Of the 104 samples, 27 included patients with mixed cancer sites, 24 focused on breast cancer, 10 on leukemia and lymphomas, 9 on lung cancer, and 34 samples on other cancer sites (e.g., colon, pancreas). The majority combined patients with early and late stages of the disease (N = 93); three focused on patients with early-stage cancer (I/II) and eight on late-stage disease (III/IV). Most (N = 85) assessed support, network size and/or marital status after cancer diagnosis; 19 examined the influence of these variables prior to cancer diagnosis. The latter were community-based cohort studies that assessed risk for different sources of mortality. For the present analyses, only data on cancer mortality were used. Forty studies reported only bivariate associations between social network and survival, 27 only multivariate associations that controlled for some confounders, and 19 bivariate as well as multivariate associations. Studies with multivariate analyses controlled for (some) confounding variables, such as age (k = 35 studies), cancer site (k = 35), gender (k = 34), stage (k = 31), SES (k = 11), medical comorbidities (k = 8), functional status (k = 7), and alcohol/tobacco use (k = 7). Thus, the controlled studies varied regarding which confounding variables were controlled. The included studies are identified in the Reference section.
A large number of different measures (N
= 22) were used for assessing perceived social support, such as the MOS Social Support Survey ([123
= 2). Social network indicators were assessed with single-item indicators (e.g., contact frequency, number of confidants; k
= 9) and multi-item scales (k
= 6), such as the Social Network Inventory [124
]). The marital status of the respondents was assessed with single-item indicators (k
The participants had a mean age of 65.9 years (SD = 6.2), 57% were women, and 14.3% were members of ethnic minorities. About 86% of the respondents were married. The average study interval between assessment of social network and of survival was 7.1 years (SD = 5.2). Intervals were longer in community-based studies (M = 12.2 years, SD = 5.5) than in clinical studies with cancer samples (M = 5.9 years, SD = 4.3; t(103) = 5.71, p < .001). During this interval, about 40.2% of the participants were deceased.
We used funnel plots to check for publication bias. With regard to associations of network size with mortality, we found that studies with below-average risk ratios were more likely to have sample sizes of 2000–4000 whereas those with above-average risk ratio were more likely to have sample sizes of 5000–6000 patients. In studies on the association of marital status with mortality, we found more studies with below-average risk ratio with 5000–7000 patients and more studies with above-average risk ratio with ≥10,000 patients. As there was no evidence for publication bias (i.e., large and significant effect sizes were not overrepresented among studies with small sample sizes), it is unlikely that small studies with nonsignificant results would be less likely to be published.
Because community-based studies on mortality usually start with people having no life-threatening disease, initially measured network variables may affect disease onset and disease whereas studies with cancer patients analyze effects of network variables on disease progression. Thus, we first checked whether the association of network variables with mortality differed between these groups of studies. Because no differences were found for social support (RR = .86, CI: .63–1.17 versus RR = .79, CI: .71–.89), network size (RR = .81, CI: .68–.95 versus RR = .78, CI: .68–.90), and marital status (RR = .87, CI: .84–.89 versus RR = .87, CI: .82–.91), both groups of studies were collapsed for the following analyses.
With regard to the first research question, we analyzed whether perceived social support, network size and marital status would be associated with cancer mortality. Separate effect sizes were computed for uncontrolled studies and for studies that controlled for one or more confounding variables (age, gender, SES, health status/cancer stage, alcohol/tobacco use). In line with our expectations, both groups of studies showed lower mortality in individuals with higher levels of perceived social support, larger social networks, and in married as compared to nonmarried respondents (). For example, the controlled risk ratio of .75 of the association between perceived social support and survival indicates that the relative risk of mortality was reduced by 25% when the level of perceived support increases by one standard deviation unit. Similarly, the relative risk for mortality was reduced in controlled studies by 20% when the size of the network increases by a standard deviation. In controlled studies, the relative risk for mortality of married respondents was 12% lower than the relative risk for mortality in unmarried persons. As indicated by overlap of the 95% confidence intervals, the effect sizes of the association of perceived social support, network size, and marital status with mortality did not differ significantly.
Association of perceived social support, network size, and being married with cancer survival.
As shown by the overlap of the 95% CI, effect sizes of uncontrolled studies and controlled studies did also not differ significantly (). Thus, we computed average effect sizes across controlled and uncontrolled studies that replicated the results of the analysis of uncontrolled and controlled effects.
We next checked whether associations of marital status with mortality would differ between never married, divorced/separated, and widowed individuals. Risk rations were computed that compare the relative risk for mortality in these groups against the relative risk in married individuals. As shown in , never married, divorced/separated, and widowed individuals had higher mortality rates than married individuals. As indicated by the non-overlap of the CI, the survival disadvantage of never married persons in controlled studies was significantly larger than the survival disadvantage of divorced/separated and widowed individuals. Similarly, when combining controlled and uncontrolled studies, never married respondents had a larger survival disadvantage than divorced/separated and widowed individuals.
Cancer mortality in never-married, divorced/separated, and widowed persons.
As shown by the Q-statistics, the size of the associations of perceived social support, network size, and marital status with mortality varied between studies, underscoring the need to identify moderators (study characteristics that moderate the size of the observed association of network variables with mortality). Thus, weighted multiple linear regression analyses were computed for analyzing the effects of moderating variables.
As cancer mortality varies by age, gender, SES, health status/cancer stage, and alcohol/tobacco use [34
], we checked whether studies controlled for all of these variables. Unfortunately, only one of the studies did. Because we could not compare high-quality studies that controlled for all relevant variables with other studies, we built a count variable that sums up the number of controlled variables. With weighted multiple regression analysis we tested whether the association of the network variables with mortality varied by the number of control variables. Interestingly, the association of network with mortality did not vary by the number of controlled variables (social support: B
= .08, β
= .23, Z
= 1.11; network size: B
= −.00, β
= −.01, Z
= −.04; marital status: B
= −.02, β
= −.11, Z
In the next step, we tested whether the association between social network and cancer survival observed in multivariate analyses would vary by the inclusion of individual control variables. As shown in , the association of social support with cancer mortality was weaker in studies that controlled for gender and SES. In addition, the association of social network with cancer mortality became stronger in studies that controlled for alcohol/tobacco use and weaker in studies that controlled for gender, functional status, and comorbidities (potential confounders). Finally, the strength of the association of martial status and mortality became stronger in studies that controlled for gender, cancer site, and comorbidity.
Moderating effects of control for confounding variables on the multivariate association of social network with cancer mortality.
In the second set of regression analyses, we tested whether the association of social network with mortality would vary by other study characteristics, such as age, gender, cancer site, and stage. Because eight studies were available on late stage of cancer, as compared to three studies on early stage, we included a dummy variable that compares studies with late-stage cancer to other studies (mixed stage, early stage). In addition, a sufficient number of studies for inclusion in the meta-analysis were available for breast cancer and leukaemia/lymphoma but not for other sites. Due to the large number of methods used for the assessment of perceived social support and network size, we could not include individual measures as predictor variables.
As shown in , we found a stronger association of perceived support and survival in studies on patients with leukaemia and lymphomas than in studies with other cancer sites. In addition, associations of network size and cancer mortality were stronger in studies with older patients and in studies with breast cancer patients. Finally, weaker associations of being married with mortality were observed in studies with late-stage cancer patients as compared to studies on early and mixed stages, and in studies with longer intervals between baseline and follow-up.
Analysis for moderating effects on the association of perceived social support, network size, and being married with cancer mortality.