HAQ-DI

HAQ-DI in Rheumatoid Arthritis Models and Cost-Utility Analysis

In economic models of rheumatoid arthritis the Health Assessment Questionnaire Disability Index usually carries the quality of life side of the model. Treatment response changes the HAQ-DI score, the score drifts upwards over the years as disability accumulates, and a mapping function turns each HAQ-DI value into an EQ-5D utility, so the shape assumed for that drift and the choice of mapping decide much of the QALY gain of a biological medicine. In NICE's multiple technology appraisal of biological disease-modifying antirheumatic drugs (TA375), the Assessment Group converted HAQ scores, together with a pain score simulated from HAQ, into EQ-5D values with a mixture model estimated on data from 16,011 patients. The name HAQ can refer to the whole Stanford Health Assessment Questionnaire, which also contains a pain scale and other dimensions, while the HAQ-DI is its disability index alone.

The Stanford outcome framework behind the questionnaire

Fries, Spitz, Kraines and Holman set out the framework in Arthritis and Rheumatism in 1980. They proposed that patient outcome in arthritis be represented in five separate dimensions, namely death, discomfort, disability, drug (therapeutic) toxicity and dollar cost, and they validated a measurement method in rheumatoid arthritis. They preferred a questionnaire technique because it was inexpensive and did not require interobserver validation.

The disability index is the part of that framework that health economists use most. Dollar cost and drug toxicity have their own place in a model as resource use and adverse events, while disability is the dimension that links a treatment's clinical effect to quality of life. Bruce and Fries reviewed the HAQ's development, content and dissemination in 2003, with reference sources for its uses, translations and validations.

How twenty questions become a score of 0 to 3

Bruce and Fries describe the disability index as 20 questions in eight categories of functioning, which together represent a comprehensive set of functional activities. The questionnaire assesses a patient's usual abilities, using their usual equipment, during the past week. For each question there is a four-level response set scored 0 (without any difficulty) to 3 (unable to do).

The scoring works in two steps. Bruce and Fries state that the highest component score in each category determines the score for that category, unless aids or devices are required. Dependence on equipment or physical assistance increases a lower score to the level of 2, which they describe as a more accurate representation of the underlying disability. The category scores are then averaged to give the index.

$$\text{HAQ-DI} = \frac{1}{K} \sum_{k=1}^{K} c_k, \qquad c_k = \max\left(m_k, a_k\right)$$

where $c_k$ is the score for category $k$, $m_k$ is the highest item score in that category (0, 1, 2 or 3), $a_k$ equals 2 when the patient reports using aids, devices or help from another person for that category and 0 otherwise, and $K$ is the number of categories answered. The index lies between 0 and 3, and a higher value means greater disability.

A second scoring rule matters for trial data. Bruce and Fries also describe a complementary scoring method that ignores the scores for aids and devices when computing the category scores, which records residual disability once equipment is in use. Studies sometimes report one method and sometimes the other, and the two give different numbers for the same patient.

HAQ-DI as the engine of rheumatoid arthritis models

Rheumatoid arthritis is a chronic condition, while trials of biological therapies last months. A cost-utility analysis therefore has to extrapolate, and HAQ-DI is the quantity most often carried forward in time. A model typically applies an initial HAQ improvement that depends on response to treatment, keeps that improvement while the patient stays on the drug, returns the score towards its earlier level when treatment stops, and then lets the score worsen gradually on later lines of treatment.

The form of that gradual worsening is a modelling choice with real consequences. In TA375, published on 26 January 2016, the committee accepted the use of data from the Early Rheumatoid Arthritis Study (ERAS) inception cohort to estimate underlying disease progression on conventional DMARDs, in preference to assuming a linear HAQ progression as earlier NICE appraisals had done. The guidance reports that Norton and colleagues used ERAS data to identify four classes of HAQ trajectory over the course of the disease. A progression rate that is too fast inflates the benefit of any treatment that halts it, so the progression assumption should be tested in sensitivity analysis alongside the response rates.

The HAQ-DI also links the model to the biological therapy sequence. Each line of treatment in a sequence model has its own response probability and HAQ improvement, so the HAQ-DI path for a simulated patient records the whole treatment history. Because the HAQ-DI accumulates over a lifetime horizon, small differences in its assumed path add up to large differences in QALYs.

Converting HAQ-DI scores into EQ-5D utilities

The HAQ-DI has no preference weights, so a score of 1.5 says nothing directly about how a year in that state is valued. Models use a mapping function to predict a health state utility value on the EQ-5D scale, and the design choices behind such functions are discussed under mapping study.

Hernández Alava, Wailoo, Wolfe and Michaud reported in 2013 an analysis of about 100,000 repeated observations from a United States observational cohort of people with rheumatoid arthritis in which HAQ, a pain visual analogue scale and the EQ-5D had all been recorded. They compared a bespoke mixture model, built to reflect the bounded and multimodal distribution of EQ-5D values, with standard multiple regression. The mixture model fitted the data very well and avoided bias and out-of-range predictions, and it identified four classes of respondents differing in disease severity. The authors concluded that linear models underestimate the QALY benefits, and therefore the cost-effectiveness, of therapies.

NICE's TA375 shows the method in use. The Assessment Group judged EQ-5D estimates more accurate when pain was included alongside HAQ, simulated a pain score for each patient from HAQ, and then estimated EQ-5D with the mixture-model approach of Hernández Alava and colleagues, based on data from 16,011 patients in the National Data Bank for Rheumatic Diseases. The group noted that this assumes all treatments affect pain in proportion to their effect on HAQ, an assumption implicit in every model that leaves pain out. In the later appraisal of upadacitinib for moderate rheumatoid arthritis (TA744), the company mapped HAQ to pain with data from its own trials, while the Evidence Review Group, describing the TA375 approach as a mapping based on National Data Bank data, preferred that approach. TA375 itself describes the pain step as a simulation from the mean and variance of pain for each HAQ score in ERAS data.

Illustrative worked example of scoring and a QALY gain

All numbers in this example are illustrative, and the utility function is a simple straight line chosen for arithmetic, not a published mapping. A patient answers all eight categories. The highest item scores in the eight categories are 1, 1, 0, 2, 1, 1, 2 and 1, and the patient reports using aids or devices in the third and fifth categories.

1. Category scores with the aids adjustment. The third category rises to 2 because aids are used, and the fifth category rises to 2 for the same reason. The adjusted category scores are 1, 1, 2, 2, 2, 1, 2 and 1, which sum to 12.

2. The index. With all eight categories answered, the HAQ-DI is 12 / 8 = 1.5. Under the complementary method that ignores aids, the category scores sum to 1 + 1 + 0 + 2 + 1 + 1 + 2 + 1 = 9 and the index is 9 / 8 = 1.125, so the choice of scoring method moves the result by 0.375.

3. Utilities before and after treatment. Suppose treatment lowers the HAQ-DI to 1.0 and the analyst's illustrative mapping is

$$u = 0.90-0.30 \times \text{HAQ}$$

where $u$ is the predicted EQ-5D utility and HAQ is the HAQ-DI score. Before treatment the utility is 0.90 − 0.30 × 1.5 = 0.45, and after treatment it is 0.90 − 0.30 × 1.0 = 0.60, a gain of 0.15.

4. QALYs and cost per QALY. If the improvement lasts two years with no discounting, the gain is 0.15 × 2 = 0.30 QALYs. An incremental cost of GBP 9,000 gives

$$\frac{\Delta C}{\Delta Q} = \frac{9000}{0.30} = 30000$$

where $\Delta C$ is the incremental cost in GBP and $\Delta Q$ is the QALY gain, so the result is GBP 30,000 per QALY gained.

The example shows two places where the HAQ-DI changes the ICER before any clinical evidence is weighed. The scoring method changes the starting score, and the mapping decides how many QALYs a given HAQ change is worth. Hernández Alava and colleagues found that linear models underestimate QALY benefits, so a linear regression fitted to the same data as a mixture model would tend to make the treatment look less cost-effective than the mixture model does.

What counts as a meaningful change in HAQ-DI

The minimal clinically important difference for the HAQ-DI depends on the setting in which it is estimated. Pope, Khanna, Norrie and Ouimet followed 225 patients with rheumatoid arthritis attending a clinical practice at two visits within a year and used the patient's own rating of change as the anchor, an anchor-based MID design. The mean HAQ-DI change was −0.09 in patients who rated themselves somewhat improved and 0.15 in those who rated themselves somewhat worse. They concluded in 2009 that the minimally important difference for the HAQ-DI in clinical practice is smaller than in randomised controlled trials.

For an economic model the point is practical. A response definition built on a trial-based threshold will class fewer patients as improved than one built on a clinic-based threshold, and the share of responders enters directly into the costs and QALYs of each treatment line.

How the HAQ-DI differs from nearby measures

The HAQ-DI is a patient-reported outcome measure of physical function, first validated in rheumatoid arthritis. Its scores are not preference-based, which separates it from the EQ-5D, a generic instrument whose value sets give utilities directly. Functional status is the wider construct of a person's ability to carry out physical, cognitive and social activities, of which the HAQ-DI captures the physical part.

The WOMAC is the nearest comparison in osteoarthritis. The WOMAC is specific to the hip and knee and has separate pain, stiffness and function subscales, while the HAQ-DI covers whole-body function across many joints in one index and leaves pain to a separate scale. Both instruments reach QALYs only through mapping, and Hernández Alava and Wailoo are among the authors of published mixture-model mappings for each. Clinical measures of disease activity, such as joint counts, record inflammation as a clinician sees it, whereas the HAQ-DI records what the patient can do.

Limitations and common misreadings

Several errors recur when the HAQ-DI is used in economic work. Most of them come from treating a functional index as if it were already a utility or a fixed quantity.

• Reading the index as a utility. A HAQ-DI of 0 means no reported difficulty, not full health on a utility scale, and the index runs in the opposite direction to a utility.
• Mixing scoring methods. The standard method raises a category score to 2 when aids or help are used, and the complementary method does not. Pooling studies that used different methods mixes two different numbers.
• Assuming linear progression. For conventional DMARDs, the TA375 committee preferred HAQ trajectories estimated with ERAS data to an assumed linear progression. A model that keeps a straight line should justify it.
• Mapping without pain. The TA375 Assessment Group judged EQ-5D estimates more accurate when pain was included alongside HAQ. A model that predicts pain from HAQ assumes every treatment changes pain in proportion to its effect on function.
• One threshold for every setting. Pope and colleagues found the minimally important difference smaller in clinical practice than in trials, so a threshold taken from one setting may not suit a model of the other.

Sources

• Bruce B, Fries JF. The Stanford Health Assessment Questionnaire: dimensions and practical applications. Health and Quality of Life Outcomes. 2003;1:20. doi:10.1186/1477-7525-1-20
• Fries JF, Spitz P, Kraines RG, Holman HR. Measurement of patient outcome in arthritis. Arthritis and Rheumatism. 1980;23(2):137-145. doi:10.1002/art.1780230202
• Hernández Alava M, Wailoo A, Wolfe F, Michaud K. The relationship between EQ-5D, HAQ and pain in patients with rheumatoid arthritis. Rheumatology. 2013;52(5):944-950. doi:10.1093/rheumatology/kes400
• National Institute for Health and Care Excellence. Adalimumab, etanercept, infliximab, certolizumab pegol, golimumab, tocilizumab and abatacept for rheumatoid arthritis not previously treated with DMARDs or after conventional DMARDs only have failed (TA375), section 4, Evidence and interpretation. Published 26 January 2016. https://www.nice.org.uk/guidance/ta375/chapter/4-Evidence-and-interpretation
• National Institute for Health and Care Excellence. Upadacitinib for treating moderate rheumatoid arthritis (TA744). https://www.nice.org.uk/guidance/ta744/resources/upadacitinib-for-treating-moderate-rheumatoid-arthritis-pdf-82611314971333
• Pope JE, Khanna D, Norrie D, Ouimet JM. The minimally important difference for the health assessment questionnaire in rheumatoid arthritis clinical practice is smaller than in randomized controlled trials. Journal of Rheumatology. 2009;36(2):254-259. doi:10.3899/jrheum.080479