1.2.4. Model inputs
1.2.4.1. Summary table of model inputs
Model inputs were based on clinical evidence identified in the systematic review undertaken for the guideline, supplemented by additional data sources as required. Model inputs were validated by the GDG. A summary of the model inputs used in the base-case analysis is provided in , , and below. More details about sources, calculations and rationale for selection can be found in the sections following this summary table.
Clinical inputs- probabilities of withdrawal and remission.
Utility weights in the model.
1.2.4.2. Treatment effects (remission and withdrawal)
The results of conventional meta-analyses of direct evidence alone make it difficult to determine which intervention is the most effective treatment. The challenge of interpretation has arisen for two reasons:
Some pairs of alternative strategies have not been directly compared in a randomised controlled trial
There are frequently multiple overlapping comparisons that could potentially give inconsistent estimates of effect.
This is particularly problematic for probabilistic analysis. To overcome these problems, a Bayesian network meta-analysis (NMA) was conducted in WinBUGS, using code and assistance provided by the NICE technical support unit.
Conventional meta-analysis assumes that for a fixed-effect analysis, the relative effect of one treatment compared to another is the same across an entire set of trials. In a random-effects model, it is assumed that the relative effects are different in each trial but that they are from a single common distribution and that this distribution is common across all sets of trials.
Network meta-analysis requires an additional assumption over conventional meta-analysis. The additional assumption is that intervention A has the same relative effect across all trials of intervention A compared to intervention B as it does across trials of intervention A versus intervention C, and so on. Thus, in a random-effects network meta-analysis, the assumption is that intervention A has the same effect distribution across all trials of A versus B, A versus C and so on.
The aim of the NMA was to calculate treatment-specific probabilities for withdrawal and remission conditional on non withdrawal. It is assumed that people who withdraw cannot go into remission, and similarly people counted as ‘a remission’ have not withdrawn due to adverse events, in other words, the two events are mutually exclusive. Treatment effects for the model had to be accounted for such that the number of withdrawals and remissions could not exceed the number of people in the trial. This negative correlation in outcomes is taken account of by carrying out a conditional logistic regression NMA.
For the network meta-analysis, treatment effects were calculated so as to reflect the clinical review as closely as possible in terms of pooling of studies. In order to obtain treatment effects for remission conditional on non withdrawal, the number of withdrawals was removed from the denominator when entering data for remission into WinBUGS. Baseline log odds of withdrawal and remission conditional on non-withdrawal were calculated using a logistic regression conducted on the placebo arms in the trials and then adjusted by the treatment specific log odds ratios calculated by the NMA using the following logic:
Let BOw, BOr∣wc, θw , θr∣wc and ORw, ORr∣wc denote the baseline odds (from the placebo arms),treatment-specific odds and treatment-specific log odds ratio for withdrawal and remission given no withdrawal respectively. Then:
And:
This approach has the advantage that baseline and relative effects are both modelled on the same log odds scale. It also ensures that the uncertainty in the estimation of both baseline and relative effects is accounted for in the model.
To reflect the populations explored in the clinical review, two separate analyses were conducted; one for people on first-line monotherapy treatment, and one for people on second-line treatment in combination with a glucocorticosteroid, having failed first-line glucocorticosteroid monotherapy.
1.2.4.3. First-line NMA
The schematic of trials compared in the first line NMA is shown in .
Network of trials compared in the first-line NMA.
In the trial network in , glucocorticosteroid treatment featured in the most trials (10) followed by budesonide (9), mesalazine (8), placebo (7), sulfasalazine (2) and azathioprine monotherapy (1). For ease of interpretation, placebo was chosen as the baseline treatment for comparisons in the NMA. The data used in the first-line induction NMA can be found in . Please note that although azathioprine monotherapy was included in the NMA for completeness, it was not compared in the cost-effectiveness analysis. First-line azathioprine monotherapy was not included because of the lack of evidence and because it has a slower response than the other treatments considered. (3 – 4 months i.e. longer than the 2 month time horizon considered)
Data for the first-line NMA.
A logistic regression was run on the placebo arms of the trials shown in ; this produced a baseline odds, BO which was used in the NMA to derive treatment-specific probabilities as described above. The baseline odds for withdrawal and remission had lognormal distributions parameterised as:
The model was run for 50,000 iterations with a burn in period of 50,000. Vague uninformative priors were combined with data-driven likelihood functions to produce posterior probability estimates. The final treatment-specific probability estimates and their associated confidence intervals can be seen in .
Probabilities from 1st line NMA.
It can be seen from , that among first-line treatments, sulfasalazine was associated with the highest probability of withdrawal- 35%- but with the 95% confidence interval ranging from 5% to 80%. Glucocorticosteroid treatment was associated with the highest probability of remission conditional on non-withdrawal- 66% with 95% confidence interval ranging from 53% to 79%. These estimates were used to parameterise treatment effects in the model; it should be noted that there is a large amount of imprecision in these estimates, particularly for withdrawal. This is often the case when calculating treatment effects for withdrawal due to small event numbers.
Model fit was assessed by calculating the total residual deviance and comparing with the number of unconstrained data points. In the withdrawals NMA, the total residual deviance was 38.6 which, when compared to 43 unconstrained data points, shows that the model fitted the data reasonably well. In the remission conditional on no-withdrawal NMA, the total residual deviance was 35.94 which again, when compared to 31 unconstrained data points, showed that the model fitted the data reasonably well. DIC statistics of 154.9 and 200.1 were calculated for the withdrawal and remission conditional on no-withdrawal NMAs respectively.
Posterior estimates of heterogeneity- between trial variance- were calculated and values of 0.29 and 0.22 were found for the withdrawal and remission trials. This shows there was a large amount of variation in treatment effects calculated from different trials.
Inconsistency in the network was assessed by fitting an ‘inconsistency model’15
1.2.4.4. Second-line NMA
The schematic of trials compared in the second line NMA is shown in .
Network of trials compared in the second-line NMA.
In the trial network in , glucocorticosteroid monotherapy and azathioprine + a glucocorticosteroid featured in the most trials (eight); methotrexate + a glucocorticosteroid featured in five trials. For ease of interpretation, glucocorticosteroid treatment- which can be thought of here as the placebo comparison- was chosen as the baseline treatment for the NMA. The data used in the second-line induction NMA can be found in .
Data for the first-line NMA.
It should be noted that since Present was a crossover trial, the denominators reported for safety and efficacy are different; this is reflected in the clinical review. To account for this, the numerator for remission given no withdrawal was not adjusted by removing withdrawals, since it is not possible to determine exactly which of the people eligible to be assessed for the remission outcome withdrew. This is a conservative approach since it means the number of people being assessed for the remission outcome is bigger and thus the overall probability of achieving remission becomes smaller. Denominators of 36 were used for both arms of the Present study for the remission outcome.
A logistic regression was run on the glucocorticosteroid monotherapy arm of the trials shown in ; this produced a baseline odds, BO which was used in the NMA to derive treatment-specific probabilities as described above. The baseline odds for withdrawal and remission had lognormal distributions parameterised as:
The model was run for 50,000 iterations with a burn in period of 50,000. Vague uninformative priors were combined with data-driven likelihood functions to produce posterior probability estimates. The final treatment-specific probability estimates and their associated confidence intervals can be seen in .
Probabilities from first-line NMA.
It can be seen from that out of the two second-line treatments, methotrexate + a glucocorticosteroid was associated with the highest probability of withdrawal- 11%- with 95% confidence interval ranging from 0% to 70%. Azathioprine + a glucocorticosteroid was associated with a higher probability of remission conditional on no withdrawal- 66% with 95% confidence interval ranging from 33% to 92%. These estimates were used to parameterise treatment effects in the model; it should be noted that there is a large amount of imprecision in these estimates, particularly for withdrawal. This is often the case when calculating treatment effects for withdrawal due to small event numbers, but the remission conditional on no withdrawal outcome is also associated with large imprecision.
Model fit was assessed by calculating the total residual deviance and comparing this with the number of unconstrained data points. In the withdrawals NMA, the total residual deviance was 20.94 which, when compared to 21 unconstrained data points, shows that the model fitted the data well. In the remission conditional on no withdrawal NMA, the total residual deviance was 23.71 which, when compared to 21 unconstrained data points, shows that the model fitted the data reasonably well. DIC statistics of 70.4 and 105.5 were calculated for the withdrawal and remission conditional on no-withdrawal NMAs respectively.
Posterior estimates of heterogeneity- between trial variance- were calculated and values of 0.67 and 1.22 were found for the withdrawal and remission trials. This shows there was a very large amount of variation in treatment effects calculated from different trials.
Posterior estimates of heterogeneity were calculated for the ‘inconsistency model’ and values of 0.61 and 1.53 were found for the withdrawal and remission trials. This shows that the incorporation of the consistency equations did not force the between trial variance to increase in the remission conditional on no-withdrawal and therefore it is unlikely that there is inconsistency within the network. There was, however, an increase in heterogeneity in the withdrawal NMA (0.61 vs 0.67), which prompted further investigation of inconsistency within the network. Since there were only three treatments in the network the Bucher method7 can be used to detect whether the amount of inconsistency in the network is statistically significant or a chance finding. A p-value of 0.46 was calculated to test the null hypothesis that there was no inconsistency in the network; this shows that there is insufficient evidence to reject the null and therefore it cannot be said that there is inconsistency.
shows how the model probabilities interact for one treatment cycle in the patient pathway. shows that from the number of people entering the model, n, the number of people entering remission nrem after one treatment cycle and the number of people failing to enter remission nno rem for a given drug can be calculated using the equations:
Patient pathway and probabilities (one treatment cycle).
1.2.4.5. Utilities
For economic evaluation, a specific measure of Health Related Quality of Life (HRQoL) known as utility is required to calculate QALYs. Utilities indicate the preference for health states on a scale from 0 (death) to 1 (perfect health). The NICE reference case36 specifies that the preferred way for this to be assessed is by the EQ-5D instrument.
A systematic search identified four studies9,21,23,54 with appropriate data to use as utility weights for the model. A description of these studies is shown in
Selecting utility weights for the model.
Utility weights derived by Stark et al54 were used due to the comparative lack of limitations and the directness of the population. In particular, the Stark data was favoured due to:
1.2.4.6. Resource use and cost
The GDG thought it likely that, as well as the treatment-specific tests and consultations, people with an exacerbation of Crohn's disease would have regular consultations and tests regardless of the drug. These consultations and tests are summarised in and the weighted average cost calculated is shown in .
Consultations and tests for people not in remission (regardless of induction treatment).
Weighted average costs of drug preparations were used in the model; weights were calculated using prescription-cost-analysis data39 in order to calculate drug costs which reflect how they are prescribed in clinical practice.
Based on the costs, weights and dosing schedule described in , a cost of £38.10 was calculated for an eight week course of prednisolone drug treatment in the base case. The total cost including tests and consultations was £362.17 ().
Prednisolone costs in the model including tests.
Based on the costs, weights and dosing schedule described in , a cost of £167.94 was calculated for an eight week course of budesonide drug treatment in the base case. The total cost including tests and consultations was £492.01 ().
Budesonide costs in the model including tests.
Based on the costs, weights and dosing schedule described in , a cost of £98.01 was calculated for an eight week course of mesalazine in the base case. The total cost including tests and consultations was £416.32 ().
Mesalazine costs in the model including tests.
Based on the costs, weights and dosing schedule described in , a cost of £44.16 was calculated for an eight-week course of sulfasalazine in the base case. The total cost including tests and consultations was £362.48 ().
Sulfasalazine costs in the model including tests.
Based on the cost and dosing schedule described in , a cost of £6.54 was calculated for an eight week course of methotrexate + a glucocorticosteroid drug treatment in the base case. The total cost including tests and consultations was £355.67 ().
Methotrexate costs in the model including tests.
Based on the cost and dosing schedule described in , a cost of £42.11 was calculated for an eight-week course of azathioprine + a glucocorticosteroid treatment in the base case. The total cost including tests and consultations was £429.34 ().
Azathioprine costs in the model including tests.
Based on the cost and dosing schedule described in , a cost of £1,056.36 was calculated for biologic drug treatment in the base case. The total cost including tests and consultations was £1,426.09 ().
Biologic costs in the model including tests.
1.2.4.7. Cost of surgery
The cost of surgery was calculated in close collaboration with the surgeon on the GDG as described below and in .
The five most common operations in Crohn's disease- pan proctocolectomy, colectomy, right hemicolectomy, small intestine resection and strictureplasty- were chosen and matched to their closest fitting OPCS and HRG codes.
Weights were calculated using HES data
41, selected OPCS codes and assuming that 10% of all operations would be associated with a ‘major complication or comorbidity’
An average cost per operation was calculated by multiplying these weights by the costs attached to selected HRG codes, and adding in a pre-operative and post-operative consultation for each operation.
So, assuming pre-operative and post-operative costs for a consultation of £94 and £57 (), and a re-operation rate of 10%, the overall cost of surgery assumed in the model was assumed to be £5,351.24.
1.2.5. Computations
The mean cost and effectiveness of the competing strategies were calculated using Microsoft Office Excel 2007.
1.2.5.1. Calculating QALYs
To calculate QALYs for a given treatment sequence, both the probability of inducing remission for each individual treatment, and the time spent in remission over the course of the model for a given treatment strategy are considered. To do this, the treatment strategy is partitioned into individual treatments and the number of weeks of remission and active disease that occur as a direct result of each treatment are calculated. These are then aggregated over the duration of the strategy and QALYs for a given strategy are calculated by multiplying the number of weeks of remission and active disease by the appropriate utility weights.
is a visual representation of how weeks of remission and active disease are calculated in the model. Note that, in the absence of data, it was assumed that for all people in whom remission is successfully induced, remission occured half-way through treatment, and people do not relapse.
Calculating weeks of remission and active disease for a given treatment strategy.
It can be seen from the diagram that QALYs are calculated as follows:
Let p1, p2, p3 and p4 be the probabilities of successfully inducing remission with treatments 1, 2, 3 and 4 respectively.
Let t1, t2, t3 and t4 be the durations of treatment in weeks associated with treatments 1, 2, 3 and 4 respectively.
Let WR1, WR2, WR3, and WR4 denote the expected number of weeks of remission associated with treatments 1, 2, 3 and 4 respectively.
Let UR and UA be the utility weights associated with remission and active disease respectively.
Let H denote the time horizon in the model, which will be equal to the length of the longest treatment strategy and is 30 weeks in the base case.
Treatment one:
Treatment two:
Treatment three:
Treatment four:
Note that the term p4 × (H − t1 − t2 − t3 − t4) is added to the equation for cases when the overall length of the treatment strategy is less than the time horizon. In the event that the duration of the treatment strategy is equal to the time horizon this term is equal to zero, since H = t1 + t2 + t3 + t4.
Then the total number of weeks of remission, WR and number of weeks of active disease WA are given by:
And the total treatment specific QALYs, Q, are calculated as:
1.2.5.2. Probabilistic analysis in the model
In the probabilistic analysis, distributions were assigned to treatment effects, utilities and, where possible, costs in order to account for the uncertainty in model inputs and capture the effect of this uncertainty on model outputs.
Treatment effects:
To capture the uncertainty in treatment effects, a sample of 1000 random sets of treatment effects was taken from the NMA using the CODA function in WinBUGS. This has the advantage of preserving the correlation between variables, which would not be accounted for if they were sampled from their individual distributions. For the probabilistic cost-effectiveness analysis, for each simulation a random set of treatment effects was chosen from the sample using random number generation.
Reference costs:
To assign a distribution to reference costs, it was assumed that they followed a lognormal distribution and used the inter-quartile range to calculate an approximate standard error on the log scale.
Let X be the cost for which a distribution is required, i.e. ln (X) ∼ Normal (μ, σ2)
Let M be the mean associated with the cost.
Let IQR be the inter-quartile range associated with the cost.
Note that for normally distributed data:
and the standard error, s, is related to the standard deviation by:
Then the standard error on the log scale can be calculated as:
And random draws can be taken from the distribution:
Utilities:
Utilities were sampled probabilistically by assigning lognormal distributions to utility decrements as described in (ref Briggs). Normal distribution parameters were converted to lognormal parameters by method of moments, as defined below:
Let E[X] and Var[X] be the mean and variance respectively, of the utility decrement U
Then the parameters of the lognormal distribution, μ and σ2 are found by:
1.2.5.3. Calculating cost effectiveness
It is possible, for a particular cost-effectiveness threshold, to express cost-effectiveness results in term of net benefit (NB). This is calculated by multiplying the total QALYs for a comparator by the threshold cost per QALY value (for example, £20,000) and then subtracting the total costs. The decision rule then applied is that the comparator with the highest NB is the most cost-effective option at the specified threshold. That is the option that provides the highest number of QALYs at an acceptable cost. For ease of computation NB is used to identify the optimal strategy in the probabilistic-analysis simulations.
Let Ct and Qt denote the mean costs and mean QALYs respectively, associated with a given treatment. Then mean net benefit NBt is calculated as:
where £20,000 per QALY represents the cost-effectiveness threshold in the NICE reference case.
The net benefit for each of the 1000 simulations in the probabilistic analysis is calculated. This allows the probability that a given treatment would be optimal, based on the number of times it has the highest net benefit, can be estimated.
However, the strategy that is optimal overall is the one that has the highest net benefit calculated using the mean costs and QALYs, where means were the average of the 1,000 simulated estimates.