Scottish Health Survey - topic report: mental health and wellbeing

2. Methodology

Data for the two-year 2012/2013 period has been used for all descriptive and regression analyses in this report. The increased sample size in using the 2012/2013 data compared to single year datasets allows for more robust analyses of results to be presented.

2.1 Descriptive analysis

Chapters 4 and 5 include results for the WEMWBS and GHQ indicators by various characteristics. For each factor, the mean WEMWBS score is reported alongside the proportion of adults scoring four or higher on GHQ12 (indicating the presence of a possible psychiatric disorder).

Data in these chapters are generally presented for all adults. Where there are significant differences by sex, these are shown in the results.

2.1.1 Age-standardisation

For each topic in the descriptive results sections, data have been age-standardised, unless otherwise stated. This ensures that comparisons between population subgroups are made on a like-for-like basis.

The socio-economic, behaviour and health condition characteristics described in the results sections each have a distinct age distribution. For example, the group of people who meet the physical activity guideline have a younger age profile than those who are not physically active. Age-standardisation enables these population subgroups to be compared, after adjusting for the effects of different age profiles. This ensures that any differences detected in mental health and wellbeing are not simply due to differences by age. In some cases, it was not appropriate to age-standardise results, for example when population sub-groups in some age bands were too small.

2.2 Logistic regression

The literature review has identified a range of socio-demographic, behavioural and health state factors associated with poor mental health and wellbeing. To explore these factors, multivariate logistic regression models were run for binary versions of each outcome measure using a reduced set of the most significant factors for each measure, as described below. The literature suggests gender specific associations with mental health for many of the factors, so the regression models were run separately for men and women.

In addition, multivariate logistic regression models were also created with the outcome or dependent variable being each of the components that make up WEMWBS and GHQ12. Due to time constraints, the same reduced sets of factors described above were used in models for the component questions under each measure. Although this is a limitation of the analysis, this has identified some components with results significantly different to the other components that make up each mental health measure, or which show considerable difference by gender.

2.2.1 Selection of most significant factors

This section describes how the significance of factors associated with measures of mental health and wellbeing, as identified in the literature review, was tested to derive reduced sets of significant variables to be included in final multivariate logistic regression models.

The methodology of variable selection used for this report is similar to that used in previous Scottish Health Survey reports.^[75] Forward selection successively adds variables that are significantly associated with the outcome measure at the 5%/95% level. Under backward selection, the least significant independent variables are removed until the remaining variables are statistically significant. A combination of forwards and backwards selection methods was used to produce a set of significant variables for the binary WEMWBS and GHQ measures, with variations in variables selected for each measure.

A potential problem with such automatic methods is that modelling can become separated from subject matter expertise. In this case, only factors identified in the literature as associated with mental health and wellbeing have been retained prior to the automatic variable selection processes. Furthermore, where bivariate analysis showed that a variable was not significantly associated with the outcome, it was not included in the corresponding regression model. Following the variable selection procedure, collinearity checks were performed on the selected independent variables, and redundant variables then removed from final models.

2.2.2 Binary outcome measures

Logistic regression models typically require the dependent or outcome variable to be a binary (two category) measure. In the case of WEMWBS, a score of less than one standard deviation below the mean has been used as a cut-off to define a low score. By this methodology, a respondent with a score of 41 or lower is classified as having a low mental wellbeing score.

WEMWBS scores can range from 14 to 70. A binary WEMWBS variable has therefore been coded such that any score between 14 and 41 is set to 1, and all other higher scores are set to zero.

A score of four or more on the GHQ12 measure has been selected as the most appropriate for identifying respondents with a possible psychiatric disorder. A binary measure was derived by setting a score of four or more to 1, and a score of 3 or less to zero.

2.2.3 Interpretation

Multivariate logistic regression estimates the independent effect of factors, while adjusting for other factors simultaneously, on the binary outcome derived from each measure of mental health and wellbeing. The value of multivariate analyses like these is being able to disentangle confounding effects, for example being able to test whether the low levels of mental wellbeing among a particular subgroup (such as carers) is explained by other demographic factors (such as the corresponding age profile).

Multivariate regression models were run on the reduced set of the most significant variables, for each binary mental health measure, on all adults of 16 years and over and then run separately for men and women.

The odds ratios of having a low WEMWBS score, or scoring four or more on GHQ12, compared to a reference group for each variable are shown in Tables 6A and 6B. In these analyses, the odds of a reference group (shown in the table with a value of 1) are compared with that of the other categories for each of the individual factors. In Table 6A, for example, an odds ratio greater than one indicates that the category in question had higher odds of scoring 1 on the dependent variable, in this case a low WEMWBS score. An odds ratio less than 1 means lower odds of having a low WEMWBS score, compared to the reference group. Odds ratios whose confidence limits span the value 1 are not significantly different to the reference category. By simultaneously controlling for a number of factors, the independent effect each factor has on the variable of interest can be established.

Multivariate logistic regression models on each of the component questions that make up WEMWBS and GHQ12 were also run, again separately by sex. The component questions for GHQ12 have previously been mapped to binary variables, allowing the same interpretation as described above to be used to test for association with independent factors. However, the component questions for the WEMWBS scale retain the five point scale (varying from 1='none of the time' through to 5='all of the time'), so ordinal logistic regression has been carried out with each component question as the dependent variable. The same independent variables used for the regression models on the aggregate measures were applied to separate models by sex.

It is important to note that the odds ratios shown in ordinal logistic regression models in this report are interpreted differently than binary logistic regression. Ordinal logistic odds ratios are interpreted as the association between the independent variable and being in a lower level of the dependent variable. In the case of the WEMWBS components, an odds ratio greater than one indicates that the category in question has higher odds of scoring a lower score on the component question than the reference category, whereas an odds ratio less than one means they had lower odds of scoring lower on the component question.

2.2.4 Missing data

The way missing data is handled can have a profound effect on the results of regression analyses.

Given the number of variables included, it was important that records which may include missing data for any one of the variables were still included in the analysis. In line with previous Scottish Health Survey reports, variables with a small number of missing values have values imputed to the category containing the largest number of cases. For variables with a large number of missing values, a separate missing value category was created and included in the analysis.^[76] Alternative imputation methods were considered to be too complex to be implemented, given the generally low volume of missing data for most variables.