Data were collected for the British household panel survey, an annual survey of individuals in private households in England, Wales, and Scotland. The design and primary aims of this survey have been described elsewhere.11,19
Only subjects who completed psychiatric assessments at two sets of interviews were included. The first set of interviews took place in 1991 (T1) and the second set (T2) 12 months later. The survey investigators complied with the ethical guidelines of the Social Research Association.19
Specific ethical approval was not sought for this secondary analysis, which was based on anonymous data supplied from the data archive of the Economic and Social Research Council in accordance with its regulations.
Common mental disorders were assessed with the general health questionnaire, comprising 12 items.20
We followed previous studies in treating common mental disorders as a single dimension.3,21
The questionnaire was scored in two ways: by designating each item as absent or present (0 or 1) according to the method of the general health questionnaire20
; and according to severity (range 0 to 3) (the Likert method). Those scoring 3 or more (out of 12) by the general health questionnaire method were classified as cases.20
Likert scores (range 0 to 36) more closely approximated a normal distribution and were used when the general health questionnaire score was treated as a continuous outcome.
To overcome likely colinearity between indices of standard of living, a poverty score comprising seven items was generated from variables previously judged to provide a comprehensive yet frugal assessment of each subject’s standard of living.11
One point was scored for each of the following: (a)
annual household income (adjusted for household size and composition19
) in the bottom fifth for region of residence (since the cost of living was expected to differ between regions); (b)
no household access to car or van, (c)
not saving from income (excluding money put by for bills but including life insurance, personal equity plans, share purchases, and holidays); (d)
fewer than four domestic household appliances from a list of nine; (e)
living in rented accommodation; (f)
overcrowded accommodation (more than two household members per bedroom); and, (g)
a home with two minor or any major structural problems such as dry rot. Where income sources could not be verified by documentary evidence missing data were imputed by the British household panel survey investigators19
using methods that minimised any tendency to overpredict associations with income.11,19
Items contributing to the poverty score were not weighted, given the absence of any rationale or method for doing so. Furthermore, cross sectional findings at T1 indicate that individual associations with the prevalence of common mental disorders differed little between items.11
Subjective financial strain at T1 was assessed by asking: “How well would you say you are managing financially these days?”, responses to which were coded as: (a) living comfortably or doing alright; (b) just about getting by; or, (c) finding it difficult or very difficult.
Potential confounding variables selected from the dataset of the British household panel survey were registrar general’s social class by head of household based on current or most recent occupation,11,19
plus marital status, education, employment, ethnic group, household size, responsibility for dependent children under the age of 16, number of current physical health problems, and region of residence.
Data were analysed in two ways. Firstly, the sample was stratified by case status at T1 and separate analyses were carried out for onset (proportion of non-cases at T1 who were cases at T2) and maintenance (proportion of cases at T1 who were also cases at T2) of common mental disorders. Secondly, to evaluate the effects of exposures at T1 on change in psychiatric morbidity between sets of interviews without imposing an arbitrary case threshold, the general health questionnaire score at T2 was treated as a continuous outcome and adjusted for the general health questionnaire score at T1.
Univariate differences were tested using non-parametric χ2
and Kruskal-Wallis tests as appropriate. Unadjusted and adjusted odds ratios and likelihood ratio χ2
tests to assess confounding, effect modification, and departure from linear trends were calculated by means of logistic regression using statistical software (Release 4.0, Stata, TX). Regression analyses were conducted using the Huber-White sandwich estimator to control for the clustering of respondents within households.22