Stability of Age Standardised Rates for Deaths at Small Geographies and Use of the 95+ Age Group Background The European Standard Population (ESP) is an artificial population structure which is used in the weighting of mortality or incidence data to produce age standardised ratesi. Eurostat, the statistical institute of the European Union, has decided to bring this population structure up to date. ONS has consulted on how to implement the change in the UK, on behalf of the Government Statistical Service (GSS) as a whole. This paper considers data for England and Wales only, but its conclusions are likely to hold for the UK as a whole. The ESP is used to calculate age standardised rates (ASRs), which allow comparison of rates, including those across populations that may have different age and sex structures. This paper focuses on the ASR for deaths. The terms ASR and ASR for deaths are used in this paper to mean age standardised rates for deaths. Different authors use different conventions and these rates may be known variously elsewhere as age standardised mortality rates – ASMRs or age standardised death rates - ASDRs. Furthermore, some authors may use ASRs to mean age specific (death) rates. Rates may be calculated for a range of events, including deaths, cancer diagnosis, cancer deaths: this paper uses the term ‘events’ when making general comments that apply to different sorts of events, and ‘deaths’ where specifically deaths are concerned. Size of different area types This document focuses on small areas; Low and Middle level Super Output Areas. Other work has been undertaken looking at larger areas. Table 1 below shows the number, minimum, maximum, and mean population size (2011) for the different types of areas (in England and Wales) as context to this work. Table 1: Mid-2011 populations by type of area Area type Lower level Super Output Area (LSOA) Middle level Super Output Area (MSOA) (1) Local authority Parliamentary constituency Clinical Commisioning Group National Park No. of Areas 34,753 7,201 348 573 211 13 England and Wales total Minimum 987 2,224 2,224 56,543 61,607 1,994 1 Population Mean Maximum 1,616 7,800 161,411 98,030 251,693 31,022 8,159 16,439 1,074,283 159,456 863,433 112,492 56,170,927 Source: ONS, mid-2011 population estimates (1) Lowest value is for Isles of Scilly, which is a 'pseudo-MSOA'. Note: all areas are for England and Wales except for CCGs (England only) Data are internally consistent. One or more whole LSOAs make up a MSOA so data can be exactly aggregated from LSOA to MSOA. Similarly, one or more whole MSOAs make up a local authority. Local authorities cover, and can be added to give, the whole of England and Wales. Parliamentary constituencies are on a different set of boundaries, but data for these areas can be added to give data for England and Wales. Similarly Clinical Commissioning Groups can be added to give totals for England (not Wales). National Parks, however, only cover part of the country. Further information on area types (referred to as geographies) is available on the ONS websiteii. Introduction This paper looks at the stability of ASRs for deaths at small geographies, specifically lower level super output areas (LSOAs) and middle level super output areas (MSOAs). In particular it looks at the advisability of using upper age groups in the calculation of rates of 85–89, and 90+, compared with 85–89, 90–94, and 95+. The ESP has an upper age of band 95+, but suggests “... caution should be applied because of the potential quality issues of the data at older ages.” Variability of Age Standardised Rates Number of deaths There is natural variability in the number of events in a particular area. For a given area, with the same underlying rate, the number of events will change from one period (year) to another due to chance alone. The variability in ASRs for deaths can be assessed by calculating the standard error for a particular areaiii. To assess the likely stability of the rate the relative standard error (RSE) can then be calculated. For example, consider an area with a stable underlying death rate: For one year the ASR for deaths is calculated as 17 per 1,000. The standard error is calculated as 3.4. The RSE is therefore 20%. The 95% confidence interval is ±1.96 x 3.4 ≈ ±6.8 or 10.2 to 23.8 deaths per 1,000. Rates based on 20 events (when summed across all age bands) typically have an RSE of about 22%, below this level of events the size of the RSE starts to increase quickly. For this reason, rates based on fewer events may be considered to be unreliable, that is natural variation will have too large an impact to give stable and reliable rates.iv This is consistent with ONS policy on standardised rates that says that rates based on fewer than 20 events across all age-bands, if quoted, should be labelled as potentially unreliable. This guideline is used to inform this paper. Population size Population estimates are subject to a degree of uncertainty. Unlike deaths, the population in an area is typically likely to be fairly stable from year to year. However, there is a degree of uncertainty in the estimates of population size. This uncertainty is higher for the more elderly age groups where estimation of the population size is more difficult; small population sizes in these age groups (relative to the number of events) means that any source of error in the estimates is likely to have a proportionately larger impact. Data used This paper is based on the following data: ONS published LSOA mid-year population estimatesv Research figures for the eldest population age groups, that is splitting the 90+ into 90 to 94 and 95+, produced by ONS in order to meet requirements for the ESPvi Calendar year deaths data from the ONS Mortality Analysis team Three years of data are used (2010, 2011, and 2012). Three years of data are used in order to give enough deaths to enable calculation of rates. In some years seasonal increases in the number of deaths may start before the end of the calendar year, in others this increase may occur later and into the next calendar year. Using three years of data also helps reduce the impact of these timing issues. LSOAs There are 34,753 lower level super output areas, LSOAs, (2011 boundaries) in England and Wales. Zero population estimates Even aggregating over three years there are still a considerable number of LSOAs that have one or more age groups with no population, particularly when the older, 95+ age band, is used (Table 2). Table 2: Number of LSOAs with zero population in a quinary age group 2010 to 2012, by sex and oldest age used Males Females 90 + 95+ 701 274 16,741 6,839 Source: Office for National Statistics ASR for deaths can still be calculated for these LSOAs. Where an age group has zero population, a crude age specific death rate for the age-group cannot be calculated (this would involve a division by zero). Instead the local authority (LA) average rate for the relevant age group has been used in the calculations. This issue does not occur where there are zero deaths, as a rate of zero can be used in the calculation of the ASR for deaths. Estimates for age groups with zero population can occur for a number of reasons; importantly, the population numbers are estimates and subject to a margin of error. Other reasons where there may be a death with no population include (but are not limited to): a death occurring before the mid-year point leaving no remaining population (population is a mid-near snapshot estimate and deaths data are for the full year); no population recorded on NHS data (used in the calculation of these estimates) in either one or other of the 90 to 95 or 95+ age groups; and deaths (e.g. in a nursing home) where someone has moved from one area to another, but where they have not been resident long enough to update their address with the NHS. Summary statistics on LSOA age standardised rates for deaths Tables 3 and 4 below show some key summary statistics for ASRs for deaths for these areas, calculated using upper age groups of (separately) 90+ and 95+, and the difference between the two. Because of the number of LSOAs with zero population estimates for some age-groups the calculations have been replicated based only on those LSOAs that have population estimates in all age groups (Table 4). To enable fair comparison LSOAs are excluded from all the calculations, even if the zero occurs only in the upper age groups, and the 90+ calculations could be made without use of LA level data. Table 3: Age standardised rates for deaths (per 1,000 population) of LSOAs in England and Wales, summary data by sex, based on upper age groups of 90+ and 95+, using revised European Standard Population 1 Average SMR Males Females 90+ 11.3 7.7 95+ Change 13.2 1.9 9.2 1.5 1 Average standard error 90+ 3.1 2.1 95+ 3.5 2.4 Change 0.4 0.3 Average1 Relative Standard Error 90+ 0.34 0.29 95+ 0.33 0.27 Change -0.01 -0.02 Number of LSOAs 34,753 34,753 2 Source: Office for National Statistics Notes 1 Averages quoted are simple unweighted means, and therefore will differ from rates calculated at the national level. 2 Including up to 5 LSOAs for males and 30 LSOAs for females where ASR is calculated as zero, which are excluded from RSE calculation. Table 4: Age standardised rates for deaths (per 1,000 population) of LSOAs in England and Wales, summary data by sex, based on upper age groups of 90+ and 95+, using revised European Standard Population Excluding LSOAs where the population for one or more age groups is estimated as zero 1 Average SMR Males Females 90+ 11.1 7.6 95+ Change 13.1 2.0 9.1 1.5 1 Average standard error 90+ 2.7 1.9 95+ 3.3 2.2 Change 0.5 0.3 Average1 Relative Standard Error 90+ 0.24 0.25 95+ 0.25 0.25 Change 0.00 0.00 Number of LSOAs 18,012 27,914 2 Source: Office for National Statistics Notes 1 Averages quoted are simple unweighted means, and therefore will differ from rates calculated at the national level. 2 Including up to 5 LSOAs for males and 30 LSOAs for females where ASR is calculated as zero, which are excluded from RSE calculation. Key points on age standardised rates for deaths and standard errors Increasing the upper age range to 95+ tends to increase the size of the ASR. This occurs because the England and Wales population has a lower proportion (by sex) of those aged 95+ than the new ESP; thus the higher death rate in the 95+ age group is given a greater weight in the calculation. An implication of this is that it matters whether the 95+ age group is used or not. Further, when comparing rates they should be compared on the same basis, i.e. all using the same upper age band. The average standard error increases when calculated using the 95+ age group. However, this is somewhat misleading as the ASR for deaths also increases. Instead the relative standard error is a better indicator. If all LSOAs are included, this actually decreases. However, if only those LSOAs where no use of LA level data are included, the RSE remains virtually unchanged. Thus using the upper age-band does not affect the average relative variability of the ASR. The use of LA level averages has an impact on conclusions drawn about LSOAs, and care needs to be taken when drawing conclusions on rates with and without the use of LA level data. Related to this, the size of the RSE for all LSOAs taken together is larger than for only those that have no zero population estimates in age groups: this is because LSOAs with larger populations overall are less likely to have a zero estimate, and will also have less variation in the number of deaths from year to year. Overall only 46% of LSOAs have 20 or more male deaths over three years, and 43% have more than 20 female deaths. This is reflected in the RSEs of 33 or 34% for males, and 27 or 29% for females (depending on whether the upper age group is used). Even if only those LSOAs where there are no zero age group estimates are used, the RSE is still around 25%. Given this reliability, it should be concluded that there is too much natural variation in the number of deaths at LSOA for reliable conclusions about the underlying ASR for deaths to be drawn – ASRs for deaths for this level of geography should be treated with great caution, and are not suitable for general usage. The use of more years would help with this reliability – but will limit the usefulness or the data in other ways. Stability of LSOAs to the upper age band The previous section shows that using the upper age band tends to increases the ASR for deaths. It also shows that there is no indication that the relative variability of the ASR changes. However, this does not show that individual LSOAs show a stable pattern of change when the upper age band is used. The following charts (Figures 1 and 2) show, for males and females respectively, the range of change to the ASR when the upper age band is used, compared to those LSOAs where LA level data are not used. Figure 1: Increase in ASR for deaths when the upper age band is increased from 90+ to 95+, males, LSOAs where no age groups have a population estimate of zero Figure 2: Increase in ASR for deaths when the upper age band is increased from 90+ to 95+, females, LSOAs where no age groups have a population estimate of zero These charts show that there is a wide range in the distribution of the increase in the ASR rates for deaths when moving to the 95+ age group. This is consistent with degree of natural variation in the number of deaths for these small geographies. This variation supports the conclusion that ASRs for deaths for LSOAs should not be used for general purposes. A point of interest is the number of zero changes (most notable in the male histogram). This reflects that there are a number of LSOAs with no deaths at aged 85+ in the three year period, meaning that the ASR is the same irrespective of which upper age band is used. Significance and multiple testing We have observed above that increasing the upper age group from 90+ to 95+ tends to increase the ASR for deaths. It is therefore reasonable to ask the questions ‘is the increase for this area significant?’ and ‘is the increase for this area significantly different from the national average?’ Given there are 34,753 LSOAs, a number of them are likely to have a ‘significant’ difference due to chance alone. Sophisticated analysis can be undertaken to deal with the issue of multiple testing. However, as a crude starting point, sophisticated consideration would only be required if the number of ‘significant’ differences across all LSOAs starts to approach 5% of areas (if using a 95% confidence interval), that is 1,738 areas. Over all the LSOAs, 672 areas for males and 893 areas for females have an apparently significant increase in ASR for deaths when moving from 90+ to 95+. These numbers are significantly less than 5% of areas. Thus any apparently significant change in any particular LSOA should be considered with caution, as chance alone may well be the reason for the reason for the apparent significance of the change. Similarly, in 209 LSOAs for males and 557 LSOAs for females, there is again an apparently significant difference from the national increase, when moving from 90+ to 95+. Again these numbers are well below the 5% level. Consequently any apparently significant change when compared to the national increase should be considered with caution, as chance alone may well be the reason for the apparent significance of the change. Population denominators As set out above there is uncertainty in the population denominators used to calculate these rates, and this uncertainty is higher for smaller areas and for the upper age groups. It is not currently possible to quantify this uncertainty well. However, this uncertainty adds further weight to the conclusion that ASRs for deaths at LSOA level are not suitable for general usage. LSOA Summary ASR death rates for LSOAs are subject to too much natural variation to make them useful for general purposes. Uncertainty in measurement of population estimates adds to the weight of evidence against use of LSOA ASR for deaths for general purposes. There is no indication that use of the higher upper age group (95+) makes the ASRs more unreliable in relative terms. Using the 95+ age group tends to increase the calculated ASR rate for deaths. The increase in ASR for deaths, in moving to the 95+ age group, is not stable across LSOAs – it has a wide variation (that includes decreases in some LSOAs). Comparisons of ASRs should be made only if the same (upper) age bands are used. MSOAs There are 7,201 Middle level Super Output Areas (MSOAs) (2011 Boundaries) in England and Wales. MSOAs are exact aggregations of one or more LSOAs. Births and deaths for MSOAs have been calculated by summing the published data (see above) for LSOAs. This aggregation was undertaken using the lookup table published by ONS Geographyvii. The calculations and analysis described above for LSOAs have been repeated, on aggregated numbers of births and deaths, for MSOAs. Zero population estimates Aggregating over three years there are no MSOAs that have age groups with zero populations when the 90+ upper age group is used. However there remain a relatively small number when the older, 95+ age band is used (Table 5). Table 5: Number of MSOAs with zero population in a quinary age group 2010 to 2012, by sex and oldest age used Males Females 90 + 95+ - 285 25 Source: Office for National Statistics Most of these zero populations are in the 95+ age group, however 3 MSOAs have zeroes in the 90 to 94 age group for males (no MSOAs have equivalent zero estimates for females). ASR for deaths can still be calculated for these MSOAs. Where an age group has zero population, a crude age specific death rate for the age-group cannot be calculated (this would involve a division by zero). Instead the local authority (LA) average rate for the relevant age group has been used in the calculations. This issue does not occur where there are zero deaths, as a rate of zero can be used in the calculation of the ASR for deaths. Summary statistics on MSOA age standardised rates for deaths Tables 6 and 7 show some key summary statistics for ASRs for deaths for these areas, calculated using upper age groups of (separately) 90+ and 95+, and the difference between the two. Although there are relatively few MSOAs with zero population estimates, for completeness the table has been replicated for only those MSOAs where there are no zero population estimates (Table 7). Table 6: Age standardised death rates (per 1,000 population) of MSOAs in England and Wales, summary data by sex, based on upper age groups of 90+ and 95+, using revised European Standard Population Average1 SMR Average 1 standard error Average 1 Relative Standard Error Number of 90+ 95+ Change 90+ 95+ Change 90+ 95+ Change MSOAs Males 10.6 12.4 1.8 1.2 1.5 0.3 0.14 0.13 -0.01 7,201 Females 7.4 8.7 1.3 0.9 0.9 0.1 0.12 0.11 -0.01 7,201 Source: Office for National Statistics Notes 1 Avera ges quoted a re s i mpl e unwei ghted mea ns , a nd therefore wi l l di ffer from ra tes ca l cul a ted a t the na ti ona l l evel . Table 7: Age standardised death rates (per 1,000 population) of MSOAs in England and Wales, summary data by sex, based on upper age groups of 90+ and 95+, using revised European Standard Population Excluding MSOAs where the population for one or more age groups is estimated as zero Average1 SMR Average 1 standard error Average 1 Relative Standard Error Number of 90+ 95+ Change 90+ 95+ Change 90+ 95+ Change MSOAs Males 10.6 12.4 1.8 1.2 1.5 0.3 0.11 0.12 0.00 6,917 Females 7.4 8.7 1.3 0.8 0.9 0.1 0.12 0.11 -0.01 7,177 Source: Office for National Statistics Notes 1 Avera ges quoted a re s i mpl e unwei ghted mea ns , a nd therefore wi l l di ffer from ra tes ca l cul a ted a t the na ti ona l l evel . Key points on age standardised rates for deaths and standard errors Increasing the upper age range to 95+ tends to increase the size of the ASR. This occurs because the England and Wales population has a lower proportion (by sex) of those aged 95+ than in the new ESP; thus the higher death rate in the 95+ age group is given a greater weight in the calculation. An implication of this is that it matters whether the 95+ age group is used or not. Further, when comparing rates they should be compared on the same basis, i.e. all using the same upper age band. The average standard error increases a little when calculated using the 95+ age group. However, this is somewhat misleading as the ASR for deaths also increases. Instead the relative standard error is a better indicator. The relative standard error is virtually unchanged, with differences of ±0.01 when comparing the use of the 90+ to 95+ upper age band; thus using the upper age band does not affect the average relative variability of the ASR. The use of LA level averages has little impact on conclusions drawn about MSOAs. However some care needs to be taken when drawing conclusions on rates with and without the use of LA level data, particularly when considering MSOAs at an individual level where LA averages are used. The RSE for males is slightly higher for MSOAs with no zero populations: MSOAs with zero populations in some age bands are likely to be smaller, with less deaths, and therefore higher relative standard errors. Only 11 MSOAs have fewer than 20 male deaths over three years, and 25 have fewer than 20 female deaths. Care should be used if individual MSOAs with fewer than 20 deaths are considered. Averaging over fewer than the three years used will increase this number of MSOAs, as well as increasing the standard errors, and is therefore not recommended. On average, MSOAs have relative standard errors for the ASR for deaths of around 12%. Natural variation in the number of deaths will therefore typically cause the rate to vary by about ± 24% (with some small variation by sex and which upper age group is used). At this level natural variation is still an issue that needs to be considered carefully when considering the rates; however a range of useful analyses becomes plausible. Stability of MSOAs to the upper age band The previous section shows that using the upper age band tends to increase the ASR for deaths. It also shows that there is no indication that the relative variability of the ASR changes. However, this does not show that individual MSOAs show a stable pattern of change when the upper age band is used. The following charts (Figures 3 and 4) show, for males and females respectively, the range of change to the ASR when the upper age band is used, compared to those LSOAs where LA level data are not used. The same horizontal axis is used as for LSOAs facilitating a distributional comparison; however the vertical axis is changed as there are fewer MSOAs than LSOAs. Figure 3: Increase in ASR for deaths when the upper age band is increased from 90+ to 95+, males, MSOAs where no age groups have a population estimate of zero Figure 4: Increase in ASR for deaths when the upper age band is increased from 90+ to 95+, females, LSOAs where no age groups have a population estimate of zero These charts show that there is some distribution of the increase in the ASR rates for deaths, when moving to the 95+ age group. However, particularly for females, most MSOAs are in a relatively narrow band. Thus there is a degree of stability in moving from the 90+ to 95+ age band. However, there is no tight relationship between the rate calculated for 90+ and that for 95+, therefore it is not sensible to mix rates calculated for 90+ and 95+. For any discrete piece of work rates should be calculated based either the 90+ or the 95+ age band, and these should not be mixed. ‘Conversion’ from one to another at MSOA level is not possible. As shown above, ASR for deaths can be used, with some caution, at MSOA level. Further, there is no additional caution needed when using the higher 95+ age band (other than for a relatively small number of MSOAs where there is zero population in some age bands). Given there is additionally some stability in moving from the use of 90+ to 95+ as the upper age band, from the perspective of natural variability in deaths, there is no reason to prefer the usage of either the 90+ or 95+ upper age-group, providing these are not mixed within the same work. However, there is a degree of judgement in this conclusion – individual users may not consider there to be enough stability in the move to the 95+ age band. Significance and multiple testing We have observed above that increasing the upper age group from 90+ to 95+ tends to increase the ASR for deaths. It is therefore reasonable to ask the questions ‘is the increase for this area significant?’ and ‘is the increase for this area significantly different from the national average?’ Given there are 7,201 MSOAs a number of them are likely to have a ‘significant’ difference due to chance alone. Sophisticated analysis can be undertaken to deal with the issue of multiple testing. However, as a crude starting point, sophisticated consideration would only be required if the number of ‘significant’ differences across all MSOAs starts to approach 5% of areas (if using a 95% confidence interval), that is 360 areas. Over all of the MSOAs, 1,125 areas for males and 1,617 areas for females have an apparently significant change in ASR for deaths when moving from 90+ to 95+. These numbers are considerably more than 5% of areas. However, only 89 areas for males, and 59 areas for females, show a change that is apparently significantly different from the national change in moving from 90+ to 95+. This is much less than 5%. Consequently any apparently significant change when compared to the national increase should be considered with caution, as chance alone may well be the reason for the apparent significance of the change. Further, moving from 90+ to 95+ does have a significant impact, reinforcing the recommendation that analysis should not mix results from usage of the 90+ and 95+ upper age. Population denominators As set out above there is uncertainty in the population denominators used to calculate these rates, this uncertainty is higher for smaller areas and for the upper age groups. It is not currently possible to quantify this uncertainty well. However, this uncertainty adds further weight to the recommendation that care should be taken in the analysis of data at this level, particularly when comparing results at the individual MSOA level. Population estimates, splitting the upper age group into 90 to 94 and 95+, are still experimental: they have not yet achieved National Statistics status. Work is continuing on their evaluation. Given uncertainty over the reliability for these estimates, a practical recommendation is therefore that the 90+ age group should be used, however it is safe to use rates based on the 95+ upper age-group if there is a sufficient reason, provided the additional uncertainty in the population denominator is taken into consideration. The different upper age bands should however, not be mixed within the same piece of work. As work continues on the population estimates at the upper age-range, it is likely that a point may be reached where there is sufficient confidence in the quality of the population denominators to facilitate a change in this recommendation to using the 95+ age group. This is in line with EU guidance that the 95+ age group should be used where the data are available. MSOA Summary ASR death rates for MSOAs are subject to natural variation and care must be taken in their usage Provided due consideration is taken of the impact of natural variation, a range of analysis of ASR death rates for MSOAs is plausible Natural variation and the implications of multiple testing mean caution should be used in interpreting results, particularly at the individual MSOA level Some additional caution is needed when using the 95+ upper age band, as some MSOAs have zero population estimates in the upper two age bands, though relatively few MSOAs have zero estimates A few MSOAs have low (under 20) numbers of deaths and additional care should be taken with these MSOAs Averaging over fewer than three years is not recommended Using the 95+ age group tends to increase the calculated ASR rate for deaths Comparisons of ASRs should be made only if the same (upper) age bands are used There is no indication that use of the higher upper age group (95+) makes the ASRs more unreliable in relative terms as a result of natural variation in the number of deaths The increase in ASR for deaths, in moving to the 95+ age group, is relatively stable across MSOAs, particularly for females; however there is no direct correlation and ‘converting’ a 90+ based estimate to a 95+ based estimate is not possible Uncertainty in measurement of population estimates (particularly at the higher ages) means additional caution should be taken, particularly when drawing conclusions at the individual MSOA level Because work is continuing on the estimation of the population at the upper ages, we currently recommend using the 90+ upper age limit, but we will change this recommendation to using the 95+ age group when the population estimates for these upper age-groups are considered to be sufficiently reliable Overall summary There is too much natural variation in the number of deaths for ASR for deaths to be used (for general purposes) at the LSOA level (Lower level Super Output Areas). ASRs for deaths may be used, with some appropriate caution, for a range of purposes at the MSOA level (Middle level Super Output Area). Currently the upper age band of 90+ should be used. This recommendation may later be changed to 95+ following further work on population estimates at these ages. Jonathan Page-Swan March 2014 Office for National Statistics i Eurostat, Revision of the European Standard Population: Report of Eurostat's task force, 2013 Edition, http://epp.eurostat.ec.europa.eu/cache/ITY_OFFPUB/KS-RA-13-028/EN/KS-RA-13-028-EN.PDF ii ONS, A Beginners Guide to Geography, web pages: http://www.ons.gov.uk/ons/guidemethod/geography/beginner-s-guide/index.html iii P. Boyle and D.M. Parkin, Cancer Registration: Principles and Methods, Chapter 11. Statistical methods for registries, http://www.iarc.fr/en/publications/pdfs-online/epi/sp95/ iv This issue is discussed here: Department of Health (New York), Rates Based on Small Numbers - Statistics Teaching Tools, http://www.health.ny.gov/diseases/chronic/ratesmall.htm v ONS, 2012 single year of age super output area mid-year population estimates: http://www.ons.gov.uk/ons/rel/sape/soa-mid-year-pop-est-engl-wales-exp/mid-2012/stb---superoutput-area---mid-2012.html ONS, 2011 mid-year LSOA estimates by single year of age: http://www.ons.gov.uk/ons/aboutons/business-transparency/freedom-of-information/what-can-i-request/published-ad-hocdata/pop/april-2013/mid-2011-lsoa-syoa-population-estimates-england-and-wales.zip ONS, Revised LSOA estimates by single year of age: http://www.ons.gov.uk/ons/publications/re-reference-tables.html?edition=tcm%3A77-285154 vi ONS, LSOA Population Estimates of the Very Old, England and Wales: Research Paper, http://www.ons.gov.uk/ons/guide-method/method-quality/specific/population-and-migration/pop-ests/lsoapopulation-estimates-of-the-very-old--research-paper.zip vii The geography lookup table is published as a zip file: https://geoportal.statistics.gov.uk/Docs/Lookups/Output_areas_(2011)_to_lower_layer_super_output_areas_ (2011)_to_middle_layer_super_output_areas_(2011)_to_local_authority_districts_(2011)_E+W_lookup.zip available from this page: https://geoportal.statistics.gov.uk/geoportal/catalog/content/filelist.page Annex 1: Technical Annex European Standard Population – weights and age groups Table 1: European standard population ages and weights Age Group (years) Standard Population 0,0 1-4 5-9 10-14 15-19 1,000 4,000 5,500 5,500 5,500 20-24 25-29 30-34 35-39 40-44 6,000 6,000 6,500 7,000 7,000 45-49 50-54 55-59 60-64 65-69 7,000 7,000 6,500 6,000 5,500 70-74 75-79 80-84 85-89 90-94 5,000 4,000 2,500 1,500 800 95+ Total 200 100,000 Source: Eurostatvii Where an upper age band of 90+ is used, the 90-94 and 95+ age groups are combined with a weight of 1,000. Calculating Age Standardised Death Rates The European Standard Population (ESP) is used to calculate direct standardised rates, in this case specifically direct age standardised death rates (ASRs) for deaths. Sex Rates are calculated separately for males and females. The same standard population is used for both sexes. Population at risk Calculations of incidence rates require an estimate of the population at risk. For this work calendar year data are used for the number of deaths, so the mid-year population estimate is used to estimate population at risk – this is a common approach. In this document the term ‘estimate of population’ is then used when expressing the formulæ. Crude age specific death rates For each age group, i, (separately by sex) calculate: Weighted age specific death rate For each age group, i, (separately by sex) calculate: The standard population (weight) is given in Table 1. Age standardised rate (ASR) for deaths Calculate, separately for each sex: This gives the rate per thousand population – the conventional expression of this rate for small areas. is the total of the standard population; for the new ESP this is equal to 100,000, as shown in Table 1, so: Where a maximum age of 90+ is used there are 20 instead of 21 age-groups, so becomes . The standard weight, population, and deaths, are then combined for the upper two groups, thus the standard weight for this 90+ age group is 800+200=1,000. The overall total for the standard weights remains changed at 100,000. Zero population estimates It is possible for the population of an age/sex group to be estimated as zero. This may be a ‘genuine’ zero or as a result of the uncertainties in the estimation method. This happens frequently at LSOA level, and occasionally at MSOA level. Where a zero occurs, the calculation of an ASR would involve division by zero. Therefore the standard approach to this situation, which was adopted for this work, is to use the rate calculated at a higher geographical level. In this case the age-specific death rate for the local authority (LA) containing the MSOA or LSOA is used in the calculation. As the use of LA level data has an impact on the calculation, this work has paid particular attention to this issue, examining results (across areas) with and without areas with zero populations (and therefore LA level rates. Zero deaths (events) in a specific age/sex group are permitted within the calculation and these occur frequently. In this case the result is an age-specific death rate of zero, which then gets included in the overall calculation of the age-standardised death rate. Calculating standard errors There is natural variation in the number of deaths in a particular age group for a particular area. Thus, even if an area was to have the same ‘underlying death rate’ from year to year, chance alone would mean that there were differing amounts of deaths. The standard error provides a measure of the natural variability. These standard errors are not a measure of how much error there is in the number of deaths recorded, nor do they provide a measure of error in the estimate of population used in the rates; they are a measure of the natural variation expected in the death rate. There are two main methods of formulaically calculating standard errors for ASRvii. The conventional approach, used for the standard errors quoted in this report, is to use the Poisson approximation for the standard error. The main alternative is to use a binomial approximation. These approximations have been found to give similar resultsError! Bookmark not defined.. For the ESP . Because of the structure of the calculations file for this work this has been implemented as: Confidence interval The standard error (SE) is approximately normally distributed – N (0, 1). The 95% confidence interval for an ASR is therefore given by Relative standard error The impact of the standard error, what it tells you about how good an estimate is, that is how much you can expect a rate to change as a result of natural variation is dependent on the size of the ASR it relates to. The relative standard error, RSE, provides a way of comparing the impact the standard error will have, irrespective of size of the ASR. Thus the RSEs can be compared for different rates – even if the size of the rate is different. The RSE is calculated as: . This may be expressed as a proportion or a percentage. Rates over a number of years It is common practice to ‘average’ rates over a number of years. In particular for death rates this can lessen the impact of the changing timing of the peaks in the excess winter deaths. Calendar year deaths data are used. In the UK context there is an increase in deaths over the winter period. In some years this may start, or be higher, ‘earlier’ than in others, if it starts before December it will have an impact in the ‘old’ year as well as the ‘new’ year. This changing timing can cause variation in the ASR, even if overall there are no more or less extra deaths over the winter period than is usual. For LSOAs and MSOAs it is also necessary to ‘average’ over a number of years in order to ensure sufficient events (deaths) to enable calculation to a degree of reliability for these small areas. It is typical to ‘average’ over three years, and this has been done for this work. In practice the calculation is undertaken by simply adding the number of deaths over a three year period, and (separately) adding the corresponding three mid-year population estimates together. It is then these three year totals that are used in the calculation. No further adjustment is required. The resultant ASR remains the rate death rate per year (expressed as per thousand population). However it is calculated (‘averaged’) over a three year period. Annex 2: Suggestions for further work Binomial method for calculating standard error This paper has been prepared using the Poisson approximation for the Standard Error to measure the variation inherent in the number of deaths (from year to year). This is the conventional approach to standardised rates. The main deterministic alternative to this is to use a binomial approximation. As referred to above these tend to produce very similar values. However a useful check would be to calculate the standard error using the binomial approximation and compare the values produced. Assessing the impact of uncertainty in population estimates and computational methods There is uncertainty in the population estimates. This uncertainty is higher for the more elderly age bands. Further the age split for the 90+ age group (into 90-94 and 95+) is currently experimental. With estimates of uncertainty it is possible to apply computational methods (such as bootstrapping) to look at the impact of these uncertainties on the calculation of ASRs. These methods also provide an alternative approach to looking at the impact of natural variation. Formal estimates of the degree of uncertainty in population estimates are the subject of research within ONS. Nevertheless it might be possible to derive either some indicative or conservative assumptions about the size of uncertainty in the population estimates that would permit the consideration of the impact of this uncertainty on ASRs.