5 STUDIES ON RECENT TRENDS IN CORONARY HEART DISEASE MORTALITY RATES: ENGLAND & WALES,
5.2 METHODS: TREND ANALYSIS USING JOINPOINT REGRESSION
This section describes the approaches I used for the trend analyses presented in this thesis. Details on specific issues relevant to each setting will be described in the methods section of each analysis.
The key question to answer is whether the rate of change varies in different time periods.
Most of the descriptive analyses on CHD trends have relied on simple qualitative trend descriptions. However, more formal trend analysis have rarely been done, except in the field of environmental CHD epidemiology129. Regression methods have been used to explain trends in terms
of risk factor contributions18; but the question of direction and speed of change in CHD epidemiology has not been formally addressed. In part, this can be attributable to the slow adoption of change- point methods in cardiovascular epidemiology. This type of time series analysis look at trends trying to identify points where the trend change and is commonly used in fields like cancer epidemiology, econometrics and statistical process control.208
One of these methods is the Joinpoint Regression Analysis approach. This was developed by the U.S. National Cancer Institute with the purpose of describing change in cancer rates in terms of pace of change, direction of that change and statistical significance.209
Joinpoint regression models explore the trend data to find points in time (“joinpoints”) defining segments where the trend has a constant pace of change. The method starts by assuming that there are no “joinpoints” thus, the entire trend has the same rate of change. Then, it iteratively adds joinpoints by performing a “grid search” (looking for points in time that minimize the error on the change of the rate) and thus identifying the years where the trend rate of change might vary, effectively identifying the “joinpoints”. Then, it formally tests whether these additional joinpoint creates segments with different rate of change. Statistical significance is tested using a Montecarlo permutation method. Permutation test method and the Bayesian Information criterion method offer two alternative goodness of fit approaches to select the optimal number of joinpoints. The Permutation test approach uses a sequence of permutation tests to select the number of joinpoint and controls the error probability of selecting the wrong model at a certain level (i.e. 0.05) using the Bonferroni multiple comparisons adjustment. The Bayesian Information criterion method (BIC) approach finds the model with the best fit by penalizing the cost of extra parameters, favouring trends with fewer segments. The models selected by BIC tend to fit the data well but are less parsimonious. The PT approach has worked well for cancer incidence and mortality data, however little experience is available in the cardiovascular literature, although the results of both approaches are generally concordant (see appendix A1). Information of the uncertainty around the rate of change and the time of the point of change in the trend can also be obtained.
Identifying patterns in a trend is thus the main goal of the approach. Figure 5-1 and 5-2 illustrate an example of this analysis. Figure 5-1 shows a rate plotted over time. The trend appears complex, with a period of decline that seems to have changed around 1988, continuing a slow decline or even a flat pattern.
Figure 5-1 An observed trend for a rate (simulated data)
A joinpoint regression analysis on this data (Figure 5-2), found two distinct periods in the trend. The point of change has been located in 1991, with a 95% confidence interval of having occurred in 1987 or 1994. This point creates two segments of different rate of change. The first segment started in 1976 and ended in 1991 with an annual percent change of - 5.3% (95% CI -6.6% to -3.9%), statistically significantly different from a 0% rate of change.
The last segment of the trend shows a “flat” pattern, starting in 1991 and ending in 2006, where the annual rate of change was 0% (-1.7% to 1.7%). Models with 0, 2, 3 or 4 joinpoints were rejected using the Bayesian Information Criterion (BIC) approach for model selection. The key strength of this technique is to avoid the detection of biased patterns when the trend is described subjectively using observer defined time intervals. The observer might bias in several ways the choice of periods for estimating summaries of the rate of change of the trend, based on prior knowledge or because the trend shows curious patterns. Joinpoint avoids this by essentially removing the observer from the selection process, instead using a formal and objective exploration of the time-series data space.
0.00 0.10 0.20 0.30 0.40 0.50 0.60 1976 1979 1982 1985 1988 1991 1994 1997 2000 2003 2006 R ate Observed Rate
Figure 5-2 Observed and joinpoint modelled rates (simulated)
In the next sections, I will use this approach to examine recent interesting phenomena observable in CHD mortality trends in a variety of different countries. Flattening of previously falling CHD mortality rates has recently been reported among young adults in the US, perhaps reflecting changes in specific risk factors. Since young adults in many other Western countries are also now showing complex risk factor trends, a similar flattening in mortality rates patterns is distinctly possible.
I will start by describing recent trends in CHD mortality in England & Wales, particularly amongst young adults, and then I will look for recent trend patterns in Australia and the Netherlands. Later I will examine recent trends in Poland and its determinants, and trends by socioeconomic status in Scotland and England.
0 0.1 0.2 0.3 0.4 0.5 0.6 1977 1979 1981 1983 1985 1987 1989 1991 1993 1995 1997 1999 2001 2003 2005 rate Observed Rate Segment 1 Segment 2