Courrier des statistiques N10 - 2023
Can we rely on non-probability sampling?
Sample surveys are based on either a probability sample or a non-probability sample. In the non-probability approach, the probability of a given individual being included in the sample generally depends on the value of the variable collected from that individual. This produces a particular error known as ‘selection bias’. In the non-probability ‘quotas’ method, this bias is limited by structuring the sample according to certain variables that explain the measured phenomenon. However, a bias remains if those variables fail to account its whole variability. In order to fully justify the method, one appeals to an assumed behaviour of individuals, known as modelling. Other non-probability selection methods exist, such as the purposive selection method – reflecting the common perception of ‘representativeness’ - or volunteer sampling, particularly developed in recent years through ‘Access panels’. In this last case, the selection bias can be significant, even considerable. Unfortunately, the bias cannot be reduced by increasing the size of the sample. Two striking examples – one relating to the vaccine uptake rate against the coronavirus, the other to the 1936 presidential elections in the United States – illustrate this phenomenon, known as the ‘big data paradox’.
Paru le :12/02/2025