1
$\begingroup$

I have a linear model $y = ax+b$ and I estimate the coefficients $a$ and $b$ in the ordinary way.

I have found out that all of my values of $x$ were systematically overestimated, and also that they have known uncertainties that were heretofore not propagated.

Using total least squares (Deming regression), I find new estimates $\tilde a$ and $\tilde b$. These estimates have a much larger uncertainty ($15\%$ greater), but overlap with the previous estimates of $a$ and $b$.

Since the distributions of the estimates overlap, a referee has said that the change is not significant, and has asked me if I can show that the increase in uncertainty is significant. How can I determine this?

Also, due to regression dilution, the estimated central value of $\tilde a$ is larger than $a$. Is there a one-sided test I can use to show that the means are different, even if their distributions overlap?

Any advice appreciated. Thanks!

$\endgroup$
1
$\begingroup$

I think you ca use the $F$ test for equivalence of variance. I.e., $H_0: \sigma^2_{a^1} /\sigma^2_{a^0} \le 1$, thus checking the estimated ratio where $$ \hat{\sigma}^2_{a^0} = \frac{MSE}{\sum(x_i - \bar x)^2} $$ against $1-\alpha$ quantaile of $F_{(n_1 -2, n_0 - 2)}$ will give you the answer.

$\endgroup$
  • $\begingroup$ This is actually what I had ended up doing so I am glad to see I got it right! $\endgroup$ – rhombidodecahedron Feb 4 '17 at 15:47

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.