Sample
We conducted a secondary analysis with three a priori questions within an ongoing study among twins (affected twin pairs, n = 51; healthy control twin pairs, n = 35) with bipolar disorder at the University Medical Center Utrecht (UMCU), The Netherlands. Of this cohort, all 51 twins with bipolar disorder (bipolar I disorder, n = 37; bipolar II disorder, n = 14) were included in the current study. The design of the study and the recruitment of the bipolar twin pairs have been described in detail elsewhere (Van der Schot et al. 2009; Vonk et al. 2007). All participants were enrolled between 2001 and 2006. There were no restrictions on duration or stage of illness for inclusion in the study, and all patients were treated naturalistically.
Demographic information is displayed in Table 1. All diagnoses were confirmed with the Structured Clinical Interview for DSM-IV (First et al. 1996) and the Structured Interview for DSM-IV Personality (Pfohl et al. 1997). Hospitalizations were confirmed through available medical records. Current mood state was assessed using the Young Mania Rating Scale (YMRS; (Young et al. 1978)) and the Inventory for Depressive Symptomatology (IDS; (Rush et al. 1996)). At the time of the study, all patients were euthymic with a YMRS score of 4 or less and an IDS score of 12 or less.
The study was approved by the medical ethics review board of the University Medical Center Utrecht, and all participants gave written informed consent after full explanation of the study aims and procedures.
Life event measures
All subjects included in the current study were interviewed with the investigator-based Bedford College LEDS (Brown and Harris 1978, 1989). The LEDS is a semi-structured interview for assessing life events and long-term difficulties in adults. It collects detailed information about the event itself, the timing of its occurrence (date) and relevant contextual information for each event. Each event is categorized into one of ten domains, consisting of education, work, reproduction, housing, money/possessions, crime/legal, health, marital/partner, other relationships and miscellaneous/death. Based on the contextual information, the threat for each event is rated via standardized rating procedures. The threat score represents the severity of the event, ranging from mild (1) to severe (4), hereby differentiating between mild life events and more stressful life events. The contextual threat is conceptualized as: ‘What most people would be expected to feel about an event in a particular set of circumstances and biography, taking no account of what the respondent says either about his or her reaction or about any psychiatric or physical symptoms that followed it’ (Brown and Harris 1989). Several studies have supported the reliability (e.g. interrater) and validity (e.g. multiple informant) of the LEDS in adults exhibiting a variety of psychiatric symptoms (Brown and Harris 1978, 1989; Ormel et al. 2001).
Only events occurring from the age of 5 years were included. All severe events were defined by the extent they were related to the bipolar disorder and to what extent they were dependent on the respondents’ own behaviour. To determine relatedness to the disorder, each severe event was rated on a three-point scale: 1) not related to psychopathology; 2) possibly related to psychopathology; or 3) clearly related to psychopathology. Only events with score 1 were included for further analyses. To determine if life events occurred independent of will or influence of the respondents’ own behaviour, each severe event was rated on a seven-point scale: 1) completely independent; 2) nearly independent; 3) possible influence, however, very unlikely; 4) physical illness; 5) cooperation or agreement with external situation; 6) likely neglect or carelessness; and 7) intentional choice. Events rating 1 to 5 were included for further analyses. Each life event was dated per year. Age was then calculated for each event.
All interviewers and raters were trained by MH, who was trained by G.W. Brown and T.O. Harris, who developed the LEDS. The interviews were conducted at the participant’s home or at the UMCU. Events were rated by two independent raters who had not been involved in the interviews. A panel consisting of the four raters (including SK and MH) reached consensus on the events that raised rating problems.
Statistical analysis
Life event load
Life event load represents the sum of the threat scores of the life events occurring in each year.
We calculated three different life event load measures: (1) cumulative load (CL), i.e. the life event load at a particular point in time (year Y) calculated as the sum of the life event load in year Y and all preceding years; (2) cumulative load excluding events possibly or clearly related to the bipolar disorder (CL-NoBP); and (3) cumulative load including only independent events, thus excluding events possibly or clearly dependent on the respondents’ own behaviour (CL-I).
Next, the life event load before the first or since the last admission was calculated. After each admission, life event load was reset to zero and was calculated as described above. The cumulative life event load in the year preceding the admission was used for analysis.
Decay model
Previous studies showed a decay effect, implying that the presumed effect of life events diminishes over time, e.g. the death of a close relative that occurred 3 or 4 years before admission has less impact compared to the same event 1 year before admission (Hillegers et al. 2004). We will investigate which decay model statistically fits the data best. To explore the degree to which the effect of life events diminishes over time, a time-specific life event load variable was calculated for every year and subjected to an exponential decay function. We tested four models; in model I, we tested the purely cumulative effect, and in models II to IV, the decay function implied a 25%, 50% and 75% loss of effect per year, respectively. The decay function yielding the best model fit (−2× log-likelihood) will be used for all further analysis.
Andersen-Gill model
The Andersen-Gill model (A-G model), an extension of the standard Cox proportional hazard model for recurrent events, accommodates censored data and time-dependent covariates (Fleming and Harrington 1991; Therneau and Grambsch 2000).
Data for the A-G model are structured such that for each individual, intervals at risk are defined by variables describing the start and end times of each year of age. An event variable is coded as ‘1’ for admission and ‘0’ for no admission.
The A-G approach follows the usual assumption of the Cox model that the hazard or risk ratio is proportional over time and more specifically that the risk of being admitted is unaffected by earlier admissions. Time-dependent covariates, such as the cumulative load of life events or the number of previous admissions, may be used to relax the latter assumption. The hazard ratio represents the proportionate change in the ‘admission’ rate due to a unit change in the respective covariate, in this case the cumulative life event load.
Andersen-Gill model: interaction effect
The presence of an interaction effect will be tested by integrating an interaction function in the A-G model, testing the effect of the interaction between the number of admissions and the cumulative load between the admissions in the best-fitted decay model, also known as a kindling effect (Post 1992).