As far as we know, published data describing repeated haemoglobin measurements from children over extended periods are very scarce. Only one analytic method for such data has been proposed heretofore 
. We previously showed that neither haemoglobin level at birth nor maternal anaemia were associated with children Hb level in the first18 months, but the occurrence of a malaria attack during follow-up, male gender and sickle cell trait carriage were associated with a lower children haemoglobin level in the first18 months. Children living in a polygamous family and with good feeding practices had a higher Hb level in the first18 months of age than the others. The present analysis, using latent class models, confirms these results, strengthening the importance of these risk factors, which have been discussed elsewhere 
. However, this approach, one among several population mixture models, provided a better understanding of the natural history of children haemoglobin levels by identifying new determinants of haemoglobin levels (i.e. placental malaria and the newborn anaemia) as well as the role played by mother’s anaemia as a predictor of belonging to the low latent trajectory. We also showed that the negative effect of sickle cell trait was more strongly marked within the low trajectory. In our preceding analysis using mixed models, we found that newborn anaemia and placental malaria were associated with a decreased haemoglobin level in the first 18 months of life. However, these associations were not significant (p
0.104 and p
0.190 respectively). This result strengthens the fact that taking into account the existence of a mixture of populations could help to identify covariates with significant but probably complex effect that cannot be easily identified under the hypothesis that all subjects belong to the same population. Finally, the innovative LCA suggests that the data are compatible with two levels of haemoglobin over time, high and low, concerning two-thirds and one-third of the sample, respectively.
Latent class analysis has been used extensively in criminology and behavioural research 
, less to date in public health research 
. It is an extension of a mixed model and assumes the presence of and identifies latent groups of individuals who share a particular developmental trajectory of some attribute, thereby allowing a better understanding of the pattern of change in the variable of interest 
. The other strength of this approach is that it reveals a predictor of belonging to the low haemoglobin group along with the effects of all other risk factors. Furthermore, it allows studying the effect of variables within trajectories and underlined the negative influence of sickle cell trait during infancy. Latent class analysis forms a part of mixture modeling, a widely applied data analysis approach used to identify unobserved heterogeneity in a population. Mixture modeling involved several techniques with potential differences 
. We provided our model equation to take these potential differences into account and to make our results easily and clearly understood, despite technical considerations, with no risk of miss understanding. One limitation of our approach could be that each group has the same variance structure rather than a variance-covariance structure among time points.
Multiple indices have been described to identify clusters in Latent Class Analysis. To date, there is not common acceptance of the best criteria for determining the number of classes in mixture modeling, despite various suggestions 
. Among them, several simulation studies indicated that the BIC performs well 
. More recently, Nylund et al. found that the Bootstrap Likelihood ratio test (BLRT) presented the best performances and that the second best index was the BIC. They also showed that the BLRT, does have its disadvantages such as the increased computation time 
, but this should not affect the fact that it is more powerful and accurate. Finally, the authors strongly recommended using the BIC as the first step. Here, we followed this recommendation, together with the criteria proposed by Andruff 
based on the percentage of population included in each group. However, despite the fact that they are widely used in the literature 
, as they are partly based on the percentage of the sample within the trajectories, this choice is somewhat subjective. To take this point into account, we also used the Gateaux derivative method to determine the number of groups required to achieve the largest possible likelihood. This method has confirmed our choice of two trajectories for the changes in haemoglobin level, suggesting that the three-point solution may represent a local maximum of the log-likelihood.
Finally, to take into account the complexity and the possibilities of these population mixture models, it could be of great interest to pursue this exploration by using more indices to define the number of clusters and by adding a variance-covariance structure among time points. This could be proposed by using and comparing several modelling mixture methods in a near future for such complex data.
In the present study, maternal anaemia at delivery seems to be predictive of belonging to the low trajectory but has no direct effect on the changes in haemoglobin level over time. The association we found between the effect of maternal anaemia at delivery, and the absence of direct effect on the evolution of haemoglobin level during infancy, could apparently seems contradictory. Indeed this trajectory loses Hb over time, whereas the higher trajectory does not. One possible explanation could be that the association between maternal anaemia at delivery and haemoglobin level during infancy could be (at least partly) mediated through newborn anaemia. Indeed, inserting both maternal anaemia and newborn anaemia in the same model could result in collinearity, which could explain the apparent absence of effect of maternal anaemia on children haemoglobin evolution. However, maternal anaemia remained not significantly associated (p
0.14) with the evolution of children haemoglobin progression even when included in a multivariate model without newborn anaemia. Furthermore, using the Bayesian Informative Criterion as criteria, the model with newborn anaemia was selected when compared to the model with maternal anaemia. Using haemoglobin level of mother at delivery as quantitative variable did not change these results (data not shown). Hence, this pattern of results can be interpreted as the fact that maternal anaemia stops having a negative effect at birth but that children born of an anaemic mother are probably disadvantaged during infancy. The mother’s anaemia could also be interpreted as an indirect marker of a woman’s and/or a family’s disadvantage because of poor socioeconomic status that could be associated with inadequate nutrition during infancy. The importance of adequate nutrition is illustrated by the positive effect of nutritional status on the haemoglobin progression during infancy.
The mechanisms by which a newborn’s anaemia at birth can affect haemoglobin level over time have not yet been clearly described. A period of rapid growth, especially during infancy, results in substantial demands for iron. In developed countries where breastfeeding is not common, mothers often use children formula fortified with iron in order to supplement the children’s needs. This is not necessary with breastfeeding, which provides high concentrations of highly absorbable iron. In developing countries where all children are routinely breastfed during the first year of life, we could expect that breastfeeding compensates iron requirements and corrects anaemia. However, the interactions between iron intakes and stores are complex and it has been shown that exclusive breastfeeding at 4 months of life was protective of iron status and of iron deficiency-anaemia at 6 months, compared with children receiving early complementary food 
. In our study population, breastfeeding is far from exclusive, and at 4 months of age only 18% of children were exclusively breastfed (data not shown). Moreover, the mothers from our sample are could be iron depleted. Put together our results show that both maternal anaemia at delivery and newborn anaemia at birth are associated at different level, with haemoglobin progression in children during the first 18 months of age. These two risk factors could interact with each other and furthermore interact with feeding practices and have to be taken into account to define preventive strategy.
In this study, placental malaria was associated with a low haemoglobin level in the first18 months of life. This association was not significant during our first analysis 
. This association has already been described by Redd et al
. (1994), who have shown that placental malaria was associated with anaemia around 2 months of age 
. In cases of iron deficiency during pregnancy, expression of placental transport proteins for iron increases, allowing a greater transport of iron to the foetus 
. In case of placental malaria, a thickening of the trophoblastic basement membrane has been described, damaging the placenta’s active transfer capacity 
. It can therefore be assumed that placental malaria reduces the transfer of iron from the mother to her children, increasing the newborn’s iron deficiency. However, a more indirect explanation can be proposed that is consistent with the effect of placental malaria on haemoglobin level over time. Indeed, some authors have hypothesized that placental malaria is associated with an immune tolerance phenomenon 
. Children born of infected placenta are more susceptible to malaria infection. As both the number of malaria attacks during the follow-up and placental malaria were independently significantly associated with the level of haemoglobin in the first18 months, we can argue that children born of mothers with placental malaria are more susceptible not only to malaria but also to other infections, as we have recently shown in this same cohort 
. These children, frailer and more often infected, have a higher risk of developing anaemia.
Our study also described a negative effect of sickle cell trait among trajectories. As this result was obtained during multivariate analysis with both variables (i.e. malaria and sickle cell) we can argue that the sickle cell trait effect is independent of the protection against clinical malaria classically described 
. Our hypothesis is that this effect could be explained by an intrinsic role of sickle cell trait in anaemia, even for heterozygous individuals. To our knowledge, one study has described an association between anaemia and sickle cell trait 
but this association has not been confirmed to date. In addition, it has been described that haematuria, both microscopic and macroscopic, is one of the most frequent complication of sickle cell trait 
. In our sample no macroscopic haematuria was found and we did not search a microscopic haematuria. However although we could hypothesize that children with sickle cell trait experienced microscopic haematuria, we cannot explain clearly the different effect of this variable in each group.
According with our protocol the children found anaemic were treated as proposed by the recommendation of the Ministry of Public Health of the Republic of Benin. They received haematinics which was prescribed by the nurses working in the public dispensaries of the area. The team involved in the follow-up was different and the population was monitored equivalently and independently of the Hb levels or of they life conditions. Furthermore, as the families involved in this program were very similar we do not think that a “cohort effect” could represent a limitation of the study.
In conclusion, this study, using latent class analysis models, shows that the occurrence of a malaria attack during follow-up, male gender, sickle cell trait carriage, children living in a polygamous family, children with good feeding practices, newborn anaemia, placental malaria were associated with haemoglobin level in children and that maternal anaemia was a predictor of a low haemoglobin level trajectory in children in their first 18 months of life. Latent class approach could be applied more frequently to analyse longitudinal data when the existence of groups with distinct pattern of evolution is suspected. This assumption cannot be adequately explored with mixed models. The prevalence of anaemia during pregnancy and in the newborn is very high in developing countries, 40% and 61%, respectively, in our study in Benin. There is a need to increase actions that target the prevention of maternal anaemia as well as placental malaria and newborn anaemia.