Scale has a central role in the modelling of an urban heat island (UHI). In the present study, the effects of the spatial extent of explanatory variables and modelling approach were explored in Turku (179 000 inhabitants), SW Finland, based on the average temperatures recorded by 50 temperature loggers over a 2 yr period. First, the optimal monthly circle of influence around temperature measurement points was determined for 6 urban and 6 non-urban explanatory variables. Second, the independent effect of the intercorrelated urban variables was tested using hierarchical partitioning (HP), a method for tackling multicollinearity. Third, the relative importance of the strongest urban variable was explored with 6 non-urban variables utilizing HP and multiple linear regression (LR). The mode of the optimal monthly buffer size was 1000 m for the urban variables and 2000 m for the non-urban variables. The optimal buffer size of the urban variables was largest in spring and smallest in autumn and winter. A clear seasonal trend was not observed with the non-urban variables. Based on HP, the strongest urban variable had a more independent effect on temperatures than the non-urban variables. The importance of the 'best' urban variable was also revealed in LR. Of the non-urban variables, the importance of the Landsat ETM+ derived 'greenness' variable was highlighted in HP and 'water bodies' in LR. In conclusion, the circle of influence of the explanatory variables varied substantially. Scale should be considered before doing multivariate analyses in UHI studies. Methodologically, the HP technique complements traditional UHI studies in multivariate model settings.