OUP user menu

Explaining structural change in cardiovascular mortality in Ireland 1995–2005: a time series analysis

Richard Layte, Sinead O’Hara, Kathleen Bennett
DOI: http://dx.doi.org/10.1093/eurpub/ckq100 597-602 First published online: 4 August 2010

Abstract

Background: Deaths from diseases of the circulatory system and the seasonality of deaths from these causes fell sharply between 1995 and 2005 among older age groups in Ireland. We examine whether a structural break occurred in deaths from circulatory causes in Ireland between 1995 and 2005 and test whether this can be statistically accounted for by cardiovascular prescribing during the same period controlling for weather trends. Methods: Grouped logit time series models were used to identify if and at which quarter a structural break occurred in Irish circulatory deaths between 1995 and 2005. Data on cardiovascular prescribing and temperature within the quarter were entered into the trend-break model to account for the structural break. Results: Controlling for temperature, β-blocker, ace-inhibitor and aspirin medications rendered the structural break indicator insignificant among all age groups for men. Diuretic, statin and calcium channel blocker medications could not account for the break point for men aged 75–84 years. β-Blocker, aspirin and calcium channel blocker medications account for mortality trends among all age groups among women. Ace inhibitor and statin could not account for trends amongst women aged 65–74 years and nitrates and diuretics did not account for trends for any age group. Conclusions: Cardiovascular prescribing accounts for the trend break in circulatory mortality among men and women aged ≥65 years after 1999 in Ireland but the effect of prescribing is lower for women. β-Blocker, ace inhibitor and aspirin medications were more successful than statin, diuretic and nitrates in accounting for the trends.

  • cardiovascular disease, mortality, older people, prescribing, time series

Introduction

As elsewhere in the EU15, all cause death rates have been falling steadily in Ireland since the 1970s, but the rate of improvement has been particularly pronounced since the turn of the 21st century. Between 1996 and 1999 all cause death rates in Ireland fell by 5.2%, but by 26.2% between 2000 and 2004. Death rates from circulatory causes fell by 29.6% between 2000 and 2005. Although there is a long term downward trend in all cause mortality in Ireland, the pronounced decrease since the turn of the century is largely explained by decreases in circulatory and respiratory causes, especially amongst the older age groups who account for the majority of Irish deaths. Although the steep fall in respiratory death rates may also be related to the processes discussed in this article, we confine our analyses here to trends in circulatory deaths rates and leave trends in respiratory deaths to a future paper.

Previous analyses1 have shown that a significant proportion of the fall in mortality rates can be explained by a decrease in excess winter mortality, that is, the size of the increase in mortality in winter months over summer months each year.2–4 The IMPACT CHD model,5–7 which was applied to the Irish context,8 suggested that ~44% of the decline in Irish CHD deaths between 1985 and 2000 can be attributed to improvements in the uptake of treatments, particularly secondary prevention (18%) and treatments of chronic angina (8.4%). The majority of the decrease in mortality between 1985 and 2000 is attributed to changes in lifestyle factors but it is important to note that all the main population risk factors for cardiovascular disease—smoking, physical activity, obesity, diabetes and population blood pressure, have all worsened in Ireland since 2000 with the exception of population cholesterol and so cannot explain the abrupt change in mortality trends.9

The publication of the Irish Cardiovascular Strategy10 provided a structured approach to the primary and secondary prevention of cardiovascular disease in Ireland. This included prescribing protocols for GP’s, contributing to a steep increase in the volume of cardiovascular drugs prescribed, with β-blockers increasing by 109%, ace inhibitors by 89% and statins by 206% between the summer of 1999 and the end of 2003. The increase in the volume of drugs prescribed was aided by the change in eligibility rules for medical cards in the third quarter of 2001 providing free primary care and pharmaceuticals to all those aged ≥70 years. Until 2001, only those eligible on the basis of a means test or involvement in specific government training schemes had access to free primary care. After July 2001, ~30% of the population were covered by a medical card.

A review of clinical trials of cardiovascular drug efficacy during the period from 1970 onwards11 concluded that pharmaceutical agents played a major role in the prevention of atherosclerosis and its consequences and that the introduction of new classes of cardiovascular drugs often has positive population health consequences. A series of papers by Lichtenberg12–14 has also shown that the introduction of new classes of cardiovascular medication reduce other types of medical expenditure, hospital expenditure and, most importantly in the context of this article, death rates from circulatory causes. Cardiovascular medications are frequently used in combination and it is likely that the impact of a new class on death rates and hospitalizations may be due, to some extent, to the combined therapy.

This article is innovative in adopting a more direct approach than has been applied in previous research8 to try account for the timing and nature of the change in mortality trends. To do this, we use time series analysis to test whether the increased levels of prescribing of CHD medications can account for the timing and extent of the change in cardiovascular mortality.

Methods

Data

Mortality data were obtained from the Central Statistics Office (CSO) on the number of deaths from circulatory disease [ICD-9-CM codes 390–459], by quarter, over the period 1995–2005, broken down by sex and age group (65–74, 75–84, ≥85 years, and an all ≥65 years group). Population totals for each age and sex group were also obtained from the CSO to calculate the death rate per 1000 population.

Temperature data were obtained from Met Eireann, the Irish Meterological Service, to test and control for the possibility that a change in weather patterns in the form of warmer winter temperatures contributed to the fall in mortality.

Data for our main hypothesis of the effect of cardiovascular medicines were obtained for the period from January 1999 to December 2005 from the Primary Care Reimbursement Service (PCRS), which does not include data on prescriptions paid for out of pocket or via a primary health insurance scheme, and is only relevant to those living in the Eastern Regional Health Authority (ERHA). The ERHA region covers ~one-third of the Irish population (the Health Board regions were superseded by the Health Services Executive regions in 2005). Nonetheless trends in this region should be representative of trends in the Republic of Ireland overall. Research has shown that the ERHA pattern of prescribing for different chronic illnesses, including CVD, is very close to the national pattern.15 It would be possible to examine death rates from circulatory disease on the population aged <65 years. However, because our data on prescribing relate only to those in receipt of a medical card, the data become increasingly unrepresentative of population deaths in younger age groups.

Methodology

We assess the relationship between aggregate data on mortality, temperature, respiratory discharges and cardiovascular prescribing through time series analysis of 44 quarterly observations. The simplest model for this is: Embedded Image where the death rate at time t (Yt) is a function of temperature (T) and rate of cardiovascular drug use in the population (D) all at time t, plus error (ε). This could be estimated using ordinary least squares (OLS) or weighted least squares (WLS) methods. However OLS or WLS will produce predictions <0 for some values of the independent variables. Instead we use grouped logit methods which are more appropriate: Embedded Image

We thus estimate the log of the number dying from a circulatory cause in quarter t divided by the proportion not dying from this cause in the quarter. As we are using grouped data, we assume that all cases at t have the same values of the explanatory variables at t.

A central issue in time series analysis is the extent to which the data are autocorrelated, that is, a process: Embedded Image is said to be autocorrelated for where r is ≥1. A Durbin–Watson (DW) test varies between 0.38 and 0.84 for male circulatory deaths (depending on age) suggesting strong autocorrelation. However, controlling for the long term downward trend in circulatory deaths and the seasonal nature of deaths produces a DW of between 1.96 and 2.4, a more acceptable result. Examination of autocorrelation and partial correlation plots for an OLS model of circulatory death rates for men aged ≥65 years controlling for the trend in deaths and the seasonal variation shows relatively low and random autocorrelation and minimal partial correlation suggesting that our models are stationary.

Modelling strategy

We firstly establish whether and at which point there is a structural break in the death rate from circulatory causes in Irish death rates by sex and age group. We do this by fitting 128 OLS models of circulatory deaths which use alternative break points between the first quarter of 1998 and the last quarter of 2001 by each age/sex group (16 possible break points × four age groups × two sexes), which include the overall trend in deaths (a variable representing each year quarter 1995–2005 running from 1 to 44), three variables representing year quarter (omitting quarter one), a term for the interaction of the structural break point and the overall trend and three other terms interacting the year and quarter with the structural break point. We use OLS models rather than grouped logit models to facilitate the comparison of non-nested models. Use of grouped logit models would necessitate the use of an information criterion approach such as the Bayesian or Akaike Information Criterion. There is still some debate about the efficacy of such approaches.16 We first select an efficient model which meets standard specification tests. We performed the same specification tests of the OLS models as used for the grouped logit models and found low and random autocorrelation and minimal partial correlation. We then used a Chow test to identify break points in the time series for each age/sex combination.17

Having established a break point, we fit a variable representing each drug to the basic model of death rate trend, seasonal pattern and break point, controlling for temperature fitted as the lowest temperature at time t. If the quantitative variable expressing the cardiovascular drug effect can render the variable representing the structural change (i.e. the break point) in death rates insignificant and reduce its coefficient, we then take this as giving support to the hypothesis that the new trend pattern can be accounted for by cardiovascular prescribing trends. Cardiovascular medications are often used in combination and ideally the role of the different combinations would be tested here. Unfortunately, the limited number of degrees of freedom available in the aggregate data mean that this is not possible.

The time series models described above assume that the rate of cardiovascular prescribing in quarter t will be associated with death rates in that quarter. This is appropriate for some medications (e.g. nitrates) but others such as statins have a longer term effect. To test for this, we fit values for statin lagged by between one and ten quarters.

Results

Circulatory death rates

Figures 1 and 2 show male and female circulatory death rates between 1995 and 2005 for four age categories: 65–74, 75–84, an ≥85 years, and all ≥65 years. Male mortality is generally higher than female mortality across all age groups. The first quarter of 1999 exhibited the highest rates of mortality in the oldest age group, with males recording 34.3 deaths per 1000 of the population aged ≥85 years compared to slightly lower 28.1 deaths per 1000 of the female population aged ≥85 years. Post-2000, the seasonal pattern has modified, with less amplitude in the differences between Quarter one and three. Previous research1 commented on the strong seasonality effects in circulatory mortality in Ireland.

Figure 1

Male circulatory death rates by age group, year and quarter 1995–2005

Figure 2

Female circulatory death rates by age group, year and quarter 1995–2005

A Chow test shows the break point which accounts for the most variance in the mortality data for males occurs in Q4 1999 for those aged 64–74 and ≥85 years (as well as all ≥65 years), and in the last quarter of 2000 for those aged 75–84 years. For women, the break point occurs around a year later in the third and fourth quarters of 2000.

The top panel of table 1 shows the results for grouped logit models of circulatory deaths by sex and age group. This shows that for men aged 65–84 years, there is a significant structural break in the time trend plus a structural break in the seasonality trend for those aged 75–84 years. For men aged ≥85 years, there is evidence only of a structural break in the seasonality accounting for the overall decrease in mortality rates for circulatory disease. Transformed, the results for men aged ≥65 years show a 0.82% fall in deaths/000 per quarter 1995–2005, increasing by 0.5% per quarter after Q4 1999. They also show a strong seasonal variation with 25% less deaths/000 in Q3 than Q1. Healy4 found a 21% increase in winter mortality for Ireland using data from 1988 to 1997 using a slightly different measure, the coefficient of seasonal variation in mortality (CVSM). There is a 9% reduction in Q3 seasonality post Q4 1999. For women the structural break in the main time trend is significant for all age groups, whereas there is evidence for the break in seasonality only for women aged 75–84 years. Seasonality among women is just as pronounced as among men with 25.2% less deaths in Q3 than Q1 although there is a smaller reduction in this proportion after Q1 2000 of 6.8% relative to 9% among men.

View this table:
Table 1

Grouped Logit Models of Circulatory Death Rates by sex and age group

MaleFemale
Aged 65–74Aged 75–84Aged ≥85Aged ≥65Aged 65–74Aged 75–84Aged ≥85Aged ≥65
βSig.βSig.βSig.βSig.βSig.βSig.βSig.βSig.
Modelling Cardiovascular deaths controlling for quarter, structural break and quarterly interactions
    Trend−0.0150***−0.0070***−0.0070**−0.0080***−0.0150***−0.0090***−0.0050***−0.0050***
    Q2−0.1740***−0.2000***−0.2210***−0.1990***−0.0930**−0.2050***−0.1960***−0.1860***
    Q3−0.2150***−0.3140***−0.3470***−0.2900***−0.2150***−0.3230***−0.2840***−0.2900***
    Q4−0.0530n.s−0.1240***−0.1610***−0.1150***−0.0670n.s−0.1490***−0.1120***−0.1320***
    Trend_Int−0.0040*−0.0050***−0.0040n.s−0.0050***−0.0050**−0.0050***−0.0030*−0.0050***
    Q2_Int0.0710n.s0.0890*0.1250**0.0890***0.0260n.s0.1020**0.0200n.s0.0630**
    Q3_Int0.0450n.s0.1300***0.1010*0.0860***0.0150n.s0.1070**0.0410n.s0.0660**
    Q4_Int0.0360n.s0.0460n.s0.1200*0.0630**−0.0420n.s0.0880**0.0320n.s0.0610**
    Constant−5.1630***−4.2280***−3.3450***−4.5870***−5.9150***−4.5780***−3.5100***−4.8350***
    Psuedo R20.00470.0030.00260.0030.00510.00340.00220.0021
    N5071600258160052320081764005680800384 0400120440010725600
Modelling Cardiovascular Medications Controlling for Minimum Temperature in Quarter
    β-Blocker−0.0055**−0.003n.s−0.0006n.s−0.0037***−0.0016n.s−0.0059*−0.0001n.s−0.0037**
    Trend_int−0.0005n.s−0.0019n.s−0.0002n.s−0.0015n.s−0.004n.s−0.0004n.s−0.0019n.s−0.002*
    Ace-Inhibitor−0.0058**−0.0029n.s−0.0008n.s−0.004**−0.0017n.s−0.0039n.s−0.0015n.s−0.0025n.s
    Trend_int−0.0005n.s−0.0021n.s−0.0002n.s−0.0015n.s−0.0052*−0.0017n.s−0.0023n.s−0.0027**
    Diuretic−0.0031n.s−0.0004n.s−0.0013n.s−0.0028n.s0.0036n.s0.0016n.s0.0013n.s0.001n.s
    Trend_int−0.002n.s−0.0032**−0.0001n.s−0.0023**−0.0055***−0.0031**−0.0022*−0.0036***
    Statin−0.0029**−0.0014n.s−0.0003n.s−0.0018***−0.0015n.s−0.0027***−0.0005n.s−0.0018***
    Trend_int−0.0009n.s−0.0022*−0.0002n.s−0.0018*−0.0035*−0.0006n.s−0.0016n.s−0.002**
    Aspirin−0.0029**−0.001n.s.−0.0006n.s.−0.0019***−0.0008n.s−0.002n.s.−0.0002n.s.−0.0016n.s.
    Trend_int−0.0003n.s.−0.0024n.s.0n.s.−0.0014n.s.−0.004n.s.−0.0011n.s.−0.0018n.s.−0.002*
    Calcium−0.0064n.s.−0.0012n.s.−0.0018n.s.−0.0051*0.0021n.s−0.0059n.s.−0.0001n.s.−0.0042n.s.
    Trend_int−0.0019n.s.−0.0031**−0.0002n.s.−0.0023**−0.0052n.s.−0.0016n.s.−0.0019n.s.−0.0026**
    Nitrates0.0113*0.006n.s.−0.0027n.s.0.0057n.s.0.0196n.s.0.0191**0.0095n.s.0.0163**
    Trend_int−0.0021n.s.−0.0031***−0.0005n.s.−0.0026***−0.005***−0.0031**−0.0022*−0.0038***
  • n.s, Not Significant

  • *P < 0.05; **P < 0.01; ***P < 0.001

Among the oldest age groups (≥85 years), the overall downward trend was stronger with a 0.68% reduction among men per quarter and 0.54% among women, a fall which decreases by 0.37% among men and 0.26% among women after the break point. Seasonal variation is higher among the oldest age group with men experiencing 29.3% fewer deaths in Q3 compared to Q1 and women 24.7%. These seasonal fluctuations decrease by 10.6% among men and 4.2% among women after the break point.

Temperature

The suggestion that milder winters1 may have had an effect on the fall in death rates is not borne out by the temperature data, which does not reflect the decline witnessed in mortality post 1999. Fitting temperature variables into our model, alongside the trend and structural break variables, yields mainly insignificant results across all age groups and disease categories.

Cardiovascular drug prescribing

Figure 3 shows the prescribing rates for different CHD medications for the period 1995–2005 for both men and women (shown as rates per 1000) for particular types of medications. Across almost all medications there is a gradual increase from the beginning of the series followed by a steepening increase after 1999, and a further change in the prescribing trend post-2001. This coincides with the timing of the provision of medical cards to all those aged ≥70 years in the third quarter of 2001, irrespective of means.

Figure 3

Cardiovascular prescribing for persons aged ≥65, 1995–2005. Source: Primary Care Reimbursement Service, Eastern Region Health Authority Data

Modelling the impact of CHD prescribing

The second half of table 1 shows that the introduction of the variables representing cardiovascular prescribing (controlling for temperature) has an important impact, both in terms of the effects of the drugs themselves and their impact on the significance and coefficient for the terms representing the trend-break point in the time series. The impact of the drug variables is highest among men with the drug terms for β-blocker, ace inhibitor and aspirin rendering the trend-break variable insignificant across all age groupings. Diuretic, statin, calcium channel blocker and nitrates do not render the trend-break variable insignificant for men aged 75–84 years. Among women, β-blocker, aspirin and calcium channel blocker render the trend-break terms insignificant for all age groups. Ace inhibitor and statin fail to make the trend break insignificant for those aged 65–74 years and diuretic and nitrates fail to do so for the trend in any age group.

The impact of the drug effects on the trend-break coefficient varied by drug and sex. For men, aspirin and β-blocker produced the largest fall in the trend-break coefficient at 82% followed by statin (77%). Diuretic, calcium channel blocker and nitrate medications reduced the trend-break coefficient by 62, 63 and 58%, respectively. The variable representing rates of ace inhibitor use produced the lowest fall in the coefficient at 25% largely because of its lack of impact for the 65–74 year age group among men. Among men aged 75–84 years, ace inhibitor produced a 45% fall in the trend-break coefficient, rising to 78% among men aged ≥85 years. The mean reduction in the trend-break coefficient masks substantial variation between the age groups. The mean fall in the coefficient for men aged 65–74 years and 75–84 years is 50 and 49%, respectively, but this rises to 92% among men aged ≥85 years.

Among women, the medication variables produce substantially smaller falls in the female trend-break variables with statin producing the largest fall at 47% followed by ace inhibitor (42%), β-blocker (41%), aspirin (38%), calcium channel blocker (23%), nitrate (12%) and diuretic (8%). As with the male results, the female results vary by age group with the largest fall occurring among women aged 75–84 years at 63% followed by women aged ≥85 years at 28%. This proportion falls to <1% among women aged 65–74 years.

Values of the rate of statin prescribing lagged between one and ten quarters were tested to examine longer term effects. Overall the lagged variables performed less well than the parameters measuring prescribing rates at quarter t with significance and parameter estimates falling as the degree of lag increased.

Discussion

This article was motivated by the pronounced fall in mortality from circulatory causes observed around the turn of the last century in Ireland. Our estimates suggest that male deaths from circulatory causes were already falling by 3.3% a year until the fourth quarter of 1999 at which point the rate proportionate fall increased to 5.3% a year. Among women the extent of change was not as large but the first quarter of the year 2000 saw the yearly fall in circulatory deaths essentially double from 2.1% to 4%. Before the fourth quarter of 1999/first quarter 2000 circulatory deaths in the third quarter of the year were typically 25.2% lower than in the first quarter among both men and women. After this point, however, the peak summer fall in mortality dropped to 18.4% among men and 20.2% among women. This article has shown that the structural break in the trend can be statistically accounted for by changes in the pattern of cardiovascular drug prescribing in Ireland after 1998. Our analyses show that variables representing the structural break across different age/sex groups can be rendered insignificant, and a high proportion of their coefficient accounted for by quantitative variables measuring the increase in prescribing of cardiovascular drugs between 1995 and 2005. Variations in this effect across the sexes and between drugs deserves some comment.

Just as the fall in circulatory deaths was largest among men, our models also showed that the impact of circulatory drugs was more significant among men with drug variables more likely to render the trend-break variable insignificant and reducing its coefficient by a larger amount. Whilst there is no evidence to show that the cardiovascular medications examined in this article are less effective among women there is some evidence internationally to suggest that women may be prescribed such medications at a lower level of risk than among men.18

In relation to the impact of the medications, among men, β-blocker, aspirin and ace inhibitor medications were most effective at rendering the structural break variable representing mortality trends insignificant and reducing its effect. Among women, all the trend-break variables remained significant when modelled alongside variables representing all the cardiovascular drugs and across all age groups. The rather small effect for statin in this context is difficult to understand given its efficacy in randomized control trials. This could have been because evidence shows that statin contributes to longer term cardiovascular disease prevention. We explored this by fitting terms for the statin prescribing rate lagged by between one and ten quarters but found no improvement in the performance of the variable.

It may be that the overprescribing of statin to individuals at a lower level of risk may lead to lower levels of effect than would be expected from trials. Adherence to therapy may also be a factor. Irish research shows that ~30% of the patients enrolled in a disease management programme for heart failure had episodes of non-persistence with therapy of varying lengths.19

Our research clearly had a number of drawbacks. Ideally, longitudinal data on circulatory mortality alongside individual specific data on prescribing would be used for this analysis. In its absence, we use aggregate data and time series methods within which we have a limited number of degrees of freedom to examine the hypotheses at issue. One of the main drawbacks of this method is that we cannot test for independent effects of different medications and combinations thereof without experiencing unacceptable levels of autocorrelation. This is unfortunate as cardiovascular medications are frequently used in combination and it may be that the combined therapy is a more important factor than any one component. The use of aggregate data also introduces the possibility that what we believe are individual level associations between probability of death and cardiovascular medications are in fact the result of another process entirely that is not observed in our data. We cannot exclude the possibility of such an ‘ecological fallacy’ by identical models to those estimated in this article by using osteoporosis rather than CVD medications found that the drug parameters did not impact on the trend-break variable in both the male and female models. This is not conclusive but does give some reassurance that the associations we have observed are plausible.

Secondly, our data on cardiovascular medications measure prescribing patterns only for medical card holders over the age of 64 years living in the Eastern Region, which represents ~one-third of our population. This partial coverage is not as problematic as it would be among the younger population as 67% of those aged ≥65 years had access to a medical card in the late 1990s rising to 87% after 2001 with 100% coverage of those >70. Ideally data on prescribing for the total population would be used although this is not available at present.

A factor which we have not examined in this article but may have played a role is the wider availability of the influenza vaccine after 1998. Data on the extent of take up of vaccine among older age groups is not available from the PCRS at present, but it would be very useful to be able to control for this factor in future analyses.

Acknowledgements

The authors would like to gratefully acknowledge Denis Conniffe for his contribution to the methodology used in this article and Sean Lyons for his insightful comments on earlier drafts. The authors would also like to thank the participants of seminars at the Economic and Social Research Institute and the Royal College of Surgeons in Ireland.

Conflicts of interest: None declared.

Key points

  • Death rates from circulatory causes and the seasonality of these deaths fell sharply around 1999–2000 in Ireland for older age groups.

  • Between summer 1999 and the end of 2003 cardiovascular prescribing increased dramatically in Ireland among older age groups due to changed practices and eligibility for free care.

  • Controlling for temperature, β-blocker, ace-inhibitor and aspirin medications statistically account for the structural break in death rates among men.

  • β-Blocker. Aspirin and calcium channel blocker medications statistically account for the structural break in death rates among women.

  • The impact of cardiovascular prescribing on death rates was weaker for women than men.

References

View Abstract