Follow-up duration, missing data and recording bias in routine primary- care pregnancy records: a data-quality case study of spontaneous abortion in 89,772 SINA records from Mashhad, Iran

Background Routine primary-care information systems are attractive for pregnancy research, but their records are generated during care, and pregnancies that end early leave fewer records. We assessed how follow-up duration, missing data and comparison-group definition affect risk-factor estimates for spontaneous abortion in the SINA system of Mashhad University of Medical Sciences, Iran. Methods We analysed 89,772 unique pregnancy records from the project year 1397 (March 2018–March 2019) (5,561 with recorded spontaneous abortion; 84,211 with live fetal status). A naive analysis using all recorded variables was compared with a structure-aware analysis restricted to entry-time variables. We evaluated supplement, laboratory and psychosocial-score records as markers of follow-up, compared three missing-data strategies for BMI, bounded the effect of the heterogeneous comparison group with restrictions to women with advanced or limited follow-up, quantified centre-level recording completeness, and classified the stability of associations across seven analyses. Results In the naive analysis iron appeared almost completely protective (OR 0.005) and folic acid strongly harmful (OR 9.49); 85% of women with only folic acid recorded had an abortion versus 0.5% with only iron. Adding supplement records to an entry-data model raised the cross-validated AUC from 0.713 to 0.937. BMI was missing in 2.4% of cases but 18.2% of comparison women, and odds ratios for one living child ranged from 2.85 to 3.52 across missing-data strategies. Only diabetes (OR 1.56, 95% CI 1.23–1.98) and university education (0.81, 0.72–0.90) were stable in direction and size across all analyses. Age < 20 years (2.27, 2.02–2.55), recurrent abortion (1.92, 1.63–2.27), stillbirth (1.38, 1.07–1.79) and living children were stable in direction, but their size depended strongly on how the comparison group was defined, and living children were opposite in direction to previous studies. Age ≥ 40 years, obesity, hard job and smoking changed direction across analyses, and the association of underweight (1.30, 1.15–1.48) depended on assumptions about missing BMI. Conclusions Without dates, gestational age and final pregnancy outcome, routine records support only a limited set of entry-time risk factors; variables recorded during follow-up give spurious associations. We propose rules for analysts and minimum data requirements for SINA.