World Bank Document

6.5 Physical Infrastructure vs. Human Capital?
A critical question faced by developing countries, particularly low-income countries, is 
where they should prioritize their scarce resources and what to prioritize – should they use the 
resources on physical infrastructure (i.e., road, electricity) or use to improve social welfare (e.g., 
health care) or use it to strengthen their long-term human capacity (e.g., education). How to 
optimize the allocation of resources between increasing human capital versus or strengthening 
physical capital? Earlier literature such as Samuelson (1954), Arrow and Kurz (1970) provide 
textbook guidance on the optimal theory of public expenditure. The question is not limited to 

---

32 
public expenditure only – it is an issue of overall investment in physical as well as human 
infrastructure. Some existing studies, such as Jimenez (1994), Devarajan et al. (1996), also point 
out this research gap. There might not be a generic answer to this question. The optimal mix of 
investments between the physical infrastructure and human capital could vary across countries, 
depending on their specific situations. It would, therefore, be interesting to investigate this 
variation and provide insights to policy makers.
6.6 Long-Run Impacts of Infrastructure Investment
Panel data studies such as Calderón and Servén (2010) pick up the combined effects of 
infrastructure adjustment processes over time for countries with differing initial economic and 
institutional levels of development. Such studies can be used to assess statistically both the impact 
effects of changes in infrastructure development and longer-term effects. However, such research 
does not in itself provide insight on how infrastructure affects development over the longer-term.
Experimental studies (e.g., Lenz et al. 2017) can provide insight on short-term impacts of 
infrastructure changes for certain users but have less to say about the longer-term impacts across 
the economy. Long-term structural models can provide that information, but such models are more 
complex to build and estimate, and the literature contains only a few exceptions (see, e.g., Roberts 
et al. 2012).