Experimental Design
We will conduct a survey experiment. The participants will be randomly assigned to one of the five groups – four treatment and one control – using a computer from the platform where the survey will be distributed. Those in treatment groups will be presented with a vignette:
(i) Treatment 1: a friend sharing her positive birth experience using midwife-led care
(ii) Treatment 2: a friend sharing her negative experience using midwife-led care, which resulted in an emergency c-section.
(iii) Treatment 3: a gynaecologist doctor sharing positive experiences using midwife-led care
(iv) Treatment 4: a gynaecologist doctor sharing negative experiences using midwife-led care, which may result in an emergency c-section
(v) Treatment 5: general information on breast cancer – this is the control group.
The main hypotheses we will test is whether a woman’s access to information from social contacts can influence her perception of the midwife-led care model. Our main hypotheses is that there are differences in perceptions about the midwife-led model among women who are exposed to information from their social contacts compared to those without information (control group).
Other hypotheses we will test include: (i) the information source effect - i.e., there are differences in perceptions about the midwife-led model among individuals who receive information from friends and medical professionals compared to the control group, and (ii) the direction of the social contact’s influence - i.e., whether there are differences in perceptions about the model among individuals who are exposed to positive or negative birth experiences from social contacts.
Our proposed sample consists of 2,079 women aged 24–55 years, residing in Italy and recruited via online panels. Of these, 250 will participate in a pilot study (50 per arm), leaving 1,829 women for the main study, allocated across five treatment arms of approximately 366 respondents each. Although this falls short of the 700 respondents per arm recommended by Haaland et al. (2023) for information experiments, we argue that our sample remains sufficiently powered to detect effect sizes commonly observed in this literature.
Following Haaland et al. (2023)—who show that information experiments typically generate large effects on beliefs but smaller effects on stated preferences or behavioural outcomes, and that null results are frequent—we adopt their recommendation that such studies should maintain at least 80% power to detect a treatment effect of 0.15 standard deviations. To account for multiple hypothesis testing across four planned pairwise comparisons, we apply a Bonferroni correction, yielding an adjusted significance level of α = 0.0125. Under this correction, our sample of approximately 366 respondents per arm achieves 80% power to detect effect sizes of 0.17 standard deviations or larger. Effect sizes between 0.15 and 0.17 standard deviations are consistent with the small-to-moderate treatment effects widely documented in the information-experiment literature, supporting both the plausibility and policy relevance of effects detectable within our power parameters.
To ensure representativeness, we employ stratified random sampling with re-weighting, benchmarking the distribution of key observable characteristics—including age, education, employment status, and macro-areas indicators—against official ISTAT population data. We adopt the entropy-balancing strategy recommended by Grembi et al. (2024) to achieve covariate balance with respect to known population characteristics, applying regional weights derived from age and educational shares. We incorporate these weights into both baseline estimation models and population-level inference, thereby ensuring that our results are generalisable to the broader population of women in Italy while preserving sufficient statistical power to detect effect sizes between 0.15 and 0.17 standard deviations at the adjusted significance level of α = 0.0125.
Data collection will capture key observable characteristics, including age, education, employment status, and macro-area indicators. Pre-treatment questions will collect demographic information (age, marital status, education, employment), pregnancy status, and awareness of alternative maternity care models, as well as factors influencing model choice. To mitigate demand effects, we will incorporate a social desirability index. Post-treatment questions will measure respondents' perceptions of pregnancy, childbirth, and the midwife-led care model. Quality-control measures will include monitoring survey completion times and implementing a post-treatment attention check that requires respondents to recall information provided before treatment (e.g., prior pregnancy history), in order to identify and exclude inattentive respondents.