Predictors of CD4-for-age after ART for 914 children with asymptotic CD4 reconstitution.

<p>Results are expressed relative to the reference category (boy, aged 6 y, with weight-for-age <i>z</i>-score −2.2 and randomised to first-line ART strategy Arm A [standard of care two-NRTI+NNRTI throughout]). Analysis was carried out on the natural logarithm scale (ln[CD4/expected CD4]), but results are reported both on the natural scale and back-transformed to the linear scale (CD4/expected CD4). Estimated effect sizes of covariates on the baseline (intercept) and the long-term (asymptote) CD4-for-age are additive on the natural log scale, but multiplicative on the linear scale—for example, a 1-y increase in age is associated with a −0.08 unit lower ln(CD4/expected CD4) at ART initiation or, equivalently, a CD4/expected CD4 lower by a factor of exp(−0.08) = 0.92. Similar effect sizes were seen in univariable models. After adjusting for the factors shown, there was no additional predictive value on intercept of sex, WHO disease stage 3 or 4, or first-line ART or monitoring strategy; on asymptote of WHO disease stage 3 or 4, weight-for-age <i>z</i>-score, or monitoring strategy, or on <i>c</i> of age, sex, WHO disease stage 3 or 4, or ART or monitoring strategy (likelihood ratio tests, <i>p</i>>0.01; further details are given in <a href="http://www.plosmedicine.org/article/info:doi/10.1371/journal.pmed.1001542#pmed.1001542.s003" target="_blank">Table S2</a>). Although weight-for-age was not selected in the backwards selection, likelihood ratio tests to assess additional predictors on top of the final multivariable model suggested weight-for-age might be a predictor for <i>c</i> (<i>p</i> = 0.003 in likelihood ratio test; estimate 9.31×10⁻⁴ higher per one-unit increase in weight-for-age [95% CI: −7.01×10⁻⁵, 1.93×10⁻³], i.e., faster proportionate recovery speed in those with higher weight-for-age at ART initiation). However, other effect size estimates were essentially unchanged, and examining closely related covariate models suggests this apparent predictive value may be the result of overfitting. Thus, the final backwards selection is presented as the final model, as described in the <a href="http://www.plosmedicine.org/article/info:doi/10.1371/journal.pmed.1001542#s3" target="_blank">Methods</a>.