Prediagnostic self-rated health and survival in women diagnosed with cancer in the Norwegian Women and Health Study: a prospective cohort study

Abstract

Objectives

Self-rated health (SRH) is a well-established predictor of all-cause mortality. However, its predictive value in patients with cancer remains unclear. We aimed to elucidate the relationship between prediagnostic SRH and mortality among patients with cancer using data from the Norwegian Women and Health (NOWAC) Study.

Design

A prospective cohort study.

Settings and participants

This study included 26 405 women from the NOWAC cohort who were diagnosed with cancer between 1992 and 2020. Subgroup analyses focused on the most common cancer types among Norwegian women: breast (n=8299), colorectal (n=3653) and lung and bronchial (n=2428) cancers.

Primary and secondary outcome measures

Prediagnostic SRH was assessed using a single-item measure with four response alternatives and categorised into ‘very good’, ‘good’ and ‘poor’ SRH. We used flexible parametric survival analysis and competing risk regression models to evaluate the association between SRH and all-cause and cancer-specific mortality after adjusting for age, physical activity, Body Mass Index, smoking, alcohol consumption and education.

Results

Poor prediagnostic SRH was associated with increased all-cause mortality among long-term cancer survivors. For specific cancer sites, the adjusted HRs and 95% CI for poor versus very good SRH were 1.29 (95% CI: 1.06 to 1.58) for breast cancer and 1.50 (95% CI: 1.20 to 1.88) for colorectal cancer. These associations were more pronounced among long-term survivors than short-term survivors. No association was observed between prediagnostic SRH and mortality for patients with lung cancer. For cancer-specific mortality, prediagnostic SRH predicted mortality from all cancers, as well as from colorectal cancer.

Conclusions

Prediagnostic SRH is a significant predictor of mortality in patients with cancer, particularly among long-term survivors. These findings indicate the potential of SRH as a predictive tool for mortality, underscoring the importance of integrating SRH assessments into identifying individuals at higher risk of mortality and highlighting the potential benefits of public health interventions aimed at improving overall health, but further studies are required to assess the effect of such interventions on SRH.

STRENGTHS AND LIMITATIONS OF THIS STUDY.

• This study included a large sample size and a prospective design.

• Reliable cancer and mortality data were obtained from national registries.

• The use of repeated waves of questionnaires facilitated the use of the most recent measurements of exposure and covariates before cancer diagnosis.

• The possibility of misclassification of cancer-specific deaths as deaths from other causes, including cardiovascular events related to treatment or the cancer itself, cannot be ignored.

• All covariates, except for age, were self-reported, and this could have introduced bias due to potential misclassification.

Introduction

Self-rated health (SRH) is a robust predictor of all-cause mortality, as consistently demonstrated across many community and cohort studies, independent of numerous known confounders.1,4 SRH is a widely used health measure in population studies, owing to its simplicity and ability to capture information about an individual’s health that objective and clinical health assessments may overlook.^{5 6} It likely captures an individual’s projected health trajectory, incorporating bodily sensations of diagnosed and undiagnosed conditions,6,8 as well as health behaviours,39,11 mental health,^{8 12} physical functioning and coping skills.^{13 14} Its validity as a comprehensive and widely capturing measure has been demonstrated by its consistent association with mortality, as well as morbidity^{5 15 16} and use of health services.¹⁷

Despite the well-established relationship between SRH and mortality in general populations, research focusing on SRH and mortality within specific patient groups, particularly among patients with cancer, remains scarce. This knowledge gap highlights a critical area where SRH could be clinically useful in cancer prevention and mortality reduction. SRH can be a cost-effective, patient-focused tool for the early identification of individuals at increased risk of mortality, providing opportunities for timely interventions; however, before such interventions can be identified, a better understanding of the predictive value of SRH for mortality in patients with cancer is required.

Using data from the Norwegian Women and Health (NOWAC) study, we previously observed that lower SRH was associated with increased all-cause mortality among Norwegian middle-aged to older women (aged 41–70 years at baseline), independent of socioeconomic status and modifiable lifestyle factors.¹⁸ We found that the predictive value of SRH for mortality was notably attenuated for women who received a cancer diagnosis during follow-up compared with that for women who remained cancer-free. This suggests that the predictive value of SRH for all-cause mortality seems to be altered by intermediate events, such as a cancer diagnosis; therefore, a more nuanced analysis of specific subpopulations is needed.

Few studies have explored prediagnostic SRH as a predictor of mortality within specific patient groups, such as patients with cancer. A study that included 424 401 participants from eight European and US cohort studies reported that SRH predicts cancer-specific mortality in older individuals (aged >60 years).³ However, whether SRH was measured before or after cancer diagnosis was not known, as the date of diagnosis was not recorded, and cancer information was restricted to the cause of death obtained from death certificates. In a prospective study involving 480 patients with breast cancer aged 65 years and older, Clough-Gorr et al reported that prediagnostic SRH collected retrospectively after breast cancer surgery predicted all-cause mortality during a 7 year follow-up.¹⁹ Similarly, in a retrospective cohort of 521 patients diagnosed with multiple myeloma (MM), Nabulsi et al found that prediagnostic SRH was associated with all-cause, but not MM-specific, mortality among individuals aged 65 years and older.²⁰

These studies demonstrated the potential of SRH as a predictive tool for mortality among patients with cancer but were limited by small sample sizes from limited populations, short follow-up periods and variability in the timing of SRH measurement relative to the cancer diagnosis.

The present study used a large population-based cohort with comprehensive follow-up data to elucidate the relationship between SRH assessed before diagnosis and survival outcomes across various cancer types. Our primary objective was to explore the associations between prediagnostic SRH and both long-term and short-term mortality following cancer diagnosis, specifically for the most common cancer types among Norwegian females: breast, colorectal and lung cancer.²¹ By examining both all-cause and cancer-specific mortality, we sought to clarify the potential of prediagnostic SRH as a predictive tool for mortality in patients with cancer, supporting its development as a practical clinical instrument. Our goal is to offer insights that can guide the incorporation of SRH into clinical practice and public health strategies for improved cancer outcomes.

Methods and materials

Study sample selection

The NOWAC study is a national cohort study representative of middle-aged and older Norwegian women. Since 1991, 172 472 women have participated in up to four waves of data collection. In the current study, we included 31 098 women from the NOWAC study who received a cancer diagnosis during follow-up, identified by linkage to the Cancer Registry of Norway. We excluded women with a prevalent cancer diagnosis (n=2949) and those with <2 years of follow-up from enrollment (n=54). We extracted data from up to three waves of follow-up questionnaires. For prediagnostic SRH and information on covariates, we used the latest measurement before cancer diagnosis. For participants who received a cancer diagnosis less than a year after the latest measurements, we used the second-latest measurements before the cancer diagnosis to avoid bias from reverse causation. We then excluded those with missing information on prediagnostic SRH (n=1197) or who received their cancer diagnosis at death (n=493), leaving 26 405 women available for analysis (figure 1). Among those who died during follow-up, 100 had missing information on the cause of death and were omitted from the analysis of cancer-specific death, resulting in a study sample of 26 305 women for the analysis of cancer-specific mortality. The participants were followed from the date of cancer diagnosis until the first endpoint of death, emigration or the end of the study follow-up period (31 December 2020).

Figure 1. Flowchart of analytical study sample inclusion. N, number of observations; SRH, self-rated health.

Endpoint information

We obtained information on cancer diagnosis through linkage to the Cancer Registry of Norway, with the date of cancer diagnosis and stage at diagnosis using the International Classification of Diseases, V.10 (ICD-10).²² We obtained the year of birth from the Norwegian Population Registry to determine the age at enrolment and the date of emigration for those who emigrated during the follow-up period. We obtained information on the date and cause of death (ICD-10) from the Norwegian Cause of Death Registry.

Main exposure

In the NOWAC questionnaires, SRH was included as a single item, worded ‘In general, would you rate your overall health as’ with four response alternatives: ‘very poor’, ‘poor’, ‘good’ and ‘very good’. In four questionnaires, the SRH item was divided into two separate items assessing self-rated physical health and self-rated mental health, rather than one global SRH item. From these questionnaires, we included participants who reported the same value for physical and mental health as their overall SRH (n=3157). The SRH variable was categorised into ‘very good’, ‘good’ and ‘poor’, which was a combination of ‘poor’ and ‘very poor’ owing to very few responses in the ‘very poor’ category.

Covariates

Covariates were selected a priori based on available data and published associations with SRH and mortality. We identified physical activity, Body Mass Index (BMI), smoking, alcohol consumption and education as potential covariates. We evaluated the contribution of each covariate to the regression coefficients for the main effect of SRH on mortality in univariate analyses and included the ones that imposed at least 5% change.²³ Using a backwards elimination approach, we continued the selection process to remove variables whose influence on the SRH-mortality association was not significant (p>0.05); all the variables that were originally identified were included in the final models after using this approach.

Participants were asked to rank their total physical activity level on a 10-item scale from 1 (lowest) to 10 (highest). We grouped physical activity level into three categories: 1–4 indicating low physical activity, 5–7 indicating moderate physical activity and 8–10 indicating high physical activity.

Self-reported body height in centimetres and body weight in kilograms were also collected. BMI (kg/m²) was computed from body weight and height and then categorised according to the WHO’s categories as underweight (<18.5 kg/m²), normal weight (18.5–24.9 kg/m²), overweight (25–29.9 kg/m²) and obese (≥30 kg/m²).²⁴

Smoking was reported as never, former or current smokers, with the average number of cigarettes smoked daily during predefined age intervals. We generated a variable for smoking metrics that combined smoking status with total pack years in categories of ≤14, 15–19 and ≥20 pack years.

Alcohol consumption was self-reported as volume and frequency of specified alcoholic beverages and calculated to grams of ethanol (EtOH) per day. Alcohol consumption was divided into teetotallers, <10 g EtOH/day and ≥10 g EtOH/day.

The participants reported their educational attainment as the total number of years of education. Education length was categorised according to the Norwegian school system before its reform in 1997 into four groups: ≤9 years, 10–12 years, 13–16 years and ≥17 years.

Cancer sites were grouped as C50 (breast cancer), C18–C20 (colorectal cancer), C34 (lung and bronchial cancer) and others, using the ICD-10.²² Information on the cancer stage was obtained through linkage to the Cancer Registry of Norway and was categorised as local, regional, distant or other/unknown.

Statistical analysis

Prediagnostic SRH and all-cause mortality

For the entire cohort of 26 405 participants diagnosed with cancer, we assessed the association between prediagnostic SRH and all-cause mortality using flexible parametric survival analysis with restricted cubic splines.^{25 26} This method allowed us to calculate age-adjusted and multivariable-adjusted HRs (aHRs) with 95% CIs. The use of Cox proportional hazards models initially revealed interactions with follow-up time. To address the likely violation of the proportional hazards assumption, we included SRH as a time-varying covariate with 3 degrees of freedom (df), and the baseline hazard function was modelled with 9 df using splines. The number of knots for the baseline hazard function and the time-varying covariate was selected by minimising the Akaike Information Criterion for models incorporating 1–9 df and by comparing the model incorporating 9 df for the baseline hazard function with the simpler model incorporating 3 df, which is the recommended default.²⁷ We plotted combined HRs with 95% CIs for all-cause mortality during the follow-up period. The estimates from each imputed dataset were combined using Rubin’s rule²⁸ to adjust for variability within and between the imputed datasets. We visualised the HRs with 95% CIs from the spline function and predicted HR with 95% CIs for mortality by prediagnostic SRH level at 5, 10, 15 and 20 years postdiagnosis.

To supplement the survival analyses with a visualisation of the baseline hazard function, we generated Kaplan–Meier plots for relative survival over follow-up for all-cause mortality across levels of SRH for the entire sample, as well as for subgroups of women diagnosed with breast, colorectal and lung cancer.

Subgroup analyses for specific cancer sites

Subgroup analyses were performed for women diagnosed with breast, colorectal or lung cancer. We used Cox proportional hazards regression models to calculate HRs with 95% CIs for the association between prediagnostic SRH and all-cause mortality. We computed separate estimates for short-term (up to 5 years after diagnosis) and long-term mortality.

Cancer-specific mortality

We evaluated cancer-specific mortality using competing risk regression, specifically by applying the Fine–Gray subdistribution hazard model.²⁹ Cancer-specific mortality was defined as death due to the same type of cancer as that at the initial diagnosis, whereas competing events were defined as death from other causes. We obtained separate estimates for short-term (up to 5 years after diagnosis) and long-term mortality.

Multiple imputation of missing covariates

The accumulation of missing covariate information in our study sample led to a significant proportion of observations with at least one missing covariate (13%). In certain subgroups, this proportion was even higher, reaching up to 23%, with a higher rate of missing data among cases than among non-cases (online supplemental table 1). To address this issue, we employed multiple imputations using chained equations under the missing-at-random assumption to estimate the missing values for education, physical activity, smoking status, pack years, alcohol consumption and BMI.

Ordinal logistic regression was used to impute education, smoking status and BMI. Predictive mean matching with five donors was used to impute physical activity, pack years and alcohol consumption. The imputation model included the following predictors: SRH, age at baseline, age at cancer diagnosis, cancer diagnosis (categorised as shown in online supplemental table 2), outcome (death during follow-up) and the Nelson–Aalen cumulative estimator, which accounts for the time from diagnosis to death while considering censoring events.³⁰ We generated 20 imputed datasets, each undergoing 10 iterations during the burn-in phase to improve the stability and accuracy of the imputation results.³¹

To ensure that the imputation produced reasonable values, we inspected the proportion of observed, imputed and completed values for categorical variables and kernel density plots for continuous variables.^{32 33} These plots are provided in online supplemental table 3 and online supplemental figure 1.

Adjustments

All models were adjusted for age at diagnosis. The multivariable models were adjusted for physical activity level, BMI, smoking habits, alcohol consumption and educational attainment. For subgroup analyses of specific cancer sites, cancer stage at the time of diagnosis was included as a stratification variable. For cancer-specific mortality analyses, we stratified the analyses according to the stage of cancer diagnosis.

All analyses were performed using Stata MP V.18.0 (StataCorp, College Station, Texas, USA).

Sensitivity analyses

We conducted sensitivity analyses with fewer adjustments for a representative subset of models to assess the robustness of our findings and the relative contribution of the selected covariates to our primary analyses. Specifically, we tested the following models: (1) all-cause mortality and total cancer, (2) all-cause mortality among women with breast cancer, (3) all-cause mortality among women with lung cancer, (4) cancer-specific, short-term mortality among women with colorectal cancer and (5) cancer-specific, long-term mortality among women with colorectal cancer. To determine whether the timing of SRH assessment relative to the cancer diagnosis influenced the predictive ability of prediagnostic SRH for mortality, we also ran the selected models with the full set of covariates with additional adjustment for the time between assessment of prediagnostic SRH and cancer diagnosis.

Patient and public involvement

Patients were not involved in the design and planning of this study.

Results

Baseline sample characteristics

At baseline, 7143 women (27.1%) reported very good SRH, 16 879 (63.9%) reported good SRH and 2383 (9.0%) reported poor SRH. Participants with lower SRH ratings were, in general, increasingly older and less physically active, more likely to be overweight or obese and current smokers and had fewer years of education than those with good SRH. Data on the distributions of the selected variables by SRH category are represented in table 1.

Table 1. Description of the study sample from the NOWAC cohort (N=26 405).

Variable	Very good SRH	Good SRH	Poor SRH
	n=7143	n=16 879	n=2383
Age at cancer diagnosis (years)
35–44 (%)	163 (2.3)	244 (1.4)	29 (1.2)
45–54 (%)	1127 (15.8)	2132 (12.6)	234 (9.8)
55–64 (%)	2897 (40.6)	6500 (38.5)	906 (38.0)
65–74 (%)	2440 (34.2)	6293 (37.3)	957 (40.2)
75–84 (%)	450 (6.3)	1522 (9.0)	238 (10.0)
85–94 (%)	66 (0.9)	188 (1.1)	19 (0.8)
Physical activity level (1–10)
Low (1–4) (%)	1014 (14.2)	4650 (27.5)	1250 (52.5)
Moderate (5–7) (%)	4105 (57.5)	9251 (54.8)	812 (34.1)
High (8–10) (%)	1746 (24.4)	2004 (11.9)	135 (5.7)
Missing (%)	278 (3.9)	974 (5.8)	186 (7.8)
BMI
Underweight (<18.5 kg/m2) (%)	90 (1.4)	257 (1.5)	67 (2.8)
Normal weight (18.5–24.9 kg/m2) (%)	4499 (63.0)	8412 (49.8)	997 (41.8)
Overweight (25.0–29.9 kg/m2) (%)	2069 (29.0)	5717 (33.9)	766 (32.2)
Obese (≥30 kg/m2) (%)	397 (5.6)	2233 (13.2)	513 (21.5)
Missing (%)	88 (1.2)	260 (1.5)	40 (1.7)
Smoking metrics (status with pack years)
Never smoker (%)	2636 (36.9)	5021 (29.7)	604 (25.3)
Former smoker with ≤14 pack years (%)	2421 (33.9)	4689 (27.8)	563 (23.6)
Former smoker with 15–19 pack years (%)	215 (3.0)	625 (3.7)	88 (3.7)
Former smoker with ≥20 pack years (%)	211 (3.0)	679 (4.0)	146 (6.1)
Current smoker with ≤14 pack years (%)	806 (11.3)	2347 (13.9)	313 (13.1)
Current smoker with 15–19 pack years (%)	308 (4.3)	1148 (6.8)	180 (7.6)
Current smoker with ≥20 pack years (%)	510 (7.1)	2271 (13.5)	463 (19.4)
Missing (%)	36 (0.5)	99 (0.6)	26 (1.1)
Alcohol consumption
Teetotaller (%)	1126 (15.8)	3491 (20.7)	767 (32.2)
<10 g EtOH/day (%)	5001 (70.0)	11 438 (67.8)	1400 (58.7)
≥10 g EtOH/day (%)	905 (12.7)	1574 (9.3)	180 (7.6)
Missing (%)	111 (1.6)	376 (2.2)	36 (1.5)
Education length (years)
≤9 (%)	1146 (16.0)	3491 (27.4)	863 (36.2)
10–12 (%)	2195 (30.7)	5686 (33.7)	762 (32.0)
13–16 (%)	2255 (31.6)	3766 (22.3)	400 (16.8)
≥17 (%)	1213 (17.0)	1751 (10.4)	188 (7.9)
Missing (%)	334 (4.7)	1057 (6.3)	170 (7.1)
Cancer diagnosis
Breast cancer (ICD-10; C50) (%)	2517 (35.2)	5133 (30.4)	649 (27.2)
Colorectal cancer (ICD-10; C18–C20) (%)	1014 (14.2)	2338 (13.9)	301 (12.6)
Lung and bronchial cancer (ICD-10; C34) (%)	385 (5.4)	1686 (10.0)	357 (15.0)
Other (%)	3227 (45.2)	7722 (45.7)	1076 (42.8)

BMI, Body Mass Index; EtOH, ethanol; ICD, International Classification of Diseases; N, number of observations; NOWAC, The Norwegian Women and Health Study; SRH, self-rated health.

Cancer outcomes and survival analyses

The distribution of cancers was as follows: breast cancer, 31.4% (n=8299); colorectal cancer, 13.8% (n=3653); lung or bronchial cancer, 9.2% (n=2428) and other cancers, 45.5% (n=12 025). At diagnosis, 43.6% (n=11 519) of participants had localised tumours, 24.3% (n=6425) had regional metastases and 15.8% (n=4161) had distant metastases. For the remaining 16.3%, the stage at diagnosis was classified as other or unknown.

Of the 26 405 participants diagnosed with cancer during follow-up in the NOWAC study and included in our study sample, 8644 (32.7%) died during the follow-up period. The analysis covered 2 125 575 person-months. Of these deaths, 6655 (77.0%) were classified as short-term deaths occurring within 5 years of the date of diagnosis. The total analysis time for short-term survival was 1 094 351 person-months.

Prediagnostic SRH and all-cause mortality

The Kaplan–Meier survival estimates for the baseline hazard function for all-cause mortality in the entire sample of women diagnosed with cancer are displayed in figure 2. Kaplan–Meier plots for the baseline hazard function for all-cause mortality by cancer site are presented in online supplemental figure 2.

Figure 2. Kaplan–Meier survival curves for the baseline hazard function by self-rated health category.

Combined aHRs with 95% CIs for the association between prediagnostic SRH and all-cause mortality over the follow-up time are displayed in figure 3. The aHRs varied over time. For the entire cohort, the aHRs declined immediately after diagnosis, followed by a progressive increase throughout the remainder of the follow-up period (figure 3). For the entire cohort, the combined aHR for poor versus very good SRH was 1.54 (95% CI: 1.25 to 1.84). The predicted HRs with 95% CIs for the association between prediagnostic SRH and mortality after 5, 10, 15 and 20 years after diagnosis are presented in table 2.

Figure 3. Combined HRs with 95% CIs for all-cause mortality during 29 years of follow-up for poor and good versus very good self-rated health (SRH) from 20 multiple imputations. HRs were calculated with restricted cubic splines with SRH as a time-varying covariate, and adjusted for age at cancer diagnosis (A) and age, physical activity level, Body Mass Index, smoking, alcohol consumption and length of education (B).

Table 2. Associations between prediagnostic SRH and all-cause mortality among women diagnosed with cancer during follow-up in the NOWAC study (N=26 405).

Survival time	Good versus very good SRH			Poor versus very good SRH
	HR	95% CI		HR	95% CI
5 years
Age-adjusted	1.07	0.99	1.15	1.34	1.19	1.49
Multivariable	0.96	0.88	1.03	1.08	0.95	1.20
10 years
Age-adjusted	1.23	1.10	1.37	1.83	1.54	2.12
Multivariable	1.10	0.98	1.22	1.48	1.24	1.71
15 years
Age-adjusted	1.30	1.13	1.47	2.04	1.66	2.42
Multivariable	1.16	1.01	1.31	1.65	1.34	1.96
20 years
Age-adjusted	1.34	1.14	1.53	2.20	1.72	2.66
Multivariable	1.19	1.02	1.37	1.78	1.40	2.16
Median HR for the entire follow-up period
Age-adjusted	1.26	1.10	1.42	1.91	1.54	2.27
Multivariable	1.12	0.98	1.26	1.54	1.25	1.84

N, number of observations; NOWAC, The Norwegian Women and Health Study; SRH, self-rated health.

Prediagnostic SRH and all-cause mortality by cancer site

Among the 8644 participants who died during the follow-up period, 16.5% (n=1422) were initially diagnosed with breast cancer, 14.5% (n=1250) with colorectal cancer, 19.1% (n=1649) with lung or bronchial cancer and 48.9% (n=4223) with other cancers. The detailed distribution of deaths (cancer-specific and non-cancer-specific) across cancer sites and SRH categories is presented in online supplemental table 4. For all-cause mortality during the entire follow-up (long-term) among women with breast cancer, the aHRs were 1.09 (95% CI: 0.96 to 1.23) for good SRH and 1.29 (95% CI: 1.06 to 1.58) for poor SRH. Among women diagnosed with colorectal cancer, the aHRs were 1.12 (95% CI: 0.98 to 1.29) for good SRH and 1.50 (95% CI: 1.12 to 1.88) for poor SRH. Among women with lung and bronchial cancer, the aHRs were 1.11 (95% CI: 0.96 to 1.28) for good SRH and 1.18 (95% CI: 0.97 to 1.42) for poor SRH. The HRs with 95% CIs for associations between prediagnostic SRH and all-cause mortality by cancer site are presented in table 3.

Table 3. HRs with 95% CIs for the association between prediagnostic SRH and all-cause mortality by the most diagnosed cancer sites in the NOWAC study.

Cancer site	Good vs very good SRH			Poor vs very good SRH			All-cause deaths
	HR	95% CI		HR	95% CI
Breast cancer (n=8299)
Long-term, age-adjusted	1.20	1.06	1.25	1.61	1.33	1.94	1440
Long-term, multivariable	1.09	0.96	1.23	1.29	1.06	1.58
Short-term, age-adjusted	1.22	1.02	1.45	1.59	1.22	2.09	677
Short-term, multivariable	1.13	0.94	1.36	1.34	1.01	1.79
Colorectal cancer (n=3653)
Long-term, age-adjusted	1.18	1.03	1.35	1.65	1.33	2.03	1267
Long-term, multivariable	1.12	0.98	1.29	1.50	1.20	1.88
Short-term, age-adjusted	1.17	1.01	1.36	1.50	1.17	1.91	1000
Short-term, multivariable	1.12	0.96	1.31	1.39	1.07	1.80
Lung and bronchial cancer (n=2428)
Long-term, age-adjusted	1.21	1.06	1.39	1.42	1.19	1.69	1669
Long-term, multivariable	1.11	0.96	1.28	1.18	0.97	1.42
Short-term, age-adjusted	1.21	1.05	1.39	1.41	1.17	1.69	1564
Short-term, multivariable	1.10	0.95	1.27	1.17	0.96	1.42

Multivariable models were adjusted for age at cancer diagnosis, physical activity level, BMI, smoking habit, alcohol consumption and length of education.

BMI, Body Mass Index; n, number of observations; NOWAC, The Norwegian Women and Health Study; SRH, self-rated health.

Cancer-specific mortality

Among the 8544 participants who died during follow-up, 7317 (85.6%) died from the same cancer they were initially diagnosed with (cancer-specific deaths). Subdistribution HRs (sHRs) with 95% CIs for the competing risk analysis of SRH and cancer-specific deaths are represented in table 4. Prediagnostic SRH was associated with both short-term and long-term mortality from all cancers: the adjusted sHR for long-term cancer-specific mortality was 1.07 (95% CI: 1.01 to 1.13) for good SRH and 1.12 (95% CI: 1.02 to 1.23) for poor SRH. Prediagnostic SRH was also associated with short-term and long-term cancer-specific mortality among women with colorectal cancer. For long-term survival, adjusted sHRs for good and poor SRH, respectively, were sHR=1.16 (95% CI: 0.99 to 1.36) and sHR=1.35 (95% CI: 1.03 to 1.77). For short-term survival, corresponding sHRs were 1.19 (95% CI: 1.00 to 1.41) for good SRH and 1.35 (95% CI: 1.01 to 1.81) for poor SRH.

Table 4. sHRs with 95% CIs for the association between prediagnostic SRH and cancer-specific mortality by total cancer and the most frequent diagnosed cancer sites in the NOWAC study (N=26 305).

Cancer site	Good versus very good SRH			Poor versus very good SRH			Cancer-specific deaths
	HR	95% CI		HR	95% CI
All sites (N=26 305)
Long-term, age-adjusted	1.80	1.12	1.25	1.42	1.30	1.55	7317
Long-term, multivariable	1.07	1.01	1.13	1.12	1.02	1.23
Short-term, age-adjusted	1.20	1.13	1.28	1.44	1.30	1.58	6100
Short-term multivariable	1.09	1.02	1.16	1.19	1.08	1.32
Breast cancer
Long-term, age-adjusted	1.05	0.91	1.22	1.21	0.92	0.59	857
Long-term, multivariable	0.99	0.85	1.15	1.08	0.81	1.44
Short-term, age-adjusted	1.15	0.94	1.40	1.36	0.95	1.94	475
Short-term, multivariable	1.09	0.88	1.35	1.24	0.85	1.81
Colorectal cancer
Long-term, age-adjusted	1.20	1.03	1.39	1.46	1.13	1.89	967
Long-term, multivariable	1.16	0.99	1.36	1.35	1.03	1.77
Short-term, age-adjusted	1.22	1.05	1.44	1.46	1.10	1.94	812
Short-term, multivariable	1.19	1.00	1.41	1.35	1.01	1.81
Lung and bronchial cancer
Long-term, age-adjusted	1.13	0.98	1.29	1.19	0.98	1.44	1458
Long-term, multivariable	1.04	0.90	1.21	1.04	0.85	1.28
Short-term, age-adjusted	1.12	0.97	1.29	1.19	0.98	1.45	1401
Short-term, multivariable	1.05	0.90	1.21	1.05	0.85	1.29

All analyses are stratified by stage. Multivariable models are adjusted for age, physical activity, BMI, smoking metrics, alcohol consumption and education.

BMI, Body Mass Index; N, Number of observations; NOWAC, The Norwegian Women and Health Study; sHRs, subdistribution HR; SRH, self-rated health.

Sensitivity analyses

Results from sensitivity analyses are presented in online supplemental table 5. HRs were progressively attenuated with the increasing number of covariates. No substantial changes in estimates were observed after additional adjustment for time from baseline to cancer diagnosis.

Discussion

For all cancers, both good and poor prediagnostic SRH (ref.: very good SRH) were associated with increased all-cause mortality in women diagnosed with cancer, particularly among long-term survivors. For women diagnosed with breast and colorectal cancers, poor prediagnostic SRH was associated with short-term as well as long-term all-cause mortality. After adjusting for modifiable lifestyle factors and years of education before cancer diagnosis, poor SRH before breast cancer diagnosis was associated with a 29% higher long-term and 34% higher short-term mortality. Similarly, poor prediagnostic SRH before colorectal cancer was associated with 50% and 39% increments in long-term and short-term mortality, respectively. Prediagnostic SRH did not predict all-cause mortality among those diagnosed with lung cancer, although the estimates indicated an inverse trend between prediagnostic SRH and increased mortality. We observed no evidence of an association between prediagnostic SRH and cancer-specific mortality from breast or lung cancer. For women with colorectal cancer, prediagnostic SRH predicted both long-term and short-term cancer-specific mortality.

To the best of our knowledge, no previous study has shown that prediagnostic SRH predicts cancer-specific mortality from colorectal cancer. The predictive value of prediagnostic SRH for cancer-specific mortality among women with colorectal cancer, but not for breast or lung cancer, may reflect differences in progression, treatment and survival of the disease. Colorectal cancer is associated with greater variability in long-term outcomes, where factors, including age, comorbidities and prediagnostic health behaviours, are known to play a role in survival and may all be captured by SRH.34,37 In contrast, breast cancer is associated with high survival rates owing to early detection through mammography screening and successful and standardised treatment protocols, which may reduce the variability in outcomes by prediagnostic SRH.38,42 Lung cancer is a more aggressive disease associated with high short-term mortality, which was also reflected in our results. Kaplan–Meier curves for the survival function among women with lung cancer (online supplemental figure 3) show that apart from lung cancer diagnosed in local stages, mortality rates were equal across all SRH categories. Most women with lung cancer in our sample were diagnosed in regional or distant stages. Our analysis of baseline characteristics grouped by SRH aligned with the results of previous research. The baseline differences between participants reporting poor and good SRH aligned with known risk factors for poor health outcomes, reinforcing the predictive power of SRH. The more pronounced association between prediagnostic SRH and increased mortality in long-term survivors than in short-term survivors may be attributed to the nature and severity of cancer at diagnosis. In more aggressive or advanced-stage cancers, such as lung cancer, the immediate prognosis is likely to be influenced more by the cancer itself than by an individual’s pre-existing health, represented by prediagnostic SRH. Although no previous studies have assessed the association between prediagnostic SRH and mortality of lung cancer, worse prediagnostic health-related quality of life was associated with reduced overall survival after lung cancer diagnosis in a prospective cohort study of 6290 older Americans (≥65 years).⁴³

Our results align with the findings of Rottenberg et al, who reported that poor prediagnostic SRH was associated with higher mortality among patients with cancer, though not after excluding breast, colon and prostate cancers.⁴⁴ This suggests that prediagnostic SRH is more predictive of mortality among patients with cancers associated with earlier detection, better prognosis and longer survival. Petrick et al found that SRH among patients with cancer declined dramatically within 1 year after cancer diagnosis, especially among those diagnosed with lung, breast and colorectal cancers.⁴⁵ This could suggest that prediagnostic SRH might be less predictive of short-term mortality than long-term mortality. This discrepancy highlights the importance of considering the timing of SRH measurements relative to the cancer diagnosis and duration of follow-up when evaluating the predictive value of SRH.

Our findings support the idea that general health before cancer diagnosis, as captured by prediagnostic SRH, can influence survival outcomes, particularly among cancers associated with early detection and treatment, as well as cancer-specific outcomes. This point is supported by studies conducted by Bamia et al,³ Clough-Gorr et al¹⁹ and Nabulsi et al,²⁰ who investigated the predictive value of SRH in different cancer contexts. Bamia et al³ reported that SRH predicts cancer-specific mortality in older individuals; however, they did not specify the timing of SRH measurement relative to cancer diagnosis, and their analysis was not restricted to patients with cancer. Clough-Gorr et al¹⁹ reported that prediagnostic SRH collected retrospectively after surgery predicted all-cause mortality among patients with breast cancer; however, retrospective measurement of prediagnostic SRH is prone to recall bias and may be influenced by cancer diagnosis and surgery. Nabulsi et al²⁰ demonstrated that prediagnostic SRH was associated with all-cause mortality among patients diagnosed with MM, but their sample size was small (n=521), and the prevalence of poor prediagnostic SRH was higher in their selected patient group (32%) than in our cohort (9%). Our study offers a comprehensive analysis of the predictive ability of prediagnostic SRH in a large prospective cohort of females with incident cancer.

The timing of SRH assessment when evaluating mortality outcomes among people diagnosed with cancer must be considered to prevent mortality among patients with cancer. Two previous studies have analysed the trajectories of SRH after serious health events.^{45 46} Petrick et al⁴⁵ analysed the trajectories of SRH among patients with cancer and reported that SRH had declined dramatically 1 year after diagnosis among patients diagnosed with breast, colorectal and lung cancer. Among those who survived longer than 5 years after diagnosis, the SRH ratings returned to prediagnostic levels.⁴⁵ Foraker et al reported a similar pattern in patients diagnosed with lung cancer.⁴⁶

The HRs for medical interventions or exposures often vary over time owing to immediate or delayed effects.⁴⁷ Schoenfeld residuals for the association between SRH and mortality for the entire cohort of females diagnosed with cancer correlated with survival time (p<0.05), indicating time-dependent covariates in our models. To address this aspect, we included SRH as a time-varying covariate. Flexible parametric models allowed us to account for non-proportional hazards, providing a more nuanced understanding of how the association between SRH and mortality changes over time. We observed no time interactions for the association between SRH and mortality among participants diagnosed with breast, colorectal or lung cancer and thus proceeded with Cox proportional hazards models for site-specific analyses.

The association between prediagnostic SRH and mortality highlights the potential benefits of prediagnostic health interventions in improving long-term outcomes in cancer survivors. Health-promotion strategies focusing on modifiable risk factors, such as physical activity, smoking and alcohol consumption, could be beneficial, but further studies addressing the influence of such interventions on SRH are needed. Furthermore, including SRH in patient assessments or risk allocation models could aid in personalising the prevention and follow-up for high-risk groups, potentially enhancing survival outcomes. With the increasing global burden of cancer, the prevention of mortality among diagnosed individuals has become a pressing public health issue. Prediagnostic SRH could serve as a useful tool for identifying high-risk individuals and guiding and informing mortality prevention strategies, as the prevalence of cancer survivors is increasing owing to improved diagnostic tools and treatment regimens for the most commonly diagnosed cancers,⁴⁸ as well as the increasing incidence of early-onset cancers.⁴⁹ Assessing the SRH of primary care patients may enhance clinician-patient communication, as lower SRH may be a risk factor for adverse health outcomes, including cancer.¹⁸ Prediagnostic SRH may serve as a reliable baseline for understanding disease progression after diagnosis, as SRH collected after diagnosis may be heavily influenced by the diagnosis itself and treatment effects.⁴⁵

Methodological strengths and limitations

The main strengths of this study are its large sample size and collection of variables in a prospective design, suitable for addressing the aims of this study. Cancer and endpoint information were ascertained through linkage to national registries that are nearly complete for cancer diagnoses and all deaths.^{50 51} The external validity of the NOWAC cohort has been assessed and found to be representative of Norwegian middle-aged to older women.^{52 53}

The main limitations of this study are the use of self-reported covariates and potential misclassification of cancer-specific deaths. Several of the self-reported variables have been validated previously, namely physical activity⁵⁴ and BMI.⁵⁵ Alcohol consumption was likely underreported, potentially influencing the precision of risk estimates adjusted for alcohol.⁵⁶

Misclassification of cancer-specific deaths may have occurred, as the Norwegian Cause of Death Registry contains a notable proportion of codes that likely represent other, immediate causes than the true underlying cause of death,⁵⁷ resulting in a potential misclassification that could have affected the precision of the estimated risks.

Conclusions

Prediagnostic SRH predicted mortality in women diagnosed with cancer, particularly among long-term survivors. This finding supports the integration of general health assessments in cancer care and underscores the potential benefit of using SRH to identify individuals at higher risk of mortality. Improving population health through lifestyle interventions and reducing social disparities in health may help to prevent cancer and improve survival outcomes among those diagnosed with cancer, but further studies are required to assess the effect of such interventions on SRH.

Data availability statement

Data may be obtained from a third party and are not publicly available.