SEARCH

SEARCH BY CITATION

Keywords:

  • cancer mortality;
  • cancer surveillance;
  • Surveillance, Epidemiology, and End Results;
  • temporal projection;
  • projection methods

Abstract

  1. Top of page
  2. Abstract
  3. MATERIALS AND METHODS
  4. RESULTS
  5. DISCUSSION
  6. FUNDING SOURCES
  7. REFERENCES

BACKGROUND:

A study was undertaken to evaluate the temporal projection methods that are applied by the American Cancer Society to predict 4-year-ahead projections.

METHODS:

Cancer mortality data recorded in each year from 1969 through 2007 for the United States overall and for each state from the National Center for Health Statistics was obtained. Based on the mortality data through 2000, 2001, 2002, and 2003, Projections were made 4 years ahead to estimate the expected number of cancer deaths in 2004, 2005, 2006, 2007, respectively, in the United States and in each state, using 5 projection methods. These predictive estimates were compared to the observed number of deaths that occurred for all cancers combined and 47 cancer sites at the national level, and 21 cancer sites at the state level.

RESULTS:

Among the models that were compared, the joinpoint regression model with modified Bayesian information criterion selection produced estimates that are closest to the actual number of deaths. Overall, results show the 4-year-ahead projection has larger error than 3-year-ahead projection of death counts when the same method is used. However, 4-year-ahead projection from the new method performed better than the 3-year-ahead projection from the current state-space method.

CONCLUSIONS:

The Joinpoint method with modified Bayesian information criterion model has the smallest error of all the models considered for 4-year-ahead projection of cancer deaths to the current year for the United States overall and for each state. This method will be used by the American Cancer Society to project the number of cancer deaths starting in 2012. Cancer 2012; © 2012 American Cancer Society.

For more than half a century, the American Cancer Society (ACS) has published the number of cancer deaths expected to occur in the current year for the total United States and for individual states.1, 2 These estimates have traditionally been available in January in order to provide the contemporary burden of cancer in that particular year. However, in the past few years, these estimates were published as late as June because of delays in the release of the final mortality data for the most recent data year from the National Center for Health Statistics, which are used in the computation of the estimates. The timing of release of final mortality data for the United States is largely contingent on receipt of information on all deaths registered in all states. The delay in receipt of data from even 1 state may result in delays in the release of national data.

The estimated number of cancer deaths in the current year have been based on a statistical model (the state-space model3) since 2004 that projects 3 years ahead of the most recent data year using the actual number of cancer deaths beginning in 1969. However, the estimated deaths in 2011 were based on a 4-year-ahead projection using the actual number of cancer deaths from 1969 to 2007,4 because of a protracted delay in the release of the final mortality data for 2008.

In this article, we compare the accuracy of the state-space model and 4 additional models for 4-year-ahead projection of cancer deaths to the current year, nationally and in each state, in order to determine the best method of projection. The additional models are: 1) the Bayesian state-space model,5 2) 2 versions of the Joinpoint regression model (based on the permutation test and modified-BIC criteria for selecting the number of joinpoints),6 3) the Nordpred model,7 and 4) the vector autoregressive model via Hilbert-Huang transform.8

Based on the historical data, we used all the methods to project the number of deaths 4 years ahead to 2004, 2005, 2006, and 2007 using data through 2000, 2001, 2002, and 2003, respectively. Because it is not obvious if a longer historical series necessarily produces more accurate predictions, we tried all methods with both the full historical series (using data starting in 1969), and a shorter series using the last 15 years of historical data. These predictive estimates are compared to the observed number of deaths that occurred for all cancers combined, and 47 cancer sites at the national level and 21 cancer sites at the state level. In the following sections, we describe the data used for this evaluation, develop criteria for evaluating the performance of the models, describe the 5 models, and summarize the results of our testing. Finally, we select a method to be used for projecting cancer mortality counts by the American Cancer Society for 2012 and after.

MATERIALS AND METHODS

  1. Top of page
  2. Abstract
  3. MATERIALS AND METHODS
  4. RESULTS
  5. DISCUSSION
  6. FUNDING SOURCES
  7. REFERENCES

Data

We obtained cancer mortality data recorded in each year from 1969 through 2007 for the United States overall and for each state from the National Center for Health Statistics. The cancer mortality data can also be retrieved from the Surveillance, Epidemiology, and End Results (SEER)9 database of the National Cancer Institute. For the national data, we used all cancer sites combined and 47 individual cancer sites. For the state-level data, we selected 21 individual cancer sites and all the cancer sites combined. Table 1 shows the cancer sites we used. The associated International Classification of Diseases codes can be viewed from SEER Web site (http://seer.cancer.gov/codrecode/1969+_d09172004/index.html).

Table 1. Cancer Sites Used in the Analysis
National LevelState Level
Oral cavity and pharynxOral cavity and pharynx
TongueEsophagus
MouthStomach
PharynxColon and rectum
Other oral cavityLiver and intrahepatic bile duct
EsophagusPancreas
StomachLarynx
Small intestineLung and bronchus
Colon and rectumMelanoma of the skin
Anus, anal canal, and anorectumBreast
Liver and intrahepatic bile ductUterine cervix
Gallbladder and other biliaryUterine corpus
PancreasOvary
Other digestive organsProstate
LarynxUrinary bladder
Lung and bronchusKidney and renal pelvis
Other respiratory organsBrain and other nervous system
Bones and jointsHodgkin lymphoma
Soft tissue (including heart)Non-hodgkin lymphoma
Melanoma of the skinMultiple myeloma
Other nonepithelial skinLeukemia
Breast 
Uterine cervix 
Uterine corpus 
Ovary 
Vulva 
Vagina and other genital, female 
Prostate 
Testis 
Penis and other genital, male 
Urinary bladder 
Kidney and renal pelvis 
Ureter and other urinary organs 
Eye and orbit 
Brain and other nervous system 
Thyroid 
Other endocrine 
Hodgkin lymphoma 
Non-hodgkin lymphoma 
Multiple myeloma 
Leukemia 
Acute lymphocytic leukemia 
Chronic lymphocytic leukemia 
Acute myeloid leukemia 
Chronic myeloid leukemia 
Other leukemia 
Other and unspecified primary sites 

The comparison is based on 4-year-ahead projection using data from 1969 to 2007. For full available years data points, we used mortality data from 1969 to 2000 to predict cancer deaths in 2004, data from 1969 to 2001 to predict cancer deaths in 2005, data from 1969 to 2002 to predict cancer deaths in 2006, and data from 1969 to 2003 to predict cancer deaths in 2007. To have a more thorough comparison among these different methods, we use various lengths of years of historical mortality data to project 4 years ahead. In addition to using the entire time span of the data beginning in 1969, we consider a shorter sequence of years to project future mortality because: 1) mortality data have been constantly improving in quality, and so a reliance on older more error-prone data may not be justified, and, 2) we noticed that the trend for larger cancer sites is approximately linear in recent years, and it was not clear if older data would add more signal or noise to the prediction. In particular, we chose 15 years of data for projection, because 15 years is the minimum number on which the Nordpred method can be run (see description of Nordpred below). For the shorter series, we used 15 years of mortality data from 1986 to 2000 to predict cancer deaths in 2004, data from 1987 to 2001 to predict cancer deaths in 2005, data from 1988 to 2002 to predict cancer deaths in 2006, and data from 1989 to 2003 to predict cancer deaths in 2007. For all the methods, we compared the result using both the 15-year and all available years historical series.

Method of Evaluation

To measure the accuracy of the prediction counts, the prediction error is estimated for each year by computing the predicted minus the observed value. We then find the ratio of the prediction error to the observed count, adding 0.5 to the denominator to avoid numerical error when the observed count is zero. The ratio is the percentage error of the prediction, compared to the observed count:

  • equation image

Finally, the average absolute relative deviation (AARD) is computed as the average of these error ratios over all the cases considered in the data. More specifically, assume equation image is the predicted mortality for specific scenario s, s = 1,…,S, and that θs is the true observed count. Then

  • equation image

For example, when calculating AARD for all-sites mortality at the state level, the observed data are death counts for each site-state combination from 2004 to 2007. There are 8088 cases across the states and cancer sites, and for each case, there are observed deaths from 2004, 2005, 2006, and 2007. Therefore, S = 8088(4) = 32352.

The AARD is interpreted as the average percent deviation from the true value (number of deaths) relative to the true value. This measure attempts to take into account the relative differences in observed mortality counts among different cancers and/or geographic areas as we attempt to assess the extent to which the estimates deviate from the observed. We also obtain the AARD for each method considered and for categories of the mortality counts, from rarest to most common, both at the national (US) and the state level. For example, the AARD for the category with less than 400 deaths (“<400”) is the average of percent deviation over many types of cancer in sparsely populated states or rare cancers in any state for the period from 2004 to 2007.

Other measures of discrepancy are also obtained to compare the methods. The maximum absolute relative deviation is a measure of the worst prediction error seen for each method over the categories of cancer types or numbers of deaths. The mean relative sums of squares deviation is similar to the AARD, only deviations are squared before computing the average, resulting in applying higher weights to larger deviations. The root mean square error (RMSE) is an estimate of variability of estimates about the true value. The normalized RMSE is the RMSE expressed as a fraction of the mean. The average rank of the relative sums of squares is the average rank of deviations among the methods. (For the technical details, please see a technical report at http://surveillance.cancer.gov/reports/review/appendix_cancercounts.pdf, section A.5). These measures have similar results to the AARD; therefore, in this study, we used AARD as the default measure for comparison. Instead of declaring a single “winner” for each evaluation, we consider methods that are within 15% of the winner using the described metrics.

Statistical Projection Methods

State-space method (SS-F and SS-15)

We use SS-F to represent the state-space method using full data and SS-15 for the state-space method using 15-year data. The state-space method has been used to project the cancer mortality counts 3 years ahead to the current calendar year since 2004.3 The method uses a local quadratic model to obtain projections of the time series of mortality counts 4 years into the future. The method proceeds as follows. First, to account for the uncertainty in measuring the observed mortality counts, the count at any point of time is assumed to be a realization of a random phenomenon, which has a normal distribution. This is the so-called measurement equation. The choice of normal distribution is made for the purposes of mathematical simplicity, and is a reasonable one when the mortality counts are high (such as when dealing with data at the national level).

Next, we modeled the average trajectory (ie, the trajectory obtained by joining the averages of these normal distributions) using a quadratic trend. Because mortality trends can be reasonably expected to change over time due to policy changes and scientific advances, we also assumed that the quadratic trend is not fixed but instead varies locally. Such a locally varying quadratic trend is obtained by representing the year-to-year variation of trend in the form of a so-called state-space model,10 whereby one relates the parameters (or state) of the model at a particular point to those of the previous point through a set of so-called transition equations. These transitions are assumed to have a stochastic component, and as before, errors in these transitions are assumed to have normal distributions.

The measurement equation, combined with the transition equations, completes the specification of the state-space model. An iterative procedure called the Kalman filter11 can then be used to estimate the trend and project it into the future. The Kalman filter, however, requires that the variances in the measurement and transition equations be known. For our case, we estimated them using the method of moments, coupled with an additional step, whereby the estimates are also adjusted to minimize the sum of squares of prediction errors for the years where observed data are available. We refer to this step as the “tuning” process, in the sense that the parameters of the model are “tuned” to make the predictions close to the actual values. The whole-model fitting and projection is done using the statistical software R.12

Bayes state-space method (BSS-F and BSS-15)

As an alternative method of projection of time series counts, we explored the use of dynamic generalized linear models, which were fitted using the Bayesian paradigm.13 We will henceforth refer to this overall method as the Bayes method. The method proceeds as follows: Because the observed values are yearly mortality counts, we assumed that for any year, the observed mortality count is a realization of a random variable, taken for our purposes to have the Poisson distribution. The parameters of the Poisson distributions are assumed to be unknown and vary smoothly from year-to-year according to a random process. In particular, because the Poisson parameter has to be positive, we assume that the logarithm of the parameter (henceforth called the state) in any year is a normal perturbation around the state of the previous year, with the amount of variation in this perturbation constant from year to year. All these can be used to put together the likelihood function, which contains all the information on the unknown model parameters from the data. Combined with prior information on the unknown parameters in the form of distribution of the initial state (at time t = 0) and the variance of the year-to-year transition of the states, the likelihood is used to obtain the posterior distribution of the parameters. Although the full posterior distribution is unavailable in a closed form, the univariate (and sometimes multivariate) conditional distributions are easy to sample from. We used a combination of various Markov chain Monte Carlo methods to generate observations from a Markov chain, that converges to the posterior distribution. Convergence was achieved within almost 100,000 iterations. After convergence, an additional 100,000 iterations of this chain were used in the generation of 4-year-ahead predictions, picking out every 100th iteration to reduce the autocorrelation. The generated predictions are estimates of the posterior of the corresponding predictive distributions. Detailed discussion of the Bayes state-space methods with particular emphasis on modeling time series of counts in epidemiology can be found in Schmidt and Pereira.5

We use BSS-F to represent the Bayes state-space method using full data, and BSS-15 for the Bayes state-space method using 15-year data.

Joinpoint regression (JP-Perm-F, JP-Perm-15, JP-MBIC-F, JP-MBIC-15)

Joinpoint regression has been applied to cancer trend analysis to summarize trend changes by using segmented line regression.6 Cancer incidence and mortality rates are modeled as a function of time, which is composed of piecewise linear segments. The model is fitted by the least squares method at a given number of change-points, called joinpoints, and then the number of joinpoints is estimated. The Joinpoint software (http://surveillance.cancer.gov/joinpoint) provides the point and interval estimates of the slope parameters, which are equivalent to the annual percent change (APC) rates, as well as the point and interval estimates of the joinpoints themselves. To select the number of joinpoints, the first version (1.0a) of Joinpoint software applied the permutation test procedure, where the permutation distribution of the test statistic is used to calculate the P value to see if the data show enough evidence to add more joinpoints. Because version 1.0a was released in 1998, Joinpoint has been continuously updated to improve its accuracy and efficiency and to include additional features.

One of the updates is the addition of 2 other model selection procedures: Bayes information criterion (BIC) and modified Bayes information criterion (MBIC). The BIC was added as a faster alternative to the permutation procedure, which is computationally expensive, and the MBIC was proposed to improve the performance of the BIC that tends to overfit the model. In this study, we used Joinpoint 3.5.1 and implemented both the permutation procedure and the MBIC. The Joinpoint program allows a user to change model-fitting specifications, which include the maximum number of joinpoints, the minimum number of observations from a joinpoint to either end of the data, and the minimum number of observations between 2 joinpoints. In our comparison, we used a maximum of 2 joinpoints, a minimum of 4 observations from a joinpoint to either end of the data (including the joinpoint), and a minimum of 4 observations between 2 joinpoints (inclusive) for 15 years of data (called JP-Perm-15 and JP-MBIC-15 for permutation-based and MBIC-based, respectively). For full data, we used a maximum of 5 joinpoints, a minimum of 4 observations from a joinpoint to either end of the data (including the joinpoint), and a minimum of 4 observations between 2 joinpoints (called JP-Perm-F and JP-MBIC-F for permutation-based and MBIC-based, respectively).

Vector autoregressive model via Hilbert-Huang transform (VAR-F, VAR-15)

Spectral analysis is a complementary tool for analyzing time series. Fourier transform, wavelet transform, and some other spectral analysis methods have been developed to analyze stationary time series data. These techniques are widely used by many scientists in the fields of electrical engineering, physics, among many others. Recently, the Hilbert-Huang Transform (HHT) based on empirical mode decomposition (EMD) was developed for nonlinear and nonstationary processes.8 The advantage of HHT-EMD is that it does not require a set of prespecified functions. Instead, it uses a set of adaptive intrinsic mode functions (IMF) derived from the time series data itself. Since its inception, the HHT-EMD methods have been applied to image processing, biomedical research, health monitoring, and atmospheric engineering.14

To project the mortality count, we apply the EMD to decompose mortality data into several IMFs and then apply multivariate time-series technique to these IMFs for projection. Among many multivariate time series models, vector autoregressive analysis (VAR) is a standard instrument for forecasting procedure. VAR is a multivariate version of autoregressive model. Although univariate time-series models may be useful for describing short-term correlation, the multivariate time-series models could give a better description of the underlying structure of the time series and give a better forecast. An R software package “vars” is used for obtaining the forecast counts.

We used VAR-F and VAR-15 to represent the vector autoregressive model via the Hilbert-Huang transform method using full data and 15-year data, respectively.

Nordpred (NP) method (NP-F and NP-15)

Time trends in incidence and mortality have also been modeled using the age–period–cohort (APC) model.15 The data used in APC model consists of a set of age-specific counts tabulated for several periods of time. The Nordpred model is an age–period–cohort model, and it has been applied to model the mortality and incidence counts.7 Interval widths for age and period are typically assumed to be equal. The “Nordpred” model used in this analysis is essentially Model B1 of Møller et al.7 Model B1 assumes mortality counts follow a Poisson distribution with linear effects of age, period, and cohort linked to the expected rate through a fifth root function. The period effect is further modeled as a linear regression on period, with the regression coefficient being the drift parameter. In this analysis, predictions involved only 25% of the drift regression parameter for the first future period estimated rates and no drift for the second future period estimated rates.

To fully implement the Nordpred method in this study, the age and period variables information of the death counts are provided from the SEER database, and the age and period variables are then grouped into 5-year intervals. We used 18 five-year age categories (0-4, 5-9, …, 85+), and 7 or 3 five-year periods are used depending on whether the data used is full data or 15-years data, respectively. Age and period effects were for 5-year groupings. Five-year predictions were translated to single-year estimates using linear interpolation. We use NP-F and NP-15 to represent the Nordpred model method using full data and 15-year data, respectively.

RESULTS

  1. Top of page
  2. Abstract
  3. MATERIALS AND METHODS
  4. RESULTS
  5. DISCUSSION
  6. FUNDING SOURCES
  7. REFERENCES

Table 2 shows a summary statistics for comparison of 12 methods (including variations of the 5 basic methods) applied to the national and state data. For the national data, averaged over the 347 combinations of cancer sites and sex, JP-MBIC-15 has the smallest AARD (0.065). Any method with AARD within 15% of this value will be considered to be within the best method circle, and thus could be a candidate for being selected as the best method. With a minimal AARD of 0.065, we obtain a threshold of 0.0748 for defining “best methods”. In addition to JP-MBIC-15, the methods NP-15, VAR-15, and BSS-15 fall into this category. These methods are shown in bold in the first row of Table 2.

Table 2. AARD of the 4-Year-Ahead Projection From the Methods Compared for All Cancer Sites Combineda
 NP-FSS-FJP-Perm- FVAR-FJP-MBIC- FBSS-FNP- 15SS- 15JP-Perm- 15VAR- 15JP-MBIC- 15BSS- 15
  • AARD indicates average absolute relative deviation; BSS, Bayesian state-space method; JP-MBIC, joinpoint method with modified Bayesian information criterion; JP-Perm, joinpoint method with permutation test; NP, Nordpred method; SS, state-space method; VAR, vector autoregressive analysis method.

  • a

    Methods with AARD within the threshold are in bold, and the winner method with the minimum AARD is in bold and italics.

National level0.094.0850.0750.10.0770.0750.0720.1180.0760.0710.0650.071
State level0.170.1760.1850.1970.1830.1590.1720.3050.1740.1850.170.162

For the state-level data, averaged over 8076 combinations of cancer sites, sex, and states, BSS-F has the smallest AARD (0.159). Adding 15% of 0.159, we obtain a threshold of 0.183. The methods BSS-F, NP-F, SS-F, NP-15, JP-Perm-15, JP-MBIC-15, and BSS-15 all fall below the threshold. The methods that are within the threshold are shown in bold in row 2 of Table 2.

Taking into consideration both the national and state-level data, only NP-15, JP-MBIC-15, and BSS-15 are the remaining possible choices because they fall below the respective thresholds at both the national and state level. Notice that although BSS-F has the smallest AARD in the state level, it is not chosen because it is not in the group falling below the 15% threshold in the national data.

Cancer sites with larger mortality are considered more common. The last row in Table 3 shows the AARD from all the methods compared for cancer sites with mortality counts greater than 200,000 (200K) cases. For these cancer sites with larger count, the national and state-level large cancer sites are overlapped with each other. The method SS-F has the smallest value (AARD = 0.011), and the methods JP-MBIC-15 (AARD = 0.015), JP-Perm-15 (AARD = 0.015), and BSS-15 (AARD = 0.012) are not too far behind.

Table 3. AARD of the 4-Year-Ahead Projection From the Methods Compared for the National Dataa
SizebNP-FSS-FJP-Perm- FVAR-FJP-MBIC- FBSS-FNP- 15SS- 15JP-Perm- 15VAR- 15JP-MBIC- 15BSS- 15
  • AARD indicates average absolute relative deviation; BSS, Bayesian state-space method; JP-MBIC, joinpoint method with modified Bayesian information criterion; JP-Perm, joinpoint method with permutation test; NP, Nordpred method; SS, state-space method; VAR, vector autoregressive analysis method.

  • a

    The cancer sites are categorized by the average number of death counts over 2004 to 2007. Methods with AARD within the threshold for the category are in bold, and the winner method with the minimum AARD in the category is in bold and italics.

  • b

    The cancer sites are categorized by number of observed mortality count.

<4000.1080.1110.1070.1020.0830.1010.0940.2360.1040.0820.0830.094
401-10000.1390.120.1030.1340.1070.1090.1080.1420.0990.1070.0820.105
1001-50000.0710.0810.0830.0890.0830.0550.0630.1090.070.0630.0760.049
5K-10K0.0660.0540.0380.0630.0450.0630.0440.070.0380.0670.0450.058
10K-20K0.0670.0540.040.0610.0310.0670.0460.0630.0370.0490.0350.079
20K-200K0.0910.0580.0670.1440.0470.0440.0530.0910.050.0690.0390.04
>200K0.0690.0110.0160.0290.0180.0170.0210.0440.0150.0310.0150.012

To further consider which method to use, we group the cancer sites by the number of deaths observed. This is to decide whether one method is better depending on the rarity of cancer death. Tables 4 and 5 show the summary of the AARD from difference sizes considered when the methods are applied to national and state-level data, respectively. At the national level, we group the cancer sites into 7 categories (<400, 401-1000, 1001-5K, 5K-10K, 10K-20K, 20K-200K, and >200K). At the state level, the cancer sites are grouped into 12 categories (<25, 26-50, 51-100, 101-200, 201-300, 301-400, 401-1000, 1001-5000, 5K-10K, 10K-20K, 20K-200K, >200K). In each category, we first find the minimum AARD among the 12 methods and then find a threshold by adding 15% of the minimum to it. Any method that falls below the threshold is considered a candidate for the best method.

Table 4. AARD of the 4-Year-Ahead Projection From The Methods Compared For The State Dataa
SizebNP-FSS-FJP-Perm- 15VAR- 15JP-MBIC- 15BSS- 15NP- 15SS- 15JP-Perm- 15VAR- 15JP-MBIC- 15BSS- 15
  • AARD indicates average absolute relative deviation; BSS, Bayesian state-space method; JP-MBIC, joinpoint method with modified Bayesian information criterion; JP-Perm, joinpoint method with permutation test; NP, Nordpred method; SS, state-space method; VAR, vector autoregressive analysis method.

  • a

    The cancer sites are categorized by the average number of death counts over 2004 to 2007. Methods with AARD within the threshold for the category are in bold, and the winner method with the minimum AARD in the category is in bold and italics.

  • b

    The cancer sites are categorized by number of observed mortality count.

<250.3980.4050.3970.4330.3890.3340.40.7650.410.4230.3940.34
26-500.2050.1960.2050.2290.2060.1840.2080.3340.2030.2270.1920.186
51-1000.1470.1540.1660.1730.1690.1460.1490.2710.1430.1550.140.149
101-2000.1150.1210.1360.1420.1430.1170.1170.1830.120.1280.1180.119
201-3000.0930.1040.1110.1140.1250.1030.0940.1540.10.1010.1010.105
301-4000.0770.0910.0870.1080.1010.090.0790.1380.0830.0930.0850.091
401-10000.070.0830.0780.0920.0870.0810.070.1160.0760.0810.0770.083
1001-50000.0430.0540.0510.0720.0530.0530.0420.0710.050.0530.0550.056
5K-10K0.0330.0380.0320.0560.0370.0440.0320.0540.0350.0480.0460.048
10K-20K0.0270.030.0240.0510.020.0450.0260.0430.0250.0370.030.049
20K-200K0.0360.040.0290.0760.0330.040.0340.0690.0330.0480.0320.039
>200K0.0250.0110.0160.0290.0180.0230.0230.0440.0150.0280.0150.035
Table 5. Frequency of the Methods That Fall Below the Threshold
 NP-FSS-FJP-Perm- FVAR-FJP-MBIC- FBSS-FNP- 15SS- 15JP-Perm- 15VAR- 15JP-MBIC- 15BSS- 15
  1. AARD indicates average absolute relative deviation; BSS, Bayesian state-space method; JP-MBIC, joinpoint method with modified Bayesian information criterion; JP-Perm, joinpoint method with permutation test; NP, Nordpred method; SS, state-space method; VAR, vector autoregressive analysis method.

US011022101144
State855035808375
All8660579094119

Across the groups, the number of times the methods with AARD fall below the respective threshold is calculated and summarized in Table 5. The joinpoint regression modified BIC (JP-MBIC-15) method performs best at the national level and combined nation-state level. For the JP-MBIC-15 method at the national level, 4 of 7 groups are below the threshold; at the state level, 7 of 12 groups are below the threshold. Overall, JP-MBIC-15 has 11 of 19 groups falling below the threshold, and it outperforms all other methods and is our final choice for mortality projection.

Figure 1 shows the trends of AARD for all the methods we compared at the national level. Except for the SS-F method, all other methods show an increasing trend over the year-ahead projections. For example, they all have larger AARD for 4-year-ahead projection than 3-year-ahead projection except the SS-F method, as expected. Notice that SS-F is the current method that has been used by the American Cancer Society for projecting the upcoming year mortality count since 2004. Figure 1 shows that SS-F performed worse than many methods we considered.

thumbnail image

Figure 1. Trends of AARD at the national level are shown for all the methods we compared.

Download figure to PowerPoint

An important finding from the comparisons is that using the full time span of the data does not have an advantage over shorter time series in improving the accuracy of prediction. Figure 2 also shows that with the same method, use of 15 years of data tends to perform better than 30 years of data. This is not always true at the state level where, for example, BSS-F is better than BSS-15. Yet in general, Figure 2 shows a similar pattern to Figure 1.

thumbnail image

Figure 2. Trends of AARD at the state level are shown for all the methods we compared.

Download figure to PowerPoint

DISCUSSION

  1. Top of page
  2. Abstract
  3. MATERIALS AND METHODS
  4. RESULTS
  5. DISCUSSION
  6. FUNDING SOURCES
  7. REFERENCES

The comparison of variations of 5 methods of projecting the number of cancer deaths indicates that the JP-MBIC-15 model is better than any of the other methods at the national level for 4-year-ahead projection. At the state level, BSS-15 has a slightly smaller AARD than VAR-15, NP-F, and JP-MBIC-15. We could have used BSS-15 for state-level data in projection, but we felt that JP-MBIC-15 performed reasonably well (with 15% of BSS-15). Because we felt that it would be better to use a single method if possible, our decision was to use JP-MBIC-15 at both the state and national levels in Cancer Facts & Figures 2012 and Cancer Statistics, 2012.

Nordpred has been used in the Nordic countries and United Kingdom to predict cancer incidence and mortality.7, 16 It performs relatively well for the state-level data. For the state-space and Bayes state-space methods, their strengths are that they are both flexible and can easily adapt to changing trends. However, the state-space method is better suited for common cancers, because it makes assumption of normality for the measurement equation. Both the VAR models and Bayesian state-space models are stochastic model-based methods that tend to capture sudden changes in trends and perform better in shorter time series and noisy data. In the companion article by Zhu et al17 about incidence rate projection, we found that these 2 methods perform the best among many others in incidence count projection.

The Joinpoint regression model was developed by the National Cancer Institute. The method has been used to evaluate cancer trends in the United States. The model needs to determine the number and locations of the joinpoints by permutation tests. Because of the computational burden of the permutation tests, a modified BIC test is incorporated in finding the joinpoints. The JP-MBIC-15 method is useful to identify changes in the trend, and is sensitive to changes in trends near the end of the series. This makes it suitable for short-term exploration.

The fluctuations in mortality are usually relatively attenuated compared to changes in incidence, thus the best methods recommended for incidence and mortality are different. Even though JP-MBIC-15 is determined to be the best overall method, it could be outperformed by other methods in some instances. Both the frequency of cancer deaths for a specific cancer site/geography combination and the length of the projection (ie, 1, 2, 3, or 4 years) play a major role in accuracy of the projection. Nevertheless, JP-MBIC-15 outperforms the method previously used by the ACS, SS-F, for 4-year-ahead projections. The AARD at the national level for 4-year-ahead projections using JP-MBIC-15 is 0.065, whereas that for SS-F is 0.085.

One motivation to reexamine the methodologies for mortality projections was that ACS was forced to move from a 3-year to a 4-year projection because of delays in the arrival of the data. Despite the concern that 4-year projections would be significantly worse than 3-year projections, the 3-year AARD for the previously used model (SS-F) was larger than the 4-year AARD for the selected model (JP-MBIC-15). At the national level, the 3-year-ahead projection for SS-F methods has AARD = 0.092 and the 4-year-ahead projection for JP-MBIC-15 has AARD = 0.065, whereas at the state level, the 3-year-ahead projection for SS-F methods has AARD = 0.178 and the 4-year-ahead projection for JP-MBIC-15 has AARD = 0.17.

Because the methods used to project these counts ahead in time are applied to each state separately, differences in the size or age structure of the population or underlying incidence rates between states are irrelevant. That is, none of the projection methods borrow strength from or assume a correlation of results with neighboring states; each state projection is independent of the others. Therefore, it is only necessary to assume that there are no sudden changes to the state population characteristics at some point in time.

The Nordpred method attempts to decompose the pattern of change in the deaths over time into changes due to age, period, and cohort effects, then projects the period effects ahead in time, holding age effects constant. The other methods treat each state's number of deaths as a time series of counts to be projected ahead based on prior trends in changes in the counts, regardless of the reason for those changes. If, for example, a state's population is not growing in size but is becoming younger over time, we would expect the number of cancer deaths to decrease over time, becuase there are fewer older residents who are more likely to die from cancer. This age effect is but one of many that may affect the number of cancer deaths; the methods, with the exception of Nordpred, do not attempt to identify each component of change in the counts, but simply attempt to model the change in the number of deaths that have resulted from an aggregate of causes.

FUNDING SOURCES

  1. Top of page
  2. Abstract
  3. MATERIALS AND METHODS
  4. RESULTS
  5. DISCUSSION
  6. FUNDING SOURCES
  7. REFERENCES

No specific funding was disclosed.

CONFLICT OF INTEREST DISCLOSURE

Dr Ghosh's work is supported by National Institutes of Health (NIH) contract HHSN261201000671P. Dr Kim's work is supported by NIH contract HHSN261201000509. Dr Pickle's work is supported by NIH contract HHSN261201100094P. The other authors made no disclosure.

REFERENCES

  1. Top of page
  2. Abstract
  3. MATERIALS AND METHODS
  4. RESULTS
  5. DISCUSSION
  6. FUNDING SOURCES
  7. REFERENCES
  • 1
    American Cancer Society. Cancer Facts & Figures 1953. New York, NY: American Cancer Society; 1953.
  • 2
    American Cancer Society. Cancer Facts & Figures 1961. New York, NY: American Cancer Society; 1961
  • 3
    Tiwari RC, Ghosh K, Jemal A, et al. A new method of predicting US and state-level cancer mortality counts for the current calendar year. CA Cancer J Clin. 2004; 54: 30-40.
  • 4
    Siegel R, Ward E, Brawley O, Jemal A. Cancer statistics, 2011: the impact of eliminating socioeconomic and racial disparities on premature cancer deaths. CA Cancer J Clin. 2011; 61: 212-236.
  • 5
    Schmidt AM, Pereira JBM. Modelling time series of counts in epidemiology. Int Stat Rev. 2011; 79: 48-69.
  • 6
    Kim HJ, Fay MP Feuer EJ, Midthune DN. Permutation tests for joinpoint regression with applications to cancer rates. Stat Med. 2000; 19: 335-351.
  • 7
    Møller B, Fekjaer H, Hakulinen T, et al. Prediction of cancer incidence in the Nordic countries: empirical comparison of different approaches. Stat Med. 2003; 22: 2751-2766.
  • 8
    Huang NE, Shen Z, Long SR, et al. The empirical mode decomposition and the Hilbert spectrum for nonlinear and non-stationary time series analysis. Proc R Soc Lond A. 1998; 454: 903-995.
  • 9
    National Cancer Institute, Surveillance, Epidemiology, and End Results. http://seer.cancer.gov/. Accessed on October 11, 2011.
  • 10
    Harvey AC. Time Series Models, 2nd ed. Cambridge, MA: MIT Press; 1993
  • 11
    Kalman RE. A new approach to linear filtering and prediction problems. J Basic Eng. 1960; 82: 35-45.
  • 12
    R Development Core Team. R: A language and environment for statistical computing, reference index version 2.13. Vienna, Austria: R Development Core Team; 2011. http://www.R-project.org.
  • 13
    West M, Harrison J. Bayesian Forecasting and Dynamic Models. New York, NY: Springer; 1997.
  • 14
    Wu ZH, Huang NE. A study of the characteristics of white noise using the empirical mode decomposition method. Proc R Soc Lond A. 2004; 460: 1597-1611.
  • 15
    Holford TR. The estimation of age, period and cohort effects for vital rates. Biometrics. 1983; 39: 311-324.
  • 16
    Olsen AH, Parkin DM, Sasieni P. Cancer mortality in the United Kingdom: projections to the year 2025. Br J Cancer. 2008; 99: 1549-1554.
  • 17
    Zhu L, Pickle LW, Ghosh K, et al. Predicting US- and state-level cancer counts for the current calendar year: part II: evaluation of spatiotemporal projection methods for incidence. Cancer. 2011; 117: 1100-1109.