Predicting hemoglobin levels in whole blood donors using transition models and mixed effects models
- Kazem Nasserinejad^{1}Email author,
- Wim de Kort^{2},
- Mireille Baart^{2},
- Arnošt Komárek^{3},
- Joost van Rosmalen^{1} and
- Emmanuel Lesaffre^{1, 4}
https://doi.org/10.1186/1471-2288-13-62
© Nasserinejad et al.; licensee BioMed Central Ltd. 2013
Received: 5 December 2012
Accepted: 17 April 2013
Published: 2 May 2013
Abstract
Background
To optimize the planning of blood donations but also to continue motivating the volunteers it is important to streamline the practical organization of the timing of donations. While donors are asked to return for donation after a suitable period, still a relevant proportion of blood donors is deferred from donation each year due to a too low hemoglobin level. Rejection of donation may demotivate the candidate donor and implies an inefficient planning of the donation process. Hence, it is important to predict the future hemoglobin level to improve the planning of donors’ visits to the blood bank.
Methods
The development of the hemoglobin prediction rule is based on longitudinal (panel) data from blood donations collected by Sanquin (the only blood product collecting and supplying organization in the Netherlands). We explored and contrasted two popular statistical models, i.e. the transition (autoregressive) model and the mixed effects model as plausible models to account for the dependence among subsequent hemoglobin levels within a donor.
Results
The predictors of the future hemoglobin level are age, season, hemoglobin levels at the previous visits, and a binary variable indicating whether a donation was made at the previous visit. Based on cross-validation, the areas under the receiver operating characteristic curve (AUCs) for male donors are 0.83 and 0.81 for the transition model and the mixed effects model, respectively; for female donors we obtained AUC values of 0.73 and 0.72 for the transition model and the mixed effects model, respectively.
Conclusion
We showed that the transition models and the mixed effects models provide a much better prediction compared to a multiple linear regression model. In general, the transition model provides a somewhat better prediction than the mixed effects model, especially at high visit numbers. In addition, the transition model offers a better trade-off between sensitivity and specificity when varying the cut-off values for eligibility in predicted values. Hence transition models make the prediction of hemoglobin level more precise and may lead to less deferral from donation in the future.
Keywords
Blood donations Hemoglobin level Longitudinal data Panel data Transition models Mixed effects models Prediction Kalman filterBackground
Blood transfusion is an essential part of modern healthcare which helps save millions of lives each year. Since blood is a unique resource for which an artificial substitute has yet to be found, blood donations are in great need. However, occasionally donation cannot be accepted. There may be several reasons for the ineligibility of a blood donor for donation, a common reason being low hemoglobin level of the donor [1, 2]. A hemoglobin (Hb) level of 8.4 mmol/l (135 g/l) and 7.8 mmol/l (125 g/l) for men and women, respectively, is widely accepted as the lower cut-off value of eligibility for donation [2–5]. While donors are asked to return for donation after a suitable period, a relevant proportion of blood donors are temporarily deferred from donation each year due to low Hb levels [2]. Rejection of donation may demotivate the candidate donor and implies inefficient planning of the donation process [6, 7]. Hence, it is important to predict the future Hb level to improve the planning of donors’ visits to the blood bank. Prediction models for low Hb level deferral have been developed previously [5, 8].
The main goal of this paper is to illustrate the use of two well-known longitudinal models in predicting the future Hb level after a visit to the blood bank. An adequate prediction will help the blood bank to apply appropriate interventions (e.g. postponing the next invitation) for blood donation when the Hb value falls below the cut-off value. Prediction is based on models developed using historical data of Hb levels obtained from Sanquin Blood Supply in the Netherlands. More specifically, in this paper we examine the predictive performance of the transition (autoregressive panel data) model and the mixed effects model.
Methods
Data
The data have been obtained from Sanquin Blood Supply, which is the only blood product collecting and supplying organization in the Netherlands. In this paper, we analyze newly registered whole blood donors whose first visit to the collection centers occurred in the period between January 1, 2007 and December 31, 2009 and have donated at least twice during this period. Whole blood is a term used in transfusion medicine for a standard blood donation as opposed to plasma and platelet donation. The data were collected from 16,158 newly registered whole blood donors (54.6% women). The reason for selecting this set of blood donors is that they constitute a relatively homogeneous group that did not donate prior to establishing the Sanquin database. We excluded donors who had missing values for the Hb level, and the data of the remaining 15,625 donors were used in the analyses.
Several factors are known to be associated with the Hb level and hence may be used as predictors for Hb level, i.e. gender [9], age [9], and body mass index (BMI) [10, 11]. In this study, we take into account the effect of gender and age in our models, but we decided to ignore the effect of BMI due to the fact that the BMI was not recorded for approximately 40% of donors. Also, based on a pilot study we found that the impact of BMI on Hb level is secondary. The season in which the visit takes place also affects the Hb level, namely in a warm season Hb level is lower on average [12, 13]. Here season is used as a binary covariate, i.e. cold season (= 0) includes fall and winter and warm season (= 1) includes spring and summer. Male and female donors have different Hb profiles, therefore we analyzed the data for men and women separately. Inter-visit intervals differ between donors, in our data set the median inter-visit interval for male donors is 72 (inter-quartile range: 29 − 92) days and for female donors it is 93 (inter-quartile range: 25 − 131) days. In principle, varying intervals between visits require continuous-time models, but these models are beyond the scope of this paper. Therefore we decided to ignore this feature of the data, and we used the sequential number of the visit rather than the actual time of the visit. We also take into account the status of the previous visit (donation or deferral) as a binary covariate in the prediction model. Since no donations have been made prior to the first visit, the value of donation at previous visit (DPV) for the first visit is defined to be ‘no donation’.
This research has been performed with the approval of the ethical advisory council of the Sanquin Blood Supply Foundation. Moreover, all donors have given their consent by stating that part or all of their donations can be used for research aiming at improving the blood supply chain. Our ethical advisory council includes members of both Sanquin and non-Sanquin affiliations. This committee includes members with the background training and experience required for such ethical committees.
Statistical analysis
Since successive Hb levels on the same subject are correlated, we need to employ statistical models that can take this correlation into account. For this purpose, we applied two well-known models, namely the transition model and the mixed effects model. However, we commence with a multiple linear regression model as a benchmark to show the capability of transition and mixed effects models. These statistical analyses were performed in R version 2.15.2 [14] using the stats package for the multiple linear regression models, the nlme package for the mixed effects models, the KalmanLike and the mle functions in the stats4 package for the transition models, and the mixAK and pROC packages to draw profile and ROC curve plots. We used a significance level of α = 0.05 and no correction for multiple testing was implemented.
Multiple linear regression model
where y _{ i t } is the tth observation of the ith individual, α is an unknown constant (intercept), and the β’s are unknown regression coefficients. It is assumed that the residuals ε _{ i t } are normally distributed and mutually independent with mean zero and constant variance, i.e., ${\epsilon}_{\mathit{\text{it}}}\stackrel{\text{iid}}{\sim}N(0,{\sigma}_{\u03f5}^{2})$. Due to the fact that this model cannot take into account the intra-subject correlations and the previous Hb levels, it is only presented as a benchmark model to show the capability of transition and mixed effects models.
Transition model
where y _{ i t } is the tth observation of the ith individual, α is an unknown constant, and the β’s are unknown regression coefficients, y _{ i t − r } and (A g e _{ i t − r },S e a s o n _{ i t − r },D P V _{ i t − r }) are rth lagged response and covariates, respectively and γ _{ r } is the corresponding coefficient of the rth lag. Classically it is assumed that the residuals ε _{ i t } are normally distributed and mutually independent with mean zero and constant variance, i.e., ${\epsilon}_{\mathit{\text{it}}}\stackrel{\text{iid}}{\sim}N(0,{\sigma}_{\u03f5}^{2})$. In a transition model with order q, the predicted values depend on q lagged previous observations; however, to calculate the predicted value using equation 2, there are not enough previous observations for the first few visits of a donor. We employed the method of maximum likelihood via a linear quadratic estimation (Kalman filter) algorithm to estimate the parameters in the transition model. This algorithm enables us to calculate the exact likelihood function, which includes the distribution of the first few observations of each donor [16–18]. As a result, the maximum likelihood estimation also includes the information of donors who have made fewer visits than the order of the transition model.
Linear mixed effects model
where α is an unknown constant, the β’s are regression coefficients (fixed effects) and the b _{ i0} is the random intercept. The random intercept b _{ i0} can be viewed here as the deviation of the ith subject-specific mean of Hb levels from the population mean of Hb levels. It is assumed that b _{0i } and ε _{ i t } are normally distributed and mutually independent with mean zero and different constant variances, i.e., ${b}_{0i}\sim N(0,{\sigma}_{b0}^{2})$, and ${\epsilon}_{\mathit{\text{it}}}\sim N(0,{\sigma}_{\u03f5}^{2})$[21]. Furthermore, in the random intercept model the correlation between two observations of a subject is constant and is equal to the intra-class correlation given by $\rho =\frac{{\sigma}_{b0}^{2}}{{\sigma}_{b0}^{2}+{\sigma}_{\u03f5}^{2}}$[19, 20].
where α is an unknown constant, the β’s contains population-specific parameters. b _{ i } = (b _{ i0},b _{ i1}) contains subject-specific parameters (intercept and the effects of age) describing how the evolution of the ith individual deviates from the average evolution in the population, and where the residual component ${\epsilon}_{\mathbf{i}}={({\epsilon}_{i1},\dots ,{\epsilon}_{i{n}_{i}})}^{\prime}$ is a vector containing the common error components, with ε _{ i } ∼ N (0,Σ _{ i }). In this paper, we assumed that ${\Sigma}_{i}={\sigma}^{2}{I}_{{n}_{i}}$, so that, conditional on the values of the random effects, a person’s measurements of the Hb level are independent. However, additional correlation among the errors can be accommodated by allowing for a more general covariance structure (e.g., autoregressive) in the model. It is assumed b _{ i } has a bivariate normal distribution with mean zero and a diagonal covariance matrix, so that ε _{ i } and b _{ i } are mutually independent. To estimate the parameters in the mixed effects models we employed the method of restricted maximum likelihood (REML). We applied an empirical Bayes method (EB) to predict a person’s random intercept and slope based on his/her all previous observations [20].
We used a likelihood ratio test to choose between the mixed model with random intercept and the mixed model with random intercept and slope. In this case, the likelihood ratio test statistic for testing a random slope in the model is a mixture of chi-squared distributions with 1 and 2 degrees of freedom [19].
Note that the linear mixed effects model is based on quite different assumptions than the transition model. In principle, if one model is correct, the other model must be wrong. However, in practice we never know the truth and in fact it is possible that both models are wrong. Despite this, we can still check which of the two models performs better in predicting the Hb level.
Prediction performance
To avoid a too optimistic assessment of the model predictions by using the data twice, i.e. for model building and parameter estimation as well as model evaluation, we have randomly divided the data set (n=15,625 donors) into two parts: a training data set consisting of all observations of 7,709 donors and a validation data set consisting of all observations of the remaining 7,916 donors [22]. The models are estimated using the training data set, and the model predictions are evaluated using the validation data set. We used a dynamic prediction approach in the sense that to predict Hb level at a visit we used the observations of all previous visits, therefore for each visit we updated our prior information. Since no prior information is available for the first visit, the predicted values are based only on the gender and age of the donor and the season in which the visit takes place.
where ${\u0177}_{\mathit{\text{it}}}$ and y _{ i t } are the predicted and observed values, respectively and N _{ t } is the total number of subjects at occasion t. MSPE _{ t } is a well-known measure to evaluate prediction. The MSPE values are calculated for the validation data set only.
We also computed the sensitivity and specificity of the predicted values for assessing the eligibility for donation in the validation data set. Specifically, we computed the proportion of individuals that are correctly predicted to be eligible for donation based on the clinical cut-off value (i.e. an Hb level of at least 8.4 mmol/l and 7.8 mmol/l for men and women, respectively). However, one may also optimize the cut-off value for the predicted values to obtain a receiver-operating characteristic (ROC) curve. In this ROC curve, the state variable is a dichotomous variable indicating whether the Hb level is below the clinical cut-off value of 8.4 mmol/l for men or 7.8 mmol/l for women; the test variable is the predicted value ${\u0177}_{\mathit{\text{it}}}$. Varying the cut-off value for the predicted value will change the sensitivity to detect that a donor will be eligible; however the assessment of donors’ eligibility is based on the clinical cut-off value, which is not changed in the ROC analysis. We calculated the area under the curve (AUC) to compare the models. The difference in the AUCs between the models was tested using a bootstrap technique [25, 26] that takes into account the correlation between the areas that is induced by the paired nature of the data.
Results
Descriptive statistics of the training and validation data sets
Data set | Gender | #Donor | #Deferral | #Cold Season | Age: Mean (SD) | Visit: Med (IQR) |
---|---|---|---|---|---|---|
Training data set | Male | 3610 | 769 (4.58%) | 10213 (50.05%) | 34.57 (12.9) | 5 (3) |
Female | 4306 | 1596 (9.62%) | 10387 (49.71%) | 32.66 (12.8) | 5 (1) | |
Total | 7916 | 2365 (7.08)% | 20600 (49.88%) | 33.53 (12.9) | 5 (2) | |
Validation data set | Male | 3449 | 688 (4.27%) | 9781 (49.95%) | 34.28 (12.6) | 5 (3) |
Female | 4260 | 1729 (10.41%) | 10341 (49.54%) | 32.77 (12.8) | 5 (2) | |
Total | 7709 | 2417 (7.38%) | 20122 (49.74%) | 33.45 (12.7) | 5 (2) |
Parameter estimates (standard errors) of the models estimated using the training data set for male donors
Parameter | Model LR | AR(1) | AR(2) | AR(3) | AR(4) | AR(5) | Model LME |
---|---|---|---|---|---|---|---|
intercept | 9.6448 | 9.6309 | 9.6441 | 9.6560 | 9.6617 | 9.6633 | 9.6719 |
(0.0142) | (0.0206) | (0.0231) | (0.0243) | (0.0246) | (0.0247) | (0.0243) | |
Age | -0.0045 | -0.0043 | -0.0044 | -0.0045 | -0.0047 | -0.0047 | -0.0049 |
(0.0003) | (0.0005) | (0.0006) | (0.0006) | (0.0006) | (0.0007) | (0.0006) | |
Season(Warm) | -0.0627 | -0.0615 | -0.0681 | -0.0699 | -0.0693 | -0.0694 | -0.0698 |
(0.0089) | (0.0074) | (0.0066) | (0.0066) | (0.0067) | (0.0067) | (0.0067) | |
DPV | -0.0610 | -0.0469 | -0.0350 | -0.0385 | -0.0440 | -0.0474 | -0.0636 |
(Donation) | (0.0092) | (0.0089) | (0.0079) | (0.0074) | (0.0072) | (0.0072) | (0.0068) |
γ _{1} | — | 0.5158 | 0.3685 | 0.3053 | 0.2746 | 0.2630 | — |
— | (0.0061) | (0.0068) | (0.0076) | (0.0082) | (0.0087) | — | |
γ _{2} | — | — | 0.2888 | 0.2080 | 0.1766 | 0.1621 | — |
— | — | (0.0078) | (0.0087) | (0.0084) | (0.0091) | — | |
γ _{3} | — | — | — | 0.2207 | 0.1730 | 0.1581 | — |
— | — | — | (0.0095) | (0.0104) | (0.0109) | — | |
γ _{4} | — | — | — | — | 0.1488 | 0.1257 | — |
— | — | — | — | (0.0123) | (0.0129) | — | |
γ _{5} | — | — | — | — | — | 0.0829 | — |
— | — | — | — | — | (0.0167) | — |
Parameter estimates (standard errors) of the models estimated using the training data set for female donors
Parameter | Model LR | AR(1) | AR(2) | AR(3) | AR(4) | AR(5) | Model LME |
---|---|---|---|---|---|---|---|
intercept | 8.2737 | 8.2394 | 8.2555 | 8.2678 | 8.2698 | 8.2702 | 8.2832 |
(0.0123) | (0.0164) | (0.0180) | (0.0186) | (0.0187) | (0.0187) | (0.0181) | |
Age | 0.0042 | 0.0044 | 0.0042 | 0.0040 | 0.0040 | 0.0040 | 0.0037 |
(0.0003) | (0.0004) | (0.0005) | (0.0005) | (0.0005) | (0.0005) | (0.0005) | |
Season(Warm) | -0.0347 | -0.0405 | -0.0415 | -0.0413 | -0.0415 | -0.0415 | -0.0411 |
(0.0078) | (0.0062) | (0.0060) | (0.0062) | (0.0061) | (0.0061) | (0.0062) | |
DPV | -0.1106 | -0.1411 | -0.1273 | -0.1307 | -0.1335 | -0.1346 | -0.1387 |
(Donation) | (0.0079) | (0.0075) | (0.0067) | (0.0064) | (0.0063) | (0.0063) | (0.0060) |
γ _{1} | — | 0.4669 | 0.3457 | 0.3012 | 0.2878 | 0.2830 | — |
— | (0.0062) | (0.0067) | (0.0074) | (0.0080) | (0.0084) | — | |
γ _{2} | — | — | 0.2573 | 0.1963 | 0.1793 | 0.1693 | — |
— | — | (0.0080) | (0.0088) | (0.0089) | (0.0099) | — | |
γ _{3} | — | — | — | 0.1742 | 0.1486 | 0.1360 | — |
— | — | — | (0.0100) | (0.0112) | (0.0121) | — | |
γ _{4} | — | — | — | — | 0.0831 | 0.0623 | — |
— | — | — | — | (0.0157) | (0.0182) | — | |
γ _{5} | — | — | — | — | — | 0.0681 | — |
— | — | — | — | — | (0.0264) | — |
AIC, BIC, and MSEP values for different models for both genders based on the training data set
Male donors | Female donors | |||||
---|---|---|---|---|---|---|
Model | AIC | BIC | MSPE | AIC | BIC | MSPE |
Linear Regression | 37087.8 | 37127.2 | 4.14 | 35968.9 | 36008.6 | 2.29 |
Mixed Effects | 30524.3 | 30571.6 | 2.90 | 30058.0 | 30113.6 | 1.75 |
AR(1) | 32051.0 | 32098.3 | 3.07 | 31559.1 | 31606.7 | 1.81 |
AR(2) | 30936.4 | 30991.6 | 2.85 | 30664.7 | 30720.3 | 1.73 |
AR(3) | 30471.9 | 30535.0 | 2.78 | 30375.1 | 30438.7 | 1.71 |
AR(4) | 30342.5 | 30413.4 | 2.78 | 30341.7 | 30413.2 | 1.72 |
AR(5) | 30321.4 | 30400.2 | 2.79 | 30325.1 | 30404.5 | 1.72 |
Discussion
In this article, we presented transition models with different numbers of autoregressive terms and mixed effects models (a mixed effects model with random intercept for male donors and a mixed effects model with random intercept and random slope based on age for female donors), as plausible models to account for the dependence among subsequent Hb levels within a donor and as models to predict the future hemoglobin level. Based on the results for the validation data set, we showed that the transition model and the mixed effects model have almost the same predictive accuracy at the first few visits of a donor; however, for longer time series the transition model offers somewhat better predictions. To give an idea of the predictive performance, we have computed the ROC curve. Our results confirm that the transition model shows a small but significant improvement in the AUC compared to the mixed effects model.
Both the transition and the mixed effects models use the data of a person’s previous observations for making predictions. In the transition model only the last q observations are used for prediction the current response. However, in the mixed effects model, the empirical Bayes method for estimating a persons random effects uses all previous observations. Therefore, the mixed effects model requires more historical information than the transition model. Since the transition model is convenient in practice and needs less historical information compared to the mixed effects model, blood banks may use this model to predict the future hemoglobin level of a candidate and to determine which candidates should not be invited for the next donation.
Our approach of using transition or autoregressive models is quite novel in biomedical research, however in other fields such as econometrics, autoregressive modeling is a very well-known technique for tackling correlated financial phenomena and time series problems [27].
We do not claim that our final model is optimal; further research is needed to arrive at a better prediction model. First, the data set used in this paper is unbalanced in the sense that the time intervals between visits vary considerably, though this was not taken into account here. Second, there are more factors that are possibly associated with Hb level than those which we have investigated in this study, such as physical activity [28], race [29], nutrition [30] and smoking status [11, 31]. Finally, the ultimate purpose of the prediction exercise is not the prediction of the future Hb value, but rather to determine the best time for the donor to return for donation. Hence, prediction models for Hb levels after blood donation should focus on the optimal timing of future donations, instead of on predicting future Hb levels. We are currently investigating such models.
Conclusion
In this study we compared transition models and mixed effects models for predicting the Hb level in whole blood donors. The results showed that the transition model provides a somewhat better prediction than the mixed effects model, especially at high visit numbers. We believe that our paper shows the capabilities of using longitudinal models for prediction and that our findings may help reduce the number of deferred candidate in the blood banks.
Declarations
Acknowledgements
This study was made possible by a grant provided by Sanquin Blood Supply Foundation in the Netherlands. The authors would like to acknowledge Sanquin for providing the grant for this study. Further, thanks goes to the referees for their valuable comments.
Authors’ Affiliations
References
- Gómez-Simón A, Navarro-Núñez L, Pérez-Ceballos E, Lozano ML, Candela MJ, Cascales A, Martínez C, Corral J, Vicente V, Rivera J: Evaluation of four rapid methods for hemoglobin screening of whole blood donors in mobile collection settings. Transfus Apher Sci. 2007, 36 (3): 235-242. 10.1016/j.transci.2007.01.010.View ArticlePubMedGoogle Scholar
- Guide to the preparation, use and quality assurance of blood components, European Committee (partial agreement) on Blood Transfusion (CD-P-TS): Recommendation No. R(95) 15, 15th Edition. Strasbourg: Council of Europe; 2010Google Scholar
- Tong E, Murphy WG, Kinsella A, Darragh E, Woods J, Murphy C, McSweeney E: Capillary and venous haemoglobin levels in blood donors: a 42-month study of 36 258 paired samples. Vox Sanguinis. 2010, 98: 266-267.View ArticleGoogle Scholar
- Radtke H, Polat G, Kalus U, Salama A, Kiesewetter H: Hemoglobin screening in prospective blood donors: comparison of different blood samples and different quantitative methods. Transfus Apher Sci. 2005, 33 (1): 31-35. 10.1016/j.transci.2004.11.004.View ArticlePubMedGoogle Scholar
- Baart AM, de Kort WL, Moons KG, Vergouwe Y: Prediction of low haemoglobin levels in whole blood donors. Vox Sanguinis. 2011, 100: 204-211. 10.1111/j.1423-0410.2010.01382.x.View ArticlePubMedGoogle Scholar
- Halperin D, Baetens J, Newman B: The effect of short-term, temporary deferral on future blood donation. Transfusion. 1998, 38 (2): 181-183. 10.1046/j.1537-2995.1998.38298193102.x.View ArticlePubMedGoogle Scholar
- Custer B, Chinn A, Hirschler NV, Busch MP, Murphy EL: The consequences of temporary deferral on future whole blood donation. Transfusion. 2007, 47 (8): 1514-1523. 10.1111/j.1537-2995.2007.01292.x.View ArticlePubMedGoogle Scholar
- Baart AM, de Kort WL, Atsma F, Moons KG, Vergouwe Y: Development and validation of a prediction model for low hemoglobin deferral in a large cohort of whole blood donors. Transfusion. 2012, 52 (12): 2559-2569. 10.1111/j.1537-2995.2012.03655.x.View ArticlePubMedGoogle Scholar
- Yip R, Johnson C, Dallman PR: Age-related changes in laboratory values used in the diagnosis of anemia and iron deficiency. Am J Clin Nutr. 1984, 39 (3): 427-436.PubMedGoogle Scholar
- Micozzi MS, Albanes D, Stevens RG: Relation of body size and composition to clinical biochemical and hematologic indices in US men and women. Am J Clin Nutr. 1989, 50 (6): 1276-1281.PubMedGoogle Scholar
- Skjelbakken T, Dahl IM, Wilsgaard T, Langbakk B, Løchen ML: Changes in haemoglobin levels according to changes in body mass index and smoking habits, a 20-year follow-up of a male cohort: the Tromso Study 1974–1995. Eur J Epidemiol. 2006, 21 (7): 493-499. 10.1007/s10654-006-9032-y.View ArticlePubMedGoogle Scholar
- Hoekstra T, Veldhuizen I, van Noord PA, de Kort WL: Seasonal influences on hemoglobin levels and deferral rates in whole-blood and plasma donors. Transfusion. 2007, 47 (5): 895-900. 10.1111/j.1537-2995.2007.01207.x.View ArticlePubMedGoogle Scholar
- Kristal-Boneh E, Froom P, Harari G, Shapiro Y, Green MS: Seasonal changes in red blood cell parameters. Br J Haematol. 1993, 85: 603-607.View ArticlePubMedGoogle Scholar
- R Development Core Team: R. a language and environment for statistical computing. [http://www.R-project.org]
- Diggle PJ, Heagerty P, Liang KY, Zeger S: Analysis of Longitudinal Data 2nd ed. 2002, New York: Oxford University PressGoogle Scholar
- Kalman RE: A new approach to linear filtering and prediction problems. Trans ASME–JBasic Eng. 1960, 82: 35-45.Google Scholar
- Harvey AC, Phillips GDA: Maximum likelihood estimation of regression models with autoregressive-moving average disturbances. Biometrika. 1979, 66: 49-58.Google Scholar
- Tsay R: Analysis of Financial Time Series, 3rd Edition. 2010, New York: John Wiley & SonsView ArticleGoogle Scholar
- Verbeke G, Molenberghs G: Linear Mixed Models for Longitudinal Data. 2000, New York: SpringerGoogle Scholar
- McCulloch CE, Searle SR: Generalized, Linear, and Mixed Models. 2001, New York: John Wiley & SonsGoogle Scholar
- Molenberghs G, Verbeke G: A review on linear mixed models for longitudinal data, possibly subject to dropout. Stat Model. 2001, 1: 235-269. 10.1191/147108201128195.View ArticleGoogle Scholar
- Kohavi R: A study of cross-validation and bootstrap for accuracy estimation and model selection. Proceedings of the 14th International Joint Conference on Artificial Intelligence - Volume 2 (IJCAI’95), Vol. 2. 1995, San Francisco: Morgan Kaufmann Publishers Inc, 1137-1143.Google Scholar
- Akaike H: Information theory and extension of maximum likelihood principle. Second International, Symposium on Information Theory Pctrov B Csáki F (cd). 1973, Budapest: Akadćmiai KiadóGoogle Scholar
- Schwarz G: Estimating the dimension of a model. Ann Stat. 1978, 6: 461-464. 10.1214/aos/1176344136.View ArticleGoogle Scholar
- Robin X, Turck N, Hainard A, Tiberti N, Lisacek F, Sanchez J, Müller M: pROC: an open-source package for R and S+ to analyze and compare ROC curves. BMC Informatics. 2011, 12: 77-View ArticleGoogle Scholar
- Hanley JA, McNeil BJ: A method of comparing the areas under receiver operating characteristic curves derived from the same cases. Radiology. 1983, 148: 839-843.View ArticlePubMedGoogle Scholar
- Earnest A, Chen MI, Ng D, Sin LY: Using autoregressive integrated moving average (ARIMA) models to predict and monitor the number of beds occupied during a SARS outbreak in a tertiary hospital in Singapore. BMC Health Serv Res. 2005, 5 (1): 36-10.1186/1472-6963-5-36.View ArticlePubMedPubMed CentralGoogle Scholar
- Beard J, Tobin B: Iron status and exercise. Am J Clin Nutr. 2000, 72 (2 Suppl): 594S-597S.PubMedGoogle Scholar
- Johnson-Spear MA, Yip R: Hemoglobin difference between black and white women with comparable iron status: justification for race-specific anemia criteria. Am J Clin Nutr. 1994, 60 (1): 117-121.PubMedGoogle Scholar
- Brussaard JH, Brants HA, Bouman M, Löwik MR: Iron intake and iron status among adults in the Netherlands. Eur J Clin Nutr. 1997, 3: S51-S58.Google Scholar
- Kristal-Boneh E, Froom P, Harari G, Ribak J: Seasonal differences in blood cell parameters and the association with cigarette smoking. Clin Lab Haematol. 1997, 19: 177-181. 10.1111/j.1365-2257.1997.tb00003.x.View ArticlePubMedGoogle Scholar
- The pre-publication history for this paper can be accessed here:http://www.biomedcentral.com/1471-2288/13/62/prepub
Pre-publication history
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.