Abhijit Banerjee provides a good overview (in this article) of the difficulties of finding what works in development economics. He describes work by Paul Glewwe and others that found schools with certain science education materials showed greater learning. Was it the materials that caused the learning? It appears it was not after they selected a group of similar schools and gave half the materials and left half as a control group, they found no difference in learning between those with and without charts. This led them to conclude that those communities with charts before their experiment, were ones that had a good environment for schooling already.
In my own work I have been studying a program that gives money to families if they send their children to school, often referred to as a conditional cash transfer. The program I study in Nicaragua had a randomized component that is the program was given to half the communities (a treatment group), which is compared to a similar control group. The program increased elementary school enrollment from 70% to 95%. Unrelated to this program many communities, in both control and treament groups offered a free meal at school, these programs were not randomized. The number of communities offering food at school dramatically increased from 25% to 60% from the year 2000 to 2001. We might expect that it had something to do with the cash transfer program, which started in 2001, but treatment communities and control communities were equally likely to offer food at school. Interestingly in 2001 communities with school feeding programs had school enrollment rates about 20% higher then those that did not. However, further inspection shows that those communities that added a school feeding program in 2001 after not having one in 2000, had enrollment rates 20% higher in 2000 compared to those that did not have one in either year. So it turns out that offering a food program is related to how many children attend school. When initial enrollment rates are controlled for the impact of the program on school enrollment disappears. Without the 2000 data we might think that food programs are just as effective as cash transfers, when they are likely not.
This is why experiments in economics work so nicely at determining causation.