The study protocol was approved by the Regional Ethical Review Board in Uppsala, Sweden. All participants provided written informed consent. Participants between 16 and 18 years of age needed written consent from their caregivers. The study was conducted in accordance with the principles expressed in the Declaration of Helsinki.

This study was part of a follow-up of Swedish tourists afflicted by the Indian Ocean tsunami on December 26, 2004, and previous findings from the data herein have assessed the impact on mental health with regard to exposure severity [1] and social support [10]. The Swedish authorities registered 13,638 adult residents repatriated from Southeast Asia between December 27, 2004 and January 15, 2005. The 21 counties in Sweden were approached to allow for the survey to be sent to residents in their respective county. Residents in the 10 counties that agreed to participate were identified and the questionnaire was distributed in March 2006. The survey was based on a Norwegian follow-up of tsunami survivors [32] and adapted to Swedish in collaboration with the Department of Medical Epidemiology and Biostatistics and with the Center for Family and Community Medicine, Karolinska Institutet, Sweden. Only variables of relevance to the current study are reported.

The questionnaire was sent to 10,501 individuals ≥16 years of age (77% of those registered by the authorities upon repatriation). Of 4,932 returned questionnaires there were 4,910 that could be retained for analysis. There were 385 individuals who rejected participation per e-mail or telephone, of whom 59% stated not being in an afflicted region as their reason, and 1% stated that they could not cope with filling out the questionnaire. Nonresponse was weakly associated with younger age and living in a large city (data not shown).

In order to select a sample that most likely had experienced a potentially traumatic stressor, respondents were included in this study if they had been exposed to any of a number of exposure criteria: The direct exposure group were those who experienced threat to life, and/or were caught by the waves, and/or were severely injured in the disaster; indirect exposure were those who had not experienced the above, but experienced death of, or life threat to close ones; and vicarious exposure were those who had not been exposed to any of the above but witnessed horrifying events (i.e., deceased people, survivors with severe injuries, people searching for others among corpses).

The Crisis Support Scale (CSS) [33] was used to assess levels of current support in relation to the event from friends and relatives. The CSS is composed of seven items with a seven-point scale ranging from never (1) to always (7), with the items being 1. availability of others; 2. contact with others with similar experiences; 3. confiding in others; 4. emotional support; 5. tangible support; 6. negative response; and 7. overall satisfaction with support. Item 7 has previously been analysed separately from the summated score of the other items, with item negative response reversed [33]. Also, a composite score for item 1–4 has been used, with negative response analysed separately [23]. A factor analysis found that the CSS contained two factors where items 1, 3, and 4 loaded on one factor, negative response was a single item on the second factor, and contact with others loaded on both factors [21]. In this study, the Cronbach’s alpha for items 1–6 was.73. Items contact with others and negative response had low item-total correlations, r = .35 and.11, respectively. Omitting these two items increased the internal consistency for the four remaining items to Cronbach’s alpha = .83. Thus, the mean score of items 1, 3, and 4 was analysed as a composite measure of perceived emotional support, and the remaining items were analysed as contact with others, negative support, and overall satisfaction with support.

Cohabitation was coded as either living with spouse/partner or being single/widow/widower. Educational status was coded as university degree or not. Age was grouped into adolescents (16–19 years), young adults (20–29 years), middle-aged (30–49 years; reference group), aged (50–65 years), and retired (≥65 years).

Ordinal regressions [34] were used to estimate the associations between demographic and exposure characteristics and social support with a separate regression for each aspect of social support. The ordinal odds ratio summarises the average change in odds for moving across any binary social support threshold, and is valid over all thresholds. The assumption that odds ratios are homogenous across thresholds was satisfied by collapsing the two lowest scoring options (after reversal of negative support) into one category. In addition, an assessment of binary logistic regression odds ratios for each threshold of each dependent variable revealed small differences in direction and on average, which indicated homogenous odds ratios across thresholds. To assess joint effects of predictor variables, first-order interaction terms were added to the regression equations and subsequently removed if p>.05.

A sensitivity analysis was performed by duplicating the regression analyses without interaction terms on a new dataset including the original data and imputed cases. The predictors and outcomes for the set of imputed cases were given frequency distributions that reflected the distributions in the respondents. The bivariate associations between predictors and outcomes were set to be unrelated (ρ = 0). The odds ratios with corresponding 95% confidence intervals (CIs) are reported, together with Likelihood ratio R² (R_L²) and predictive efficiency (τ_p). The τ_p indicates the reduction in discrete classification error [35]. Analyses were performed with PASW Statistics 18 (SPSS Inc., Chicago, IL).

Of the 4,910 respondents there were 3,536 (72%) who fulfilled the exposure criteria. There were missing data for social support, from 1% for emotional support to 4% for tangible support. An age of ≥65 years was associated with higher probability of missing values (Table 1). Forty participants (1%) had missing values for all social support items and were removed from further analyses.

For demographic characteristics of the sample, see Table 1. The average age was 42 years (SD = 14). The majority of participants (95%) had been raised in Sweden, 2% in the Nordic countries, and another 2% in the remaining European countries. There were 42% who had experienced life threat to themselves or to close relatives, 28% had been swept away by the waves, and 11% was bereaved of close friends or relatives in the disaster. The items were skewed except for contact with other survivors (Figure 1). The sum of items 1–6 with the item negative response reversed was calculated for comparison purposes; the CSS total had a mean of 30.7 (SD = 7.56) and a range of 6 to 42.

Each social support item was the outcome variable in five separate ordinal regression models. The model fit was statistically significant for all models, indicating better fit than the intercept only model, whereas the coefficients of determination showed a very modest reduction in classification error (Table 2). The models were reanalysed adjusting for additional stressful life events after the tsunami, but this adjustment produced negligible differences in model parameters (data not shown).

The differences in social support among groups were small (Table 2). Men were less likely than women to perceive both positive and negative support. Cohabiting participants were more likely to experience social support than those in a single-household. Having a university degree was associated with more positive social support, but not with lesser negative support. For age, no overall pattern emerged, although participants who were ≥50 years of age were less likely to perceive negative support than participants in the reference group. Disaster exposure was associated with lower overall support and more negative support.

First-order interaction terms were then examined. Interactions terms with p<.05 were found only in the regressions on emotional and negative support. For emotional support, only the interaction Cohabitation × Gender was statistically significant, Wald χ² = 15.5, p<.001, suggesting that living in a single-household was associated with less emotional support for men but not for women (Figure 2). With the interaction included, the parameters and fit indicia were minimally altered (data not shown) whereas the main effect of gender was augmented, OR = 0.47, 95% CI [0.36, 0.62], p<.001, and the main effect of cohabitation disappeared, OR = 0.98 [0.80, 1.19], p = .83. For negative support, the regression coefficients for Age × Gender indicated that younger age was associated with more negative support, Wald χ² = 6.86, p = .009. The main effects of age and gender were slightly attenuated by inclusion of the interaction term and the fit indicia were slightly improved (data not shown). The Age × Gender interaction is illustrated in Figure 3.

Based on a study of the nonrespondents in highly similar survey of tsunami registrants in a neighboring Scandinavian country [36] we assumed that 50% of the nonrespondents were eligible for this study (i.e., fulfilled the exposure criteria). Hence, we imputed a set of 2,475 cases for which the predictors were unrelated to the outcomes. Regression analyses with the original sample and the imputed nonrespondents combined showed that statistically significant odds ratios with a 95% CI within the range of 0.96–1.04 in the original analyses now became nonsignificant (i.e., indirect vs. direct exposure for emotional support; men vs. women, and 50–64 vs. 30–49 years for tangible support; partner vs. single-household for contact with others and for satisfaction with support).

Posttraumatic stress during the first year after the disaster could have influenced the impact of gender and disaster exposure on social support. Previous findings suggest, however, that social support influences distress up to 18 months after disaster, whereas the opposite is true 18–24 months post disaster [18]. Therefore, we chose to not control for current posttraumatic stress since it may be caused by both the predictor and outcome variables, which would increase risk for bias [42].

Moreover, the present study fails to provide conclusions as to whether groups of disaster survivors might differ in social support during the first time after the event. Governmental bodies in Sweden arranged for various psychosocial interventions during the first six months after the tsunami. The findings herein would thus reflect to some extent the effects of these interventions. However, a long-term follow-up of Vietnam veterans demonstrated that current, but not homecoming, emotional sustenance predicts both the development and the maintenance of PTSD [43]. Hence, the current findings may be relevant with respect to the maintenance of negative psychological consequences after disasters.

The sum score of CSS found in this study is comparable with findings within the same time after the event [44], [45]. Assessments of social support with single items do not necessarily result in a decreased precision in measurement [23], and analyses of aggregate measures that contain items reflecting diverse factors are not recommended [46]. Nevertheless, the measurement of social support by few or single items yields limited variation in social support that may have attenuated the differences in social support. In addition, the suboptimal distributions among items (see Figure 1) indicate that the CSS is in need of further refinement. Further studies with appropriate measures of social support may elucidate greater variation among groups of survivors.

The proportion of missing data for participants who were ≥65 years old was substantial, which raises concerns with regard to the representation particularly in this age group. Moreover, there was a low response rate. Based on other studies [36] and on the reasons given herein for not participating we assumed a 50% oversampling in the nonrespondents. The sensitivity analysis indicated that the associations found herein generally were robust even if predictors and support variables were unrelated in the group of nonrespondents. Although any sensitivity analysis of nonresponse can only assume the actual values in nonexistent data, assuming a zero association between predictors and outcomes in nonrespondents provides a fairly conservative but straightforward assessment of nonresponse bias. Based on this analysis it seems reasonable to conclude that any bias induced by nonresponse could attenuate the associations between predictors and outcomes, but would need to be very large to alter the direction of these associations.

The association between demographic characteristics and social support in disaster survivors are thus far equivocal. As of yet, there are too few studies that focus on social support as an outcome variable to pinpoint the origins of differences among studies. They may reflect variations in the disaster context, such as in the disaster characteristics, the impact of the disaster on the social support systems, as well as temporal fluctuations in social support. In this study, educational status, gender, and cohabitation were associated with several dimensions of social support and could serve as proxies for targeting groups that may benefit more from promotion of social support by psychosocial interventions after disasters. No firm conclusions can be drawn from this study alone;, however, and the low predictive efficiency raises concerns about whether demographic and exposure characteristics reliably can predict support in future events. Nonetheless, the present study provides a rare attempt of examining predictors for socials support and adds to the literature a starting point from which further studies can establish how to best inform more tailored approaches to psychosocial interventions.