Methods

Data sources
Analytical techniques

Data sources

The analysis of associations between the dissolution of a marriage or common-law relationship and a new episode of depression over a two-year period among those who were free of depression at baseline is based on data from the National Population Health Survey (NPHS).  The NPHS, which began in 1994/1995, collects information about the health of Canadians every two years.  It covers household and institutional residents in all provinces, except persons living on Indian reserves, on Canadian Forces bases, and in some remote areas.

In 1994/1995, 20,095 respondents were selected for the longitudinal panel.  The response rate for this panel was 86% or 17,276 respondents.  Attempts were made to re-interview them every two years.  The response rates for subsequent cycles, based on these 17,276 individuals, are: 92.8% for cycle 2 (1996/1997); 88.3% for cycle 3 (1998/1999); 84.8% for cycle 4 (2000/2001); 80.5% for cycle 5 (2002/2003); and 77.4% for cycle 6 (2004/2005).  This analysis uses the cycle 6 longitudinal “square” file, which contains records for all responding members of the original panel whether or not information about them was obtained in all subsequent cycles.

More detailed descriptions of the NPHS design, sample and interview procedures can be found in previously published reports.^20-22

Top of Page

Analytical techniques

NPHS respondents who met the following criteria were used for this analysis: aged 20 to 64 at baseline interview; living with a partner or living common-law or married at baseline; living in a private residence (baseline and follow-up); provided complete data on the depression modules (baseline and follow-up); and were not classified as having depression (baseline).

The analysis of the association between marital dissolution and depression was based on data from cycles 1 to 6 (1994/1995 to 2004/2005) of the NPHS.  For this analysis, “pooling of repeated observations,” combined with logistic regression, was used.  Pooling of repeated observations results in increased cell sizes for respondents who have experienced marital or cohabitating union dissolution, and thereby reduces the probability that a lack of statistical power is responsible for non-statistically significant associations.²³  Use of the design-based bootstrapping technique for repeated observations ensured that the variance was not underestimated by eliminating the problem of dependence among observations derived from the same individual.^24,25 

The analysis used five cohorts of pooled observations.  Individual respondents for whom the requisite data were available were considered at baseline and follow-up in each two-year interval:  1994/1995 to 1996/1997 (cycle 1 to 2); 1996/1997 to 1998/1999 (cycle 2 to 3); 1998/1999 to 2000/2001 (cycle 3 to 4); 2000/2001 to 2002/2003 (cycle 4 to 5); 2002/2003 to 2004/05 (cycle 5 to 6).  The first cycle in each of the two-cycle intervals served as the baseline, and the next cycle, the  follow-up.  For each baseline year, all respondents who were married or cohabiting and who had not had a major depressive episode in the previous 12 months were selected.  They were considered to have experienced a marital breakdown if, in the follow-up interview two years later, they reported that they were divorced, separated or single.

 
Marital status, depression and most control variables were assessed at both cycles.  At the end of each two-year interval, marital status was assigned one of two values:  remained married or became divorced/separated/single.  Depression was dichotomously categorized as not depressed or depressed at follow-up (see Definitions).  Each eligible respondent could contribute as many as five records.  For this analysis, 7,614 respondents contributed 25,329 records; 1,037 records were excluded because of depression at baseline.

The variables entered into the multivariate model, which were selected based on the literature and availability in the NPHS, were change in household income, change in social support, change in number and/or presence of children in the household, change in employment status, history of depression, education, and age. 

Preliminary analysis revealed that some characteristics of respondents who were excluded because of depression before the baseline interview differed from those of respondents who were retained in the analysis (Appendix A, Table A).  For example, respondents who were excluded were slightly younger, more likely to be female, less educated, and generally less likely to be employed at baseline and follow-up.  These exclusions likely weakened the observed association between marital dissolution and depression.

Weighted cross-tabulations were used to examine the association of marital dissolution with the selected control variables:  change in household income, social support, presence and number of children, and employment status.  Decisions to collapse certain categories of control variables were guided by the distribution of responses and by sample sizes. 

The relationship between marital dissolution and two-year incident depression was examined using unadjusted and adjusted logistic regression.  Unadjusted odds ratios were estimated for marital dissolution in relation to depression.  Because previous research has suggested that the consequences of marital dissolution may differ between men and women,^{6,9,11,14,26-28} preliminary logistic regression models were run for depression to test for interaction effects between marital status and sex.  Most of the previous studies compared results from separate sex-specific models but failed to use interaction analysis to assess the observed differences between the sexes.²³  Following interaction testing, unadjusted and adjusted odds ratios were calculated for each sex (Table 1).  Before exclusions, the data were weighted to represent the target population in 1994/1995.  Coefficients of variation on estimates and confidence intervals on odds ratios were calculated using the bootstrap technique, which accounts for survey design effects and dependence between observations from the same respondent.^24,32,33  Results at the p < 0.05 level were considered significant.