The accuracy of general practitioner workforce projections
 Malou Van Greuningen^{1}Email author,
 Ronald S Batenburg^{1, 2} and
 Lud FJ Van der Velden^{1}
DOI: 10.1186/147844911131
© Van Greuningen et al.; licensee BioMed Central Ltd. 2013
Received: 19 February 2013
Accepted: 2 July 2013
Published: 16 July 2013
Abstract
Background
Health workforce projections are important instruments to prevent imbalances in the health workforce. For both the tenability and further development of these projections, it is important to evaluate the accuracy of workforce projections. In the Netherlands, health workforce projections have been done since 2000 to support health workforce planning. What is the accuracy of the techniques of these Dutch general practitioner workforce projections?
Methods
We backtested the workforce projection model by comparing the expost projected number of general practitioners with the observed number of general practitioners between 1998 and 2011. Averages of historical data were used for all elements except for inflow in training. As the required training inflow is the key result of the workforce planning model, and has actually determined past adjustments of training inflow, the accuracy of the model was backtested using the observed training inflow and not an average of historical data to avoid the interference of past policy decisions. The accuracy of projections with different lengths of projection horizon and base period (on which the projections are based) was tested.
Results
The workforce projection model underestimated the number of active Dutch general practitioners in most years. The mean absolute percentage errors range from 1.9% to 14.9%, with the projections being more accurate in more recent years. Furthermore, projections with a shorter projection horizon have a higher accuracy than those with a longer horizon. Unexpectedly, projections with a shorter base period have a higher accuracy than those with a longer base period.
Conclusions
According to the results of the present study, forecasting the size of the future workforce did not become more difficult between 1998 and 2011, as we originally expected. Furthermore, the projections with a short projection horizon and a short base period are more accurate than projections with a longer projection horizon and base period. We can carefully conclude that health workforce projections can be made with data based on relatively short base periods, although detailed data are still required to monitor and evaluate the health workforce.
Keywords
Health workforce Projections Evaluation AccuracyBackground
One of the major challenges in healthcare systems worldwide is that of managing the health workforce to meet the demands of an accessible and effective health service. Shortages and imbalances of healthcare personnel are a major concern of health policymakers, professional bodies and patient organizations [1–5]. Health workforce planning is an important instrument to prevent shortages and oversupply within the healthcare workforce [6–9]. An increasing number of countries apply different types of health workforce planning. Recently, Matrix Insight [10] conducted a study that provides an overview of health workforce planning in the European Union and shows a large variation across countries. Thirteen European countries, including the Netherlands, engage in modelbased workforce planning, all of which use some form of supplyside projections.
Health workforce projections require accurate and comprehensive information and careful accounting of stocks and flows of human resources for health [10]. In most settings, the results and methods of workforce projections are not monitored and evaluated regularly and, consequently, it is difficult to assess whether workforce planning has been successful and projections are accurate. This implies that shortcomings and room for improvement are difficult to identify [10]. For the feasibility and further development of workforce projections in rapidly changing health systems, it is important to evaluate the accuracy of projections and their techniques [11]. The increasing dynamics of the health workforce – through mobility [12], reduction of working hours, the ageing workforce, increasing number of female physicians, changing division of labour – implies that projecting the future workforce could become more difficult [13–16].
The accuracy of the Dutch simulation model
A simulation model had been developed in 2000 to support health workforce planning in the Netherlands. This model calculates the required number of health professionals in training to advise the Ministry of Health on the adjustment of the inflow numbers per year, to balance the supply and demand and to prevent a shortage or an oversupply of health professionals in the future [17–20].
The launch year is the year of the latest data used as a basis to make a projection and the target year is the projection year. Other terms used in this article are the projection horizon, which is the interval between launch year and target year, and the base period, which is the period of data the projection is based on (the interval between base year and launch year, with the base year being the year of the earliest data).
In the Netherlands, general practitioners (GPs) are of high importance as they provide primary health care 24 hours a day, 7 days a week and are the ‘gatekeepers’ of the healthcare system [13]. Additionally, there is much data available about the Dutch GP, because the Netherlands Institute for Health Services Research (NIVEL) administers a GP database, which provides longitudinal information about all Dutch GPs regarding gender, age, position, moment of firsttime accreditation, etcetera since 1975 [14, 19, 24, 25].
The question of this article is: what is the accuracy of the current model for Dutch GP workforce projections? To answer this question, we will conduct a posteriori projections to backtest the current workforce projections and compare the projected expost number of GPs with the observed number of GPs. In practice, the Dutch GP workforce is projected with a base period of 15 years. Longterm data are used to prevent base data being influenced by fluctuations. Specifically, we will compare the accuracy of projections based on 15year base periods and based on 5 or 10year base periods to investigate if a shorter base period is as accurate as a 15year base period.
There is no standard for workforce projection horizon lengths, but in European countries, a 10year projection horizon is common [26, 27]. In the Netherlands, it is common to make projections with horizons of 10 and 15 years, because of the relatively long period of physician training. Furthermore, it takes one or two whole years to adjust the inflow in training, because of the decisionmaking process [20]. The accuracy of different lengths of projection horizons is tested.
 1.
The longer the projections, the lower the accuracy of the Dutch GP workforce projection model is.
 2.
The shorter the base period, the lower the accuracy of the Dutch GP workforce projection model is, because short base periods could be influenced by fluctuating data.
 3.
The accuracy of the Dutch GP workforce projection model will be highest when the lengths of the base period and the projection horizon are similar. Hypothesis 3 is not dependent on hypotheses 1 and 2.
Methods
Backtesting (or hindcasting) is the process of evaluating a strategy, theory, or model by applying it to historical data. A key element of backtesting that differentiates it from other forms of historical testing is that backtesting calculates how a strategy would have performed if it had actually been applied in the past. This requires the backtest to replicate the conditions of the time in question in order to get an accurate result. In this article, the Dutch GP workforce projection model is backtested [33, 34] by comparing a posteriori projections with the observed number of GPs in the target years. The projections of the GP workforce are made using the current version of the workforce simulation model and historical GP workforce data retrieved from the NIVEL GP database. The only way we can evaluate the current model is by using historical data to generate new projections. Original projections are not available to evaluate the performance of the simulation model.
All data and assumptions used in the projections are – depending on the length of the base period – based on 5year averages from preceding periods (0 to 5 years, 0 to 10 years and 0 to 15 years back), except for the inflow in training. This inflow is not based on an average of historical data, but the observed inflow in training is used to test the accuracy of the modelling techniques. The reason for this is that the workforce simulation model actually has influenced the inflow in GP training in the past – as its results are taken into account by the Ministry and stakeholders in their decision about GP training inflow in the Netherlands [20]. Hence, the observed inflow in training is used in the a posteriori projections to exclude past interference of policy decisions with regard to training inflow. Using inflow projections made in the past would obviously blur the method of backtesting as applied in this study.
nGP = number of GPs; nOUT = number of outflow; nIN = number of inflow; T = target year; X = projection horizon; Y = base period; ϵ = projection error.
The total estimated supply of GPs in the future $\left(\hat{n}G{P}_{T,X,Y}\right)$ is calculated using the GPs in stock in the launch year (nGP _{ T,X }), minus the estimated outflow $\left(\hat{n}\mathit{OU}{T}_{T,X,Y}\right)$, plus the estimated inflow $\left(\hat{n}I{N}_{T,X,Y}\right)$ of GPs in the years between launch and target year (T − X → T), based on a specific base period (T − X − Y → T − Y). For example, to predict the number of GPs in 2011 (e.g. 12 000), the number of GPs in stock in 2006 is used (e.g. 10 000). The estimated outflow between 2006 and 2011 (e.g. 2 000) is subtracted from the 2006 GP number and the estimated inflow between 2006 and 2011 (e.g. 4 000) is added to the 2006 GP number to predict the 2011 number. The estimated outflow and inflow numbers are based on observed data between 2001 and 2006.
The estimated inflow $\left(\stackrel{\u02c6}{n}I{N}_{T,X,Y}\right)$ is composed of several parts: the inflow from abroad and its labour market return, and the inflow from Dutch training and its return on training and labour market return. For example, the estimated inflow between 2006 and 2011 (e.g. 4 000) is calculated by multiplying the inflow from abroad between 2006 and 2011 (e.g. 250) with the labour market return of this inflow (e.g. 80%) and then add the inflow from Dutch training (e.g. 4 200) multiplied by the return on training (e.g. 95%) and its labour market return (e.g. 85%).
Several sources provide information for the projections. This is mainly the NIVEL GP database, which provides information about the GP stock [24]. Other sources are the training institutions and the Medical Accreditation Committee, which provide data for elements of the model, such as return on training [20].
The GP database is administered according to Dutch privacy legislation. The privacy regulation was approved by the Dutch Data Protection Authority. According to Dutch legislation, approval by a medical ethics committee was not required for this kind of data collection.
Calculating the projection errors
The accuracy of the a posteriori GP workforce projections is backtested for three different projection horizons (X) and three different base periods (Y). By comparing the results of the projections with the observed number of GPs (for the target years 1998 to 2011), the mean absolute percentage errors (MAPE) are calculated. The MAPE is a summarizing measure to express the error during a certain period of time and ignores the direction of error. It has been used frequently in evaluations of population forecast accuracy [23, 35, 36].
nGP = number of GPs; T = target year; X = projection horizon; Y = base period; ϵ = projection error.
The first target year, 1998, is determined by the first year of available data, 1968, and the sum of 15 years of base data and a 15year horizon.
Years and time periods on which observations are based and which are used to calculate the projections’ accuracy
5year base period  10year base period  15year base period  MAPE_{x}  

5year projection horizon  Base years  1988  2001  1983  1996  1978  1991  
Base periods  1988 → 1993 –  1983 → 1993 –  1978 → 1993 –  
2001 → 2006  1996 → 2006  1991 → 2006  
Launch years  1993  2006  1993  2006  1993  2006  
Projection horizons  1993 → 1998 –  1993 → 1998 –  1993 → 1998 –  
2006 → 2011  2006 → 2011  2006 → 2011  
Target years  1998  2011  1998  2011  1998  2011  
MAPE_{x,y}  1.9 (14 tests)  3.0 (14 tests)  4.2 (14 tests)  3.0 (42 tests)  
10year projection horizon  Base years  1983 1996  1978  1991  1973  1986  
Base periods  1983 → 1988 –  1978 → 1988 –  1973 → 1988 –  
1996 → 2001  1991 → 2001  1986 → 2001  
Launch years  1988  2001  1988  2001  1988  2001  
Projection horizons  1988 → 1998 –  1988 → 1998 –  1988 → 1998 –  
2001 → 2011  2001 → 2011  2001 → 2011  
Target years  1998  2011  1998  2011  1998 – 2011  
MAPE_{x,y}  4.9 (14 tests)  7.0 (14 tests)  9.1 (14 tests)  7.0 (42 tests)  
15year projection horizon  Base years  1978  1991  1973  1986  1968 – 1981  
Base periods  1978 → 1983 –  1973 → 1983 –  1968 → 1983 –  
1991 → 1996  1986 → 1996  1981 → 1996  
Launch years  1983  1996  1983  1996  1983  1996  
Projection horizons  1983 → 1998 –  1983 → 1998 –  1983 → 1998 –  
1996 → 2011  1996 → 2011  1996 → 2011  
Target years  1998  2011  1998  2011  1998  2011  
MAPE_{x,y}  9.0 (14 tests)  12.4 (14 tests)  14.9 (14 tests)  12.1 (42 tests)  
MAPE_{y}  5.3 (42 tests)  7.5 (42 tests)  9.4 (42 tests) 
Analyses
Analyses were made using STATA 12 software. The first and second hypotheses were tested using the KruskalWallis equalityofpopulations rank test and the twosample Wilcoxon ranksum (Mann–Whitney) test. The first test was used for testing the difference between the percentage errors of three different horizon lengths (first hypothesis) and three different base period lengths (second hypothesis). The second test was used to test which of the three horizon and three base period lengths differ significantly. The third hypothesis was tested using the twosample Wilcoxon ranksum (Mann–Whitney) test. To test this hypothesis, the difference between, on the one hand, percentage errors of projections with similar horizon and base period lengths and, on the other hand, the percentage errors of projections with different horizon and base period lengths were tested.
Results
The MAPEs that resulted from the analyses range from 1.9% to 14.9%. This means that, on an average of 8 801 GPs in the period 1998 to 2011, the projection error equals 167 to 1311 GPs. These numbers show a large range and are equal to one third to 2.5 times the size of the yearly inflow in GP training (almost 500 persons on average started the training every year between 1998 and 2011).
Accuracy by length of projection horizon
To test the first hypothesis (the longer the projections, the lower the accuracy) the MAPE_{X} of projections with 5, 10 and 15year horizons are compared and the differences were tested. Table 1 shows that projections with a 5year horizon have a higher accuracy than projections with a 10year horizon, which subsequently have a higher accuracy than projections with a 15year horizon.
The differences were significant according to the KruskallWallis test (χ^{2} = 75.669; P = 0.0001). According to the twosample tests (Wilcoxon), projections with 5 and 10year horizons are different (z = 5.122; P = 0.0000), as well as projections with 10 and 15year horizons (z = 5.896; P = 0.0000) and 5 and 15year horizons (z = 7.497; P = 0.0000). Consequently, hypothesis 1 is confirmed. This is in accordance with earlier research [11, 22, 28–30]. In the present study, projections with a 5year horizon are twice as accurate as projections with a 10year horizon and four times as accurate as projections with a 15year horizon.
Accuracy by base period length
The second hypothesis (the longer the base period, the higher the accuracy) is tested by comparing the MAPE_{Y} of projections based on 5, 10 and 15year base periods (0 to 5 years, 0 to 10 years and 0 to 15 years before the launch year). Table 1 shows that projections with a 5year base period have a higher accuracy than projections with a 10year base period, which subsequently have a higher accuracy than projections with a 15year base period. The differences were significant according to the KruskallWallis test (χ^{2} = 15.826; P = 0.0004). According to the twosample tests (Wilcoxon), projections with 5 and 10year base periods are different (z = 2.246; P = 0.0247), as well as projections with 5 and 15year base periods (z = 3.865; P = 0.0001). Projections with 10 and 15year base periods are not different (z = 1.923; P = 0.0544). In conclusion, hypothesis 2 is not confirmed, because projections with a shorter base period are not less accurate.
Accuracy by similarity of projection horizon length and base period length
The MAPE_{X,Y} of projections with three horizons based on 5, 10 and 15year base periods are compared to test the third hypothesis (the accuracy will be highest when the lengths of the base period and the projection horizon are similar). Table 1 shows that for every projection length, the projections with a 5year base period have a higher accuracy than projections based on a 10year period, which subsequently have a higher accuracy than projections based on a 15year period. The differences between the errors of two groups were tested: projections with similar horizon and base period lengths and projections with different horizon and base period lengths. According to the twosample tests (Wilcoxon), the errors of the two groups are not different (z = 0.391; P = 0.6960). Consequently, the accuracy is not highest when projection horizon length and base period length are similar and the third hypothesis is not confirmed.
Discussion
The goal of this article was to evaluate the accuracy of the techniques of Dutch GP workforce projections by backtesting projections and comparing the a posteriori projections with the observed number of GPs in 1998 to 2011. Another goal was to test three hypotheses about the accuracy of different projection horizon and base period lengths.
According to the results of the present study, the projections with a short projection horizon and a short base period are more accurate than projections with a longer horizon and base period.
The Dutch health workforce projections usually have projection horizons of 10 and 15 years. According to the results, projections with a 5year horizon are however the most accurate. This is in accordance with the results of studies regarding the accuracy of population projections [11, 22, 28–30]. Large errors in supply projections could cause an imbalance between supply and demand, and as a result major adjustments in training inflow would be needed. To minimize the errors in projections with a longer horizon, it is recommendable to monitor the workforce continuously and to execute projections frequently. In practice, it is not feasible to execute projections with a shorter projection horizon, because there would only be a short period to match supply and demand. Dramatic fluctuations in yearly training inflow would be needed to reach a balance between supply and demand. It is undesirable to adjust the inflow number in training by large numbers each year, because this would be practically impossible for training institutions, for example.
Dutch GP workforce projections that are carried out to advise the government are based on a 15year base period. According to the results of the present study, projections with a 5year base period are more accurate than those with a 10 or 15year period. Consequently, a base period of 5 or 10 years also seems extensive enough to make reliable projections. It seems that the GP workforce of today is different from the past GP workforce and, therefore, we can conclude that base periods containing not only recent data but also older data are less representative for GPs in the target year. According to projections based on base periods including older data, we expected the GPs to leave the workforce at an earlier age than was observed. Current GPs stay in the workforce longer.
We can thus carefully conclude that health workforce projections can be made with data based on relatively short periods and less data, although detailed data are required to monitor and evaluate the health workforce [37].
The accuracy of the projections varies per year and there seems to be a trend towards more accurate projections in more recent years for all base period lengths. Hence, forecasting the size of the future workforce did not become more difficult between 1998 and 2011, as we originally expected [13–15]. This trend could be explained by two things. First, it seems that the GP workforce of 1980 to 1990 is less similar to the workforce of 1990 to 2000 than the workforce of 1990 to 2000 is similar compared to the workforce of 2000 to 2010. In other words, the GP workforce changed more extensively between 1980 and 1990 and 1990 and 2000 than between 1990 and 2000 and 2000 and 2010. Second, the Dutch GP workforce has become larger. Data based on a larger base population size have more stable averages than data based on smaller populations.
The errors of the Dutch GP workforce projections range from 1.9% to 14.9%. This is a large range, which illustrates the importance of doing projections with different projection horizon lengths and base period lengths.
The projection errors are mainly caused by bias and not by variance [38]. The variance is low, because data of all Dutch GPs is used to make projections. The projection error is mostly bias, caused by differences between the past GP workforce and the current and future GP workforce.
From a data availability perspective, it may be possible that there is significant scope for more countries to engage in modelbased health workforce planning than is currently the case, and for countries already engaging in such planning to extend the reach of their current models, which was also concluded from the Matrix Insight report [10].
However, the successful application of a model similar to the Dutch workforce projection model is dependent on the health workforce planning system of a country. The output of the Dutch projection model is the required inflow in specialized training per year to balance the supply and demand for health professionals in the future [17–20]. Hence, the height of inflow in specialized training is the ‘adjustment component’ of the Dutch health workforce. In other health workforce planning systems, other parts of the planning system are possibly used as the ‘adjustment component’, such as postponing retirement or increasing the return on training (Figure 1). In Belgium, for example, the inflow in initial medical training (not specialized training) is the ‘adjustment component’ [39]. Future research is needed to investigate which type of health workforce planning fits with which type of healthcare system [40].
Limitations
This study has several limitations. First, in the present study, we backtested the current GP workforce projection methods a posteriori. There are other methods to analyse the accuracy of workforce projections, which we did not use. For example, we did not evaluate the current projections by comparing the results of GP workforce projections that were done in the past, with the actual observations. This second method seems simpler, but with this method we would not evaluate the current model, but older versions of it. The only way we can evaluate the current model, is by using old data to generate new projections. This is because future numbers are not yet known.
Second, the present study was limited to testing one healthcare profession in the Netherlands: general practitioners. In practice, the model is used for all types of medical and allied health professionals, as the model is designed as ‘one size fits all’. This implies that the backtesting of projections is possible for all types of health professionals. However, for most of them there is less data available and therefore it is more difficult to backtest.
Third, the accuracy of the demand side of the Dutch health workforce simulation model was not tested, because of a lack of data. However, this should be a topic of future research on the accuracy of the Dutch health workforce planning system [18].
Fourth, testing the accuracy of workforce projections can be done disaggregated by several factors, such as gender, region, cohort or type of GP. Although this would have been an interesting exercise, we limited this study to the total supply of GPs. It would be an interesting case for future research.
Conclusions
According to the results of the present study, forecasting the size of the future workforce did not become more difficult between 1998 and 2011, as we originally expected. Furthermore, the projections with a short projection horizon and a short base period are more accurate than projections with a longer projection horizon and base period. We can carefully conclude that health workforce projections can be made with data based on relatively short base periods, although detailed data are still required to monitor and evaluate the health workforce.
Abbreviations
 GP:

General practitioner
 MAPE:

Mean absolute percentage error.
Declarations
Acknowledgements
Initial analyses of the accuracy of our general practitioners workforce projections and previous work on the projections themselves were funded by the Dutch Advisory Committee on Medical Manpower Planning (Capaciteitsorgaan).
Authors’ Affiliations
References
 Askildsen JE, Baltagi BH, Holmas TH: Wage policy in the health care sector: a panel data analysis of nurses' labour supply. Health Econ. 2003, 12: 705719. 10.1002/hec.836.View ArticlePubMedGoogle Scholar
 Commission of the European Communities: Green paper on European Workforce for health. 2008, Brussels: Commission of the European Communities COM, 725Google Scholar
 Correia I, Veiga P: Geographic distribution of physicians in Portugal. The Journal of Health Economics. 2010, 11: 383393. 10.1007/s1019800902088.View ArticleGoogle Scholar
 OECD: The looming crisis in the health workforce. 2008, Paris: OECD Health Policy Studies, OECD PublishingView ArticleGoogle Scholar
 Simoens S, Villeneuve M, Hurst J: Tackling nurse shortages in OECD countries. 2005, Paris: OECDView ArticleGoogle Scholar
 Bloom BS: Health manpower planning. Health Serv Res. 1980, 15: 6768.PubMed CentralGoogle Scholar
 Dreesch N: An approach to estimating human resource requirements to achieve the Millennium Development Goals. Health Policy Plan. 2005, 20: 267276. 10.1093/heapol/czi036.View ArticlePubMedGoogle Scholar
 Maynard A, Walker A: The physician workforce in the United Kingdom: issues, prospects and policies. 1997, London: Nuffield TrustGoogle Scholar
 Yett DE, Drabek L, Intriligator MD, Kimbell LJ: Health manpower planning. Health Serv Res. 1972, 7: 134147.PubMedPubMed CentralGoogle Scholar
 Matrix Insight : EU level collaboration on forecasting health workforce needs, workforce planning and health workforce trends. A feasibility study. 2012, European Commission: BrusselsGoogle Scholar
 O'BrienPallas L, Birch S, Baumann A, Tomblin Murphy G: Integrating workforce planning, human resources, and service planning. Human Resources for Health Development Journal. 2001, 5: 216.Google Scholar
 Wismar M: Health professional mobility and health systems: evidence from 17 European countries. 2011, Copenhagen: WHOGoogle Scholar
 Schäfer W: The Netherlands: health systems review. 2010, Copenhagen: World Health OrganizationGoogle Scholar
 Van den Berg M: Workload in general practice [PhD thesis]. 2010, Amsterdam: GVO drukkers & vormgevers B.V.  Ponsen & LooijenGoogle Scholar
 Dussault G, Franceschini MC: Not enough there, too many here: understanding geographical imbalances in the distribution of the health workforce. Human Resources for Health. 2006, 4: 1210.1186/14784491412.View ArticlePubMedPubMed CentralGoogle Scholar
 Olsen KR: Organisational determinants of production and efficiency in general practice: a populationbased study. Eur J Health Econ. 2013, 14 (2): 26776. 10.1007/s1019801103681.View ArticlePubMedGoogle Scholar
 Advisory Committee on Medical Manpower Planning: The 2010 recommendations for medical specialist training in medical, dental, clinical technological and related educational as well as further training areas. 2011, Utrecht: Advisory Committee on Medical Manpower PlanningGoogle Scholar
 Smits M, Slenter V, Geurts J: Improving manpower planning in health care. Proceedings of the 23rd Bled eConference ‘eTrust: Implications for the individual, enterprises and society’. Edited by: Pucihar A. 2010, Bled: Slovenia, 144154.Google Scholar
 Van der Velden LFJ, Hingstman L: The supply of general practitioners in the Netherlands. Morbidity, performance and quality in primary care: Dutch general practice on stage. Edited by: Westert GP, Jabaaij L, Schellevis FG. 2006, Oxford: Radcliffe Publishing, 257264.Google Scholar
 Van Greuningen M, Batenburg RS, Van der Velden LFJ: Ten years of health workforce planning in the Netherlands: a tentative evaluation of GP planning as an example. Human Resources for Health. 2012, 10: 2110.1186/147844911021.View ArticlePubMedPubMed CentralGoogle Scholar
 De Beer J: Transparency in population forecasting. Methods for fitting and projecting fertility, mortality and migration, [PhD thesis]. 2011, Amsterdam: Amsterdam University PressGoogle Scholar
 Rayer S: Population forecast errors. A primer for planners. Journal of Planning Education and Research. 2008, 27: 417430. 10.1177/0739456X07313925.View ArticleGoogle Scholar
 Smith S, Sincich T: Evaluating the forecast accuracy and bias of alternative projections for states. International Journal of Forecasting. 1992, 8: 495508. 10.1016/01692070(92)90060M.View ArticlePubMedGoogle Scholar
 NIVEL databases of health professionals. (http://www.nivel.nl/beroepenindegezondheidszorg). Date of access: September 25^{th} 2012
 van Dijk CE: Changing the GP payment system: do financial incentives matter? [PhD thesis]. 2012, Utrecht: LABOR Grafimedia BVGoogle Scholar
 Dussault G, Buchan J, Sermeus W, Padaiga Z: Investing in Europe's health workforce of tomorrow: scope for innovation and collaboration. Assessing future health workforce needs. 2010, Geneva: WHOGoogle Scholar
 Sermeus W, Bruyneel L: Investing in Europe's health workforce of tomorrow: scope for innovation and collaboration. Summary report of the three Policy Dialogues. 2010, Leuven: Catholic University LeuvenGoogle Scholar
 Shapiro JR: Modeling the supply chain. 2001, Pacific Grove, CA, USA: DuxburyGoogle Scholar
 Evaluating population projections. The importance of accurate forecasting. 2007, New York: Esri
 Ahlburg DA, Lutz W: Introduction: the need to rethink approaches to population forecasts. Frontiers of population forecasting. A supplement to Vol. 24. Edited by: Lutz W, Vaupel JW, Ahlburg DA. 1998, New York: The Population Council, 114.Google Scholar
 Smith S, Sincich T: On the relationship between length of base period and population forecast errors. J Am Stat Assoc. 1990, 85: 367375. 10.1080/01621459.1990.10476209.View ArticlePubMedGoogle Scholar
 Alho J, Spencer BD: The practical specification of the expected error in population forecasts. Journal of Official Statistics. 1997, 13: 203225.Google Scholar
 Dowd K: Backtesting stochastic mortality models: an expost evaluation of multiperiodahead density forecasts. North American Actuarial Journal. 2010, 14: 281298. 10.1080/10920277.2010.10597592.View ArticleGoogle Scholar
 Dowd K: Evaluating the goodness of fit of stochastic mortality models. Insurance Mathematics and Economics. 2010, 47: 255265. 10.1016/j.insmatheco.2010.06.006.View ArticleGoogle Scholar
 Ahlburg D: A commentary on error measures: error measures and choice of a forecast method. International Journal of Forecasting. 1992, 8: 99111. 10.1016/01692070(92)900107.View ArticleGoogle Scholar
 Keilman N: How accurate are the United Nations world population projections?. Population and Development Review. 1999, 24: 1541.View ArticleGoogle Scholar
 Diallo K, Zurn P, Gupta N, Dal Poz M: Monitoring and evaluation of human resources for health: an international perspective. Human Resources for Health. 2003, 1: 310.1186/1478449113.View ArticlePubMedPubMed CentralGoogle Scholar
 Hastie T, Tibshirani R, Friedman J: The elements of statistical learning. Data mining, inference, and prediction, Springer Series in Statistics. 2008, New York: Springer,Google Scholar
 Artoisenet C, Deliege D: Medical workforce in Belgium: assessment of future supply and requirements. Louvain Med. 2006, 125: 421.Google Scholar
 Kuhlmann E, Batenburg R, Groenewegen PP, Larsen C: Bringing a European perspective to the health human resources debate: a scoping study. Health Policy. 2013, 110: 613. 10.1016/j.healthpol.2012.11.002.View ArticlePubMedGoogle Scholar
Copyright
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.