Article Text

Ocular abnormalities in a cohort of children born prematurely: effects of selection bias and possible confounding
  1. Centro de Ciências Biomédicas
  2. Universidade Federal de Uberlândia
  3. Uberlândia, MG, Brazil
  1. Av Teresina 1415, 38405–384 Uberlândia, MG, Brazil.

Statistics from

Request Permissions

If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.

Editor,—Studying children born prematurely Pennefather and colleagues1 showed that the prevalence of several ocular abnormalities at the follow up examination at 2 years differed significantly between children who belonged to families who attended follow ups as a routine (group 1) and those who were classified as very reluctant for assessment (group 3). These differences were of relatively high magnitude, with relative risks (actually odds ratios) varying from 5.54 for strabismus to 10.91 for cicatricial retinopathy of prematurity. The authors claim that they used multiple logistic regression in their statistical analysis, but if the figures shown in their study are adjusted odds ratios they are, according to my calculations, identical to the crude odds ratios, which means that there were no confounders to any of the studied associations. This apparently contradicts the findings of the study of Campbell and colleagues,2 quoted by the authors, that level of prematurity of the children and age and marriage status of the mothers were correlated with non-attendance. If these variables are also associated with the ocular abnormalities of Pennefatheret al's study, and not intermediate variables between the exposures and outcomes of interest, they are confounders, and should have been adjusted for in the multivariate analysis.3 This point needs clarification.

Very interesting was the finding that the overall prevalence of abnormalities was similar between the total cohort (13.4%) and group 1 (11.3%). This small difference is explained because the proportion of losses to follow up that would have occurred under routine conditions was relatively small (9.5%), and did not have an important impact. It is an empirical demonstration that in cohort studies, for obtaining valid relative risk estimates it is very important to keep losses to follow up to a minimum, thereby minimising the role of selection bias.4