Essay On Randomization

Introduced by Fischer in 1935, randomised experiments are often considered the gold standard of social science experiments (Rosenbaum, 2002). Starting the 1960s they have been progressively used in social science, mostly in clinical trials but also in areas of economics, such as welfare reforms (Moffitt, 2004). Nonetheless, they also contains inherent limitations and difficulties in the way they are implemented, enough for Beck to downgrade them as “the bronze standard” (2005, p1).
I will first briefly cover what are randomised experiments and how they support the goal of social science experiments by allowing for proper casual inference. Next, I will discuss their main inherent challenges and why it is sometime not possible to use randomisation

Non truly random procedures will reintroduce selection bias and make casual inference difficult. Indeed random is different from haphazard or uncontrolled: sampling participants based on their or day of birth, or order of arrival to the lab is not a sound idea. Those variables are not produced solely by chance and we can suspect unobserved confounding covariates in both groups.
In terms of randomisation design the choice of level of randomisation can raise some issues: it is sometime not possible to randomise at the individual level, in which case one should turn to cluster randomisation. While cluster randomisation allow for proper inference, this might increase the overall cost of the experiment as this requires a much larger total sample size (Duflo, 2009). On the other hand individual randomisation might leads to a contagion between the control and intervention group, and estimated effects that are lower than their true population counterpart, such as demonstrated by Miguel and Kremer (2004) in the case of deworming drugs. Some of the challenges of simple randomisation might appear when using a small sample size (N<100). In this case, to ensure the the randomisation lead to equal group size or an equal distribution of underlying covariate one might need to use block and stratified