Intended for healthcare professionals


Neonatal risk factors for cerebral palsy in very preterm babies

BMJ 1997; 314 doi: (Published 31 May 1997) Cite this as: BMJ 1997;314:1624

Time oriented analyses of risk are useful

  1. Elizabeth N Allred, Instructor in neurology, Harvard Medical Schoola,
  2. Olaf Dammann, Research fellow in neurology, Harvard Medical Schoola,
  3. Karl Kuban, Assistant professor of neurology, Tufts Medical Schoola,
  4. Alan Leviton, Professor of neurology, Harvard Medical Schoola,
  5. Marcello Pagano, Professor of statistical computing, Harvard Medical Schoola
  1. a Neuroepidemiology Unit, Children's Hospital, 300 Longwood, Boston, MA 02115, USA
  2. b Department of Obstetrics and Gynaecology, St Michael's Hospital, Bristol BS2 8EG

    Editor—Deirdre J Murphy and colleagues recently reported a case-control study comparing 59 children with cerebral palsy with 234 without.1 One of their key messages was “Neonatal pneumothorax, sepsis, and transfusion are associated with preterm cerebral palsy independently of adverse antenatal factors.” We are not convinced that they have shown an independent effect.

    To identify the contributions of exposures and characteristics occurring at one time independent of the contributions of exposures and characteristics that occur later, we have performed time oriented logistic regression analyses in which risk factors are ordered in a temporal pattern. The earliest occurring predictors and covariates of the outcome are entered first and are not displaced by covariates occurring later.2 An odds ratio significantly different from 1 for the variable occurring later suggests that it contributes risk information that supplements the information provided by the earlier factor for which one adjusts.

    The authors might have done something similar, but it is not clear which of the six antenatal variables remained in each final model (with no more than six variables, including the neonatal variable and gestational age). Entering variables into logistic regression models, as Murphy and colleagues did, might have allowed neonatal variables to displace highly interrelated antenatal variables already in the model. Antenatal variables might thus lose their significance and be dropped from the model after correlated postnatal variables are entered. For example, maternal infection, a documented antecedent of cerebral palsy in this sample, might produce its effect either directly or through postnatal phenomena such as hypotension. Maternal infection is not adequately adjusted for if it is displaced by the variable for postnatal hypotension.

    The problems posed by the analytic strategy of the authors are exacerbated by the authors' selection of variables for multivariable analyses only if univariable associations with the outcome of interest had P values of ≤0.05. This approach can inappropriately exclude confounder variables from further analysis. Dales and Ury suggested using P values of <0.25, or even no significance testing at all.3

    In conclusion, the authors' methods of sequential processing and variable selection may have limited their opportunity to achieve what they wanted. We invite them to clarify what they did and what they found.


    1. 1.
    2. 2.
    3. 3.

    Author's reply

    1. Deirdre J Murphy, Specialist registrarb
    1. a Neuroepidemiology Unit, Children's Hospital, 300 Longwood, Boston, MA 02115, USA
    2. b Department of Obstetrics and Gynaecology, St Michael's Hospital, Bristol BS2 8EG

      Editor—Elizabeth N Allred and colleagues highlight the difficulties of using logistic regression techniques to identify independent associations, particularly in the context of observational data where temporal relations are of interest. In our study the strategies available to us were to design logistic models on the basis of clinical plausibility, statistical strength, or a combination of both. In the first instance we selected the antenatal variables associated with preterm cerebral palsy identified from an earlier study.1 We selected variables both on the basis of significance and if an interaction with neonatal events seemed plausible. For example, an odds ratio for antepartum haemorrhage of 1.5 (95% confidence interval 0.8 to 2.6) was selected as an antenatal variable of interest despite the 95% confidence interval including the null value as clinically this factor could be related to postnatal hypotension or the need for neonatal blood transfusion.

      The earliest multivariate analyses involved a single antenatal factor, a single neonatal factor, and in each case gestational age (the strongest potential confounder). A pattern emerged identifying a small number of antenatal factors (maternal infection, chorioamnionitis, pre-eclampsia, and caesarean section without labour) that altered the associations between preterm cerebral palsy and neonatal events. We then built on the single factor approach by adding in a second antenatal factor exploring all the possible combinations of the aforementioned variables. This then led to the final models including a maximum of six variables where in fact all the variables had P values of ≤0.05 but the opportunity had been explored of incorporating other variables that did not reach this level. As Allred and colleagues advise, we entered the antenatal factors first in the model, followed by the mode of delivery and then the neonatal factors. We hope that this clarifies some of the issues raised.


      1. 1.
      View Abstract