Adverse perception of cough in patients with severe asthma: a discrete choice experiment

Background Asthma symptoms adversely impact quality of life in particular in those with poor disease control. Commonly used patient-reported measures for asthma used to assess asthma control often inadequately capture the impact of cough, despite evidence that cough is one of the most bothersome symptoms for patients with asthma. This study aims to improve our understanding of how patients with asthma perceive cough to better understand its clinical impact. Methods A discrete choice experiment (DCE) was performed in two distinct adult asthma populations; those with severe asthma as defined by Global Initiative for Asthma (GINA) step 4/5 classification and those with moderate asthma (a GINA steps 2 or 3 classification of asthma severity). Results Choices were highly dominated by the cough attribute in the symptoms complexes; 48.4% of patients with severe asthma and 31.3% with moderate asthma consistently chose the alternative with the lowest level of cough. Furthermore, cough predominance was found to be significantly associated with severity of asthma (p=0.047). Patients with moderate asthma were not willing to accept any additional symptoms to reduce cough from severe to mild. However, these patients were willing to accept mild breathlessness, mild sleep disturbance, severe chest tightness and severe wheezing to remove coughing altogether. Conclusions Patients with asthma prefer to have less cough and are willing to accept greater levels of other symptoms to achieve this. Additionally, asthma severity may influence an individual's perception of their symptoms; cough is a more important symptom for patients with severe asthma than those with a milder disease.


Discrete Choice Experiment Methodology and Models
In the MNL specification, the deterministic component of utility (the random component of the utility function follows a type I extreme value distribution) for respondent n and alternative i in choice task t (out of 8) is written as: V int = β BreathL1 BreathL1 int + β BreathL2 BreathL2 int + β SleepL1 SleepL1 int + β SleepL2 SleepL2 int + β TightL1 TightL1 int + β TightL2 TightL2 int + β WheezeL1 WheezeL1 int + β WheezeL2 WheezeL2 int + where, as an example, CoughL1 int is set to 1 if alternative i contains the Cough level 1 (and is set to 0 if alternative i has a Cough level other than 1), and where β CoughL1 is the associated marginal utility coefficient, which is to be estimated.
Equation 1 shows the utility individual n will receive if they select either of the first two alternatives, whereas Equation 2 shows the utility individual n will receive through the selection of the 'Don't know' option (displayed as alternative 3, in this case). The attributes were entered as dummy variables in order to allow us to capture any non-linear preference structure for these attributes, where the 0 level was used as the baseline (i.e. the sensitivity for absence of symptom was fixed to zero). Notably, since the baseline was set to level 0 for each attribute, it would be sensible to expect all of the level 1 and level 2 coefficients to be negative, as it is improbable for a patient to prefer experiencing symptoms to no symptoms.
For example, it is unlikely that a patient would prefer Cough Level 2 (A lot of coughing with restricted activities) to Cough level 0 (No coughing). If a coefficient (e.g., β CoughL2 ) is found to be significant, this means that patients' preferences for that level is significantly different to the baseline of level 0.
The specification above assumes that preferences for the different symptom attribute levels are the same for all respondents. As we are interested in whether preferences for cough vary across patients, we can revise our model specification to allow for differences in sensitivities by specific demographics/characteristics. Consider for example, a model, which elicits preference differences between male and female respondents. For each of the cough levels (other than the baseline 0), we thus estimate a base coefficient, along with offsets for the separate groups (male vs female). This specification is shown in Equation 3, where, for example, Δ CoughL1;Female shows the shift in the utility for Level 1 Cough for a female respondent relative to a male respondent. The shift parameter represents the difference in preferences between the two groups; where a value of 0 would mean that the two groups have the same preference.
The MNL models estimated are described in Table E1 below. In the primary MNL model (model 1), all patients are assumed to have the same preferences for each of the attributes.

ACQ-5 model
Preferences for cough allowed to vary by asthma control; ACQ-5 score > 1.5 compared to ACQ-5 score ≤ 1.5 (baseline). 4 Age model Preferences for cough allowed to vary by age; age > 50 years compared to age ≤ 50 years (baseline).

Estimation of scale factors
As study respondents were recruited from two distinct asthma populations, it is important to determine whether any differences in preferences found are caused by true preference differences or differences in their associated scale factors 13 . Scale heterogeneity (also referred to as heteroskedasicity 14 ) refers to heterogeneity in the variance associated with the random component of utility, ε. Thus, we estimate one set of coefficients, β and an additional scale coefficient for the second primary care population, µ PC . The estimation of a scale model was performed as described by Swait and Louviere (1993) 13 . The test statistic retrieved, λ A = 38.04, is significant at the 5% significance level; we therefore conclude that the two groups have different preferences, and thus should be modelled separately, rather than performing a grouped analysis with all participants.

Multinominal Logit (MNL) models
For the purposes of quality control and to ensure that patients were engaged when completing the questionnaire, patients who answered any of the choice scenarios irrationally were not included for data analysis (table E2). Given that the two groups (severe asthma vs mild/moderate asthma) needed to be estimated separately (see scale analysis above), over-parameterisation was a methodological concern (i.e., estimating too many parameters). Therefore, as level 1 Chest tightness and level 1 Wheeze were found to be not significant in any of the preliminary models, for these two attributes level 0 and level 1 were combined. Table E2: Distribution of the attribute levels for each scenario in the discrete choice experiment The distribution of attribute levels for each of the scenarios are shown in Table 2. For the purposes of quality control and to ensure that patients were engaged when completing the questionnaire, scenarios 1 and 3 were included to assess for rational choice behaviour. Namely, the scenarios were set up so that one alternative was an "obvious" better choice in terms of symptom burden. For example, in scenario 1 (as shown in Table 2), patients should always prefer week B to week A. Patients who answered scenarios 1 and 3 irrationally were not included for data analysis.