shows the descriptive statistics for the sample. The bottom row tabulates the number of participants with a cumulative GHQ score included in the analyses at each phase (data on phase PT and phase PT+1. shows the probability of being a GHQ case at phase PT+1 according to cumulative GHQ score and number of measurements of cumulative GHQ score available at phase PT. The probability of future GHQ caseness increased in a dose-response manner with the number of times a participant had previously been a GHQ case. For example, there were 3028 participants who had 3 measurements of the cumulative GHQ available at phase 3 and had a cumulative GHQ score of 0 at that time. These participants had a 10% probability of being a GHQ case at the subsequent follow-up. In contrast, of the 466 participants with 3 measurements of the cumulative GHQ by phase 3 and a cumulative GHQ score of 3 at that point in time, 59% were GHQ cases at the next phase. The gradient was steeper with fewer measurement points, which is to be expected; the ratio of high GHQ score per measurement changes as the number of measurement points increases. For instance, a cumulative GHQ score of 3 with 3 measurement times would indicate GHQ caseness at each phase up to that point, while the cumulative score of 3 with 6 measurement times would indicate GHQ caseness only in half of the follow-up phases up to that point.
Figure 1 Probability (P) of being a GHQ case at Phase PT+1 according to cumulative GHQ score and number of measurements of the cumulative GHQ score available at phase PT. The error bars are 95% confidence intervals. The table below the figure shows the number (more ...)
Next we examined whether the risk markers assessed at phase PT predicted GHQ caseness at phase PT+1 (, Model 1). The probability of GHQ caseness was higher in women, low SES groups, in single individuals, in those of non-White ethnicity, those undertaking low levels of physical activity, the smokers and the obese. Model 2 of shows the effect of each risk marker among participants with a cumulative GHQ score of 0 (main effect) and the proportional change in this effect associated with a one unit increase in the cumulative GHQ score (interaction effect). The interaction effect indicates whether the association between the risk markers and future GHQ depended on the prior cumulative GHQ score. For example, the main effect of sex shows that among participants with a cumulative GHQ score of 0, women have a 38% increased odds of becoming a GHQ case at the next phase, compared with men. However, the significant interaction effect indicates that this increased odds decreases by a factor of 0.95 for each unit increase in the cumulative GHQ score. For participants with a cumulative GHQ score of 6, for example, the odds ratio for women compared to men of being a GHQ case at phase 8 is only 1.01 (=1.38 × (0.95)6).
Predicting the Probability of GHQ Caseness at Study Phase PT+1 by Sociodemographic and Behavior-Related Risk Markers Assessed at Phase PT.
There were statistically significant interactions for 5 of the 8 risk markers. The interaction effects of are illustrated in and which plot the estimated probabilities of being a GHQ case for each risk marker according to the cumulative GHQ score. Note that these lines have a slightly different shape than the lines in , because in and the predictions are calculated over all measurement times combined. Low SES, non-White ethnicity, being single, and alcohol abstinence predicted future GHQ caseness more strongly in individuals with a longer history of GHQ caseness than in individuals with no such history. The nature of these GHQ-history-dependent association was similar to the main effects presented in Model 1, with one exception: compared to being married, divorce was associated with a lower risk of future GHQ caseness among individuals with a high cumulative GHQ score although among individuals with a cumulative GHQ score of 0, divorce was associated with 16% higher risk of future GHQ caseness (Model 1, Table 3).
Figure 2 Model-predicted probability (P) of GHQ caseness at phase PT+1 as a function of sociodemographic risk markers and cumulative GHQ score at phase PT. The error bars are 95% confidence intervals. See the interaction effects of table 3 for statistical details. (more ...)
Figure 3 Model-predicted probability (P) of GHQ caseness at phase PT+1 as a function of health-related risk markers and cumulative GHQ score at phase PT. The error bars are 95% confidence intervals. See the interaction effects of table 3 for statistical details. (more ...)
The opposite pattern was observed for sex, as women had a higher risk of GHQ caseness than men among participants with no or only a short history of GHQ caseness but this sex difference was not observed among participants with a long history of GHQ caseness. The influence of physical activity, smoking, and obesity on the probability of being a GHQ case at the next phase was not modified by GHQ history. When all the statistically significant risk markers were mutually adjusted in a single multivariate model, all the interaction effects remained statistically significant with the exception of ethnicity (OR=1.09, CI=0.98-1.22; data not shown).
To examine potential bias due to selective attrition, we created a dichotomous variable indicating whether the participant had data at phase PT+1 (0=data, 1=no data). We then examined whether this probability was dependent on GHQ caseness at phase PT and cumulative GHQ score at phase PT in separate models, also including sex, age, and period. Increased probability of attrition was predicted by GHQ caseness (OR=1.14, CI=1.06-1.22, p<0.001) but not by cumulative GHQ score (OR=1.01, CI=0.98-1.05, p=0.51), indicating that GHQ cases were more likely than non-cases to be lost during the follow-up. This might have led to some underestimation of the cumulative effect of prior GHQ caseness on future risk, because the number of participants with high GHQ scores was lower than it would have been without selective attrition. When the interaction effects shown in were fitted in a subsample with complete data at all of the study phases (n=4496 participants, 26976 observations), the results closely resembled the main results, with the exception of a non-significant interaction effect for ethnicity in the complete-case analysis (OR=1.09, CI=0.97-1.22; data not shown), suggesting that the interaction effects were not substantially affected by attrition bias.