A few things caught my eye that might be more than just minor quibbles. Here’s what I’d ask the authors:
1. Your peer uninformed decision-making instrument seems to violate the exclusion restriction.
It’s the leave-out proportion of people in the same province and age class who make financial decisions without consulting reliable sources. But that same peer behavior likely correlates directly with local economic norms, social capital, or even regional financial fragility (e.g., if everyone around you is financially fragile, you might be too, regardless of your own knowledge). So how can you argue it affects financial fragility only through your own financial knowledge? This looks like a classic reflection problem; and if it’s invalid, your IV identification falls apart.
2. Your exogeneity tests (Amemiya-Lee-Newey) show p-values near 0.94, which is suspiciously perfect.
That suggests overidentification restrictions are almost exactly satisfied. In my experience, that often happens when instruments are weak in a specific way or when the test lacks power due to small sample or model misspecification. Given your first-stage F is huge (190), the instruments are strong, so the overid test should have some bite. A near-1 p-value makes me wonder: are your instruments actually correlated with the error term in a compensating way? Did you check sensitivity to dropping one instrument at a time?
3. In your dynamic model, the lagged fragility effect (AME ~0.03) is tiny compared to initial fragility (~0.22).
That implies almost no true state dependence once you control for initial conditions; but then you claim strong evidence of genuine state dependence. With only 4 waves and an unbalanced panel, your Wooldridge-style initial conditions estimator is known to be sensitive to misspecification of the distribution of unobserved heterogeneity. How do you rule out that the lagged coefficient isn’t just picking up remaining serial correlation in the errors rather than causal persistence?