A fundamental assumption in quantitative genetics is that traits are controlled by many loci of small effect. Using genomic data, this assumption can be tested using chromosome partitioning analyses, where the proportion of genetic variance for a trait explained by each chromosome (h2c), is regressed on its size. However, as h2c-estimates are necessarily positive (censoring) and the variance increases with chromosome size (heteroscedasticity), two fundamental assumptions of ordinary least squares (OLS) regression are violated. Using simulated and empirical data we demonstrate that these violations lead to incorrect inference of genetic architecture. The degree of bias depend mainly on the number of chromosomes and their size distribution and are therefore specific to the species; using published data across many different species we estimate that not accounting for this effect overall resulted in 28% false positives. We introduce a new and computationally efficient resampling method that corrects for inflation caused by heteroscedasticity and censoring and that works under a large range of data set sizes and genetic architectures in empirical data sets. Our new method substantially improves the robustness of inferences from chromosome partitioning analyses.