3319.0.55.001 - Multiple Cause of Death Analysis, 1997-2001  
ARCHIVED ISSUE Released at 11:30 AM (CANBERRA TIME) 22/07/2003   
   Page tools: Print Print Page Print all pages in this productPrint All

CONTENTS

1. Introduction
2. Aims of multiple cause of death analysis
3. Methods of analysis
4. Limitations of multiple cause of death analysis
5. Conclusion
6. Appendixes



1. INTRODUCTION

1. Mortality data are published annually by the Australian Bureau of Statistics in Deaths Australia (cat. no. 3302.0) and Causes of Death Australia (cat. no. 3303.0). Detailed data cubes and additional tabulations are also available from ABS on request, to support mortality analysis.

2. To date, mortality analysis has primarily focused on underlying causes of death (UCD). The underlying cause is the disease or condition which led directly to the death. These underlying causes are coded in accordance with the International Classification of Diseases (currently ICD revision 10). However, death certificates contain more information than just the underlying cause of death. The certificates also include significant diseases and/or conditions which led or contributed to death (i.e. associated causes). Multiple cause of death (MCD) statistics, therefore, refer to statistics which include both underlying and associated causes of death. Since 1997, multiple cause of death data have been compiled for Australia.

3. Single underlying cause of death analysis is useful and relatively simple to tabulate and interpret. It remains the primary tool for most researchers, and is adequate for examining many conditions. However, it provides a somewhat limited picture of mortality and health of the population. In particular, it does not provide complete information about conditions which frequently appear as a multiple cause of death, but which rarely appear as an underlying cause. In these situations, multiple cause of death statistics can provide a more comprehensive and ultimately insightful view of mortality patterns. Overseas health researchers are increasingly using multiple cause of death analysis (e.g. Goldacre, Roberts and Griffith, 2003, Jin et al., 2003, Frenzen, 2003, Mannino, Brown and Giovino, 1997, Mannino et al., 1998), and exploring the implications for actuarial practices (Stallard, 2002).

4. Multiple cause of death analysis will become increasingly important in Australia. As Australia's population ages, conditions more commonly found in elderly persons (particularly chronic or long-term illnesses) are increasingly important causes of death. The presence of certain chronic conditions often increases the risk of developing other associated long-term conditions (for example, the presence of diabetes increases the risk of developing heart disease (Senes and Britt, 2001, pp. xi-xii)). Hence multiple cause of death data are increasingly useful in analysis of Australia's changing health profile. The Australian Institute of Health and Welfare is currently undertaking research on selected multiple causes of death, and has expressed its intent to begin using multiple cause of death data in its publication 'Trends in Deaths' (Dunn, Sadkowsky and Jelfs, 2002, p. 4).

5. It should be noted that analysis of multiple cause of death data is a complex task and it is important for the researcher to understand the clinical relationships between diseases, and hence, possible relationships between causes of death. As well, researchers should be aware of known data quality issues relating to cause of death data (see Explanatory Notes in Deaths Australia (cat. no. 3302.0) and Causes of Death Australia (cat. no. 3303.0)).


1.1 CONTENTS OF THIS PAPER

6. This paper explores potential uses of multiple cause of death data by giving examples of several methods of analysis. Section 2 provides a brief introduction to the aims of MCD analysis. Section 3 outlines various methods of examining MCD data. These include ratios MCD:UCD as well as more detailed methods which measure the association between two co-occuring causes of death. Limitations to MCD analysis are considered in section 4, before some concluding remarks in section 5. The Appendixes include a glossary of terms and other technical information. A list of references can be found at the end of the paper.

Back to top



2. AIMS OF MULTIPLE CAUSE OF DEATH ANALYSIS

7. The ABS analyses and disseminates multiple cause of death data to:
  • provide more comprehensive information on and identification of a range of illnesses/conditions which contribute to the death process, but which do not often appear as the underlying cause of death (e.g. Alzheimer's, diabetes, pneumonia)
  • provide better documentation on multi-morbid associations and the strength of associations between conditions which led to death (for example by examining the frequency of associations between diseases such as diabetes and ischaemic heart disease)
  • provide a method of more comprehensively analysing specific causes of death (e.g. examining the nature of injury for external causes).

8. In addition, analysis of multiple cause of death data may assist in identifying problems with the process of recording and coding cause of death information (Lindahl and Johansson, 1994).

Back to top


3. METHODS OF ANALYSIS

3.1 RATIOS MCD:UCD

9. Ratios of multiple causes of death to underlying causes of death can be used to identify which diseases are relatively more important as multiple rather than underlying causes. Two slightly different ratios can be calculated and these are described below. In order to understand the difference between the two ratios, it is important to distinguish between the number of times a cause or group of causes appears on death records and the number of deaths where the cause (or group of causes) is mentioned. For example, if a death record contained three different codes from the range of codes belonging to "Diseases of the circulatory system" (ICD-10: I00-I99), then there are three causes (within circulatory disease), but only one death.

10. In ABS Causes of Death, Australia, 2001 (cat. no. 3303.0), a ratio has been calculated for a specified cause or group of causes, by taking the number of times the specified cause occurs on death records and dividing by the number of deaths where the cause was selected as the underlying cause. This gives a measure of the relative frequency with which a cause is mentioned anywhere on death records compared with the number of times that cause has been selected as the underlying cause of death.

11. Other researchers (Pavillon et al,1994, Wilkins et al,1997, Israel et al,1986) have calculated a similar ratio by taking the number of deaths with any mention of the cause, and dividing by the number of deaths where the cause was the underlying cause. Similarly, this ratio provides information about the nature of a condition and the role which it plays in the chain of events leading to death (Israel et al, 1986).

12. The following table compares these two ratios using data from the ABS deaths collection 1997-2001. Both of these ratios show the relative importance of causes (or groups of causes) as multiple causes of death relative to their importance as underlying causes. For specific causes with only a narrow range of possible codes (eg Diabetes ICD-10: E10-E14), the two ratios are likely to produce the same results since one death record generally only has one code within the range. In contrast, for broad groups of causes eg Diseases of the Circulatory System (ICD-10: I00-I99), the ratio of MCD mentions to UCD deaths produces higher results than the ratio of MCD deaths to UCD deaths, since one death record may have many codes within the specified range of codes.


TABLE 1 - COMPARISON OF MCD:UCD RATIOS FOR SELECTED CAUSES OF DEATH

Number of times the specified cause occurs on
death records
(Multiple causes)
Number of deaths
with any mention of specified cause
(MC deaths)
Number of deaths where specified cause is
underlying cause
(UC deaths)
Ratio of
multiple causes
to UC deaths
Ratio of
MC deaths
to UC deaths

No
No
No
Ratio
Ratio
Scepticemia (A40-A41)
32,718
32,673
4,216
7.8
7.7
Influenza and pneumonia (J10-J18)
90,349
90,207
11,804
7.7
7.6
Atherosclerosis (I70)
16,519
16,514
2,859
5.8
5.8
Angina pectoris (I20)
1,395
1,393
258
5.4
5.4
Diabetes mellitus (E10-E14)
49,060
49,012
14,939
3.3
3.3
Diseases of the circulatory system (I00-I99)
639,108
371,532
255,739
2.5
1.5
Ischaemic heart disease (I20-I25)
242,995
190,524
138,120
1.8
1.4
Leukaemia (C91-C95)
8,395
8,308
6,649
1.3
1.2
Malignant Neoplasms (C00-C97)
236,320
198,803
177,046
1.3
1.1
Intentional self harm (X60-X84)
13,190
12,727
12,712
1.0
1.0
Transport accidents (V01-V99)
10,155
10,153
10,054
1.0
1.0

Calculated from ABS Deaths data 1997-2001.


13. In general, for both of these ratios, high values are often associated with long-term, chronic conditions which increase the risk of death in association with other diseases. Such conditions include atherosclerosis, and diabetes. Large ratios are also typical of some acute conditions which do not often initiate the sequence of death, such as angina pectoris, as well as diseases which often appear as complications, such as pneumonia and septicaemia (Israel et al.,1986, pp. 170-171).

14. For conditions with high ratios, underlying cause analysis on its own will not give a true picture of the number of deaths involving these conditions. In these situations it makes sense to also look at multiple cause data. Researchers who have explored this approach include Israel, Rosenberg and Curtin, 1986, Tardon et al., 1995, Mannino, Brown and Giovino, 1997 and Mannino et al., 1998. For example, in the above table Influenza and pneumonia have a ratio of MC deaths to UC deaths of 7.6, indicating that this group of causes is involved in nearly eight times as many deaths as would be indicated by underlying cause of death data.

15. In contrast, low ratios (close to 1) are typical for causes such as cancer, suicide, transport accidents, ischaemic heart disease and leukemia, as they are generally the condition which initiated the train of events leading to death (Israel et al.,1986, pp. 170-171). Although multiple cause of death data are especially useful for providing information for those causes with large MCD:UCD ratios, multiple cause of death data also allows useful analysis for some causes with low MCD:UCD ratios. For example, both suicide and transport accidents have a ratio of MC deaths to UC deaths of 1, indicating that they are almost always the underlying cause, rather than an associated cause of death. For these underlying causes, specific details about the nature of injury or poisoning are also recorded on the death record, as associated causes, using ICD-10 codes from the Chapter "Injury, poisoning, and certain other consequences of external causes" (S00-T98). By analysing these associated causes, much useful information can be obtained about the types of injury sustained from different external causes of injury.


3.2 DESCRIPTIVE ANALYSIS

16. While examining the absolute numbers of multiple causes of death provides some information, there are other important analytical uses of multiple cause data. MCD data can also be used to examine the links between the different causes listed on a death certificate. Ultimately, these methods are of the greatest value to health researchers.

17. A simple way to investigate the association between causes of death is to calculate the percentage of deaths involving a particular cause which also involve another specified cause. Tables of this nature are included in the ABS publication Causes of Death Australia (cat. no. 3303.0) for selected underlying and multiple causes of death and additional detailed tables are available. More detailed descriptive analysis is used to cross-classify underlying external cause of death (ICD-10: V01-X49) by nature of injury (as multiple cause) (ICD-10: S00-T98). Another area where this form of analysis might be particularly useful is cross-classifying instances of suicide (as a multiple cause) by mental health problems, as set out in table 2. From this table it is clear, for example, that the proportion of suicides where Mood(affective) disorders are also mentioned is substantially higher for females than males.

TABLE 2 - SUICIDES INVOLVING SELECTED CAUSES OF MENTAL AND BEHAVIOURAL DISORDERS, by sex and age at death, 1997-2001
15 - 24
25 - 44
45 - 64
65 +
MALES
FEMALES
MALES
FEMALES
MALES
FEMALES
MALES
FEMALES

PERCENTAGE
All mental and behavioural disorders (F00-F99)
15.8
19.7
19.0
26.0
18.7
25.2
12.1
22.7
Mental and behavioural disorders due to psychoactive substance use (F10-F19)
10.3
8.8
10.0
8.6
7.1
4.2
2.7
2.2
Schizophrenia, schizotypal and delusional disorders (F20-F29)
2.2
2.6
2.6
3.1
1.0
2.9
0.5
1.4
Mood [affective] disorders (F30-F39)
3.8
9.1
8.3
14.5
11.5
18.2
8.7
19.2

NUMBER
All Suicides
1,627
385
4,832
1,165
2,361
691
1,240
365

Calculated from ABS Deaths data 1997-2001.



3.3 RATIO OF OBSERVED TO EXPECTED JOINT FREQUENCIES

18. In isolation, however, descriptive analysis is relatively uninformative when dealing with chronic causes of death. These causes of death affect far larger numbers of people, and each person is frequently affected by more than one cause. Two chronic multiple causes of death may thus occur together frequently simply due to the high prevalence of these diseases amongst the population. For example, nearly 11% of deaths (20,939 deaths, 1997-2001, all ages) involving ischaemic heart disease also involve malignant neoplasms (cancer). However, over 30% of all deaths involve malignant neoplasms. The number of deaths involving both of these causes is actually much smaller than would be expected if the two causes were occurring independently in the deaths population. Thus, in addition to percentage analysis, it is important to test the strength of association between causes.

19. Strength of association can be measured by the ratio between the actual number of deaths for which both causes are mentioned and the number of deaths that might be expected under the assumption that the conditions were independent (Stallard, 2002, p. 3). Hence an observed to expected ratio higher than 1 indicates that two causes may have a medically significant link. The method used to calculate observed to expected ratios can be found in Appendix V.

Observed to expected ratioIndicates
<1
Causes occur together less often than expected
1
Causes occur together as often as expected
>1
Causes occur together more often than expected

20. High ratios of observed to expected joint frequencies tend to occur when one of the conditions is a background risk or significant associated condition for the other (for example diabetes is a risk factor for heart disease), or when one of the conditions is a lethal sequelae of other, chronic conditions (for example septicemia frequently develops in patients with chronic liver disease) (Stallard, 2002, p. 85). In contrast, pairs of conditions which are each a major UCD, normally have low ratios of observed to expected joint frequency. In the above example, the observed to expected ratio involving malignant neoplasms and ischaemic heart disease was only 0.35 (i.e. the two conditions occurred together much less frequently than would be expected if the two conditions were independent).

21. Tables of ratios of observed to expected joint frequencies provide an ideal way of quickly identifying conditions which have a potentially significant association. Differences in age and sex frequently have an impact on the strength of association figures, and such breakdowns should be used whenever possible. The following table is an example of how observed to expected ratios can be presented in a table. In each case, two multiple cause of death numbers are being compared.

TABLE 3 - THREE LEADING MULTIPLE CAUSES OF DEATH,
(diabetes mellitus (E10-E14), ischaemic heart diseases (I20-I25) and cerebrovascular disease (I60-I69)), by age and sex, 1997-2001
Age
Diabetes and Ischaemic
Heart Diseases
Ischaemic Heart Diseases
and Cerebrovascular Disease
Diabetes and
Cerebrovascular Disease

OBSERVED JOINT FREQUENCY (NUMBER)

Males45 - 54
660
163
119
55 - 64
1,755
622
404
65 - 74
4,251
2,330
1,340
75 - 84
5,021
4,664
2,075
85 +
1,825
2,634
1,016

Females45 - 54
264
53
67
55 - 64
803
244
206
65 - 74
2,370
1,269
835
75 - 84
4,350
4,059
2,108
85 +
2,983
5,351
1,864


EXPECTED JOINT FREQUENCY (NUMBER) (a)

Males45 - 54
303
285
69
55 - 64
930
835
240
65 - 74
2,560
2,955
867
75 - 84
3,408
6,473
1,583
85 +
1,446
4,143
823

Females45 - 54
60
79
42
55 - 64
279
278
137
65 - 74
1,231
1,615
617
75 - 84
2,866
6,477
1,824
85 +
2,289
9,319
1,660


RATIO OBSERVED TO EXPECTED

Males45 - 54
2.18
0.57
1.73
55 - 64
1.89
0.74
1.68
65 - 74
1.66
0.79
1.55
75 - 84
1.47
0.72
1.31
85 +
1.26
0.64
1.23

Females45 - 54
4.41
0.67
1.60
55 - 64
2.88
*0.88
1.51
65 - 74
1.92
0.79
1.35
75 - 84
1.52
0.63
1.16
85 +
1.30
0.57
1.12

* test for independence not significant at p=0.01 level.
(a) Rounded to nearest whole number.
Calculated from ABS Deaths data 1997-2001.

22.There are a number of other options in presenting tables of ratios of observed to expected joint frequencies. For example, age standardised figures can be used, particularly when making comparisons between years (Stallard, 2002). Another possibility is to examine one cause in depth, by examining its association with several, more specific causes. Examples of both of these techniques are given in Appendix I.

3.3.1 Significance testing

23. Each ratio of observed to expected frequencies is calculated from two-way frequency tables involving two causes of death. Expected frequencies are calculated under the assumption that the two causes of death are independent. In this situation simple tests can be used to test whether the two causes of death are independent (Pavillon et al.,1994). The null hypothesis is that the two causes are independent. If the test result is significant, the null hypothesis is rejected and there is evidence to support the alternative hypothesis that the two causes are not independent. For most combinations of causes presented in this paper, tests for independence are statistically highly significant, indicating that the two causes are not independent. However in some cases, especially when results are presented for specific age and sex groups, the test for independence is not significant. Where the test result is not significant at the p=0.01 level the corresponding ratio of observed to expected frequencies will be flagged with an asterisk.


3.4 OTHER METHODS FOR MEASURING STRENGTH OF ASSOCIATION

24. Many researchers internationally use the ratio of observed to expected joint frequencies to measure strength of association between causes. However, it is not the only method which can be used.

3.4.1 Proportionate Mortality Ratios

25. Some researchers have adapted the Proportionate Mortality Ratio (PMR), a technique frequently used in occupational epidemiology, for use with multiple cause of death data. In fact, as adapted to multiple cause analysis, this technique is mathematically very similar to the ratio of observed to expected joint frequency method. For example, in his examination of deaths due to gastroentiritis of unknown etiology (GUE), Frenzen defined the PMR of GUE combined with another condition
          "as the observed number of GUE deaths associated with the condition divided by the expected number of GUE deaths associated with the condition. The expected number of deaths was determined by multiplying the number of GUE deaths by the proportion of all deaths associated with the specified condition." (Frenzen, 2003, p. 443.)

3.4.2 Odds Ratios

26. Other researchers have measured associations between causes using odds ratios (OR). This is an adaptation of a technique used for evaluating case/control situations, for example in examining the results of drug trials. ORs were used by Wilkins et al. to study the association of selected causes of death with heart diseases, in a paper written for Statistics Canada. They defined the OR for heart disease and another condition as
          "the odds of diseases of the heart being mentioned on a death certificate, given the mention of another specific cause, divided by the odds of diseases of the heart being mentioned on a death certificate, given that the other specific cause was not mentioned." (Wilkins et al., 1997, p. 20.)

27. Although they are different measures, ORs are similar to ratios of observed to expected in that a result greater than one indicates that there may be a link between the two conditions. However, where diseases are associated, the calculated value of the OR is larger than the ratio observed:expected, so care must be taken in interpreting results from the two methods. For example, using the same data (ABS Deaths data 1997-2001, all ages, persons), ischaemic heart diseases and diabetes have a ratio observed:expected of 1.69, but an OR of 2.58 . The ratio of observed to expected can be interpreted to mean the two diseases are occurring together 1.69 times more than expected (assuming independence) in death records. The OR result can be interpreted to mean that the odds of ischaemic heart diseases occurring on a death record with a mention of diabetes were 2.58 times as great as the odds of ischaemic heart diseases occurring on a death record with no mention of diabetes.


Back to top



4. LIMITATIONS TO MULTIPLE CAUSE OF DEATH ANALYSIS


4.1 DATA QUALITY


28. As with UCD analysis, MCD analysis is subject to limitations due to data quality. The issues involved however, are not entirely the same. Stallard has argued that recording multiple causes eliminates the problem of deciding which single condition should be considered the underlying cause, because sequencing is no longer an issue (Stallard, 2002, p. 70). On the other hand, multiple cause analysis encounters the additional problem of some conditions present at death not being listed at all. For example, research in the United States has shown that death certificates rarely record tobacco use disorder (ICD-9 305.1) on certificates which also list lung cancer, despite a well established link between tobacco use and lung cancer (Mannino et al., 1998, p. 164).

29. As with other statistical collections, data quality depends on the procedures being followed at every stage of the collection and processing of the data. There are a number of possible sources of error and these can be minimised by the application of quality assurance measures. In particular, the quality of multiple cause of death data is highly dependent on the contribution that doctors and certifiers make in completing the Medical Certificate of Cause of Death. It is important that doctors and certifiers understand the need for accurate completion of death records and appreciate that the resulting statistical data is used widely by many organisations in health research. For this reason, the ABS regularly produces a booklet with guidance for doctors in order to encourage accurate recording of causes of death. The booklet is entitled Cause of death certification Australia: A booklet for the guidance of medical practitioners in completing Medical Certificates of cause of death, and was most recently produced in 2001. As well, the ABS conducts a process whereby death certificates containing inadequate information are queried, and amended if more accurate information can be provided by the doctor or certifier who completed the original certificate.


4.2 SELECTION OF DATA TO BE PUBLISHED

30. As the examples given above demonstrate, a vast array of multiple cause of death cross tabulations could be examined and/or published. Selectivity on the part of the researcher is thus impossible to avoid. However, this judgement can only be made on the basis of what seems most statistically important. To the health professional, results obtained in this way may seem obvious, while the medically trained eye may well see important connections which do not obviously stand out from the statistical point of view.


4.3 CORRELATION IS NOT CAUSATION

31. It is important to be aware that just because two conditions or diseases are associated in a statistical sense, this does not mean that there is a cause-effect relationship between the two diseases. For example, using 1997-2001 ABS Death data, ischaemic heart diseases and asthma have an observed to expected ratio of 1.22 (amongst all ages), indicating a small positive association between these diseases. In interpreting results such as this, a number of factors should be considered.

32. As a general rule, it is important to consider whether alternative explanations such as chance, bias, or confounding could have contributed to a statistical association (Hennekens et al., 1987, p. 30). For example, one possible explanation is that both these diseases (ischaemic heart diseases and asthma) have an even stronger association with some other confounding factor, for example diabetes.

33. Moreover, in order to make judgements about cause and effect relationships between diseases, researchers require evidence from a number of different sources (not just multiple cause of death data). Knowledge of disease processes and an understanding of the biological plausibility of any hypothesised causal link between two diseases would also be important.


Back to top



5. CONCLUSION

34. Multiple cause of death records are a data source which is currently being under-utilised in Australia. In particular, MCD data can reveal significant information about the strength of association between different causes of death, for example by examining observed to expected ratios, as outlined in this paper. Although multiple cause of death analysis is not appropriate in all situations, it can provide a more complete picture of mortality for causes which are frequently not recorded as the underlying cause of death. This includes many chronic conditions and diseases which appear as complications, which are major causes of death. As Australia's population ages, these conditions are becoming relatively more prevalent. Multiple cause of death analysis will therefore become an increasingly useful tool in health and demographic research.

5.1 Acquiring multiple cause of death data

35. More detailed cause of death information is available upon request from the ABS. This information can comprise standard or customised tabulations (by hardcopy or electronic media). Unit record files are available to approved users upon application. Generally, a charge is made for providing information upon request. For more information, contact:

Peter Burke (07) 3222 6069
Email: peter.burke@abs.gov.au


5.2 Comments about this paper

36. Questions or feedback about the contents of this paper are welcomed, and should be directed to:

Chris Gordon (02) 6252 7318
Email: chris.gordon@abs.gov.au


Back to top