Sample size and statistical analysis

The sample size was calculated to detect a 20% difference in NEC morbidity rate after using breast milk cell fractions, assuming an 11% NEC morbidity rate [23]. We used the sample size calculation for comparing proportions [24], considering a 5% alpha-type error rate and a statistical power of 80%. After correcting the 15% sample volume loss in each group, a sample size of 80 participants was needed for each group. Neonates were 1:1 randomly assigned to the case and control group using a computer-generated block randomization sequence of variable block-sized [25], stratified for birthweight of 1800 g. According to the stratified block randomization method, patients were divided into two equal groups: a) neonates with birthweight less than 1500 g and b) neonates with birthweight 1500–1800 g.

Where for the sample size N, Z_α/2 was the critical value of the normal distribution at α/2 (considering a 5% alpha-type error rate, the critical value is 1.96), Z_β was the critical value of the normal distribution at β (considering a statistical power of 80%, the critical value is 0.84) and p₁ and p₂ were the expected sample proportions of the two groups [24].

Categorical variables were presented as numbers (percentage) and compared using chi-square and Fisher exact test. Numerical variables are reported as mean ± standard deviation, and the Kolmogorov-Smirnov test was used to evaluate the distribution. Means of numerical variables were compared using an independent group T-test if the data were normally distributed; otherwise, the Mann–Whitney U-test was used. We fitted binary logistic regression analysis to evaluate risk factors associated with in-hospital mortality. In this study, all numerical variables had non-normal distribution. Variables with P-value < 0.1 in the univariate model (including gestational age, birthweight, and 1-min APGAR score) were considered as possible confounders and entered the multivariable model. We employed a standard entry method to adjust these models for possible confounders. The Hosmer–Lemeshow’s test was used to evaluate the goodness of fit for logistic regression models. Data were analyzed using SPSS version 23, and P-value < 0.05 is considered statistically significant.