Osteoporos Int (2010) 21:21252134 DOI 10.

1007/s00198-010-1189-2

ORIGINAL ARTICLE

Functional data modelling approach for analysing and predicting trends in incidence ratesan application to falls injury
S. Ullah & C. F. Finch

Received: 23 December 2009 / Accepted: 12 January 2010 / Published online: 4 March 2010 # International Osteoporosis Foundation and National Osteoporosis Foundation 2010

Abstract Summary Policy decisions about the allocation of current and future resources should be based on the most accurate predictions possible. A functional data analysis (FDA) approach improves the understanding of current trends and future incidence of injuries. FDA provides more valid and reliable long-term predictions than commonly used methods. Introduction Accurate information about predicted future injury rates is needed to inform public health investment decisions. It is critical that such predictions derived from the best available statistical models to minimise possible error in future injury incidence rates. Methods FDA approach was developed to improve longterm predictions but is yet to be widely applied to injury epidemiology or other epidemiological research. Using the specific example of modelling age-specific annual incidence of fall-related severe head injuries of older people during 19702004 and predicting rates up to 2024 in Finland, this paper explains the principles behind FDA and demonstrates their superiority in terms of prediction accuracy over the more commonly reported ordinary least squares (OLS) approach. Results Application of the FDA approach shows that the incidence of fall-related severe head injuries would increase by 2.32.6-fold by 2024 compared to 2004. The FDA predictions had 55% less prediction error than traditional OLS predictions when compared to actual data.

Conclusions In summary, FDA provides more accurate predictions of long-term incidence trends than commonly used methods. The production of FDA prediction intervals for future injury incidence rates gives likely guidance as to the likely accuracy of these predictions. Keywords Accidental falls . Functional data analysis . Incidence . Injuries . Prediction . Trend

Introduction Injuries resulting from falls in older people are a major public health concern in terms of morbidity and mortality, and the associated cost of health and social services for older people has increased with concurrent rises in life expectancy [17]. It is expected that there will be a significant increase in the size of the older population in most countries over the next 20 years, with a coincident increase in the number of falls [2, 4, 8], and this will escalate the size of the burden that injurious falls will place on both individuals and their communities [9]. Although fall-related injuries are a recognised major public health concern, until recently relatively little has been known about trends in falls injury rates [911]. Recent increased interest in modelling falls injury rates [24, 12 14] and projections [1, 8, 10, 1518] has largely been driven by the need for good data to inform government policy and planning in relation to fundamental changes in welfare policy needed in response to the projected population demographic shifts [5, 19]. Importantly, such models and predictions will only be useful in the long term if they are accurate, based on good quality data and generated through application of appropriate statistical methods.

S. Ullah (*) : C. F. Finch School of Human Movement and Sport Sciences, University of Ballarat, Mt Helen, VIC 3353, Australia e-mail: s.ullah@ballarat.edu.au

2126

Osteoporos Int (2010) 21:21252134

Much injury epidemiology is largely descriptive in nature, and it has only been recently that sophisticated statistical modelling approaches have been applied and developed to handle complex injury data [20, 21]. To date, most falls injury incidence trend predictions have been based on the ordinary least squares (OLS) methods [8, 11, 17, 18]. Poisson and negative binomial regressions have recently become more commonly used to model and predict injury incidence [3, 4, 22]. Common problems associated with these regression model prediction approaches arise because there are strong parametric assumptions underlying the statistical models, most notably that the trend in incidence rates and the resources used to treat these injuries will remain unchanged into the forecast period [23, 24]. Another approach has been to use Poisson regression modelling age, period and cohort effects [25, 26]. However, there is ongoing debate of applying ageperiodcohort models with the high sensitivity of the projections to the most recent changes in cohort effects [23, 2729]. The introduction of stochastic methods for predicting rates in other health areas has been shown to have significant advantages for better understanding trends, risk factor relationships and the effectiveness of preventive measures [23, 30 32]. A major advantage of stochastic methods is that they can measure prediction uncertainty through the estimation of prediction intervals of future injury incidence. An important milestone has been the methodology proposed by [31, 33] for the modelling and extrapolating of long-term trends in mortality rates used to predict US mortality rates to 2065. More recently, functional data analysis (FDA) approach has started to receive attention, particularly in demographic and medical applications [23, 30, 34, 35]. To our knowledge, FDA has not been widely applied to epidemiological studies for the prediction of incidence rates, including studies relating to injury incidence associated with falls in older people. Using the specific example of fall-related head injuries in older people, this paper shows how to apply FDA to describe time trends of age-specific incidence rates and their future projections. The modelled incidence rates are discrete observations for specified ages and calendar year and predictions are made into the future. The essence of FDA is to express these discrete values in the form of a function, and then to draw statistical information from the resultant collection of functional data by applying multivariate data analysis approaches. This approach improves the likely accuracy of these predictions, compared to current methods in the literature such as the more common OLS method. In applying FDA, there are a number of key statistical issues that need to be addressed, and this paper provides guidance and examples for: useful graphical

tools for effective data representation; making no parametric assumptions about the function or the observed data; ensuring the functions are smooth to enable modelling of the datas dynamic behaviour; handling complexity and interdependencies between variables; and deriving prediction intervals to assess the likely accuracy of projections. In the specific application area of modelling fall-related head injury, this study: (a) models age-related changes in fall-related head injury incidence time trends using FDA; (b) predicts future age-related changes in these incidence rates using a recently developed prediction method [30]; and (c) estimates the accuracy of predictions for future fallrelated head injuries.

Methods Description of data Both fatal and non-fatal head injuries can be a serious consequence of falls in older people. A recent study in Finland showed that the secular trend in the incidence of these injuries in people aged 80+ years is rising [16]. That study also compared actual, observed fall-related head injury incidence rates to previous predictions of those rates by the same authors [8]. The earlier study adopted the common OLS method for projecting long-term trends from 1970 to 1995. The presence of published predictions [8] and data on actual observations published years later [16] prompted this comparison of two predictive approaches (OLS and FDA) for estimating future incidence trends in fall-related head injuries in Finland. The actual data underpinning the published analysis [16] was obtained directly from the first author of that study (Pekka Kannus, UKK Institute for Health Promotion Research, Tampere, Finland). The data originated from the Finnish National Hospital Discharge Register and cases were identified through relevant International Classification of Diseases primary and secondary diagnoses and E-code causes of injury. A severe fall-related head injury was defined as a head injury as a consequence of a fall from standing height of 1 m and that resulted in hospitalisation. Data obtained was the annual number and incidence of fallrelated severe head injuries during the period 19702004 in people aged 8084, 8589 and 90+ years. All incidence rates were age-adjusted by direct standardisation using the mean population of persons aged 80+ years between 1970 and 2004 as the standard population. The age-specific injury incidence was calculated in 5-year age groups and was expressed as the number of cases per 100,000 persons per year. A more detailed description of the data has been published elsewhere by [8, 16].

Osteoporos Int (2010) 21:21252134

2127

Functional data analysis Let yt(x) denote the observed fall-related head injury incidence for age x in year t. Then we assume the following models yt x ft x s t x"t x
K X k 1

particularly relevant because it leads to functional data that can be analysed with FDA [3840]. Functional principal components The smoothed curves (or functional observations) were then modelled by FDA to estimate the functions representing the ageincidence relationship, (i.e. the basis functions), according to the method of Hyndman and Ullah [30]. As all parameters on the right-hand side of Eq. 2 are unobservable, fitting the model using the OLS method is impossible. To overcome the situation, functional principal components (FPC) decomposition was applied to the smoothed incidence curves ft(x) to estimate the basis functions fk(x) which minimise the mean integrated squared error (MISE)
n Z 1X MISE e2 t x dx; n t 1

ft x mx

bt;k fk x et x

where ft(x) is an underlying head injury incidence smooth function of x observed with error, t(x) are independently and identically distributed standard normal random variable and t(x) allows the variance to change with age and year according to the nature of the data. The second equation describes the dynamics of ft(x) evolving through time. In this equation, (x) is the mean of smooth incidence curves ft(x) across years and et(x) is the model error. The age component fk(x) is a set of orthogonal basis functions or principal components which modifies the main time trend according to whether change at a particular age is faster or slower than the main trend (and in the same or opposite direction). The model assumes that fk(x) is invariant over time. The time component, t,k, are time series coefficients which capture the overall time trend in ft(x) at all ages. The model makes no assumptions about the functional form of the trend in t,k. Smoothing Within each age group, the annual fall-related severe head injury incidence rates were considered as a function of age, centred around the mid-point of the age group. The age-specific injury incidence rates for each year were plotted and the incidence rates defined as age incidence curves. The age-specific injury incidence rates for each year were smoothed using penalised regression splines [30] to estimate the ageincidence curves. Since the incidence in the Finnish data increased rapidly with age, the incidence curves were assumed to increase monotonically with age. To capture the trends in incidence, constrained penalised regression splines [36, 37] with a monotonic constraint applied to represent the smooth curves were used according to the method of Hyndman and Ullah [30]. The gam() function in R from the mgcv package was used with the penalised regression splines to represent the curves. The use of penalised regression splines has a number of advantages in that: (a) the smoothness conditions can easily be adopted to the nature of the incidence data analysed, thereby reducing the noise in the incidence curves; and (b) the underlying process generating the ageincidence curves is then continuous and smooth. This continuity property is

where et x ft x ft x denote the fitted error from Eq. 2. The basis functions show the shape of the variability of the incidence curves. The extracted functions can be easily interpreted because the basis functions are defined in the same domain as the original incidence curves. The predictions of each incidence curve are sensitive to the number of basis functions. A higher number of basis functions does not necessarily give better forecasts or improve the prediction accuracy [34]. Although there are many ways for selecting the basis functions, the method used [30] minimises the MISE.

Prediction framework The FPC method described in the previous section also estimated the time series coefficients of the model. The values of the coefficients form a time series, with one value representing each year of data, and standard time series methods can be used to model and predict these time series coefficients. Although a model incorporating a random walk with drift is almost exclusively used for predicting rates in other health areas [31, 41], an exponential smoothing state space model was found to be statistically adequate for the current dataset and was therefore used. Suppose we have injury incidence data up to year t = n, and we wish to estimate future incidence rates to year n ^ h; yt x for t n 1; :::::; n h and all x. Let b n;k ;h denote ^n;h x denote the the h-step ahead prediction of n+h,k. Let y h-step ahead prediction of yn+h(x) and let f^n;h x denote the h-step ahead prediction of fn+h(x). Then x n;h x fn;h x m y
K X k 1

b n;k ;h fk x

2128

Osteoporos Int (2010) 21:21252134

Following [30], we can give the following expression for the variance of prediction 2 Vn;h x Varynh xjI; f s m x vx s 2 t x 5 where I y t x; t 1; :::::::; n denotes all observed data, unh;k Var bnh;k jb1;k ; :::::::::::; bn;k can be obtained 2 from the time series model, s m x (the variance of the x) can be obtained from the smoothing smooth estimate m 2 method used and v(x) is estimated by averaging e t x for each x. The averages of multi-step-ahead predictions and the corresponding standard deviation of the prediction errors were used to construct the prediction intervals [42, 43]. n;h x za y q Vn;h x 6
K X k 1

unh;k f2 k x

incidence rates increased with age and time. As described by Kannus and colleagues [16], over the 35-year period, the relative increases in severe fall-related head injury incidence rates were 160% for 8084-year-olds, 134% for 8489-year-olds and 132% for the 90+-year-olds. There was a large increase in incidence rates over the last decade in all age groups, though this was most marked in the 90+-year age group. The characteristics of these agegroup-specific incidence rates clearly indicate the appropriateness of the FDA approach for making predictions based on it. Firstly, there is a high degree of variability of incidence rates at older ages, and secondly the general increase in incidence rates over time is not uniform across all age groups. Functional data analysis The 5-year age-group incidence rates were converted into functional data by estimating a smooth curve through the observations, taking the centre of each age group as the point of interpolation. Examination of the comparatively stable data shown in Fig. 2 suggests that the incidence rates vary fairly smoothly and regularly within each year. The average incidence curve is shown in Fig. 3a. As expected, the average incidence rates sharply increase with age. In fitting the model, the model diagnostics suggested that the first three basis functions (or ageincidence relationships) were sufficient to describe the injury incidence curves. These three functions explained 97.5%, 2.0% and 0.5% of the total variation, respectively, and can therefore be taken as being jointly adequate for describing the incidence data. Because the second and third basis functions collectively explained only 2.5% of the total variation, they are less important than the first one. The first basis function (Fig. 3b) indicates that the increase in incidence has been faster in 90+-year-olds. The first coefficient (Fig. 3c) indicates a fairly steady increase in fall-related severe head injuries over time. Prediction 20 years Figure 4 shows the predicted age-specific fall-related severe head injury incidence rates for 20052024 with 95% prediction intervals, together with the observed rates for 19702004. The predicted incidence rates are estimated to increase sharply over time, as was previously noted by [16]. The predicted incidence rates for 2024 are: 1,108 injuries per 100,000 persons [95% confidence interval (95% CI)= 686, 1,792] for 8084-year-olds; 1,573 injuries per 100,000 persons (95% CI=1,059, 2,337) for 8589-year-olds; and 1,916 injuries per 100,000 persons (95% CI=1,251, 2,936) for 90+-year-olds; representing 2.32.6-fold increases over the 2004 rates for all age groups.

where za is the standard normal quantile. Prediction accuracy The accuracy of the injury incidence prediction is evaluated by computing the integrated squared prediction error (ISPE) Z 7 ISPEn h e2 n;h xdx
x

n;h x denote the prediction where en;h x ynh x y error. In designing the accuracy measures for the future ageincidence rates, an out-of-sample test was performed [44]. An out-of-sample evaluation of prediction accuracy begins with the division of the time series set into a fit period and a test period. The fit period is used to identify and estimate an appropriate model, based on a set of the observed data for that period, and does not involve any predictions. The test period also uses observed data but this is compared to predictions arising from the model generated for the fit period and so measures the models prediction accuracy. Based on the fitting period 19701994, the FDA predictions of injury incidence rates for 19952004 were directly compared with the published actual rates for 19952004 through averaging of the ISPE [30]. All statistical analyses were performed in R version 2.10.0 [45], operating on a Windows XP professional platform.

Results Descriptive Figure 1 shows the trends in age-specific incidence rates for head injuries in Finland over the period 19702004. All

Osteoporos Int (2010) 21:21252134 Fig. 1 Age-specific incidence rates of fall-induced severe head injuries in people aged 80+ years and older in Finland, 19702004
80-84 y ears 85-89 y ears 90+ years

2129

Incidence (per 100,000 persons)

900 800 700 600 500 400 300 200 100 1970 1975 1980 1985 1990 1995 2000 2005 90+ years

80-84 years

Year

Comparison of the FDA versus OLS predictions Data on the trends in the Finnish injury incidence rates for 19701994 (fit period) were used to predict the incidence for 19952004 (test period). The FDA predictions were compared with the published predictions based on the OLS method [16] for the test period. The actual rates for the fit period from [16] are also shown in Fig. 5. The shaded area corresponds to the 95% prediction intervals for the FDA predictions. In every age group, the FDA out-of-sample prediction performed better than the OLS predictions in that it was closer to the actual values at every point. The OLS

predictions were contained within the FDA prediction intervals for the earliest years in the test period but then deviated from, and were significantly lower than, the FDA predictions for longer test periods. A further advantage of the FDA approach was that it had approximately 55% less prediction error than the OLS method.

Discussion and conclusion Functional data analysis provides a relatively novel modelling and prediction approach to estimating incidence

Fig. 2 Age-specific smoothed incidence rates of fall-induced severe head injuries in people aged 80+ years and older in Finland, 19702004

900 800 700 600 500 400 300 200 100 1970 94 92 90 88 86 1975 1980 1985
Yea 1990 r

84

82 1995 2000 80 2005

Age

Ag e

85-89 years

Inci den

p ce (

er 1

00,

000

so per

ns)

2130 Fig. 3 Principal components from the functional data analysis (FDA) for incidence of fallinduced severe head injuries in Finland: a the mean of the incidence rates, b the first basis function and c the coefficient associated with the first basis function

Osteoporos Int (2010) 21:21252134

400 380 360 340

Mean

320 300 280 260 240 220 80 81 82 83 84 85 86 87 Age 88 89 90 91 92 93 94

b 1.66
1.64 1.62 1.60 Basis 1 1.58 1.56 1.54 1.52 1.50 80 81 82 83 84 85 86

87 Age

88

89

90

91

92

93

94

c
0.4 0.3 0.2 Coefficient 1 0.1 0.0 -0.1 -0.2 -0.3

1970 1972 1974 1976 1978 1980 1982 1984 1986 1988 1990 1992 1994 1996 1998 2000 2002 2004 Year

trends. However, to date their application to epidemiological data has been limited and, to our knowledge, nonexistent in the injury literature. This paper, therefore, represents the first application of FDA to a significant injury epidemiology problem, from any context. With ageing populations worldwide and the expected increasing burden of fall-related morbidity and mortality in older people, it is becoming more important to understand the trends in all types of serious fall-related injury incidence

trends over time. Moreover, policy makers need information about predicted future injury rates to inform their decision making about public health and economic investments to reduce the burden of these injuries well into the future [19]. It is critical that such predictions are robust and based on the best available statistical modelling approaches so as to minimise possible error in the forecasts. This is also true for other areas of public health and injury prevention.

Osteoporos Int (2010) 21:21252134 Fig. 4 Predicted age-specific incidence of fall-induced severe head injuries in people aged 80+ years in Finland, 20052024: a 8084 years, b 8589 years and c 90+ years. 95% prediction intervals are shaded in the figures

2131

a
Incidence (per 100,000 persons)

3000 2800 2600 2400 2200 2000 1800 1600 1400 1200 1000 800 600 400 200 1970 1975 1980 1985 1990 1995 2000 Year 2005 2010 2015 2020 2025 2030
80-84 y ears

b
Incidence (per 100,000 persons)

3000 2800 2600 2400 2200 2000 1800 1600 1400 1200 1000 800 600 400 200 1970 1975 1980 1985 1990 1995 2000 Year 2005 2010 2015 2020 2025 2030
85-89 y ears

c
Incidence (per 100,000 persons)

3000 2800 2600 2400 2200 2000 1800 1600 1400 1200 1000 800 600 400 200 1970 1975 1980 1985 1990 1995 2000 Year 2005 2010 2015 2020 2025 2030
90+ years

The FDA approach is a natural extension of methods developed for mortality and fertility prediction that have evolved over the last two decades [30, 31, 33]. The methodology has since become widely used in demographic applications and there have been various extensions and modifications proposed [33, 35, 4648]. Somewhat surprisingly, the use of FDA in epidemiological applications has been limited to date. The FDA is effective for exploring complex multivariate functional relationships and it has the major strength of

being able to model the functional form of age-related changes in incidence rates over time and to make predictions for specific age groups [30]. In the injury context, this facilitates understanding of the different patterns of injury rate progression in different age groups. By applying it to the Finnish data, we have shown that it can provide more precise estimates of trends in fall-related severe head injuries among older people than is possible with the more common OLS method projections. Our analysis agrees with the earlier findings of [16] that the falls

2132 Fig. 5 Comparisons of the predictions of the incidence of fall-induced severe head injuries using functional data analysis (FDA) and ordinary least squares (OLS) methods. The fit period is 19701994 and the test period is 19952004: a 8084 years, b 8589 years and c 90+ years. The FDA 95% prediction intervals are shaded in the figures

Osteoporos Int (2010) 21:21252134

a 1000
FDA OLS

900 Incidence (per 100,000 persons) 800 700 600 500 400 300 200 100 1970 1975 1980
80-84 y ears

Fit period
1985 1990 Year 1995 2000

Test period
2005 2010

b1000
FDA OLS

900 Incidence (per 100,000 persons) 800 700 600 500 400 300 200 100 1970 1975 1980
85-89 y ears

Fit period
1985 1990 Year 1995 2000

Test period
2005 2010

c 1000
FDA OLS

900 Incidence (per 100,000 persons) 800 700 600 500 400 300 200 100 1970 1975 1980
90+ years

Fit period
1985 1990 Year 1995 2000

Test period
2005 2010

injury rates are expected to rise over time. However, our FDA estimates are likely to be more accurate than the published OLS predictions, especially the further ahead of time the predictions are made. A second major strength of the FDA approach is the improved modelling of inconsistent increases in incidence rates over time, particularly the high degree of variability of incidence rates at older ages. The OLS approach with autocorrelated errors previously used with this data [8, 16] was

unable to capture the unstable trend and high variability of injury incidence rates at older ages. While it is important to understand the reason for such inconsistencies, it is equally important that statistical models and the predictions they generate are able to factor them into their estimation processes. Because FDA has not been previously applied to injury data, it is possible that failure to account for this could have lead to incorrect or non-optimal epidemiological inferences about future injury outcomes. The FDA approach

Osteoporos Int (2010) 21:21252134

2133 2. Kannus P, Parkkari J, Koskinen S et al (1999) Fall-induced injuries and deaths among older adults. JAMA 281:18951899 3. Boufous S, Finch C, Lord S et al (2005) The increasing burden of pelvic fractures in older people, New South Wales, Australia. Injury 36:13231329 4. Boufous S, Finch CF, Lord SR (2004) Incidence of hip fracture in New South Wales: are our efforts having an effect? Med J Aust 180:623626 5. Moller J (2005) Current costing models: are they suitable for allocating health resources? The example of fall injury prevention in Australia. Accid Anal Prev 37:2533 6. Hindmarsh DM, Hayen A, Finch CF et al (2009) Relative survival after hospitalisation for hip fracture in older people in New South Wales, Australia. Osteoporos Int 20:221229 7. Kannus P, Parkkari J, Niemi S et al (2005) Fall-induced deaths among elderly people. Am J Public Health 95(3):422424 8. Kannus P, Palvanen M, Niemi S et al (1999) Increasing number and incidence of fall-induced severe head injuries in older adults: nationwide statistics in Finland in 19701995 and prediction for the future. Am J Epidemiol 149:143150 9. Mann E, Icks A, Haastert B et al (2008) Hip fracture incidence in the elderly in Austria: an epidemiological study covering the years 1994 to 2006. BMC Geriatr 8:35 10. Kannus P, Niemi S, Palvanen M et al (2005) Rising incidence of fall-induced injuries among elderly adults. J Public Health 13:212215 11. Kannus P, Palvanen M, Niemi S et al (2009) Rate of proximal humeral fractures in older Finnish women between 1970 and 2007. Bone 44:656659 12. Robertson MC, Campbell AJ, Herbison P (2005) Statistical analysis of efficacy in falls prevention trials. J Gerontol A Biol Sci Med Sci 60:530534 13. Lord SR (2006) Visual risk factors for falls in older people. Age Ageing 35(Suppl 2):4245 14. Boufous S, Finch C, Lord S et al (2006) The epidemiology of hospitalised wrist fractures in older people, New South Wales, Australia. Bone 39:11441148 15. Kannus P, Palvanen M, Niemi S et al (2007) Alarming rise in the number and incidence of fall-induced cervical spine injuries among older adults. J Gerontol A Biol Sci Med Sci 62:180183 16. Kannus P, Niemi S, Parkkari J et al (2007) Alarming rise in fallinduced severe head injuries among elderly people. Injury 38:8183 17. Kannus P, Niemi S, Parkkari J et al (1999) Hip fractures in Finland between 1970 and 1997 and predictions for the future. Lancet 353:802805 18. Kannus P, Palvanen M, Niemi S et al (2002) Increasing number and incidence of low-trauma ankle fractures in elderly people: Finnish statistics during 19702000 and projections for the future. Bone 31:430433 19. Finch CF, Hayen A (2006) Governmental health agencies need to assume leadership in injury prevention. Inj Prev 12:23 20. MacKenzie EJ (2000) Epidemiology of injuries: current trends and future challenges. Epidemiol Rev 22:112119 21. Bangdiwala SI (2009) Random or accidentalwhat is the goodness of the fit? Int J Inj Contr Saf Promot 16:5960 22. Rivara FP, Cummings P, Koepsell TD et al (2001) Injury control: a guide to research and program evaluation. Cambridge University Press, New York 23. Erbas B, Hyndman RJ, Gertig DM (2007) Forecasting agespecific breast cancer mortality using functional data models. Stat Med 26:458470 24. Mathers CD, Loncar D (2006) Projections of global mortality and burden of disease from 2002 to 2030. PLoS Med 3:20112030 25. Li G, Shahpar C, Grabowski JG, Baker SP (2001) Secular trends of motor vehicle mortality in the United States, 19101994. Accid Anal Prev 33:423432

of initially smoothing the data and then using these smoothed observations for the modelling and prediction estimation is a major methodological improvement over methods that simply fit linear/non-linear trends to observed incidence data, such as OLS methods. Finally, FDA approach provides prediction intervals for future incidence projections. All previous injury prediction studies have only presented point estimates [1, 8, 10, 15, 16, 49]. Although the calculation of prediction intervals is possible using OLS methods, they are likely to be inaccurate outside the range of injury incidence because it is unlikely that the linear relationship assumption would hold for future incidence [50]. The methods presented in this paper could easily be incorporated into other public health and injury incidence applications, including the modelling of known risk factors and economic evaluations. As the shape of the incidence curves varies with risk factors over time, the FDA approach enables the models and predictions to pick up subtle variations for these factors [23], such as the identified age effects. In the context of application of FDA to the incidence of fall-related severe head injuries in Finns aged 80+ years, this study provides an improved understanding of current trends (particularly in terms of the age relationships) and more accurate estimates of future incidence rates. Moreover, this paper clearly demonstrates the greater predictive accuracy of FDA over OLS approach, the former having 55% less prediction error for the same data, and the generation of prediction intervals that give information about the likely accuracy of the predictions. In summary, this paper describes FDA and its important application to epidemiology. In doing so, it establishes FDA as a preferred modelling approach for incidence trends and predictions, with the potential for many significant applications across a range of epidemiological issues.
Acknowledgements The authors wish to thank Professor Pekka Kannus, Chief Physician and Head, Injury & Osteoporosis Research Centre, UKK Institute for Health Promotion Research, Tampere, Finland for providing the data analysed in this study. Dr Shahid Ullah was supported by an Injury Trauma and Rehabilitation (ITR) Research Fellowship funded through a National Health and Medical Research Council (NHMRC) Capacity Building Grant in Population Health. Professor Caroline Finch was supported by an NHMRC Principal Research Fellowship. Conflicts of interest None.

References
1. Kannus P, Niemi S, Palvanen M et al (2000) Continuously increasing number and incidence of fall-induced, fractureassociated, spinal cord injuries in elderly persons. Arch Intern Med 160:21452149

2134 26. Wildner M, Casper W, Bergmann KE (1999) A secular trend in hip fracture incidence in East Germany. Osteoporos Int 9:144150 27. Icks A, Haastert B, Wildner M et al (2008) Trend of hip fracture incidence in Germany 19952004: a population-based study. Osteoporos Int 19:11391145 28. Holford TR (1991) Understanding the effects of age, period, and cohort on incidence and mortality rates. Annu Rev Public Health 12:425457 29. Willets RC (2004) The cohort effect: insights and explanations. Br Actur J 10:833877 30. Hyndman RJ, Ullah S (2007) Robust forecasting of mortality and fertility rates: a functional data approach. Comput Stat Data Anal 51:49424956 31. Lee RD, Carter LR (1992) Modeling and forecasting U.S. mortality. J Am Stat Assoc 87:659675 32. Lee R, Miller T (2002) An approach to forecasting health expenditures, with application to the U.S. Medicare system. Health Serv Res 37:13651386 33. Lee RD, Miller T (2001) Evaluating the performance of the Lee Carter method for forecasting mortality. Demography 38:537549 34. Hyndman RJ, Booth H (2008) Stochastic population forecasts using functional data models for mortality, fertility and migration. Int J Forecast 24:323342 35. De Jong P, Tickle L (2006) Extending LeeCarter mortality forecasting. Math Popul Stud 13:118 36. Ruppert D, Wand MP, Carroll RJ (2003) Semiparametric regression. Cambridge University Press, New York 37. Simonoff JS (1996) Smoothing methods in statistics. Springer, New York 38. Ramsay JO, Dalzell CJ (1991) Some tools for functional data analysis. J R Stat Soc Series B Stat Methodol 53:539572

Osteoporos Int (2010) 21:21252134 39. Ramsay JO, Silverman BW (2002) Applied functional data analysis. Springer, New York 40. Ramsay JO, Silverman BW (2005) Functional data analysis, 2nd edn. Springer, New York 41. Lee RD, Tuljapurkar S (1994) Stochastic population forecasts for the United States: beyond high, medium, and low. J Am Stat Assoc 89:11751189 42. Hyndman RJ, Koehler AB, Ord JK, Snyder RD (2005) Prediction intervals for exponential smoothing using two new classes of state space models. J Forecast 24:1737 43. Hyndman RJ, Koehler AB, Snyder RD et al (2002) A state space framework for automatic forecasting using exponential smoothing methods. Int J Forecast 18:439454 44. Tashman LJ (2000) Out-of-sample tests of forecasting accuracy: an analysis and review. Int J Forecast 16:437450 45. R Development Core Team (2004) A language and environment for statistical computing. R Foundation for Statistical Computing: Vienna, Austria. Available at: http://www.R-project.org 46. Renshaw A, Haberman S (2003) LeeCarter mortality forecasting: a parallel generalized linear modelling approach for England and Wales mortality projections. Appl Stat 52:119137 47. Lee RD (1993) Modeling and forecasting the time series of US fertility: age distribution, range, and ultimate level. Int J Forecast 9:187202 48. Booth H, Maindonald J, Smith L (2002) Applying LeeCarter under conditions of variable mortality decline. Popul Stud 56:325336 49. OConnor PJ (2005) Forecasting of spinal cord injury annual case numbers in Australia. Arch Phys Med Rehabil 86:4851 50. Makridakis S, Wheelwright SC, Hyndman RJ (1998) Forecasting: methods and applications, 3rd edn. Wiley, New York