Background: Tabulating annual national health indicators sorted by outcome may be misleading for two reasons. The implied rank order is largely a result of heterogeneous population sizes. Distinctions between geographically adjacent regions are not visible. Methods: Regional data are plotted in a geographical map shaded in terms of percentiles of the indicator value. Degree of departure is determined relative to control limits of a corresponding funnel plot. Five methods for displaying outcome and degree of departure from a reference level are proposed for four indicators selected from the 2004 European Perinatal Health Report. Results: Spread of indicator values was generally largest for small population sizes, with results for large populations lying mostly close to respective European medians. The high neonatal mortality rate for Poland (4.9 per 1000); high low-birthweight rates for England and Wales (7.8%), Germany (7.3%) and Estonia (4.5%); and high caesarean section rates for Italy (37.8%), Poland (26.3%), Portugal (33.1%) and Germany (27.3%) were statistically significant exceptions to this pattern. Estonia also showed an extreme result for maternal mortality (29.6 per 100 000). Conclusion: Extreme deviations from EU reference levels are either correlated with small population sizes or may be interpreted in terms of differing medical practices, as in the case of caesarean section rate. EURO-PERISTAT has now decided to use 5-year averages for maternal mortality to reduce the variance in outcome. Use of two colours in three intensities and solid fill versus crosshatching is best suited to display rate and significance of difference.
The chief objective of any collection of indicators of health for a number of countries will be to highlight and interpret differences found between countries, with the ultimate aim of suggesting possible causes associated with risk factors or health care. Current applications of international comparisons of health and health care are published on the websites of EUROSTAT1 and WHO2 and in the European Community Atlas of Avoidable Death.3 In addition, a wide range of commercially available databases and programs can be found on the Internet.
In the perinatal field, each country collects data used to derive indicators through civil registration, hospital discharge systems or specialist registers. The indicators include mothers’ and children’s health, clinical practices and maternal risk factors. Thus, in the European Union, all countries produce sufficient data to construct at least 10 core indicators, to describe their situation.4 Comparison of these indicators across countries is an essential tool for defining national policy priorities, or generating hypotheses on factors that might explain differences such as risk factors or health policies. To achieve this aim, it is essential to display data in a way that is simultaneously accurate as well as informative without being too cumbersome and voluminous to digest.
The recent claim that the validity of international comparisons is compromised by variations in the registration of births at the borderline of viability5 may hold for some publications, although not so for the European Perinatal Health Report,6 as care was taken to ensure comparability by appropriate censoring of indicators at the lower end of the gestational age distribution.7 After due consideration and suitable adjustment for methodological difficulties8 and availability and comparability issues,9 however, there remain questions of presentation of these data, e.g. how to order a table or what to display on a geographical map.
A tabular presentation is attractive because it can provide a large amount of numerical information such as rates for each country or population size and because of its naturally inherent ranking option. An alphabetical ordering such as adopted in the European Perinatal Health Report6 may be considered objective, as reporting sequence is solely a function of lexical ordering. However, this objectivity is at the expense of any form of structured data display that might help interpretation. On the other hand, the ranks of countries in terms of observed rates will be instable10,11 because absolute differences between observed rates are frequently small. An instable ranking may lead to misinterpretation of observed raw outcome rates.
Longitudinal variation is best displayed graphically. Although graphs of temporal trends can reveal valuable additional information, we cannot always expect data to be available for long time series. This is so especially for the more sophisticated perinatal indicators. These may not be available for all countries on a yearly basis.
Funnel plots of rates against population size, as pioneered by Macfarlane12 and later adopted for the National Health Service more generally by Spiegelhalter,11 are increasingly used in performance monitoring. A major advantage is the inclusion of a formal assessment of degree of departure from specified reference levels with simultaneous display of outcome–volume relationships. Funnel plots were recently proposed for monitoring and assessing improvements13,14 in surgical interventions in Germany.
The plotting of performance indicators for countries in a geographical context is possible because of the availability of graphical information software. This encourages and facilitates regional interpretation. Correlations between rates in adjacent states are immediately apparent3,15. Holland3 suggests a number of steps that might be taken in studying geographical variation. One is to see how these differences change over time; another is to map individual deaths and to investigate clustering.
In this article, we take these ideas one step further and suggest moving the information contained in funnel plots into choropleth maps to improve their legibility and enhance their interpretability. The proposed method is applied to perinatal health indicators from the European Perinatal Health Report.7
The EURO-PERISTAT6 data are available for 25 European Union (EU) member states plus Norway. Data to construct a number of indicators were compiled for the year 2004, in most cases at national population–based level. In the absence of population-based data, data were provided on a survey basis for the province of Valencia in Spain and for France. Population-based data for the Brussels and Flanders regions of Belgium, for the countries of Northern Ireland, Scotland and England and Wales jointly were available separately, thus yielding 25 European countries/regions for analysis. The data were mainly obtained from civil registration based on birth and death certificates or from other national sources, including specific perinatal databases.
Selection of indicators
To explore the effects of various forms of presentation, four indicators were chosen from the EURO-PERISTAT set of ‘core’ indicators for 2004. They were chosen because they differed with respect to their average rate and with the further requirement that comparable observations were available for the majority of countries. In ascending order of mean incidence, these are maternal mortality (ratio of number of pregnancy and childbirth-related maternal deaths to live births) with a ratio of ∼4–6 per 100 000 live births, crude neonatal mortality (deaths at 0–27 days after live birth) with a rate of ∼3–5 per 1000 live births, birthweight below 2500 g among all live and stillbirths with a rate of ∼6% and caesarean sections (all elective and emergency caesarean sections combined) with a rate of ∼15% per total number of births.
Owing to its low overall incidence, maternal mortality was compiled for the years 2003–2004 combined, except for Germany (2004), Italy (2002) and Luxembourg (2000–2004), to increase the stability of these estimates. For countries with extremely small population sizes such as Estonia and Malta, only 2-year averages were available. No data were available on maternal deaths from Cyprus, Ireland and Slovakia. No data were available for neonatal mortality for Ireland, which only provided data on early neonatal mortality (deaths at 0–6 days after birth), whereas Cyprus provided no data on the distribution of birthweight, and no data were available on mode of delivery for Greece.
Throughout, the data sets are defined by observations (ci, ri and ni) denoting country, number of cases in the numerator and denominator, respectively, with the subscript i running through all 25 countries/regions. The raw rates are computed as yi = ri∕ni with 0 < yi < 1.
In addition to a tabular summary, standard choropleth maps and funnel plots were constructed for these four selected indicators of perinatal health. An enhanced choropleth map was constructed for caesarean section rates. This indicator was chosen because of its status as a core indicator, its availability in all but one country and its good quality of documentation.
A table of denominators ni, observed rates yi and p-values indicating significance of departure from respective EU medians y0 was constructed in alphabetical order of country. The median was selected to prevent large countries from dominating the comparisons. The p-values are computed as pi = Φ−1(zi) from the inverse normal distribution function where
Standard choropleth maps
We plot the observed rates using colour coding derived from six percentiles of the distributional characteristics of the outcome measures and with upper limits at 10, 25, 50, 75, 90 and 100. We chose shades of the neutral colour blue, with dark shades corresponding to high rates, in keeping with the examples in the European Perinatal Health Report.6
The maps were constructed with the SAS GMAP16 procedure. The perimeters of country boundaries were extracted from the SAS/MAPS16 database and linked to observed rates through the country identifiers ci. Version 9.1 of SAS16 provides data at national level. Subsets of the database corresponding to the country of Scotland and the Flanders and Brussels region of Belgium were generated by straight-line dissection of existing national coordinate files. Countries for which no data were available are left blank.
Funnel plots were constructed using R17 by plotting the observed outcome against population size as a proxy measure of precision. Control limits were drawn vertically around the median rate for all countries. To assess degree of departure from the median rate y0, statistical control limits were computed from the standard normal approximation to the binomial distribution as
for α levels of 5.0 and 0.1% and 0 < y0 < 1.
Enhanced choropleth maps
Two basic options for indicating significance of difference from a reference level were considered, either by superimposing suitable symbols or by application of different shading patterns. The five variants of these two approaches are listed in table 1 and figure 1. Whereas colouring schemes A, B and C use a neutral colouring of percentile bands, schemes D and distinguish deviations above and below the reference by different colours. Schemes A and B use superimposed symbols, schemes C and E differentiate by pattern and scheme D uses bold perimeters to indicate significant differences. A significance level of 0.001 was used throughout.
Colouring schemes for enhanced choropleth maps: Schemes A, B and C use three darker intensities of blue shading above and three lighter intensities (two of which are shown) of blue below the median. Significance is indicated by appropriate symbols superimposed in schemes A and B, by solid colouring in schemes C and E and by bold perimeters in scheme D
Choices of colouring schemes for enhanced choropleth maps
Shades of blue
Shades of blue
Shades of blue
Shades of red and green
Shades of red and green
Plus and minus symbols
Up and down arrows
Intensity of shading corresponds to respective percentile band. Darker shades of blue and red colours indicate values above the median, lighter shades of blue and green colours indicate values below the median. Extreme shadings correspond to values below the 10th or above the 90th percentile.
Table 2 shows denominators, rates and significance of departure from respective EU medians for the four selected indicators at regional or country level. The denominators vary considerably according to the demographic characteristics. The population of England and Wales is ∼160 times that of Malta and ∼50 times that of Estonia and Luxembourg. The median rates are 7.3 per 100 000 for maternal deaths, 6.7% for the proportion of births below 2500 g, 2.7 per 1000 live born for neonatal deaths and 20.5% for caesarean sections.
Denominators (n), regional rates and P-values for EU member states for maternal deaths (MD) per 100 000 live births, birthweight under 2500 g (LBW) as percentage of total births, neonatal deaths (NND) per 1000 live births and caesarean sections (CS) as percentage of total births
England and Wales
1 261 190
1 529 280
The p-values indicate significance of difference from respective EU medians.
Figure 2 shows a corresponding panel of standard choropleth maps, with intensity of blue corresponding to percentile band for all four indicators. Countries for which no data were available are left blank. The percentile grading is reflected in the broader ranges at the extremes of the distributions. Some of the subranges are not contiguous owing to their automatic generation by SAS/GRAPH software16 based on the actual distribution of observed rates.
Choropleth panel plot of selected EURO-PERISTAT indicators. Colour grading is by quintiles of outcome rate. Maternal mortality is based on the period 2003–2004 for the majority of countries (see text)
Figure 3 shows the corresponding panel of funnel plots. In the case of maternal mortality, most countries fall within the control limits. Estonia stands out with eight maternal deaths in 27 028 live births between 2003 and 2004 (29.6 per 100 000 live births vs. 0.0–12.2 for the remaining countries/regions). For neonatal mortality rates, the majority of smaller countries (with <150 000 births) lie within the limits. Although there are several other countries with high neonatal mortality rates, all new member states of the European Union as well as Poland stand out due to its large population. The distribution of low birthweight shows countries largely falling either below, within or above the control limits, irrespective of population size. In the case of caesarean sections, only one, Denmark (DK) alone, lies close to the European median rate. All other countries deviate considerably from the median, again seemingly irrespective of population size. Italy stands out due to its high caesarean section rate and its large population size.
Funnel plots of indicators shown in figure 2 against total number of deliveries for European countries and regions with control limits drawn at 95.0 and 99.9%. Some of the axes exclude zero to enhance readability. For maternal mortality, the x-axis is on a wider scale due to the inclusion of the period 2003–2004 for the majority of countries (see text)
Figure 4 shows an enhanced choropleth plot for caesarean sections using scheme E from figure 1. The width of the upper percentile band is ∼5 times that of the lower band as a result of the skewed distribution, with Italy and Portugal showing comparatively high caesarean section rates. All countries except Denmark, Latvia and France show significant differences from the EU median rate of 20.5% (table 2). Twenty-two countries deviate significantly from the EU median, 11 above and 11 below.
Enhanced choropleth plot for caesarean section rates using colouring scheme E. Regions with solid-fill patterns differ significantly from the EU median rate of 20.5%. The abundance of such countries indicates a bimodal distribution with evident grouping of rates below and above the median
Standard choropleth maps
The chief advantage of choropleth maps lies in making regional patterns transparent. At a glance, one can discern the strong contrast in neonatal mortality rates between the Eastern European states and Nordic states or the high caesarean section rates in Italy, Germany and Portugal (figure 2). Displays such as these were used for reporting perinatal indicators,7 and joint inspection may help to generate hypotheses about possible common causes for rates in selected regions.
Panel plots are also well suited for comparisons across different performance indicators, which are highly correlated. For instance, one may see that the Nordic states have lower rates for caesarean sections, for low birthweight and, with the exception of Finland, for maternal mortality. A further advantage lies in the information contained in geographical size of the region. This, albeit rough, proxy measure of population size can be of help in interpreting the results.
Six categories were chosen for grading the outcome rates in terms of distribution percentiles for all performance indicators to differentiate countries with respect to the median. Although this indicates distinct categories of absolute rates, it allows no conclusion about the significance of departure of this rate from a reference value. This can be misleading, especially for small countries. It must be borne in mind that although identically coloured regions imply identical percentiles across different choropleth maps, this will not in itself indicate how far an individual rate differs from the mean or median.
This shortcoming is remedied with funnel plots. They reveal a considerable heterogeneity in spread not only between countries but also between selected indicators of perinatal health (figure 3). This strongly suggests non-random causes for the differences in perinatal health, which is particularly marked for low birthweight and caesarean sections. Presenting indicators of perinatal health in a panel facilitates the generation of hypotheses about differences in the provision and effectiveness of health care. This form of presentation also makes the fact that population size differs between units more visible. EURO-PERISTAT has now decided to collect maternal mortality for a 5-year period because of the difficulties associated with low sample sizes. This is expected to result in a lower maternal death ratio for Estonia and conversely a higher ratio for Malta, where data were available only for 2 successive years. When data are presented as rates, it is easy to forget how few maternal deaths there are.
Choice of pattern
Although initially favoured due to their widespread use, quintile grading was not considered, as this would lead to ambiguous colouring and shading in the case of countries within the central quintile deviating significantly in opposite directions. Symmetrically arranged percentiles were selected, with a focus on extreme lower and upper 10%. The use of superimposed symbols may lead to distortions and graphical difficulties for small countries such as Luxembourg. In scheme A, there is an inherent bias towards higher rates because of the larger appearance of the ‘plus’ sign, which is not present in scheme B. Although initially appealing, the discrimination of significance in terms of thickness of country perimeters has the serious disadvantage that bold black borders are less distinct for regions with darker shadings. To avoid this bias and also because of obvious problems for small regions, scheme D was dropped. In the trade-off between colour bias (scheme E) and difficulty in readily identifying countries above from those below the median (scheme C), the final choice fell on scheme E.
Interpreting an enhanced choropleth map
Figure 4 illustrates the advantage of enhancing a standard choropleth map with colouring scheme E applied to caesarean section rates. The distribution of caesarean section rates is noteworthy, as one observes that most countries have either significantly high or significantly low rates, with only Denmark, Latvia and France in between, reflecting a bimodal distribution. It is attractive to use these data to entertain a hypothesis about patterns of mode of delivery across Europe. Local choices made by pregnant women and clinical custom varying across countries as well as interactions between these factors may well account for the observed pattern.
The EURO-PERISTAT project is funded by the Public Health Programme of the European Commission, Directorate General of Public Health (Agreement 2010 13 01). The Directorate General of Public Health had no role in the collection of the data, the writing of the manuscript or the decision to submit for publication. Data collection for Estonia was supported by grant numbers EE ESF SF0130018s11 and EE ESF 8325.
Conflicts of interest: None declared.
Tabular summaries of rates, population sizes and z-scores provide a maximum of statistical information, while being unwieldy and not easy to read.
Funnel plots also contain most of this information and enable rapid identification of outliers; however, they do not provide information on regional patterns.
Choropleth plots with enhanced patterns where colour reflects polarity of a country from a reference level, intensity of colouring to indicate size of deviation and solid fill versus crosshatching to indicate significance of difference can add valuable additional information and facilitate the interpretation of regional data.
As customary geographical maps of indicators of perinatal health tend to exaggerate differences between regions, the indication of significance of such differences is expected to improve monitoring and assessment of public health policy and practice.
The EURO-PERISTAT Scientific Committee: Gerald Haidinger (Austria), Sophie Alexander (Belgium), Pavlos Pavlou (Cyprus), Petr Velebil (Czech Republic), Jens Langhoff Roos (Denmark), Luule Sakkeus (Estonia), Mika Gissler (Finland), Béatrice Blondel (France), Nicholas Lack (Germany), Aris Antsaklis (Greece), István Berbik (Hungary), Sheelagh Bonham (Ireland), Marina Cuttini (Italy), Janis Misins (Latvia), Jone Jaselioniene (Lithuania), Yolande Wagener (Luxembourg), Miriam Gatt (Malta), Jan Nijhuis (The Netherlands), Kari Klungsoyr (Norway), Katarzyna Szamotulska (Poland), Henrique Barros (Portugal), Mária Chmelová (Slovak Republic), Živa Novak-Antolic (Slovenia), Francisco Bolúmar (Spain), Karin Gottvall (Sweden) and Alison Macfarlane (United Kingdom). Further acknowledgements to contributors to the 2008 European Perinatal Health Report can be found in the accompanying supplementary file.
. Some statistical approaches to comparisons between perinatal mortality rates for small areas. In: ChalmersI, McIlwaineG, editors. Perinatal Audit and Surveillance: Proceedings of the Eighth Study Group of the Royal College of Obstetricians and Gynaecologists 1980; 274–92.
. Qualitätsvergleiche mit Funnelplots - Plädoyer für eine einheitliche Methodik (Comparison of quality by means of funnel plots a plea for a uniform methodology). Zeitschrift für Evidenz, Fortbildung und Qualität im Gesundheitswesen 2009;103(Issue 8):536-41.
. Wirksamkeitsanalyse externer Qualitässicherungsmaßnahmen anhand von Veränderungen in Qualitätskennzahlen einzelner Krankenhäuser (An analysis of the effectiveness of external quality assurance programmes using changes in quality indicators of individual hospitals). Zeitschrift für Evidenz, Fortbildung und Qualität im Gesundheitswesen 2010;104(Issue 6):503-11.
NicholasLack, BeatriceBlondel, Ashna D.Mohangoo, LuuleSakkeus, ChristineCans, Marie H.Bouvier-Colle, AlisonMacfarlane, JenniferZeitlinEur J Public Health(2013)cks176DOI: http://dx.doi.org/10.1093/eurpub/cks176First published online: 7 January 2013 cks176