Aims. The present study aimed to assess how far associations between possible explanatory variables and smoking onset depend on the use of cross‐sectional versus prospective designs. Design and subjects. Data were analysed from a three‐wave 5‐year longitudinal survey among 1063 secondary schoolchildren (12‐18‐ year‐olds) in The Netherlands. Measurements. The survey questionnaire contained items on smoking history, smoking‐specific beliefs and attitudes, social influence variables and socio‐demographic variables in each wave. Findings and conclusions. The cross‐sectional analyses showed strong associations between explanatory variables and smoking behaviour (R2 varied from 0.50 to 0.71). However, only 8% of the variance in change of smoking status from non‐smoking to regular smoking over a period of 5 years and 14% of the variance over a period of 3 years could be predicted by the model variables. Factors believed to lead to smoking may result from smoking or they may change quickly in ways that make them of low predictive value even though they may be important aetiologically.