Databases
Three databases collated at two university teaching hospitals in northeast England were used for this analysis. These databases were set up for audit or research purposes and included the MDADI as part of a battery of swallowing outcome measures. Patients were consecutively and prospectively approached, and their data were anonymised when enrolled into the databases. The focus of the databases was 1) a service evaluation of functional outcomes for minimally invasive surgery (transoral laser microscopy (TLM) or transoral robotic surgery (TORS)), 2) a research database recording swallowing outcomes for non-surgical primary treatment (chemotherapy (CRT) or radiotherapy (RT)), and 3) a feeding tube audit comparing outcomes for those receiving a reactive nasogastric tube (NGT) or prophylactic radiologically inserted gastrostomy (RIG) tube for primary or adjuvant (chemo)radiotherapy.
Questionnaire
The MDADI was routinely collected pre-treatment, three, and 12 months post-HNC treatment. Questionnaires completed at three months were extracted from these existing databases. This time point was chosen as it represented the greatest deterioration in post-treatment MDADI scores and contained a full range of the scale of responses. It therefore maximised coverage of most questionnaire items [
3,
20]. In order to conduct factor analysis on the patient responses, a wide range of scores is desirable [
22,
23]. We opted not to use data from the same patient twice (e.g. both three and 12-month time points), as their questionnaire interpretation and responses would be very similar. Each of the 20 MDADI items are rated on a 5-point Likert scale ranging from 1: “strongly agree” to 5: “strongly disagree”, except for questions abbreviated by “Conscious” and “Eat Out” (Appendix A) whose ratings are scored in reverse order ranging from 5: “strongly agree” to 1: “strongly disagree”. A composite MDADI score was generated by calculating the mean response for the 19 items (excluding the global question) making up the emotional, functional and physical subscales and multiplying the result by 20, resulting in a score ranging from 20 representing a low QOL function to 100 indicating high QOL function[
8]. All MDADI question responses were inputted into the data extraction sheet for analysis.
Factor Analysis
The data were analysed using IBM SPSS Statistics for Windows, version 24 (IBM Corp., Armonk, N.Y., USA). Multivariate factor analysis was carried out to examine relationships between multiple ordinal questionnaire items each measured on a Likert scale and collected from the 3-month MDADI questionnaire. By exploring the structure of the data in this manner, we were informed as to whether item reduction was viable.
It is recommended that at least ten completed questionnaires per questionnaire item are used for factor analysis implementation [
22,
24,
25].
For this study, factor analysis was carried out using the most recommended and utilised combinations of options [
24], i.e. (i) principal axis factoring (PAF) and (ii) principal components extraction. Principal axis factoring is, strictly, a “factor analysis” method whereas factor analysis with principal components extraction is often termed “PCA”. Although different in their mathematical derivation, factor analysis and PCA share a common aim: to identify the underlying dimensions in the data. It is, however, acknowledged that there is no guarantee that factor analysis and PCA will result in the same solution [
23,
26]. Formally, the dimensions are called “factors” (in factor analysis) and “components” (in PCA); however, for readability, we have termed both entities “factors”.
In a factor analysis, the total number of factors equals the number of items in the questionnaire. Each factor captures a proportion of the overall variance in the observed items. Factors are output in the order of how much variation they explain with the eigenvalue representing the variance explained by a particular factor. The first factor is, thus, the most important and accounts for the largest amount of variance in the data. Factors that explain the least amount of variance are discarded. In this study, factors with eigenvalues greater than 1 were retained [
27].
To allow for better differentiation of the factors, factor rotation was utilised. Orthogonal rotation results in independent factors; oblique rotation allows the factors to correlate. As there was no consensus as to whether or not the underlying factors should be related, both direct oblimin (oblique) and varimax (orthogonal) rotation were employed.
For interpretation purposes, the weights (loadings) of the items for each factor are considered, i.e. items having a large weight are used to label a factor. In a factor analysis, these loadings describe the strength of the relationship between each MDADI question and the underlying factor. Factor loadings were initially interpreted with an absolute value greater than 0.5 [
28]. Recognising that a factor loading threshold greater than 0.5 is acceptable whilst one that is greater than 0.7 is deemed good [
29], a more stringent factor loading threshold of greater than 0.7 was used to identify the top endorsed MDADI items.
The Kaiser–Meyer–Olkin measure of sampling adequacy and Bartlett’s test were generated to ensure that the criteria for a satisfactory factor analysis were met [
24].
As highlighted by a recent systematic review by Patel et al. [
30], the MDADI did not include a plan for missing data. Missingness in this study was assessed by evaluation of the percentage of questionnaire responses missing for each item. Participant anonymisation at previous enrolment to the databases precluded a statistical assessment of differences in demographic and clinical characteristics between those who submitted complete and incomplete MDADI questionnaires.
The top five endorsed items using PAF and PCA were checked for agreement and descriptive statistics for each item were generated. Cronbach’s alpha was used to evaluate the psychometric properties, reliability, and internal consistency of the MDADI, with an acceptable value ranging from 0.7 and 0.9, where higher values would suggest redundancy of items [
31]. Preliminary validity assessment was also carried out with a view to future expansion.