Measures of asthma control and quality of life: longitudinal data provide practical insights into their relative usefulness in different research contexts

Abstract

Purpose

To further our understanding of the relationships between asthma control and health-related quality of life (HRQOL) and provide insights into the relative usefulness of various measures in different research contexts. We present a conceptual model and test it with longitudinal survey data.

Methods

Participants recruited via population sampling and hospital Emergency Departments completed questionnaires every 6 months for up to 3 years. Measures included: sleep disturbance, use of short-acting beta agonists (SABA), activity limitation, urgent medical visits, hospital use, Marks’ Asthma Quality of Life Questionnaire (AQLQ-M) and the SF-36 Health Survey. Correlation analysis and multi-level models tested predictions from the conceptual model.

Results

A total of 213 people with asthma aged 16–75 years provided 967 observations. Correlations between asthma control and asthma-specific HRQOL were stronger than those between asthma control and generic HRQOL. The asthma control variables explained 54–58% of the variance in asthma-specific HRQOL and 8–25% of the variance in generic HRQOL. Activity limitation was the main contributor to between-person variation, while sleep disturbance and SABA use were the main contributors to within-person variation.

Conclusions

Sleep disturbance and SABA use may be most useful in evaluating treatment effectiveness, while activity limitation may be better when monitoring the impact of asthma in populations.

Appendix

Each of the four HRQOL scales was analysed, separately, in the following sequence and with the following models:

Step 1

One model: a random intercept model

$$ y_{ij} = \alpha + u_{i} + e_{ij} $$

for i = 1 to n participants on j = 1 to 6 occasions (the 6 × 6-monthly timepoints of the 3-year survey period), where y _ij is the QOL score for patient i on the jth occasion, α is the population mean QOL score (intercept), u _i is a random effect giving the deviation of the average QOL score of the ith patient from the population mean, and e _ij gives the deviation of y _ij from the average QOL score of the ith patient.

In this model and all subsequent models, u _i and e _ij are assumed to be independent and normally distributed, with mean zero, between-patient variance σ² _u and within-patient variance is σ² _e . Thus;

$$ u_{i} \sim N(0,\Upphi_{u}^{2} ),\,\,e_{ij} \sim N(0,\Upphi_{e}^{2} ) $$

The covariance matrix for the e _ij was modelled as unstructured (no constraints) at this stage.

Step 2

One model: five explanatory variables were added simultaneously to the model at Step 1; age, gender, residential area (capital city or regional NSW), smoking status (current smoker or not), recruitment source (community or hospital):

$$ y_{ij} = \alpha + \sum\limits_{k = 1}^{5} {\beta_{k} x_{ki} } + u_{i} + e_{ij} $$

where x _ki are the observed levels of k = 1 to 5 recruitment and socio-demographic variables for the ith participant. These variables are called ‘person level’ or ‘level 2’ effects, and there is only one value of x _ki for each participant, as measured at baseline. β _k is the regression parameter (slope) for the kth socio-demographic variable.

Step 3

Five models, each containing only one asthma control variables (with one corresponding fixed parameter), added to the model at Step 2 as level 1 (time varying) effects:

$$ y_{ij} = \alpha + \gamma_{l} z_{ijl} + \sum\limits_{k = 1}^{5} {\beta_{k} x_{ki} } + u_{i} + e_{ij} $$

where z _ijl is the observed level of the lth asthma control variable for participant i at time j. These variables are called ‘time varying’ or ‘level 1’ effects, and there was one value of z _ijl each time each participant completed a survey. Each asthma control variable was centred on its overall mean, except hospital use. γ _l is the regression parameter (a fixed slope parameter) for the lth asthma control variable.

Step 4

Five models, corresponding to the five models at Step 3 plus a person-specific random slope parameter for the corresponding asthma control variable;

$$ y_{ij} = \alpha + \gamma_{l} z_{ijl} + \sum\limits_{k = 1}^{5} {\beta_{k} x_{ki} } + u_{i} + v_{il} z_{ijl} + e_{ij} $$

where v _il is a subject-specific effect giving the deviation of the ith subject-specific slope from the population slope, γ _l , for the lth asthma control variable. v _il were assumed to be normally distributed, with mean zero, $v_{il} \sim N(0,\Upphi_{v}^{2} )$, and to be independent of u _i and e _ij . The likelihood ratio test was used to compare each model with the corresponding model at Step 3 to determine whether the random slope improved model fit.

Step 5

One model, being the model from Step 2 with all five of the asthma control variables, included simultaneously as fixed effects (i.e. five fixed parameters), plus a random slope parameter for a control variable only where this was indicated by the models at 4:

$$ y_{ij} = \alpha + \sum\limits_{l = 1}^{5} {\gamma_{l} z_{ijl} } + \sum\limits_{k = 1}^{5} {\beta_{k} x_{ki} } + u_{i} + \sum\limits_{l = 1}^{L} {\nu_{li} z_{ijl} } + e_{ij} $$

where v _1i ,…,v _Li were retained if they improved the model fit. The covariance matrix for e _ij was constrained to a variance components structure (zeros in the off-diagonals) at this stage because the large number of random effects lead to convergence problems with other covariance structures. The likelihood ratio test was used again at Step 5 to determine if all random slopes retained in the five separate models at Step 4 continued to contribute to the single model at Step 5; the model of best fit here was then the final model from Step 5;

Step 6

One model, being the final model from Step 5 plus two dummy variables for treatment (level 1 effects): (1) ICSLABA, indicating regular users of combined ICS and LABA; (2) ICS, indicating regular users of ICS alone:

$$ y_{ij} = \alpha + \sum\limits_{l = 1}^{5} {\gamma_{l} z_{ijl} } + \sum\limits_{k = 1}^{5} {\beta_{k} x_{ki} } + \beta_{6} ICSLABA_{i} + \beta_{7} ICS_{i} + u_{i} + \sum\limits_{l = 1}^{L} {\nu_{li} z_{ijl} } + e_{ij} $$

Since neither treatment effect was statistically significant for any of the four HRQOL scales, the final model was the best fit model from Step 5.