Author's response to reviews Title: Association between self-rated health and mortality: 10 years follow-up to the Pro-Saude cohort study Authors: Joanna M N Guimarães (joannaguimaraes@hotmail.com) Dora Chor (dorachor@fiocruz.br) Guilherme L Werneck (guilherme.werneck@terra.com.br) Marilia S Carvalho (carvalho@fiocruz.br) Claudia M Coeli (coeli@iesc.ufrj.br) Claudia de S Lopes (lopes@ims.uerj.br) Eduardo Faerstein (efaerstein@gmail.com) Version: 3 Date: 10 May 2012 Author's response to reviews: see over
Dear Editor, On behalf of the authors of the manuscript entitled Association between self-rated health and mortality: 10 years follow-up to the Pro-Saude cohort study, I would like to thanks the reviewers for their thoughtful comments. We have revised the manuscript to take care of their considerations. Following are answers and comments to each of the questions raised. Hope we were able to meet your requirements. Please contact us if you need any further explanations. Sincerely, Mrs. Joanna Miguez Nery Guimarães
Reviewer #1: Michaela Benzeval Major Compulsory Revisions 1. The paper acknowledges the significant existing literature on this topic, but points to the lack of evidence from Brazil as a significant factor in the novelty of the paper. This implies that the authors believe the context of the association may affect their findings. Do they expect the association to differ in Brazil form other countries? If so why, and how do their findings shed light on these ideas? There are many issues that might justify studying such relationship in our specific cultural context. First, there are few studies on the association between SRH and mortality arising from Brazil, a country with a different socioeconomic profile as compared to most developed countries where the majority of reports on this issue come from. Second, the pattern of mortality by age and cause and the profile of response to the SRH question are quite diverse; both might per se contribute to a different pattern of association between SRH and mortality. Finally, there are recognized cultural differences in SRH classification, which also might alter the relationship between mortality and SRH (Jürges H. True health vs response styles: exploring cross-country differences in self-reported health. Health Economics 2007; 16:163-78). Although we believe that these issues would justify a study like ours, our results actually reinforce the findings of most international studies. This is in fact an interesting contribution, suggesting external validity for the SRH-mortality association. 2. As noted on p 13 this is a very select population employees in technical and administrative careers in a single university, who were relatively young what does this mean for the generalisability of findings beyond this group? Although the study population has some specificity (like some of the most important cohort studies as the Nurses Study or the Whitehall Study ), we do not think this would affect the generalisability of our results on the association between SRH and mortality. Our cohort shows enough heterogeneity concerning age, gender, literacy, income and race, which allows adjustment for the effects of these variables in multivariate analysis. In addition, our results are in line with most of the studies, suggesting external validity of our findings. 3. Limitations the authors mention a key limitation of the study that only SRH was a time varying variable? Whilst they suggest that SES variables may no vary in this group, what of other factors such as disease state? A more rounded discussion of this issue is required. We would like to thank for the suggestion. Although the presence of diseases for sure will change over time, we do not believe that this would happen that importantly for SES in this specific population. This is because participants are all permanently employed (almost no risk for being dismissed), experiencing neither substantial changes in their occupational position nor in their salary along the time of the study. Therefore, we decided to run new analyses with only presence of diseases as a time-dependent covariate (besides SRH). In this new analysis, we found a stronger effect of presence of diseases in survival time (univariate analysis) both for men (HR=3.74; IC95% 2.31-6.06) and women (HR=4.64; 2.51-8.6), as compared to our previous analysis with
presence of diseases not varying over time (men: HR=2.76; 1.73-4.42 / women: HR=4.39; 2.44-7.91). In the new multivariate analysis, the adjusted effect of SRH fair/poor on the risk of mortality was somewhat lower (men: HR=2.13; 1.03-4.40 / women: HR=3.43; 1.23-9.59), but very similar to that obtained adjusting for presence of diseases not varying over time (men: HR=2.44; 1.19-4.98 / women: HR=3.55; 1.29-9.80). Considering these new results, we made the following changes in the text: Abstract, 3 rd paragraph: Results: ( ) men who reported Fair/Poor SRH showed relative hazard of death of 2.13 (CI95% 1.03-4.40) and women, 3.43 (CI95% 1.23-9.59), as compared with those who reported Very good SRH. Methods (Measurements), 4 th paragraph: All variables analyzed were collected at stage 1 only (1999), except the main exposure variable (SRH) and age, updated in 2001 (stage 2) and 2006 (stage 3), and the covariate presence of diseases, updated in 2001 but not available in 2006. Methods (Statistical Analysis), 1 st paragraph: The association between SRH and mortality was estimated using extended Cox proportional hazard models with both SRH and presence of diseases varying over time. Results, last paragraph: After adjustment (model 4), men who reported Fair/Poor SRH showed relative hazard of death of 2.13 (CI95% 1.03-4.40) and women, 3.43 (CI95% 1.23-9.59), as compared with those who reported Very good SRH. Discussion, 1 st paragraph: Men and women with Fair/Poor SRH had 2.1 and 3.4 times greater hazard of mortality, ( ) Discussion, 5 th paragraph: In our study, although inclusion of this variable contributed to reducing the strength of the association between Fair/Poor SRH and mortality (reduction of age-adjusted HR by 22.9% and 22.1% in men and women, respectively), ( ) Discussion, second-to-last paragraph: Analysis of SRH in three categories (instead of the SRH positive / SRH negative dichotomy found in most research) and also the use of both SRH and presence of diseases as time-dependent covariates make the results ( )
4. There is obvious drop out in the study; what is the effect of drop out on the findings? Since those who drop out will have their earliest SRH measure include in the analysis, this might suggest that the associations reported are weaker than one might expect (as one would anticipate those in less good health and more likely to drop out of the study) Thanks for the comment, we agree that those who drop out would probably have worse health as compared to the first SRH evaluation and higher probability of death. Including only the first SRH assessment for these participants would probably lead to underestimation of the strength of the association between SRH and mortality. We included the following comment in the discussion section (9 th paragraph): Results of our study might have been biased due to lack of complete information on changes of SRH over time for the participants. About 14% of them had only the first baseline SRH measure, and other 10% had only two SRH measures (baseline plus SRH recorded on stages 2 or 3). The potential effect of such problem on the results is unknown, but one might suppose that those who drop out would probably have worse health as compared to their earliest SRH evaluation and higher probability of death. Including only the first SRH assessment for these participants would probably lead to underestimation of the strength of the association between SRH and mortality. Minor Essential Revisions 5. I did not follow the discussion about source of mortality data in the methods (p.5) and why it was necessary to use the university records instead of national data. It became clear in the discussion (p12) this was because national data were not available from 2006 onwards. The methods need rewriting to make this clear. Thanks for the comment, we correct the text to make it clearer. Now, in the 2 nd paragraph Methods (Measurements) the text reads: It was opted to consider the information from the human resources department as a primary source with a view to: (1) permit identification of the events that occurred up to the end of the observation period (2009), since deaths occurring after 2006 were not available in the Mortality Information System (SIM); and (2) ( ) 6. P6 in describing ethnic categories the authors talk of yellow people. Is this a term commonly used? I assume they mean Asian as this is the census category provided, and personally I would find this a more appropriate term. We agree and changed the text to take care of your comment, now in Methods (Measurements), 3 rd paragraph, the text reads as follows: Those who classified themselves as asian or indigenous were ( ). 7. First line of results p 7 would read better as: The group was mostly female and predominantly. Changed as suggested.
8. P 11 last sentence did not figure AMONG causes of death for women Corrected. Reviewer #2: Davide Malmusi Major Compulsory Revisions 1.1. The authors claim in their discussion and conclusion about the importance of using more than one measure of SRH, introducing it as a time-dependent variable in survival analysis model. I suggest that, for making this claim more robust, the survival analysis is replicated using SRH baseline measure only, and HR obtained in these models are compared with those already presented. We agree that we might have gone too further in the conclusion, since it was not our objective to make comparisons between different ways to evaluate the effect of SRH on mortality. We simply decided to use SRH as a time-varying covariate based on the literature, which suggests that SRH is a dynamic measure that usually changes over time (Strawbridge & Wallhagen, 1999; Han et al., 2005; Lyyra et al., 2009). In our opinion, using a time-varying covariate would only improve our understanding of the effect of SRH on mortality, since we are using more recent data for those who have this information. Actually, the results of the analysis using only baseline information are not that different from those using SRH as a time-varying covariate, except for the fact that the association was stronger for women (for men, HR = 2.1; IC95% 1.07-4.12, and for women, HR = 6.1; IC95% 1.77-21.35). Since our objective is not to compare models with only baseline SRH with those with time-varying SRH, we prefer not to present this result and decided to remove the sentences in the conclusion that led to this misunderstanding. 1.2. I also suggest that the family income variable is recalculated, replacing the simple per capita measure with the use of some equivalence scale for household income (see for instance OECD at www.oecd.org/dataoecd/61/52/35411111.pdf). We thank for the suggestion, but we consider that the use of an equivalence scale for household income would not be appropriate for this specific study. The problem is that the denominator that we used for calculating per capita income is not the size of the household, but the number of persons that depend on the total income. In Brazil, it is common that dependants include elderly people and young adults (sons and daughters) that do not live in the same house. Therefore, applying the usual weights for those persons would not be that obvious, since they do not share the same household needs (food, electricity, etc). 1.3. I would also be very cautious with over-interpretations of the HR for fair/poor SRH being higher in women than in men. Confidence intervals are overlapping (you may want to try a model including both sexes with an interaction term), but most importantly, looking at crude data in Table 1, one realises that only 3 women were dead in the very good health group, thus providing the case for the very high value of the
relative hazard ratio. Actually, if you look at absolute difference in the probability of death of fair/poor versus very good (at baseline), this turns out to be 7% in men and 5.4% in women. We would like to thank the reviewer for pointing out this important analytical question. We decided to run separated models for men and women because there is evidence showing that the predictive ability of SRH for mortality varies by gender (e. g., Nishi et al., 2012. Sex/gender and socioeconomic differences in the predictive ability of selfrated health for mortality. PLoS One. 7(1):e30179.). We agree that including an interaction term would be an option for assessing such differences, but with sparse mortality data, the statistical power to assess interaction would certainly be low. Although the lack of the desired precision might hamper a definite inference on gender differences, we think that it is worth mentioning the differences in point estimates. In addition, after running the new models with both SRH and presence of diseases as timedependent covariates (as suggested by reviewer 1), confidence intervals were no longer overlapping (men s adjusted HR=2.13; 1.03-4.40 / women s = 3.43; 1.23-9.59). Anyway, we included a commentary in order to take care of your consideration. Now it reads: Discussion, end of 4 th paragraph: However, this difference between hazard of death for men and women must be interpreted with caution in our study, because there are few observations in some categories, (e. g., only 3 women were dead in the very good health category) generating imprecise estimates (model 4, Table 2). Minor Essential Revisions 2.1. The Background in the Abstract refers to few studies on the subject in Latin America, but the literature review in the main text is limited to Brasil. Please verify or change. Actually all studies in Latin America come from Brazil. We tried to make it clear in the Discussion section (2 nd paragraph): In Latin America the only three studies on this subject all evaluating populations of older adults come from Brazil. Two of them also encountered greater hazard of mortality for individuals with worse SRH (Maia et al., 2006; Lima-Costa et al., 2007). 2.2. The Methods in the Abstract speaks about a cohort of staff at a public university, maybe it could be specified that they are non-faculty civil servants, to promptly provide information about the external validity. We agree and made the following corrections throughout the text. Abstract, Methods section: Methods: Cox regression models were used to examine the association between SRH (Very good, Good, Fair/Poor) and mortality, over a 10 year period, in a cohort of non-
faculty civil servants at a public university in Rio de Janeiro, Brazil (Pró-Saúde Study, n=4009, men=44.4%). Introduction, end of 5 th paragraph: The aim of this study is to evaluate the association between SRH reported at three data collection stages and mortality, over a period of 10 years, among a cohort of non-faculty civil servants at a public university in Rio de Janeiro. 2.3. Please replace NAs in Table 1 footnote with Not answered. Corrected. 2.4. In Methods-Measurements you mention that test-retest reliability was performed with SRH at stage 1: could you please specify it this was done on the whole sample, a sub-sample... Test-retest reliability study was performed among individuals that were not enrolled in the cohort study, but with a similar sex, age and literacy profile. We changed the last phrase of this paragraph to make it clearer. Now it reads: Methods, Measurements section, end of 1 st paragraph: A test-retest reliability study with a one-week interval between responses was performed among individuals that were not enrolled in the cohort study, but with a similar sex, age and literacy profile. Reliability for the SRH reported at stage 1 of the study (1999) was estimated using weighted kappa (square weighting), returning a value of 0.65 (CI95% 0.54-0.72). 2.5. Please explain if you had any losses to follow-up. For all baseline participants we were able to check their outcome status (alive or dead) by checking files of the University human resources department. This is an extremely reliable system, since any death should be informed by the family to the University, because the family has the right to receive a regular pension and also funds to cover the funeral costs. Therefore, we consider implausible the possibility that an employee dies and this information remains unknown. Anyway, we agree with the reviewer s appointment and included (Discussion section, 10 th paragraph) an explanation to elucidate this question. Now, it reads: Besides, the university records system is extremely reliable, since any death must be informed by the family to the institution, due to their right to receive a regular pension and also funds to cover the funeral costs. 2.6. Why marital status is included as an intermediate characteristic, together with diseases and smoking, and not as a distal one together with schooling and income? Please reconsider. (I actually see that it was not enough significant to be retained in models)
We agree with the reviewer and the analysis was performed as suggested. In any case, marital status remained non-significant and was not retained in the model. In the Methods section, we moved marital status from the intermediate group of variables to the distal group. Methods, Statistical analysis section, 2 nd paragraph: Variables' entry into the multivariate models was determined on the hierarchical causality theoretical model, including first the distal characteristics (age, sex, color/race, schooling, income and marital status) and then the intermediate characteristics (presence of diseases, smoking, BMI and common mental disorders). 2.7. Maybe the 20% criterion for altering the effect of SRH on mortality is a too strict one, I would suggest 10%. By the way, how do you calculate the %? The value of 20% was chosen considering the principles of parsimony and statistical efficiency, since choosing a lower value would potentially lead to the inclusion of more variables in the multivariate model. It should be clear, however, that this was not the only criterion, but the level of significance of the association (p<0.05) was also considered. In any case, we performed the analysis again using the 10% criterion, and the same variables were selected. To calculate the percentage we used the following formula: [(HR 0 HR x ) / HR 0 ] * 100 where, HR 0 is the HR for AAS in a model without variable X HR x is the HR for AAS in a model in which variable X was added. 2.8. I suggest (also for identification issues) removing the specification of the causes of the four deaths from other causes (one person each). We agree and modified the text accordingly. Now the text in the Results section (5 th paragraph) reads as follows: The causes of mortality for deaths occurring up to 2006 (N=80) were distributed as follows: diseases of the circulatory system (N=24, 30%), neoplasms (N=13, 16%), diseases of the respiratory system (N=11, 14%), external causes (N=8, 10%), infectious and parasitic diseases (N=6, 8%), endocrine, nutritional and metabolic diseases (N=6, 8%), diseases of the digestive system (N=4, 5%), symptoms, signs and abnormal clinical and laboratory findings, not elsewhere classified (N=4, 5%), and others (N=4, 1% each). 2.9. For the Results in Abstract, I would replace the part starting from risk of mortality for the more accurate and clearer formulation you have in the main text, page 9, starting from men who reported.
We agree and modified the text accordingly. Now the text in the Results in Abstract reads as follows: Results: About 40% of the population changed their self-rating over the course of follow-up. After adjustment for self-reported physician-diagnosed diseases and other covariates, men who reported Fair/Poor SRH showed relative hazard of death of 2.13 (CI95% 1.03-4.40) and women, 3.43 (CI95% 1.23-9.59), as compared with those who reported Very good SRH. Discretionary Revisions 3.1. I would shade the claim, based on Figure 1, that 40% of subjects would be misclassified using the baseline SRH only. Taking into account that survival differences between Very good and Good SRH are far from significant, I would mention the values for total changes between very good/good and fair/poor categories (the four last groups in Figure 1, making up slightly less than 20% at each stage). This result aims at highlighting the patterns of change of SRH, which would support the idea that, whenever possible, including time-varying information on SRH might be useful. We thank the suggestion, but we decided to retain the original Figure. 3.2. In the background and discussion about the factors influencing SRH, you might want to incorporate references from Manderbacka Scand J Public Health 1998, or Singh-Manoux et al. J Epidemiol Community Health 2006. Thanks for the suggestion. We included the reference from Manderbacka Scand J Public Health 1998 in the Discussion section (end of 6 th paragraph). Manderbacka (1998) suggests that in addition to the medical model of health, adopting health promotion messages and "healthy" lifestyles are important factors contributing to health assessments. 3.3. Please be more specific in the mention of socio-demographic indicators (Szwarcwald et al), which ones? Age, sex...? Thanks for the suggestion, now the text (Background, 2 nd paragraph) reads as follows: ( ) socio-demographic indicators such as age, sex and education (Szwarcwald et al., 2005). 3.4. In the first line of results, please indicate the % (55.6%) of females. Corrected. 3.5. I suggest that reading of the Results section would be more straightforward by moving the second paragraph ( Of the 4009 participants.. ) next to the fourth (before the fifth, starting with The causes of mortality.. ).
Thanks, we made changes in the Results section to take care of this suggestion. Now the first three paragraphs of that section reads as follows: Due to the difference encountered in the death hazard by SRH among men and women, the analyses were performed separately by sex (Table 1). The group studied was mostly female (55.6%) and predominantly young adults (mean age = 40.1 years) (Table 1). About 40% had completed undergraduate or postgraduate education, 52% classified themselves as white, and just over 20% were smokers at the time of the study. Approximately 30% reported at least one medical diagnosis of some disease of interest. In addition, more than half the population were overweight (BMI 25) and the prevalence of common mental disorders was estimated at more than 30%. At stage 1, health was self-rated as Very good by 28.2% of participants, Good by 53.2%, Fair by 17% and Poor by 1.6%. Self-perceived health was worse among women than among men, with 21.3% reporting SRH as Fair or Poor, compared with 13.2% of the men (p<0.001). Worse SRH was also observed among the older individuals, those with less income, less schooling, who were widowed, who reported some disease, had higher BMI, were classified as positive for common mental disorders or as smokers, for both sexes (p<0.05) (data not presented). At stages 2 and 3, SRH prevalences were, respectively, 27.8% and 26.3% (Very good), 51.5% and 50.7% (Good), 17.8% and 20.2% (Fair), and 1.6% and 2.2% (Poor). As regards changes in SRH over the course of follow-up (Figure 1), 36.2% of the population changed SRH category from stage 1 to stage 2, and 38.2% from stage 2 to 3. The pattern of change was more often towards worsening health than towards improvement, and that difference was most marked from stage 2 to stage 3 (21.3% worsened and 16.9% improved) than from stage 1 to stage 2 (19.1% worsened and 17.1% improved). However, the proportion of participants whose SRH deteriorated or improved by two categories (from Very good to Fair/Poor, or the opposite) was very small: 1.2% and 0.8% from stage 1 to stage 2; and 1.1% and 1.0% from stage 2 to stage 3, respectively. 3.6. In the Discussion, page 11, maybe it s more correct to define The presence of diseases as explanatory factor more than a confounder. Whether the presence of co-morbidities should be considered a confounder or other type of explanatory factor is a controversial issue. However, since the presence of comorbidities is an independent risk factor for mortality and is also associated with SRH, we think these features make this a potential confounding variable. Actually, the presence of co-morbidities is considered one of the most important confounding variables in the studies relating SRH to mortality, as stated by Idler et al., 1990: The first and most important confounding variable to be considered must be that of prior physical health status itself. (Idler EL, Kasl SV, Lemke JH: Self-evaluated health and mortality among the elderly in New Haven, Connecticut, and Iowa and Washington counties, Iowa, 1982-1986. Am J Epidemiol 1990, 131: 91-103.). 3.7. In the Discussion, page 12, I would not say that the studies that have used SRH as a time-dependet covariate in Cox regression are few (they are five!).
We agree and changed the text to correct this. Now the text reads as follows (7th paragraph of Discussion section): Many studies have indicated that the ability of SRH to predict mortality diminishes with increasing cohort follow-up time (Grant et al., 1995; Singh-Manoux et al., 2007(a); Singh-Manoux et al., 2007(b)). This result may possibly stem in part from the use of SRH measured at the baseline alone, making it a good predictor of early mortality, but not of late mortality. The studies that have investigated the association between SRH as a time-dependent covariate and mortality using Cox regression (Strawbridge & Wallhagen, 1999; Ferraro & Kelley-Moore, 2001; Han et al., 2005; Lyyra et al., 2009; Sargent-Cox et al., 2010) are not that frequent, but the results are consistent. Reviewer #3: Martí Casals Toquero Major Compulsory Revisions: I consider you have done good analysis but you should explain the Methods and Statistical analysis sections better. You use repeated SHR measures. For this reason, I understand you use time-dependent covariates in the Cox models. The covariates may change their values over time and the form of a time-dependent covariate is much more complex than in Cox models, so it isn t clear in the abstract and methods sections. I suggest revising the time-dependent covariates in the Cox models literature. We made the following corrections in the text to take care of the recommendation: Abstract, 2 nd paragraph: Methods: Cox regression models were used to examine the association between SRH (Very good, Good, Fair/Poor) varying over time and mortality, ( ) Methods (Statistical Analysis), 1 st paragraph: The association between SRH and mortality was estimated using extended Cox proportional hazard models with both SRH and presence of diseases varying over time. Minor Essential Revisions: I also consider that this article has a relevant importance but the following minor essential revisions are needed. Methods: As I mentioned before I think or I hope you use a time-dependent covariate, which is a correct method to use. To understand better your following data I suggest adding a flow-chart figure in the manuscript. We modified the text to make these points clearer (as described above), so we consider that a flow-chart would not be necessary. Anyway, if you think that the new text is still unclear and a flow-chart is really necessary, we would be glad to include it.
Page 6, 2nd line: I recommend writing right-censored instead of censured. We agree and made the correction: Methods (Measurements), 2 nd paragraph: Participants still living in May 2009 were right-censured. In possible future studies if you want to analyze the survival through death of age, I suggest usind left-truncated method. It s necessary to update the value of age if you want to fit an interaction of age and SRH. Thanks for the suggestion. We did not have the objective to assess interaction between SRH and age in this manuscript, but we will follow your recommendation in future studies with that aim. Statistical Analysis: I understand you use two separate models but you can confuse the reader if you use a stratified Cox models. You should explain it better. To make readers sure that we are doing separate models for men and women, the first phrase of the results section now reads: Due to the difference encountered in the death hazard by SRH among men and women, the analyses were performed separately by sex (Table 1). Results: Table 1: I don t quite understand the meaning of R. The R in Table 1 corresponds to risk (Cumulative Incidence - CI). We preferred to use R instead of CI to avoid confusion with confidence interval. To take care of this recommendation we changed R for Risk in Table 1. Table 2: You could show the best models fitted directly. We prefer to maintain it as it is. We think it is important for the reader to follow the changes in HR as each covariate is included in the model.