Performance of quantile rank score tests used for hypothesis testing and constructing confidence intervals for linear quantile regression estimates (0 ≤ τ ≤ 1) were evaluated by simulation for models with p = 2 and 6 predictors, moderate collinearity among predictors, homogeneous and hetero-geneous errors, small to moderate samples (n = 20–300), and central to upper quantiles (0.50–0.99). Test statistics evaluated were the conventional quantile rank score T statistic distributed as χ² random variable with q degrees of freedom (where q parameters are constrained by H ₀:) and an F statistic with its sampling distribution approximated by permutation. The permutation F-test maintained better Type I errors than the T-test for homogeneous error models with smaller n and more extreme quantiles τ. An F distributional approximation of the F statistic provided some improvements in Type I errors over the T-test for models with > 2 parameters, smaller n, and more extreme quantiles but not as much improvement as the permutation approximation. Both rank score tests required weighting to maintain correct Type I errors when heterogeneity under the alternative model increased to 5 standard deviations across the domain of X. A double permutation procedure was developed to provide valid Type I errors for the permutation F-test when null models were forced through the origin. Power was similar for conditions where both T- and F-tests maintained correct Type I errors but the F-test provided some power at smaller n and extreme quantiles when the T-test had no power because of excessively conservative Type I errors. When the double permutation scheme was required for the permutation F-test to maintain valid Type I errors, power was less than for the T-test with decreasing sample size and increasing quantiles. Confidence intervals on parameters and tolerance intervals for future predictions were constructed based on test inversion for an example application relating trout densities to stream channel width:depth.