of economic evaluation in health care: an international perspective [in German]. Baden-Baden: Nomos Verlagsgesellschaft, 1999: 156–70Google Scholar
- 25.Rinne UK, Bracco F, Chouza C, et al. Early treatment of Parkinson’s disease with cabergoline delays the onset of motor complications: results of a double-blind levodopa controlled trial. The PKDS009 Study Group. Drugs 1998; 55 Suppl.1: 23–30Google Scholar
- 33.Rascol O, Brooks DJ, Melamed E, et al. Rasagiline as an adjunct to levodopa in patients with Parkinson’s disease and motor fluctuations (LARGO, Lasting effect in Adjunct therapy with Rasagiline Given Once daily, study): a randomised, double-blind, parallel-group trial. Lancet 2005; 365 (9463): 947–54PubMedCrossRefGoogle Scholar
- 41.Rascol O, Brooks DJ, Korczyn AD, et al. A five-year study of the incidence of dyskinesia in patients with early Parkinson’s disease who were treated with ropinirole or levodopa. 056 Study Group. N Engl J Med 2000; 342 (20): 1484–91Google Scholar
- 44.Schrag A, Jahanshahi M, Quinn N. How does Parkinson’s disease affect quality of life? A comparison with quality of life in the general population. Mov Disord 2000; 15 (6): 1112–8Google Scholar
- 45.Rinne UK, Larsen JP, Siden A, et al. Entacapone enhances the response to levodopa in parkinsonian patients with motor fluctuations. NOMECOMT Study Group. Neurology 1998; 51 (5): 1309–14Google Scholar
- 60.Jankovic J. Essential tremor: clinical characteristics. Neurology 2000; 54 (11 Suppl.4): S21–5Google Scholar
- 63.Hughes AJ, Ben-Shlomo Y, Daniel SE, et al. What features improve the accuracy of clinical diagnosis in Parkinson’s disease: a clinicopathologic study. 1992. Neurology 2001; 57 (10 Suppl.3): S34–8Google Scholar
- 70.Parkinson Study Group. Pramipexole vs levodopa as initial treatment for Parkinson disease: a randomized controlled trial. Parkinson Study Group. JAMA 2000; 284 (15): 1931–8Google Scholar
- 72.Brooks DJ, Sagar H, UK-Irish Entacapone Study Group. Entacapone is beneficial in both fluctuating and nonfluctuating patients with Parkinson’s disease: a randomised, placebo controlled, double blind, six month study. J Neurol Neurosurg Psychiatry 2003; 74 (8): 1071–9Google Scholar
- 74.Ministry of Health. Prestazioni di assistenza specialistica ambulatoriale erogabili nell’ambito del SSN e relative tariffe GU n 216 of 14/11/1996: DM 22/07/1996Google Scholar
- 75.Winter Y, von Campenhausen S, Arend M, et al. Healthrelated quality of life and its determinants in Parkinson’s disease: results of an Italian cohort study. Parkinsonism Relat Disord; 17 (4): 265–9Google Scholar
- 78.Briggs A, Claxton K, Sculpher M. Decision modelling for health economic evaluation. Oxford: Oxford University Press, 2006Google Scholar
- 79.Institute for Quality and Efficiency in Health Care. Working Paper on Modelling in health economic evaluations. Cologne: IQWiG, 2009 [online]. Available from URL: http://www.iqwig.de/download/Working_Paper_Modelling.pdf [Accessed 2010 Jun 21]Google Scholar