You don't want to know if the devices are different but which is more accurate. Just run the test vehicles over several measured distances and compare. Remember tire wear will affect the odometer measurements so that the tracker will probably be more accurate. Best, David

You could frame the problem as an ANOVA and use a variance-component model to estimate the intraclass correlation (rho) or the proportion of total variance accounted by speed reading differences between cars (basically you have a repeated measure framework where cars are the subjects and the readings, one on the odometer and one on the digital tracker respecitvely for each car, represent the occasions). If rho is close to 1 that means that all the variability is between cars and so the odometer is a very reliable replacement for the 'gold standard' but more expensive digital tracker device. On the contrary, if rho is not that close to 1, then part of the variance in the readings is explained by within-car variability. This means, in turn, that the two readings are not so consistent with each other. You can calculate a confidence interval for rho and decide if the odometer's readings are sufficiently reliable as a cheaper alternative to the digital tracker. Compared with a t-test on mean reading difference for paired data, this approach relaxes the assumption of independence between readings taken on the same vehicle.

You have the differences of digital-odometer per vehicle, where I assume that the odometer (O) values are your "gold standard" where you want to compare the digital tracker values (D) against. Showing these differences D-O vs. O will tell you a good part of the story. This is actually almost a Blant-Altman-plot (the latter would use (D+O)/2 x-coordinate). The further analysis depends on what you see there. Is there a constant offest? Is there a trend? Is there a non-linear relationship? Is the ratio D/O constant/independent of O (-> plot D/O vs. O)?

If the traveled distances are all very similar (and typical in practice), then an analysis of the dependency of the correctness on the travelled distance (as you did with the Bland-Altman plot) is not required (and not really possible or sensible to do). Instead, you may simply estimate the mean difference. As you have to make a decision (substitute the odometers?) you may want to define a loss function and integrate the product of the loss and the posterior density of (D-O) or (D/O) to get the expected loss. If this is negative (net win) then you should decide to switch to digital trackers (as this is expected to save money - or in whatever unit you give the loss function). I don't see how anything else (including t-tests and ANOVAs, p-values in general) would help you make a reasonable decision.