4

I fit the model y = a + b * x. And 95% CI for estimates of a and b are (a1, a2), (b1, b2), respectively.

If we have a new observation x0, then the estimated response is a_hat + b_hat * x0. And the 95% CI is (a1 + b1 * x0, a2 + b2 * x0)

Case 1, now given y = 0, I want to estimate the responding x. I guess the confidence interval for x is (a1 / b2, a2 / b1). Am I right?

Case 2, if we fit a model y = a + b * x + c * z + d * x * z , a_hat<0, b_hat,c_hat,d_hat>0, Given y = 0, how to calculate the confidence interval for x + z?

kjetil b halvorsen
  • 63,378
  • 26
  • 142
  • 467
OMG
  • 171
  • 4
  • 2
    For these questions to be answerable you need to supply more information, including (at a minimum) the *probabilistic* terms in each model and the method by which the parameters are being estimated. – whuber Jan 18 '14 at 22:53
  • 1
    For OLS regression, it is not the case that the 95% CI for $\hat y$ is $(a_l + b_lx_0,\ a_u + b_ux_0)$. – gung - Reinstate Monica Jan 18 '14 at 23:32
  • Yea. We still need error term. I think this problem is looking for a calibration interval. – OMG Jan 18 '14 at 23:37
  • What is a "calibration interval" in this context? – whuber Jan 18 '14 at 23:38
  • The 95% confidence interval for the estimated X will be given by the points where the `y=Y`horizontal line intersects the lower and upper confidence bands of the regression line – OMG Jan 18 '14 at 23:45
  • 1
    @whuber: in this context, calibration is inverse prediction. See p145 of Seber&Lee Linear Regression Analysis (2ed) – kjetil b halvorsen Nov 23 '18 at 18:17

1 Answers1

1

You seem to be asking about inverse prediction, also known as calibration. In short, if you have a linear regression model $$ Y=\beta_0+\beta_1 x +\epsilon $$ and you estimate it based on some selected values $x_1, x_2, \dotsc, x_n$. Then, you want to use the estimated model (and estimated parameters $\hat{\beta_0},\hat{\beta_1}$ to find the value of $x$ that will give some choosen value of $Y$, say $y_0$, and a confidence interval for that $x$.

This is discussed in Seber and Lee, page 145, for instance, with many references. There is a dedicated R package on CRAN ,investr. And it has been discussed on this site Error bars, linear regression and "standard deviation" for point

(I will try to come back to post a fuller account)

kjetil b halvorsen
  • 63,378
  • 26
  • 142
  • 467