Measurement model for gestational age from four sources of information, study in Londrina, Brazil

INTRODUCTION:  Several methods are employed to define gestational age (GA). From the fact that the measurement error is intrinsic to any measure, the objective is to create a latent variable to represent GA based on the four sources of measurements using structure equation modeling.

METHODS: The study is a population-based case-control where 437 cases are preterm mothers living in Londrina (PR, Brazil), Jun/06-Mar/07, and 402 controls preterm births obtained from a random sample according to the distribution of births per hospital. The last menstrual period (LMP) was obtained through interviews with mothers. The GA reviewed by an obstetrician(GAo), pediatrician(GAp) and USG<20 was obtained from hospital records. The continuous latent variable GA is obtained using structural equation models with maximum likelihood estimator using the software MPlus.

RESULTS:  The percentages of information in all observed individuals were: 91.9% (by LMP), 90.5% (obstetrician), 60.4% (pediatrician), 50.3%(USG<20). The four measures of GA were present in 30.4% of events. The factor loadings in the model for cases are quite high: LMP (0.804,p<0.001); GAo (0.999,p<0.001); GAp (0.949,p<0.001); USG<20 (0.996,p<0.001) and show that the correlation between them are high. The model fit was satisfactory (TLI=0.988, CFI=0.998, RMSEA=0.040). For the controls, the result of the factor loadings: LMP (0.402,p<0.001); GAo (0.997,p<0.001); GAp (0.627,p<0.001); USG<20 (0.981,p<0.001) shows that LMP and GAp differs from GAo and USG<20 but the difference is larger for LMP. The model fit was satisfactory (TLI=0.997, CFI=0.999, RMSEA=0.039). There was greater agreement between the USG<20 and GAo. The agreement between LMP and USG<20 was higher among controls than cases. The agreement between the LMP and USG<20 is higher among preterm infants.

CONCLUSIONS: Although all four measures are subject to measurement errors, the use of latent variables creates a variable for GA without measurement error that can be used in regression analyses. LMP estimative were higher in cases.