Design
This was a cross-sectional study with questionnaire and clinic data each collected at one time point.
Participants
Patients aged ≥16 years of age and diagnosed with severe asthma as defined by the ERS/ATS guidelines were invited to participate [12]. Participants were recruited from six UK severe asthma centres and were excluded if they were diagnosed with another condition that significantly contributed to their respiratory health, e.g. lung cancer, heart disease or chronic obstructive pulmonary disease.
Patient reported outcome measures:
Severe Asthma Questionnaire (SAQ)
The SAQ consists of 16-items scored from 1 to 7, with a higher score indicating better quality of life. The mean of the 16 items is calculated to provide the SAQ score. The SAQ also contains a separate, Borg-type scale ranging from 0-100 and based on the Global Quality of Life Questionnaire [13] which provides the SAQ-global score [8].
Asthma Control Test (ACT)
The ACT consists of five asthma symptom and medication use items, which are totalled to provide an indication of asthma control. The sum of the five items, scale – 5, is calculated to give the ACT score, with a higher number indicating better asthma control [14].
Asthma Control Questionnaire-6 (ACQ-6)
The ACQ consists of six items, five concerning asthma symptoms and one on daily use of rescue bronchodilator. Patients respond to these items on a 0-6 scale (0 = no impairment, 6 = maximum impairment). The mean of the six items is calculated to provide the ACQ-6 score with a lower number indicating better asthma control [15].
EQ-5D-5L and mood measurement
The EQ-5D-5L consists of 5 items scored from 1 to 5 with a higher score indicating greater impairment, and a 0-100 visual analogue score, the EQ-5D VAS [16]. Index scores are calculated using the 2012 value set for England [17] and these index scores are presented here. For this study we used item 5 of the questionnaire as a proxy measure of mood. Participants indicate the degree to which they feel “anxious or depressed” on a five point scale of severity.
Clinical data
Clinical data included body mass index (BMI) and asthma severity as measured by the following items: GINA treatment step, spirometry (forced expiratory volume in 1 second (FEV1) and FEV1 % predicted), prednisolone dose (mg/day), health care utilisation in the last 12 months including number of hospital admissions, emergency department visits and exacerbations requiring oral corticosteroids (OCS). An estimate of cumulative OCS exposure (mg/year) was calculated by multiplying a patient’s maintenance OCS dose by 365 and adding an estimated use of OCS following each exacerbation. British Thoracic Society and GINA guidelines suggest that 40mg of prednisolone for 7 days should be prescribed for the treatment of exacerbations [18]. This equates to 280mg of OCS per exacerbation.
Procedure
Patients with severe asthma at five specialist treatment centres were approached for recruitment to this study. Questionnaires were completed in clinic once written informed consented was given. Spirometry was conducted either at the time of questionnaire completion or the most recent within the previous 6 months. Participating sites collected either ACT or ACQ data as a measure of asthma control for this study as per their normal clinical practice. The same data collected for a previous study [8] from a sixth specialist centre were also included for analysis.
Ethical Approval
This study received ethical approvals from the Research Ethics Committee/Health Research Authority (REC reference: 19/WA/0011, IRAS project ID: 250167) and was sponsored by University Hospitals Plymouth NHS Trust. Data from a previous study received ethical approval number 16/NE/0188, IRAS ID: 207601) [8].
Statistical analysis
Exploratory factor analysis (EFA) is a statistical procedure that can be used to make inferences about underling causal structures. The procedure is based on the assumption that correlations between variables is due to a common cause, referred to mathematically as a factor (i.e., causal factor) and psychologically as a construct (i.e., psychological construct.) In the case of patient reported outcomes, the constructs are dimensions of meaning that are responsible for the way patients interpret and respond to the individual items of a questionnaire. People can use many different dimensions of meaning to evaluate their outcomes, so the aim of the technique is to identify the main dimensions that drive response to individual items. Factor solutions that achieve a ‘simple structure’ [19,20] indicate that those main dimensions have been identified and therefore provides a good description of the underlying dimensions of meaning used to interpret the items of a questionnaire. However, people interpret any item of a questionnaire by using one or more dimensions of meaning, and so discovery of the main dimensions of meaning is aided if the items tend to be specific to different meaning dimensions. Factor analysis of patient reported outcomes is therefore a way of exploring the meaning of a questionnaire but that exploration depends on the items of the questionnaire. The meaning of simple structure and the rationale for choosing the factor parameters for this analysis are described below.
There are two main forms of data extract: principal component analysis and factor analysis. Principal component analysis is a simpler and older form of analysis that became popular when computers were slower and is the default option in many statistical packages. Principal component analysis is a method of data reduction only, it does not distinguish between unique and shared variance and therefore does not identify causal factors (psychological constructs). The method risks overestimating variance. Factor analysis analyses only shared variance and in so doing provides information about underlying causal structures, it does not inflate estimates of variance and for most purposes is the recommended form of extraction [20]. We used factor analysis rather than principal component extraction because we wanted to identify causal constructs and estimate variance, and we used principal axis factor analysis as a commonly used type of factor analysis [20].
EFA is an exploratory tool that provides choice in the numbers of factors to be extracted. When used for subscale construction in HRQoL, the primary determinant of factor number and hence subscale number is a number that is both theoretically plausible and clinically useful. If that number produces a simple structure (see later), then that number can be accepted as the final solution. If that number fails to produce a simple structure, then alternatives should be considered. In our case, a plausible and useful number based on content is that there should be three factors, corresponding to activity, emotion and extra-pulmonary symptoms.
There are several driven methods of determining factor number that can be used in addition to the primary, theoretical determination, but these methods typically produce different results and are therefore advisory only [19]. The eigenvalue is a measure of variance explained, and because of the way factors are extracted eigenvalues decrease with the number of factors extracted. The default setting in many statistical packages is to select the number of factors with eigenvalues greater than one (the Kaiser-Guttman rule) [21]. Because eigenvalues increase with the number of items analysed this method provides limited information and is widely held to be the least useful data driven method of advising on factor number [19,20]. However, the overall pattern of all eigenvalues is useful not only by providing data for another, widely recommended test of factor number, the scree test. The scree test requires inspection of the eigenvalues to determine the point at which eigenvalues reduce in a similar way – the analogy is with the scree at the bottom of a cliff.
Once the number of factors is set, principal axis factoring coupled with rotation provides a solution capable of interpretation. The technique of rotation can be done either by forcing the factors to be uncorrelated (called orthogonal rotation, e.g., varimax) or allowing the factors to be correlated (called oblique rotation, e.g., oblimin, promax), each type of orthogonal or oblique rotation having slightly different properties. Orthogonal rotation should be used only when uncorrelated factors are predicted on theoretical grounds or when there is evidence from an earlier oblique rotation that the factors are largely uncorrelated. Varimax (i.e., orthogonal) rotation became popular through its use in psychology where there was a theoretical requirement for personality factors to be uncorrelated [22], but this form of rotation is often used incorrectly in situations where factors may be correlated. In the present case, factors are predicted to be correlated as the three content derived domains of the SAQ all form part of the overall HRQoL. Promax and oblimin are commonly used forms of oblique rotation, promax being computationally simpler than oblimin, oblimin being the preferred form [20] and that which was used here.
EFA produces a factor matrix where each item of a questionnaire has a value, called a loading, on each of the factors. The item loadings vary between -1 and 1 and can be considered equivalent to correlations between the item and the artificial variable represented by the factor. We adopted the convention that items that load at or greater than 0.3 should be allocated to that factor [19,20]. Orthogonal rotations produce only one factor matrix whereas oblique (i.e., correlated) rotations produce two matrices, the structure matrix and the pattern matrix. The pattern matrix expresses the relationship between items and a factor after removing the effect of the correlations between the factors, and therefore provides a clearer picture of the separation of items between factors, should that be the case, compared to the alternative, the structure matrix. However, by removing the correlations between factors, only the factor loadings of the structure matrix but not of the pattern matrix can be considered equivalent to a correlation with an artificial variable. Therefore, in order to interpret the pattern matrix it is necessary to know the degree of correlation between the factors produced by the rotation. These factor correlations are reported separately from the pattern matrix, and are similar but not identical to subscale correlations because factor correlations are based on response to weighted items whereas subscales are based on unweighted items [19].
Rules for sample size for EFA have largely disappeared because sample size depends to some extent on the data though a common rule of thumb is a ratio of 10:1 participants to items [19]. Adequacy of sample size can be checked statistically. The solution provided by any EFA depends on the correlation matrix between the variables. Differences in that matrix resulting from low correlations and small sample sizes can produce large differences in solution, i.e., factor instability. The Kaiser-Meyer-Olkin measure of sampling adequacy provides a way of measuring the level of factor stability. The Kaiser-Meyer-Olkin varies between zero and one, values above 0.8 indicating that the factor solution is likely to be stable, and above 0.9 highly stable. However, if sample size allows, factor stability can be checked by separate analysis of subgroups. In the analysis conducted here, we examined factor solutions for males and females separately, a technique that also checks that males and females interpret every item in the same way.
The aim of an EFA, as a statistical tool, it to find a solution where there is a simple structure to the data. Simple structure is summarised as “item loadings above .30, no or few item crossloadings, no factors with fewer than three items” [20] . Validation of HRQoL subscales has an additonal requirement, that the subscales so produced are both theoretically plausiable in terms of content as well as clinically useful. An EFA solution producing 10 subscales may achieve simple structure but is unlikely to have much clinical use. Cross-loading items (i.e., where the loading is > 0.3 on more than one factor) indicate either that response to the item is affected by more than one construct, or that the solution provides a poor fit for data. Either way, the presence of cross-loading items is undesirable and absence of all but a bare minimum of cross-loading items is a primary requirement for construct validation of subscales of a HRQoL questionnaire [20]. Validation requires a “clean” factor matrix, namely one where there is good separation between loadings for every item. An item that loads 0.32 on one factor and 0.25 on another is a poor item. Items with loadings of 0.32 and 0.01 and 0.25 and 0.60 are acceptable, but the goal is for the largest possible separation.
Although a HRQoL questionnaire may fail to provide validated subscales according to the criteria described above, the overall scale score can still be used. It is almost inevitable that all the items of HRQoL questionnaires will load on the first unrotated factor. This is because, in general, HRQoL deficits in a population increase with severity and so the first factor unrotated factor is simply a severity factor. An HRQoL item must by definition be related to health and it would be unusual if an item failed to correlate with overall severity. Subscale construct validation by EFA is more demanding as it requires specificity of items to constructs, rather than specificity to severity.
Following EFA, subscales were constructed on the basis of the factor loadings by taking the mean of items loading on any factor. The relationship between the subscales and other variables was examined using Pearson correlations. EFA and correlations were conducted using SPSS version 25. Tests of difference between correlations were carried out using Psychometrica (https://www.psychometrica.de/correlation.html).