OUP user menu

Widening educational differences in cancer survival in Norway

Håkon Kravdal
DOI: http://dx.doi.org/10.1093/eurpub/ckt082 270-275 First published online: 27 June 2013


Background: All-cause and cause-specific mortality have long been known to be associated with various indicators of socio-economic status, and social gradients have been shown also for cancer survival. In recent decades, several studies have reported increasing social differentials in mortality rates. This study aims to investigate the development with respect to cancer survival, which has not been done before. Methods: Discrete-time hazard regression models for cancer deaths among women and men diagnosed with cancer 1970–2007 at age 30–89 were estimated, using register data encompassing the entire Norwegian population. The analysis was based on >200 000 cancer deaths during over 2 million person-years of exposure among >440 000 individuals diagnosed with cancer. Results: There has been an increasing advantage for women of all educational categories when compared with those with only compulsory schooling. No such widening of the educational gap has appeared with respect to cancer survival among men. Conclusions: Increasing educational differentials in health at the time of diagnosis, health behaviour and cancer treatment seem plausible, and would to some extent accord with the increasing social gaps in all-cause or cause-specific mortality rates that have been reported in other studies. Also, it is not impossible that such trends in the educational gradients in health and treatment are stronger for women than for men, though such sex differences have not been indicated in mortality studies. There is no obvious explanation for the complete absence of change in the education effects among men.


It has long been recognized that higher socio-economic status, whether measured as educational level, income or occupation, is associated with lower all-cause mortality.1 Furthermore, mortality from many specific causes follows the same pattern.1,2 In recent decades, it has been shown that cancer survival is influenced by socio-economic status as well: among cancer patients, the chance of dying (from cancer) is, for example, highest for the less educated.3,4

In many countries, the socio-economic differences in mortality seem to have increased during the past decades.5 Such a development has been observed even in Nordic countries,5–10 in spite of persistent political emphasis on social solidarity and free public health care, and obviously gives rise to some concern.

The objective of this study is to see whether there has been a similar trend in social inequalities in cancer survival in Norway, measured by educational differentials. Such an analysis has not been undertaken earlier, in any country. The study is based on register data that cover the last four decades and encompass the entire Norwegian population. In total, >440 000 men and women with 13 common cancer forms are included in the analyses. The data allow a control for educational differentials in tumour stage at time of diagnosis. The remaining effects of education thus reflect differentials in the cancer patients’ general health status at diagnosis or health behaviour afterwards, or in treatment. Possible reasons for these differentials will be discussed, but with the data available, nothing can be done to assess their relative importance except checking the impact of adding marital status, which is among the potentially mediating variables.


All cancer cases in Norway have been registered by the Norwegian Cancer Registry from 1953 onwards.11 This study is restricted to the 441 556 women and men who were 30–89 years old when they were diagnosed with a first tumour of one of the following 13 forms in 1970–2007: stomach, colon, rectal, pancreatic, lung, breast (females only), cervical, uterine, ovarian, prostate, or bladder cancer, malignant melanoma or central nervous system tumours.

Data on marital status as of January 1 every year since 1970, date of death (if any) and dates of immigration and emigration (if any) were extracted from the Norwegian Population Register, complete from 1964 onwards, and linked to the cancer data by means of unique individual identification numbers after ethical review by the Norwegian Board of Medical Ethics. The highest educational attainment as of January 1 each year since 1980 was added from the Education Register operated by Statistics Norway. Educational attainment prior to 1980 was extracted from the 1970 census. The cause of death was obtained from the Norwegian Cause-of-Death Register.

Discrete-time hazard models were estimated; a common and convenient type of survival analysis. For each individual, a series of 3-month observations was created, starting at the time of diagnosis and ending at the end of 2007 or when the person died, had lived 10 years since diagnosis or emigrated, whichever came first. (Using one-month observations gave the same results.) Each observation included a number of variables that referred to the situation at the beginning of the 3-month interval, and the outcome variable was death from the cancer type under consideration within the 3-month interval (i.e. the so-called ‘corrected survival’ approach). If the person died from another cause, the observations were censored at that time. Observations were excluded if the person did not live in Norway at the beginning of the interval, and logistic models were estimated from the remaining observations (using the Proc Logistic in the SAS software version 9.2), separately for women and men. A total of 113 906 deaths occurred within the 894 814 person-years of observation for men. The corresponding figures for women were 91 796 deaths within 1166 316 person-years.

Models were estimated for each of the 13 cancer types, but because there are relatively few cases of some of them, the analysis was largely based on models including all cancer forms. In the latter type of model, one should obviously control for cancer site, because the less educated may tend to develop other types than the better educated, with a better or poorer prognosis. In this study, for example, relatively many of the malignancies among those who have only compulsory schooling are lung cancers, for which the prognosis is poor.

All other covariates are also categorical. Five categories were defined for educational attainment: compulsory school (10 years according to the current school system, fewer for older cohorts), lower secondary education (11 years), upper secondary education (12–13 years), tertiary education up to and including the Bachelor level (14–17 years) and higher education (18+ years). Because the latter group was quite small, especially among women and at the start of the study period, the two groups with highest education were combined. Education was measured at the beginning of the calendar year of the 3-month observation or (for observations before 1980) in 1970, i.e. up to 10 years earlier (which is unproblematic given the relatively high age of the study objects). Tumour stage was classified into localized, regional spread, distant spread and unknown.

Another factor was age at the beginning of the 3-month interval, defined as age at the end of that calendar year. It was grouped into 5-year categories, running from 30–34 to 95–99 years. The calendar year of the 3-month observation (referred to simply as ‘period’ below) was grouped similarly: 1970–74, 1975–79, 1980–84, 1985–89, 1990–94, 1995–99, 2000–04 or 2005–07. Time since diagnosis was grouped into 10 one-year intervals.

Marital status was included in some additional analyses reported on briefly. It was defined as married, never-married, widowed or divorced/separated, and referred to the beginning of the calendar year of the 3-month observation.

Because the intention was to analyse changes in the effects of educational level, models were estimated separately for 5-year periods. In addition, an interaction between education and period (minus 1990, so the main effect of education can be interpreted as the effect in 1990) was included in some models to see whether there was a significant linear trend. In these models, also the other covariates were interacted with period to ensure that a change in the education effect did not merely reflect a changing importance of some variable correlated with education.


Table 1 shows estimates from a model that includes educational level, age, year, cancer site, time since diagnosis and stage. Generally, cancer survival has improved over the decades under study, and it improves with increasing level of education, for both sexes (as also reported in earlier studies, without any attention to time trends12,13).

View this table:
Table 1

Effects [with 95% confidence intervals (CI)] of socio-demographic factors and disease characteristics on the odds of dying from cancer among men and women diagnosed with 13 types of cancer at ages 30–89 in Norway 1970–2007a, and number of deaths and exposure time in the various categories

Effect 95% CI (odds ratio)Number of deathsNumber of 3-month observations
        1970–7418693129 750
        1975–790.93 (0.90–0.96)13 518267 382
        1980–840.81 (0.79–0.83)15 069361 980
        1985–890.63 (0.61–0.65)16 357422 973
        1990–940.53 (0.52–0.55)16 337481 531
        1995–990.46 (0.44–0.47)16 377561 331
        2000–040.35 (0.34–0.36)15 934667 958
        2005–070.27 (0.26–0.28)8901472 650
        30–34 years0.53 (0.47–0.61)21612 246
        35–39 years0.67 (0.62–0.73)57431 101
        40–44 years0.71 (0.66–0.76)103849 900
        45–49 years0.81 (0.77–0.85)215978 791
        50–54 years0.85 (0.82–0.88)4060131 634
        55–59 years0.93 (0.90–0.96)7114221 477
        60–64 years111 090347 443
        65–69 years1.10 (1.07–1.13)15 956485 747
        70–74 years1.27 (1.24–1.30)19 752623 305
        75–79 years1.58 (1.54–1.62)20 990631 130
        80–84 years2.07 (2.01–2.12)17 169469 591
        85–89 years2.76 (2.68–2.85)9650229 119
        90–94 years3.43 (3.23–3.64)137542 172
        95–99 years5.64 (4.26–7.45)531899
    Cancer site
        Stomach112 992155 748
        Colon0.35 (0.34–0.36)11 95617 431
        Rectum0.40 (0.39–0.41)7729272 293
        Pancreas2.98 (2.89–3.09)836641 087
        Lung1.70 (1.66–1.74)30 937254 946
        Prostate0.26 (0.25–0.26)25 9211 370 914
        Bladder0.30 (0.29–0.31)6472462 371
        Malignant melanoma0.32 (0.31–0.34)2899270 878
        CNS1.25 (1.20–1.30)3812110 266
    Tumour stage
        Localized134 2122 058 969
        Regional spread2.53 (2.49–2.58)39 51266 3815
        Distant spread7.67 (7.51–7.83)26 351219 460
        Unknown2.08 (2.03–2.13)11 111423 311
    Educational level
        10 years160443149 4813
        11 years0.91 (0.89–0.92)27315901 854
        12–13 years0.88 (0.86–0.90)11575430 191
        14+ years0.82 (0.80–0.84)11853538 697
    Time since diagnosis
        <1 year162826704 544
        1 year0.66 (0.65–0.68)19042492 481
        2 years0.50 (0.49–0.51)9861392 300
        3 years0.40 (0.39–0.41)6108324 144
        4 years0.35 (0.34–0.36)4305270 959
        5 years0.30 (0.29–0.31)3031229 862
        6 years0.27 (0.26–0.28)2342195 878
        7 years0.23 (0.22–0.24)1701166 274
        8 years0.21 (0.20–0.23)1326140 910
        9 years0.20 (0.19–0.22)1061119 375
        1975–790.89 (0.87–0.92)10 897388511
        1980–840.78 (0.76–0.81)12 012515077
        1985–890.62 (0.60–0.64)12 878576426
        1990–940.50 (0.49–0.52)12 862633453
        1995–990.42 (0.41–0.44)13 153718431
        2000–040.33 (0.32–0.34)13 145830 083
        2005–070.27 (0.26–0.28)7629555 427
        30–34 years0.64 (0.57–0.71)37339 701
        35–39 years0.73 (0.69–0.78)1100120 196
        40–44 years0.76 (0.72–0.80)2028205 603
        45–49 years0.76 (0.73–0.79)3346305 468
        50–54 years0.82 (0.79–0.85)5084407 423
        55–59 years0.92 (0.89–0.95)7177475 825
        60–64 years19212513 062
        65–69 years1.08 (1.05–1.12)10 937535 223
        70–74 years1.26 (1.23–1.30)13 116547 429
        75–79 years1.48 (1.44–1.52)14 257521 012
        80–84 years1.86 (1.81–1.92)12 837415 136
        85–89 years2.50 (2.42–2.58)8788240 845
        90–94 years2.49 (2.34–2.66)118756 263
        95–99 years5.26 (4.22–6.55)843645
    Cancer site
        Stomach18406107 729
        Colon0.36 (0.35–0.37)13 681514 405
        Rectum0.40 (0.39–0.41)5786238 757
        Pancreas2.85 (2.74–2.96)797140 748
        Lung1.60 (1.55–1.65)12 573118 929
        Breast0.21 (0.21–0.22)17 968169 8208
        Cervix0.27 (0.26–0.28)3637313 690
        Uterus0.22 (0.21–0.23)3125381 665
        Ovaries0.37 (0.35–0.38)8996318 192
        Bladder0.48 (0.46–0.51)2709142 568
        Malignant melanoma0.20 (0.19–0.21)1845350 666
        CNS0.76 (0.72–0.79)2829161 674
    Tumour stage
        Localized121 170264 2621
        Regional spread3.09 (3.03–3.15)38 1481 335 383
        Distant spread10.82 (10.57–11.07)23 675240 187
        Unknown3.18 (3.08–3.28)6533169 040
    Educational level
        10 years154 1332 126 969
        11 years0.92 (0.90–0.93)24 9911 381 340
        12–13 years0.81 (0.78–0.84)3752309 135
        14+ years0.81 (0.79–0.84)6650569 787
    Time since diagnosis
        <1 year146 769748 677
        1 year0.68 (0.67–0.70)16 235584 018
        2 years0.50 (0.49–0.51)8702499 573
        3 years0.40 (0.39–0.41)5723439 667
        4 years0.32 (0.31–0.33)3821391 824
        5 years0.26 (0.25–0.27)2767351 631
        6 years0.22 (0.21–0.23)2039316 819
        7 years0.20 (0.19–0.22)1664285 884
        8 years0.18 (0.17–0.19)1266258 419
        9 years0.15 (0.14–0.16)970233 888
  • a: The model included period, age, cancer site, tumour stage, educational level and time since diagnosis.

In Supplementary table S1, it is shown how education affects the odds of dying from cancer in separate 5-year periods. For men, the odds relative to those among men with only compulsory schooling have been stable for all educational categories. Among women, however, the point estimates suggest that the advantage compared with those with compulsory schooling has increased for all groups.

These patterns are supported by the trend analysis. For men, there is no significant linear trend in the effect (log-odds of dying compared with that among men with compulsory schooling) for any of the educational groups, while such a trend is seen for all educational groups among women (see table 2, which as opposed to table 1 and Supplementary table S1 shows effects on log-odds rather than odds). The most favourable trend is seen for women with upper secondary education: their log-odds of dying was 0.2067 lower than among women with compulsory education in 1990 (see the main effect of upper secondary education in table 2), and this advantage increases by 0.00807 each year, i.e. 0.0807 each decade. This means that the odds of dying was 12 per cent lower for this group than for those with compulsory schooling in 1980 [1 − е(0.2067 + 0.0807) = 0.118], 19 per cent lower in 1990 [1 − е(0.2067) = 0.187] and 25 per cent lower in 2000 [1 − е(0.20670.0807) = 0.250]. The educational differentials in death probabilities have developed in much the same way as the educational differentials in the odds of dying, as the overall level of these probabilities has not changed dramatically over the study period.

View this table:
Table 2

Effects [with standard deviations (SD)] of educational level and an interaction between educational level and period on the log odds of dying from cancer among men and women diagnosed with 13 types of cancer at ages 30–89 in Norway 1970–2007a

 Effects of educational level in 1990 (i.e. main effect of education)
        10 years0
        11 years−0.09430.0081P < 0.01
        12–13 years−0.12140.0112P < 0.01
        14+ years−0.19670.0111P < 0.01
 Annual additional effects of educational level (i.e. effect of interaction between education and period-1990)b
        10 years0
        11 years0.000490.00079P = 0.53
        12–13 years0.000170.00110P = 0.88
        14+ years0.000690.00108P = 0.52
 Effects of educational level in 1990 (i.e. main effect of education)
        10 years0
        11 years−0.08690.0084P < 0.01
        12–13 years−0.20670.0184P < 0.01
        14+ years−0.20120.0143P < 0.01
 Annual additional effects of educational level (i.e. effect of interaction between education and period-1990)c
        10 years0
        11 years−0.002120.00081P = 0.01
        12–13 years−0.008070.00179P < 0.01
        14+ years−0.005070.00139P < 0.01
  • a: The models also included age, cancer site, tumour stage, time since diagnosis, interactions between all these variables and period, and a main effect of period.

  • b: When marital status was added, the effects were 0.00065 (P = 0.40), 0.00054 (P = 0.62) and 0.00105 (P = 0.33).

  • c: When marital status was added, the effects were 0.00186 (P = 0.02), 0.00781 (P < 0.01) and 0.00449 (P < 0.01).

The Supplementary table S2 displays trends in educational differences in cancer survival (i.e. education–period interaction effects, just as shown in table 2, but not the main effects) for each of the 13 cancer sites. Among men, only a few of the trends are significant or close to significant, and point estimates vary between positive and negative, often within the same cancer site. Among women, many point estimates are negative, indicating a progressively stronger advantage associated with education, but significance is attained for at least one educational group for only a few of the sites.


The results show increasing educational differences in cancer survival among women, while there has been stability among men. While a number of mortality studies have also revealed increasing social inequality—though as least as much for men as for women (see elaboration below)—no attention has been paid earlier to the corresponding trends in cancer survival.

Prognosis following a cancer diagnosis is in general determined by (i) characteristics of the tumour and especially the stage at diagnosis, (ii) host factors, i.e. general health at time of diagnosis and health behaviour (including exposures to environmental hazards) afterwards and (iii) treatment received. Tumour stage at diagnosis has in some other studies been shown to explain part of the social differences in cancer survival.4 However, as tumour stage is adjusted for in this analysis, only the two latter mechanisms can contribute to the educational differentials observed and the change in these. Below, it will first be discussed how education may affect cancer survival through health (behaviour) and treatment, and then whether there may have been any change in the importance of some of these pathways that might explain the observed trends.

Reasons for educational differences in cancer survival in general

Education is thought to be positively associated with good health and health behaviour for a number of reasons: specific skills obtained in school may increase incomes; the better educated likely have more knowledge about health and higher problem-solving capacity; they may have a longer planning horizon and feel more responsible for own health; they tend to have a larger degree of control in their jobs; and they may have a feeling of having achieved something and therefore feel better when comparing themselves with others.14

In addition to affecting health, the knowledge, higher income and attitudes following from higher education may have implications for treatment. It is not impossible that a more appropriate regimen is set up if a patient is more actively engaged and has more knowledge about the alternatives, and adherence to treatment might differ also. There is not consistent evidence of such relationships, though.15,16 The patient’s economic resources are in some settings important, but probably less so where there is a public health care system, as in Norway.

A pathway of relevance both for health (behaviour) and treatment is that the higher income and level of knowledge among the better educated affect their chance of marrying and remaining married (though differently for women and men; see below). Marital status is an important determinant of health17 and may also have an effect on cancer treatment.18

Finally, selection is involved: high education partly results from parental resources, self-discipline, intellectual endowments and a relatively good health early in life,19 which also have implications for health (behaviour) later and possibly treatment.

Potential reasons for increasing educational differences in cancer survival

It is possible that the link between education and health behaviour has changed, with implications for the pattern in cancer survival. For example, individuals who are more trained to digest new knowledge may be the first to change behaviour in response to expert advice. The increasingly negative association between education and smoking is an illustration of that.20,21

Social differentials in cancer treatment may also have increased. To the extent that there is an association between treatment adherence and education, it may have become stronger because of a growing proportion of the treatment taking place in the outpatient setting, with increasing demands for patients’ own participation. Furthermore, it seems to be a common perception among health personnel that their workload is increasing. If that is the case, it is not impossible that physicians occasionally yield to pressure about giving a potentially better treatment that would usually be considered unaffordable, and that the most well-informed would exert the strongest pressure.

Such changes would accord with the sharper social gradient in mortality that has been reported, and that may reflect increasing advantages for the higher educated with respect to health, health care and health behaviour in general (and not only of relevance for cancer). Most studies of self-rated health have shown that there are no changes over time in the social differentials,22–24 but they have been based on much smaller data sets than typically used in mortality studies.

Interestingly, the mortality studies have—on the whole—not shown the particularly strong increase of the social gradient for women that appears in the present analysis. On the contrary, only one of the Norwegian investigations points in this direction,5 while the others conclude that the increase has been most pronounced for men.6–9 Thus, it may seem that some health or treatment factors with particularly beneficial effect on cancer survival have become more positively associated with especially women’s education over time than what is the case for the health and treatment factors that are important for general mortality.

One possible explanation for an increasing educational gradient for women in particular (though not necessarily of relevance only for cancer survival and not all-cause mortality) is that the effect of education on (household) income and marital status has changed more for women than men, with implications both for health (behaviour) and treatment. Some decades ago, high education was negatively associated with marriage among women, while there is a non-negative or even positive association today (as there has always been among men).25,26 However, inclusion of marital status changed the results very little, for men and women alike (see notes to table 2). Furthermore, education had relatively little impact on married women’s (own) income in the past, as they often stayed home regardless of education. Today, women with high education are not only more likely to be married (to a resourceful man), but also contribute more to the household income.

One could also suspect that social gradients for some reason have increased more (less) for cancers that only occur among women (men), but this seems not to be the explanation for the observed sex differences in overall cancer survival. The same pattern appeared when only cancers that can occur in both sexes were included (not shown).

The estimates from site-specific models (see Supplementary table S2) do not give many clear clues to understanding the sex differences in the development of the educational gradients in overall cancer survival. However, there is an opposite development in the educational differences in lung cancer survival for men and women, which might indicate a more strongly increasing social gap in smoking for women, as survival from lung cancer is much influenced by smoking.27,28 Obviously, such a pattern in smoking would have broad implications for health. On the other hand, studies of social gradients in smoking in Norway have given mixed evidence, and suggest on the whole that the gap has widened no more for women than men.21,29

The educational distribution has changed markedly over time. For example, while 63% of the 1970–75 observations were in the lowest educational category and 7% in the highest, the corresponding proportions for 2005–7 were 34 and 22% (varying across age, e.g. 23 and 38% at 40–45 and 45 and 14% at 80–85). Given this educational expansion, those at the lowest level will, of course, fall increasingly behind the average, while those at the highest will come closer to it. This relative position in the hierarchy may matter for cancer survival, and in principle changes over time in the education effect could depend on whether the relative or absolute level is considered. Supplementary analysis was therefore done with an education variable defined for each sex and period as the average of the proportion having lower education and the proportion having lower or the same education (i.e. the same kind of variable as used when computing the Relative Index of Inequality). Using an interaction model such as described earlier, the main effect of this relative education was −0.258 (P < 0.0001) for men and −0.273 (P < 0.0001) for women, while the interaction effect with period was 0.001 (P = 0.62) and −0.090 (P < 0.0001), respectively. In other words, the same picture appears when a measure of relative education (taking into account the category sizes) is used.

Changes in the sizes of the categories do not have to, but can, be a result of forces also producing changes in their composition. For example, educational policies and a general economic growth have contributed to the educational expansion, and have probably also made individual economic resources a less important determinant, so that, for example, the less educated to a larger extent consists of people with little self-discipline or low intellectual endowments, which could be more important for cancer survival. However, it is not easy to see how such a change over time in the selection factors could vary between sexes. Unfortunately, such problems related to changing selection cannot be solved with the available data.


There are no obvious reasons to expect an increasing educational gradient in cancer survival for women more than for men, or even restricted to women, as is the case according to this analysis.

The results may be seen as running counter to those from mortality studies, many of which have shown the sharpest increase in the social gradient for men or little difference between the sexes. However, mortality is influenced by a number of health factors and various aspects of medical care that may play a different role for cancer survival, so the analyses are not strictly comparable.

Whatever the explanation for the sex differences, this study adds to the concerns about increasing social health inequalities. Any attempt to reverse this trend through good policies would have to be built on good knowledge about the underlying factors, including the mechanisms through which education affects (or at least is linked with) cancer survival—apparently in different ways for women and men.

Supplementary data

Supplementary data are available at EURPUB online.


Thanks to Øystein Kravdal who prepared the data for analysis and gave much advice.

Conflicts of interest: None declared.

Key points

  • All-cause and cause-specific mortality, as well as cancer survival, are well known to be influenced by socio-economic status.

  • Several studies have in recent years discovered increasing social differences over time with respect to all-cause and cause-specific mortality.

  • This study investigates changes over time in educational differences in cancer survival, which has never been done before.

  • Increasing differences are seen for women, while there is stability among men.


View Abstract