This study implemented the standardized EQ-5D-5L valuation protocol developed by EuroQol using the EQ-Portable Valuation Technology (EQ-PVT) [12]. Details of the study design were reported by Malik et al. [11], in which the descriptions of the pilot feasibility study that precedes this EQ-5D-3L valuation study were provided. Two preference elicitation methods were used, including the composite time-trade-off (cTTO) and discrete choice experiments (DCE). In the DCE task, respondents are presented with two EQ-5D-3L health states (State A and State B), and choose the health state they consider to be the best. The cTTO method is more complex. Respondents are presented with two different lives: 10 years in some EQ-5D-3L health state (Life B) and a number of years in full health, smaller than or equal to 10 years (Life A). Respondents choose which life they consider to be the best, and depending on their answer, the number of years in full health is subsequently varied. The cTTO task ends when respondents indicate they are indifferent between Life A and Life B, and a value is inferred from the number of life years in full health are traded. If respondents indicate they would prefer to die (0 years in full health in Life A), they are presented with a lead-time TTO task (LT-TTO), where Life A and Life B are preceded by an extra 10 years in full health (e.g. 10 years in full health in Life A versus 10 years in full health, followed by 10 years in some EQ-5D-3L state in Life B). The procedure of the LT-TTO task is then similar to the regular TTO task. Details of the DCE method [13] and the description of iteration process of cTTO method is available elsewhere [12, 14].

For cTTO, 28 health states being valued, grouped into 3 blocks with 10 health states in each block. The most severe health state, ‘33333’ (the most severe health problems defined by all EQ-5D-3L domains) was included and valued in all three blocks. For DCE, 60 pairs of health states grouped in 6 blocks were used.

All related documents including the interview guides, survey instructions and the EQ-PVT platform were translated into Urdu, the official language of Pakistan. The translation was done by a professional translator commissioned by the EuroQol Office and reviewed by the study team members. The ethics approval was obtained from the Health Research Ethics Committee at the Hamdard University (HREC, HUIC-091).

A moderately representative sample was recruited from the Pakistani general population. A multi-stage stratified quota sampling approach was used on the basis of respondents’ ethnicity, age, gender and religion beliefs. Adults aged 18 years and older who were able to complete the interviews were included as respondents. Those who reported to have severe mental and/or physical illnesses at the time of survey were excluded. Chan KKW et al. (2020) explored the sample size and prediction accuracy of EQ-5D-3L values and reported that the effect of sample size and threshold for the minimum number of respondents is 300–500 [15]. Moreover, Hansen et al. [16] reported that, keeping in view the cost, the expected gain in prediction accuracy from increasing sample sizes beyond 300–500 respondents is minimal and that the choice of model can compensate for a smaller sample size. Taking this into contemplation, a total of 300 participants were recruited from the three cities using population proportionate to size sampling (PPS) to allocate this sample among the five major ethnic groups of Pakistan. The total sample of 300 participants was distributed among these ethnicities according to their proportion. We considered those cities which have higher density of these ethnic groups, such as Punjabi participants from Lahore, Kashmiri and Pashtun’s participants from Islamabad, and Karachi as the area for Sindhi and Baluchi ethnic communities. Moreover, PPS was also used to allocate quota for each demographic characteristic including gender, age and religion. Data collections were conducted in the three most populous and diverse cities of Pakistan including Islamabad (Capital), Lahore (Punjab) and Karachi (Sindh). Three interviewers who were trained by experts from the EuroQol Research Foundation recruited respondents and conducted the interviews in these cities. Recruitment used several strategies, for example, the networks of the interviewers, promoting via social media, posting flyers in local markets, pharmacies and respondents’ recommendations of other potential respondents. Interviews were performed between February 2019 and August 2019.

During the data collection, performance of the interviewers and the quality of data were checked continuously by the quality assurance officers from the EuroQol Research Foundation. Low-quality interviews and/or responses were identified as follows: (1) the lead-time time trade-off (LT-TTO) was not explained by the interviewers; (2) not enough time was spent on explaining the task, using the wheelchair examples—180 s was used as a lower limit; (3) the ten cTTO health states were valued in less than 5 min; and (4) inconsistency was spotted in the cTTO ratings (e.g. if the value for the worst state ‘33333’ was not the lowest value. Data collected by interviewers that met any of these criteria in at least 40% of their interviews were discarded, as they were of suspicious quality. Details of the quality control protocol can be found in the publication by Ramos-Goni et al. [17].

Descriptive analyses were first performed to examine the respondents’ demographic information and other general characteristics. Findings were reported using the mean values, standard deviations and percentages/frequencies.

Several models were used to estimate the cTTO data. In these models, the cTTO value, on a scale between 1 and − 1, was used as the dependent variable. Ten dummy variables were constructed representing the presence of a certain level of problems on a dimension in the health states that are valued, coded as regular dummies. These dummies equal 1 if the respective level of problems is present on the respective dimension. For example, MO3 represents mobility problems at level 3, described as ‘confined to bed’. If the health state to be valued has mobility problems described as ‘confined to bed’, MO3 will equal 1, and 0 otherwise. For example, for health state 21322, the variables MO2, UA3, PD2 and AD2 equal 1, and all other variables (MO2, SC2, SC3, UA2, PD3 and PD3) equal 0, as the health state to be valued is not described by those health problems. The first of the estimated models was a random intercept model, which takes into account the nested structure of the data, that is, that responses were nested in respondents, which may cause correlation between responses within respondents. Second, a random intercept Tobit model was estimated. Tobit models account for censoring that may be present in the observed data. In the case of the cTTO, respondents cannot assign values lower than − 1 to a health state, due to the way in which the cTTO task was constructed. However, some respondents may actually be willing to assign an even lower value to some health states, which is then not reflected in the data. To this end, the Tobit model accounted for this censoring of the data at − 1. Third, as responses vary over the severity of health states, with more variation in responses present for more severe health states, the valuation data is likely to be subject to heteroskedasticity as the error term is not constant over the health domains. To account for this heteroskedasticity, we model the variance of the error term as a function of the ten dummy variables that were included as the dependent variables. Lastly, we estimated a Tobit model that corrected both for heteroskedasticity and for the censored nature of the data.

For the DCE data, we estimated a conditional logit model and a probit model. The dependent variable was a binary variable, with 1 indicating the alternative chosen in the paired comparison and 0 otherwise. Both of these models produce values on a latent scale ranging from 0 to some positive value. Therefore, they cannot be used directly to compute QALYs. The DCE data is therefore needed to be anchored onto the 0–1 scale, i.e. the cTTO data, to produce utilities scaled on the 1 (full health) and 0 (dead) QALY scale. This is done via a hybrid modelling strategy [16] in which a joint likelihood function was estimated for the cTTO and DCE data combined. Hybrid models account for the censoring and the heteroskedasticity, and the combination of both cTTO data were used.

All hybrid and cTTO-only models were estimated with and without a constant. All models were estimated using maximum likelihood estimation (MLE). The final value set was selected based on the logical consistency, significance of the coefficients of the models, model fit criteria such as Bayesian information criterion (BIC) and, mean absolute error (MAE). The MAE was computed both as a MAE over all responses, and as a comparison of the mean observed and predicted values for the 28 health states included in the cTTO health state design.