Mortality Atlas of Cancer and Other Causes.
Spain 1975-1986.
MATERIAL and METHODSDeath data have been taken from the Death Statistics Bulletins (Boletines Estad¡sticos de Defuncion), as supplied by the National Statistics Institute (Instituto Nacional de Estad¡stica, INE) on 1/2" magnetic tape. Of the variables listed in the Bulletins, analyses were run on data relating to gender, age, date of death, province of residence and underlying cause of death in line with the 9th revision of the International Classification of Diseases (ICD-9). For External Causes, the complementary classification of external causes of injuries and poisonings was used.
Although this is essentially an Atlas of Cancer Mortality, it includes 18 rubrics which do not correspond to this pathology. There are two reasons for this: 1) to take into account the possible influence and distribution of competitive causes of death (6) (e.g., lung cancer versus myocardial infarction); and 2) to ascertain mortality distribution for other causes and large groups of pathologies.
Population data, broken down by age and gender for Spain's 52 provinces, were obtained from the 1975 and 1986 municipal population registres and the 1981 and 1991 census figures. The population corresponding to each study period has been reckoned by using a polynomic regression model which gives the 1981 census a weighting double that of the remaining pivots (7).
Person years of exposure, in the calculation of age- and sex-specific rates for the entire calendar period (1975-86), were obtained by multiplying the population estimated as of 31st December 1980 (midpoint) by 12. The years under study were grouped in 3 four-year periods, with the provincial populations being estimated as of 31st December for the years 1976, 1980 and 1984.
A fundamental problem plaguing cartography is that of how to graphically depict spatial data in the most accurate and correct way, while simultaneously affording visual appreciation of geographical patterns. The problems posed range along three fronts: scales, indicators and the graph per se (colours, dimensions, etc.) (8). Not only have diverse statistical solutions been proposed to deal with the problem of indicators referring to unequal population sizes (9,10), but powerful multivariant techniques are now available which -in addition to assessing geographical differences- make it possible to study temporal shifts and detect interactions (11,12).
With respect to scales (absolute or relative), there is no preference for any one in particular. However, an accepted criterion does exist as regards the undesirability of solely representing the statistical significance of the indicator values (13). The prime aim of these maps is to detect patterns, and any representation of statistical significance could hide these.
The areas depicted are imposed by Spain's administrative grid, it proving difficult to study smaller areas since population composition at sub-provincial levels is not always accessible. Results are shown in the form of theme-specific maps, graphs and tables. For each cause and gender, 2 maps and one histograms are provided.
The first map corresponds to adjusted mortality rates for the whole period for all 52 Spanish provinces. These are rates
adjusted by the direct method using the standard European population. In this map a relative scale, having 7 classes or intervals, has been used, with provinces being grouped according to the
following criteria: after rates had been ranking in descending order of importance, 3 provinces were allocated to the most
extreme intervals, 5 to the next in, 10 to the next and 16 (the remainder) to the central interval. Class intervals were
colour-coded in three shades of red for rates exceeding that of the central interval, yellow for intermediate rates and three
shades of green for rates lower than that of the central interval.
Color scale and categories : 
The second map shows provincial time trends. Annual changes in provincial mortality have been calculated, taking the first study
period (1975-1978) as reference. For this purpose, use was made of a Poisson regression model that included an interaction term
between province and period of death. In this map, nine fixed cut-off points were selected, representing relative rises in risk
(and equivalent decreases on the logarithmic scale) of 1%, 3% 5% and 10% p.a. The colour scale denotes shades of magenta for
increases and blue for decreases.
Color scale and categories: 
Account has been taken of Poisson overdispersion in the statistical evaluation of the annual risk change (14).
Mortality distribution is shown in one histogram that presents the provinces and rates ranked in descending order. In order to interpret this histogram, attention should be paid to rate ranges and male/female differences. It was precisely with this in mind that the histograms for both sexes were placed on the same page and drawn to the same scale. Wide variability in rates is indicative of the greater role of environmental factors in the disease etiology in question. Gender-based disparities generally point to the presence of exposure differentials (Maps Guide).
The Poisson regression model is the tool best-suited to geographical studies using rates based on grouped data. In our study, the independent variables are age, year of death and province of residence and the dependent variable, deaths. Deaths in each stratum are deemed independent Poisson variables, with a mean and variance equal to the number of deaths in that particular stratum.
The AGE variable was classified into 5 levels excluding the under-20s (1 "20-44"; 2 "45-54"; 3 "55-64"; 4 "65-74"; 5 "75 and over") and the PERIOD variable into three (0 "1975-78"; 4 "1979-82"; 8 "1983-86"), the latter being included in the analysis as a single variable in order to reduce the number of interaction terms in the model. The way in which the period variable is coded allows for direct estimation of the annual relative increase or decrease. 51 indicator variables were used for the PROVINCE variable.Weighted provincial estimators were averaged to unity, which is equivalent to taking the Spanish average as the reference level . One model was adjusted for each sex.
Use has been made of a simple statistical method which enables one to answer the question of whether clustering could have occurred by chance (15,16) This method is based on pair-wise observation of contiguous provinces and measuring the differences between them in the rate rankings. For this purpose, a value D is calculated which is the average difference in the ranking of all K pairs of adjacent provinces. This gives an idea of clustering, i.e., if D is very small then provinces with similar levels of risk lie near one another..
