The association between stressful life events and breast cancer risk: A meta-analysis
Version of Record online: 18 SEP 2003
Copyright © 2003 Wiley-Liss, Inc.
International Journal of Cancer
Volume 107, Issue 6, pages 1023–1029, 20 December 2003
How to Cite
Duijts, S. F.A., Zeegers, M. P.A. and Borne, B. V. (2003), The association between stressful life events and breast cancer risk: A meta-analysis. Int. J. Cancer, 107: 1023–1029. doi: 10.1002/ijc.11504
- Issue online: 27 OCT 2003
- Version of Record online: 18 SEP 2003
- Manuscript Accepted: 23 JUL 2003
- Manuscript Revised: 18 JUL 2003
- Manuscript Received: 7 JAN 2003
- breast cancer;
- stressful life events;
Breast cancer is the most prevalent cancer in women in Western societies. Studies examining the relationship between stressful life events and breast cancer risk have produced conflicting results. The purpose of this meta-analysis was to identify studies on this relationship, between 1966 and December 2002, to summarize and quantify the association and to explain the inconsistency in previous results. Summary odds ratios and standard errors were calculated, using random effect meta-regression analyses, for the following categories: stressful life events, death of spouse, death of relative or friend, personal health difficulties, nonpersonal health difficulties, change in marital status, change in financial status and change in environmental status. The presence of publication bias has been explored, and sensitivity analyses were performed to identify heterogeneity, using calculation of the percentage of variability due to heterogeneity, meta-regression analyses and stratification. Only the categories stressful life events (OR = 1.77, 95% CI 1.31–2.40), death of spouse (OR = 1.37, 95% CI 1.10–1.71) and death of relative or friend (OR = 1.35, 95% CI 1.09–1.68) showed a statistically significant effect. Publication bias was identified in both stressful life events (p = 0.00) and death of relative or friend (p = 0.02). Sensitivity analyses resulted in the identification of heterogeneity in all categories, except death of spouse. The results of this meta-analysis do not support an overall association between stressful life events and breast cancer risk. Only a modest association could be identified between death of spouse and breast cancer risk. © 2003 Wiley-Liss, Inc.
Cancer risk might be related to stressful life events or emotional factors. Of the cancer sites that have been studied from this perspective, breast cancer—the most prevalent cancer in women in Western societies1—has received a great deal of attention.2 Approximately 50% of all breast cancer incidences can be attributed to the most important risk factors, i.e., age, family history of breast cancer, parity, early menarche, late menopause and the use of oral contraceptives.2
Many individual studies have investigated the association between stressful life events and breast cancer risk. However, these studies have produced conflicting results, varying from no association, e.g., OR = 0.95,24 up to a strong association, e.g., OR = 11.6.29 This inconsistency in results may be explained by the methodologic diversity of these observational studies, in which design, adjustment for confounding, population characteristics and the effect measures that were used are more variable than in randomised trials.3 The reviews that have been published on the relationship,2, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13 not surprisingly, reported the same variability. They tend to fall into 2 categories: (i) those that concluded there is no association because of methodologic differences2, 5, 7, 11, 12, 13 and (ii) those that weakly supported an association.4, 6, 8, 9, 10 Clinical and methodologic diversity, among studies included in reviews and meta-analyses, necessarily leads to statistical heterogeneity.3 To our knowledge, no meta-analysis thoroughly investigated sources of heterogeneity, using various methods, to explain the inconsistency and variability in previous studies.
The purpose of this meta-analysis was to identify studies between 1966 and December 2002 that examined the association between stressful life events and breast cancer risk, to summarize and quantify this association for various categories of stressful life events (i.e., stressful life events, death of spouse, death of relative or friend, personal health difficulties, nonpersonal health difficulties, change in marital status, change in financial status and change in environmental status) and to explain the inconsistency in previous results.
MATERIAL AND METHODS
Publications were identified by the first author (SFAD) through computerized Medline, Psychinfo, Cinahl and Cancerlit searches for studies published until December 2002 with no language restrictions. The key words used were a combination of stress, life events, breast, mamma, cancer, carcinoma and neoplasms. References cited in published original and review articles were examined. For inclusion in this analysis, the articles had to describe a retrospective case-control study, a prospective case-control study, a limited prospective cohort study or a prospective cohort study. Furthermore, the articles had to describe the association between stressful life events and breast cancer risk. Articles describing the association between personality traits and breast cancer risk, or the association between stressful life events and breast cancer survival, were excluded.
In compiling the database, a distinction was drawn between an article and a study. A study comprised all the analyses of a given group of research subjects. These analyses may have been described in more than one article. When the same study population was referred to in more than one article, they were considered as part of a single study. When different results pertaining to the research subjects of a single study were published in more than one article, all such studies were included, but the data were combined to reflect the fact that only one sample of subjects was involved.
Qualitative data extraction
Three reviewers extracted qualitative information from each paper. For 2 reviewers, who co-authored the present article, the original articles were blinded for authors, affiliations, journal name, publication year and acknowledgements. Differences in assessment between the blinded reviewers and the not-blinded reviewer did not appear. By means of a questionnaire, the reviewers independently assessed studies for methodologic heterogeneity on the following qualitative items: design, confounding, population characteristics and exposition. When disagreement existed, it was discussed until consensus was reached.
While defining the design, a distinction was made between case-control studies, i.e., subjects were selected according to their disease status and controls were sampled from the entire source population, and cohort studies, i.e., subjects were classified according to their exposure status and followed over time to ascertain disease incidence.14 Two different designs were defined within each category, i.e., retrospective and prospective case-control designs, and limited prospective and prospective cohort designs. A retrospective case-control design uses exposure measurements taken after disease, whereas a prospective case-control study uses measurements taken before disease.14 In a limited prospective cohort study, the number of subjects required for the analysis was limited by selecting those known to be at risk, such as women undergoing biopsy. Psychological factors in both cases and controls were evaluated prior to confirmation of benign, malignant or no breast disease under similar conditions. We stated the maximum time period between exposure and diagnosis at 6 months.4 A prospective cohort design does not have these limitations.
Potential confounders in the form of well-established risk factors ideally need to be considered in design and/or analyses. These include age, family history of breast cancer, parity, early menarche, late menopause, the use of oral contraceptives and other, less obvious, risk factors as alcohol consumption and BMI.
Within the population characteristics, a distinction was made between sources from which cases and controls were selected, i.e., hospital, population, suspicion-cohort and screening-cohort. Suspicion-cohort and screening-cohort included both patients at risk and were described as a symptomatic population, e.g., patients who experienced a breast lump, and an asymptomatic population, e.g., patients who received routine breast examination, respectively. The choice of a control group was also considered to be important; for example, subjects with benign disease share the same apprehensiveness prior to diagnosis as those found to have cancer. They are also at greater ultimate risk of developing breast cancer.7
Possible instruments for assessing stressful life events were divided in 2 groups: interview and questionnaire.
Different categories of qualitative items were included in our sensitivity analyses to identify heterogeneity between studies.
Quantitative data extraction
The first author (SFAD) extracted quantitative data, allowing her to calculate within-study odds ratios and corresponding standard errors to estimate the association between stressful life events, death of spouse, death of relative or friend, personal health difficulties, nonpersonal health difficulties, change in marital status, change in financial status, change in environmental status and breast cancer risk. The selection of these specific categories was based on the most prevalent stressful life event categories in questionnaires used in the individual studies.
When the odds ratios in a study were stratified for different experiences, e.g., end of a relationship, got married and got divorced, a within-study odds ratio for the specific category was calculated, e.g., change in marital status, using fixed-effect pooling across strata. Unfortunately, including each experience separately in the meta-analysis was not possible due to lack of data. Preferably adjusted odds ratios were extracted. When adjusted odds ratios could not be calculated, 2-way contingency tables were constructed to calculate the unadjusted odds ratios and corresponding standard errors.15
To detect publication or related biases, we explored heterogeneity in funnel plots, i.e., plots of effect estimates against their estimated precisions (reciprocal of the variance). We examined funnel plot asymmetry visually and measured the degree of asymmetry by using Egger's unweighted regression asymmetry test.16 Identification of publication bias implies the absence of small studies, with rather minor results. Because of potential heterogeneity between studies, we estimated the summary odds ratios (SOR) and corresponding 95% confidence intervals (CI) with random effect pooling, using the Stata statistical software,17 for each category. The between-study variance was estimated iteratively, using the empirical Bayes method. To explore reasons for heterogeneity, we used 3 different methods: (i) we calculated the statistic I2, i.e., the percentage of variability in point estimates that is due to heterogeneity,18 for every category of stressful life events and across specific strata based on different levels of the extracted study characteristics. Heterogeneity exists when I2 within unstratified analyses is higher than within stratified analyses; (ii) we performed meta-regression analyses, using the empirical Bayes method, by including these study characteristics as covariates in the regression model (statistical significance indicates an effect modifier); and (iii) stratification. The specific odds ratios and corresponding 95% confidence intervals (CI) illustrate statistical significance.
The search strategy revealed 29 articles.19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47 Three articles were excluded because no sufficient data could be extracted.19, 20, 31 Another 3 articles were combined in the analysis because the same study was published 3 times.21, 22, 23 Within the 24 articles22, 24, 25, 26, 27, 28, 29, 30, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47 that were included, 3 articles22, 28, 32 presented 2 studies each.
The 27 studies included comprise 10 retrospective case-control studies,24, 27, 30, 32, 33, 37, 41, 44, 45, 46, 4 prospective case-control studies,22, 28, 38, 42, 9 limited prospective cohort studies22, 25, 28, 29, 32, 34, 35, 40, 43 and 4 prospective cohort studies26, 36, 39, 47 (Table I). Almost all studies were published in English except for one French article.45 Only one study37 was not performed in a Western country; the other 26 studies were performed in Europe,22, 25, 27, 28, 29, 30, 32, 34, 35, 38, 39, 41, 42, 43, 45, 46, 47 in the United States24, 26, 36, 44 and in Australia.33, 40
|Author, year||Reference no.||Country||Design||Exposition||No. of cases||Source cases||Age cases||Source controls||Source cohort||Correction confounding||Time frame|
|Protheroe, 1999||25||UK||Lim pros coho3||Questionnaire5||106||Suspicion||61.6||Unknown||Suspicion||Yes||5 years|
|Chen, 1995||29||UK||Lim pros coho3||Questionnaire5||41||Suspicion||57||Unknown||Suspicion||Yes||5 years|
|Ginsberg, 1996||33||Australia||Retro case-cont1||Questionnaire||98||Population||Unknown||Population||Unknown||Yes||2, 10 years|
|Roberts, 1996||24||USA||Retro case-cont1||Questionnaire6||258||Population||64.8||Population||Unknown||Yes||5 years|
|Terra, 1986||45||France||Retro case-cont1||Questionnaire67||100||Hospital||45–65||Population||Unknown||Yes||5 years|
|Kvikstad, 1994||38||Norway||Pros case-cont2||Record linkage||4,491||Population||Unknown||Population||Unknown||Yes||Unknown|
|Priestman, 1985||44||UK||Retro case-cont1||Questionnaire7||100||Hospital||50||Population||Unknown||No||3 years|
|Geyer (a), 1993||32||Germany||Lim pros coho3||Questionnaire5||33||Suspicion||49.2||Unknown||Suspicion||Yes||8 years|
|Geyer (b), 1993||32||Germany||Retro case-cont1||Questionnaire5||33||Hospital||49.2||Hospital||Unknown||Yes||8 years|
|Forsen, 1991||30||Finland||Retro case-cont1||Questionnaire6||87||Hospital||<70||Population||Unknown||Yes||1, 6 years|
|Jacobs, 2000||36||USA||Prosp coho4||Interview||39||Population||Unknown||Unknown||Population||No||>20 years|
|Kocic, 1996||37||Croatia||Retro case-cont1||Interview||106||Hospital||Unknown||Hospital||Unknown||Yes||Unknown|
|Achat, 2000||26||USA||Prosp coho4||Questionnaire||219||Population||46–72||Unknown||Population||Yes||Unknown|
|Scherg, 1988||43||Germany||Lim pros coho3||Questionnaire||125||Screening||52.7||Unknown||Screening||Yes||Unknown|
|Cooper (a), 1989||22||UK||Pros case-cont2||Questionnaire||171||Suspicion||55||Hospital||Unknown||Yes||2 years|
|Cooper (b), 1989||22||UK||Lim pros coho3||Questionnaire||171||Suspicion||55||Unknown||Suspicion||Yes||2 years|
|Price, 2001||40||Australia||Lim pros coho3||Questionnaire5||239||Screening||61.3||Unknown||Unknown||Yes||2 years|
|Bremond, 1986||27||France||Retro case-cont1||Questionnaire||50||Hospital||49.3||Hospital||Unknown||Yes||5 years|
|Cheang (a), 1985||28||UK||Lim pros coho3||Questionnaire||46||Suspicion||50.5||Unknown||Suspicion||No||2 years|
|Cheang (b), 1985||28||UK||Pros case-cont2||Questionnaire||46||Suspicion||50.5||Hospital||Unknown||No||2 years|
|Greer, 1975||34||UK||Lim pros coho3||Interview||69||Suspicion||53.3||Unknown||Suspicion||No||Unknown|
|Hughes, 1986||35||UK||Lim pros coho3||Questionnaire||33||Suspicion||52||Unknown||Suspicion||No||Unknown|
|Scherg, 1987||42||Germany||Pros case-cont2||Questionnaire||75||Screening||48.7||Hospital||Unknown||Yes||Lifetime|
|Lillberg, 2001||39||Finland||Prosp coho4||Questionnaire||205||Population||55.5||Unknown||Population||No||Unknown|
|Ewertz, 1986||46||Denmark||Retro case-cont1||Record linkage||175||Population||<70||Population||Unknown||Yes||1 year|
|Johansen, 1997||47||Denmark||Prosp coho4||Record linkage||198||Population||Unknown||Unknown||Population||No||1–40 years|
|Snell, 1971||44||USA||Retro case-cont1||Interview||352||Hospital||55.5||Hospital||Unknown||No||5 years|
The studies differed in populations from which patients were selected. Seven studies selected their patients from hospitals,27, 30, 32, 37, 41, 44, 45, and 8 selected their patients from general populations.24, 26, 33, 36, 38, 39, 46, 47 Patients were also selected from a suspicion cohort (n = 9)22, 25, 28, 29, 32, 34, 35 and from a screening cohort (n = 3).40, 42, 43 The same classification was used in the selection of controls, in case-control studies and in the selection of study base in cohort studies. Seven studies selected their controls from hospitals22, 27, 28, 32, 37, 42, 44 and 7 from regular populations.24, 30, 33, 38, 41, 45, 46 In 3 studies, the study base was selected from general populations,26, 36, 39, 47, in 7 studies from a suspicion cohort22, 25, 28, 29, 32, 34, 35 and in 1 study from a screening cohort.43
Twenty studies used questionnaires to assess the effect of stressful life events,22, 24, 25, 26, 27, 28, 29, 30, 32, 33, 35, 39, 40, 41, 42, 43, 45, whereas 4 studies used interviewing techniques.34, 36, 37, 44 Three studies were population-based record linkage studies.38, 46, 47
The total number of cases across all studies was 7,666, with a mean age at diagnosis of 53.8, ranging from 45–72 years (Table I).
Both visual control and the identification of statistical significance (p < 0.10) indicated the presence of publication bias in the categories stressful life events (p = 0.00) and death of relative or friend (p = 0.02) (Fig. 1). We could not identify publication bias in the categories death of spouse (p = 0.20), personal health difficulties (p = 0.63), nonpersonal health difficulties (p = 0.94), change in marital status (p = 0.60), change in financial status (p = 0.17) and change in environmental status (p = 0.70) (Fig. 1).
Summary odds ratios
The association between stressful life events and breast cancer risk was investigated in 11 studies.24, 25, 27, 29, 30, 32, 33, 37, 39, 45 Random effect meta-regression analysis revealed a summary odds ratio, i.e., stressful life events vs. no stressful life events, of 1.77 (95% CI 1.31–2.40) (Fig. 2). Four studies described the association between death of spouse and breast cancer risk.22, 24, 43 Because we could not identify heterogeneity within this category, we estimated the summary odds ratio and corresponding 95% confidence interval with fixed effect pooling (SOR = 1.37, 95% CI 1.10–1.71). Eleven studies provided quantitative information about the association between death of relative or friend and breast cancer risk (SOR = 1.35, 95% CI 1.09–1.68),22, 24, 28, 35, 36, 41, 43, 44, 47 and 6 studies about the association between personal health difficulties and breast cancer risk24, 25, 35, 41, 43, 44 (SOR = 1.17, 95% CI 0.74–1.86). Eight studies investigated the association between nonpersonal health difficulties and breast cancer risk (SOR = 0.92, 95% CI 0.72–1.17),22, 24, 25, 28, 35, 44 and another 9 studies analyzed the association between change in marital status and breast cancer risk (SOR = 0.88, 95% CI 0.73–1.08).22, 24, 30, 38, 41, 43, 44, 46 Four studies described the association between change in financial status and breast cancer risk22, 24, 44 (SOR = 0.90, 95% CI 0.54–1.50), and finally, 6 studies described the association between change in environmental status and breast cancer risk (SOR = 1.02, 95% CI 0.87–1.21)22, 24, 35, 41, 44 (Fig. 2).
Sources of heterogeneity
Calculation of the statistic I2 for every category of stressful life events and for individual qualitative items within study characteristics resulted in the identification of heterogeneity in all categories, except death of spouse (I2 = 0%) (Table II). The percentages of variability in point estimates, due to heterogeneity, of the other categories differed from 55% in change in environmental status to 88% in change in financial status. Forest plots confirm these conclusions (Fig. 2). Sources of heterogeneity that could be identified by means of meta-regression analyses were the study characteristic design, within the category change in financial status (p = 0.04); the study characteristic correction for confounding, within the categories stressful life events (p = 0.00) and change in financial status (p = 0.00); the study characteristic selection of patients, within the categories change in marital status (p = 0.01) and change in financial status (p = 0.00); the study characteristic selection of controls, within the category change in marital status (p = 0.03) and, finally, the study characteristic exposition, within the categories change in financial status (p = 0.00) and change in environmental status (p = 0.01). Sources of heterogeneity were not identified within death of spouse, personal health problems and nonpersonal health problems. Odds ratios and corresponding confidence intervals per strata supported these findings (Table II).
|Study characteristics||Categories of stressful life events|
|Stressful life events I2 = 84||Death of spouse I2 = 0||Death of relative or friend I2 = 79||Personal health difficulties I2 = 85||Nonpersonal health difficulties I2 = 79||Change in marital status I2 = 64||Change in financial status I2 = 88||Change in environmental status I2 = 55|
|Design||(p = 0.75)||(p = 0.34)||(p = 0.11)||(p = 0.74)||(p = 0.89)||(p = 0.13)||(p = 0.04)||(p = 0.97)|
|Retrospect case-cont1||1.93 (1.13–3.31)||84||1.04 (0.55–1.96)||0||0.92 (0.65–1.30)||51||0.84 (0.39–1.83)||0||0.86 (0.35–2.11)||87||0.79 (0.59–1.07)||22||1.30 (0.73–2.32)||82||0.98 (0.73–1.31)||64|
|Prospective case-cont2||1.21 (0.39–3.70)||0||1.75 (1.06–2.88)||0||2.49 (0.86–7.33)||0||1.27 (0.47–3.40)||92||0.81 (0.59–1.12)||55||0.65 (0.28–1.49)||0||1.21 (0.80–1.83)||0|
|Limited cohort||2.46 (0.98–6.18)||81||1.43 (1.12–1.82)||0||1.73 (1.20–2.49)||0||1.09 (0.50–2.37)||61||0.94 (0.44–2.03)||46||1.60 (1.07–2.38)||67||0.59 (0.26–1.34)||0||0.95 (0.64–1.41)||14|
|Prospective cohort||1.36 (0.90–2.05)||91|
|Confounding3||(p = 0.00)||(p = 0.65)||(p = 0.16)||(p = 0.35)||(p = 0.12)||(p = 0.00)||(p = 0.27)|
|Yes||2.22 (1.39–3.56)||86||1.37 (1.09–1.71)||0||0.87 (0.60–1.30)||87||2.11 (0.75–5.94)||88||0.67 (0.28–1.57)||80||1.73 (1.42–2.10)||67||0.41 (0.24–0.71)||45||1.36 (1.04–1.79)||12|
|No||1.04 (0.90–1.20)||0||1.30 (0.92–1.83)||72||1.46 (0.83–2.55)||50||0.82 (0.46–1.45)||82||0.93 (0.50–1.74)||0||0.71 (0.53–0.96)||0||1.10 (0.97–1.26)||20|
|Selection patient||(p = 0.96)||(p = 0.30)||(p = 0.05)||(p = 0.45)||(p = 0.61)||(p = 0.01)||(p = 0.00)||(p = 0.32)|
|Hospital||2.15 (1.11–4.17)||85||1.05 (0.62–1.77)||0||0.92 (0.21–4.06)||0||0.74 (0.20–2.69)||0||0.54 (0.34–0.86)||0||1.73 (1.26–2.37)||0||0.88 (0.59–1.29)||45|
|Population||1.53 (0.53–4.42)||86||1.04 (0.55–1.96)||0||1.13 (0.78–1.63)||87||0.77 (0.17–3.40)||0||1.00 (0.27–3.59)||0||0.90 (0.84–0.97)||0||0.96 (0.65–1.41)||0||1.13 (0.98–1.30)||0|
|Suspicion||2.48 (0.95–6.46)||81||1.17 (0.55–2.48)||0||1.79 (1.22–2.60)||0||1.40 (0.62–3.18)||85||1.06 (0.57–1.95)||79||0.69 (0.49–0.95)||0||0.61 (0.47–0.80)||0||1.07 (0.80–1.42)||45|
|Screening||1.69 (1.32–2.17)||0||1.71 (0.81–3.60)||0||1.45 (1.13–1.86)||0|
|Selection control||(p = 0.27)||(p > 0.05)||(p = 0.10)||(p = 0.43)||(p = 0.93)||(p = 0.03)||(p = 0.90)||(p = 0.49)|
|Hospital||2.55 (1.22–5.34)||91||1.21 (0.40–3.70)||0||1.37 (0.99–1.89)||72||1.53 (0.58–4.05)||94||1.09 (0.38–3.10)||85||0.62 (0.45–0.84)||0||1.07 (0.41–2.79)||93||0.98 (0.73–1.34)||89|
|Population||1.49 (0.77–2.87)||76||1.04 (0.55–1.96)||0||0.84 (0.54–1.32)||0||0.77 (0.19–3.09)||0||1.00 (0.18–5.66)||0||0.90 (0.83–0.97)||0||0.96 (0.24–3.78)||0||1.18 (0.78–1.77)||0|
|Type control||(p = 0.36)|
|Healthy/normal||1.99 (1.12–3.53)||88||1.08 (0.62–1.87)||0||1.16 (0.81–1.65)||78||1.39 (0.44–4.40)||93||1.20 (0.45–3.20)||89||0.84 (0.74–0.97)||0||0.79 (0.53–1.16)||87||1.19 (1.03–1.37)||69|
|Other diseases||4.72 (0.74–30.20)||0|
|Selection cohort||(p > 0.05)||(p = 0.53)||(p > 0.05)|
|Suspicion||2.55 (0.83–7.86)||81||1.13 (0.41–3.20)||0||1.82 (1.01–3.30)||88||1.07 (0.38–3.01)||61||0.97 (0.54–1.76)||85||0.76 (0.38–1.52)||0||0.59 (0.41–0.84)||0||0.94 (0.60–1.47)||14|
|Screening||1.45 (1.13–1.86)||0||1.71 (0.69–4.21)||83||1.45 (1.13–1.86)||0|
|Exposition||(p = 0.30)||(p = 0.91)||(p = 0.71)||(p = 0.59)||(p = 0.16)||(p = 0.00)||(p = 0.01)|
|Questionnaire||2.08 (1.22–3.54)||81||1.37 (1.09–1.71)||0||1.45 (1.06–1.99)||77||1.22 (0.63–2.37)||87||1.04 (0.22–4.95)||78||0.92 (0.67–1.26)||67||0.71 (0.53–0.97)||45||1.10 (0.99–1.23)||0,2|
|Interview||1.99 (0.73–5.42)||0||1.51 (0.85–2.69)||89||0.92 (0.24–3.56)||0||0.74 (0.12–4.76)||0||0.51 (0.24–1.07)||0||1.73 (1.09–2.75)||0||0.78 (0.62–0.98)||0|
In this meta-analysis, the association between stressful life events and breast cancer risk was investigated, using qualitative and quantitative data from 27 studies. A statistically significant effect was reported in 3 categories of stressful life events: stressful life events, death of spouse and death of relative or friend. Only death of spouse showed no publication bias. Sources of heterogeneity were identified for all categories, except for death of spouse by means of calculation of the statistic I2, and for the categories stressful life events, death of relative or friend, change in marital status, change in financial status and change in environmental status by means of meta-regression analyses. The results of this meta-analysis do not support an overall association between stressful life events and breast cancer risk. Only a modest association could be identified between death of spouse and breast cancer risk, being the only category without publication bias and without sources of heterogeneity.
The summary odds ratios (SOR) and corresponding 95% confidence intervals (CI) were estimated with random effect pooling. In contrast to fixed effect pooling, random effect pooling is suitable when summarizing and quantifying psychological variables, such as stressful life events, in which heterogeneity might be expected. Measurement of the statistic I2 was considered more relevant than the results of the test for heterogeneity that is commonly presented in meta-analyses, i.e., the Q-test, which has poor power with few studies and inappropriately high power with many studies. The principal advantage of I2 over Q is that comparisons can be made across meta-analyses of different sizes.18
The results from sensitivity analyses suggested that the study characteristic design, showed a statistically significant interaction within the categories death of relative or friend, change in marital status and change in financial status. It appeared that the summary estimates of retrospective studies were somewhat lower than for prospective studies and not always statistically significant. This contrast might be a consequence of selection bias because patients diagnosed with breast cancer might have interpreted stressful life events in a different way than patients not diagnosed with breast cancer. The study characteristic correction for confounding showed a statistically significant interaction within the category stressful life events. The summary estimate of the adjusted studies was somewhat higher than the estimate of the not-adjusted studies. The most important confounders, i.e., well-established risk factors such as age, family history of breast cancer and early menarche, therefore should be considered in design and/or analyses in future studies. Selection of patients showed a statistically significant interaction for the categories change in marital status and change in financial status. One might expect that especially suspicion may have influenced the summary estimate because it distinguishes a symptomatic at-risk group of patients. However, such a distinction could not be found. Finally, the study characteristic exposition showed a statistically significant interaction for the categories change in financial status and change in environmental status. It appeared that the summary estimates for questionnaire, i.e., the instrument of exposition, were somewhat higher than for interview. This might have been a consequence of the influence of the interaction between an interviewer and the patient within a specific interview situation.
To complete these general findings, a translation of the results of this meta-analysis into daily clinical practice would be justified. An increase in summary odds ratios, for the association between various categories of stressful life events and breast cancer risk, seems to run parallel with an accumulation in severity of these stressful life events. For example, the categories death of spouse and death of relative or friend appear to be more influential than the categories change in marital status and change in financial status. The biologic explanation of the overall association and these additional findings might be that stress disturbs various areas of the immune system and that impaired immune system function predisposes to malignant growth.4 In contrast to previous studies, clinicians can use these general findings to take an unambiguous and consistent position with regard to women with breast cancer.
- 1Borstkanker: kort en bondig. Bilthoven: RIVM, 2001., , , .
- 12Psychosocial stress as a precursor to breast cancer: a review. Curr Psych Res Rev 1986; 5: 268–80., , .
- 14Modern epidemiology, 2nd ed. Philadelphia: Lippincott-Raven Publishers, 1998. 67–75., .
- 17Stata Corporation. Stata statistical software: release 7.0 [program]: College station, TX: Stata Corporation, 2002.
- 27Psychosomatic factors in breast cancer patients: results of a case control study. J Psychosom Obstet Gynaecol 1986; 5: 127–36..
- 28Psychosocial factors in breast cancer. Dissert Abs Int 1992; 52: 5567..