Suppose we want to estimate a model:

$Y_i = \beta_0 + \beta_1X_i + \epsilon_i\tag{1}$

where the $Y_i$ are averages or rates per capita for geographical regions or zones, so that if $Z_i$ is the regional aggregate variable of interest and $P_i$ is population:

$Y_i = Z_i / P_i\tag{2}$

If all the variables are accurately measured, and if there exist $β_0^*$ and $β_1^*$ such that for all i:

$E[Y_i - \beta_0^* - \beta_1^*X_i] = E[\epsilon_i] = 0\tag{3}$

then OLS estimation of model (1) should yield an unbiased estimate of $\beta_1$.

Suppose however that estimates of $P_i$ are only approximate. This is in practice likely since the values of $P_i$ will almost certainly be obtained or derived from census data, leading to several possible sources of error:

  1. error in the original census data;
  2. extrapolation from census data using assumptions on population trends where data are required in respect of a later date;
  3. the regions used in the model may not correspond to census areas, eg concentric rings around a central point. An example is a travel cost valuation study reported in Herath (1999) (1) where populations in the range 5,000 to 35,000 for concentric ring zones are all stated in exact multiples of 5,000, suggesting that the figures are only very approximate.

The errors in $P_i$ will obviously ‘infect’ the values of $Y_i$, but not in a straightforward way, since the absolute effect on $Y_i$ of a given error in $P_i$ will depend on the size of $Z_i$ (and if $Z_i = 0$ there will be no error in $Y_i$).

Question: Given errors in $P_i$, will OLS estimation of model (1) yield an unbiased estimate of $\beta_1$, and if not, under what additional conditions would the estimate be unbiased?


  1. Herath, G (1999) Estimation of Community Values of Lakes: A Study of Lake Mokoan in Victoria Australia Economic Analysis & Policy 29(1) Table 1 p 37

2 Answers 2


What about assuming a multiplicative error process and then using logs?

Say the data generating process were a little different: $Y_i = \beta_0 \cdot X_i^{\beta_1}\cdot E_i$

If $Y_i = Z_i / P_i$ in truth but all we could really observe was:

$\hat{P}_i = P_i \cdot \Gamma_i$, where $P_i$ was the true value and $\Gamma_i$ was a strictly positive measurement error.

We' be estimating the following equation in practice:

$\hat{Y}_i = Z_i / \hat{P}_i = \beta_0 \cdot X_i^{\beta_1}\cdot E_i$

take the logs of both sides (lower case letters are logs of individual variables):

$y_i = \beta_0 + \beta_1 \cdot x_i + \epsilon_i - \gamma_i$

If we define a variable $\xi_i = \epsilon_i - \gamma_i$ then we can have an equation $y_i = \beta_0 + \beta_1 \cdot x_i + \xi_i$ that looks very much like the one you wrote above.

If $\xi_i$ satisfies the same relationships with the log variables that $\epsilon_i$ does with the level variables it seems like everything should still work.


Just to complement @Bkay's answer, "measurement error" in the dependent variable is relatively "harmless" -what hurts is error in measurement in the regressors.

If we have measurement error in the dependent variable, what we need to assume in addition, in order to preserve unbiasedness, is that this error is independent of the regressors. If we can reasonably assume that (and usually we can), then @BKay 's answer shows that the effect is just a transformation of the error term of the regression. It may affect the variance, but not the parameter estimates.

On the contrary, if we have measurement error in the regressors, then they stop being strictly exogenous to the error term, and unbiasedness is lost.

  • $\begingroup$ What concerns me is that this is a setting where you want to infer individual responses from average responses from sub-populations. But we know from Simpson's Paradox that this is often wrong. So the result from @AdamBailey equation 3 is actually quite restrictive and I wasn't sure that adding the LHS measurement error which is usually harmless at the individual level would similarly be generally harmless at the sub-population error given the desired inference. $\endgroup$
    – BKay
    Feb 1, 2015 at 10:44
  • $\begingroup$ @BKay Possibly I didn't formulate my (3) correctly. What I meant to imply is that the model meets all the conditions of the Classical Linear Regression Model, other than homoscedasticity (that exception has to be made because models of the type described are almost inevitably heteroscedastic since, other things being equal, $Var[Y_i]$ will be smaller for regions with larger populations). $\endgroup$ Feb 1, 2015 at 21:27
  • $\begingroup$ @Bkay In any setup, whether the measurement error of the dependent variable is harmless or not, depends on whether it can be considered stochastically independent of the regressors. So one looks at the regressors and makes an assessment of the issue. In Adam's question the regressors are not identified at all, so we cannot go further than the general principle just stated. $\endgroup$ Feb 1, 2015 at 21:57
  • $\begingroup$ @AdamBailey Regarding induced heteroskedasticity, it depends on what are considered reasonable assumptions on the error term, i.e. how we believe that the error measurement emerges, and what does this may imply for its stochastic behavior. $\endgroup$ Feb 1, 2015 at 22:01
  • 1
    $\begingroup$ @Bkay Random coefficient models need not be hierarchical, nor do they necessarily need panel data. I recall for example the Hildreth-Houck model. As for what we recover when running estimations, this is indeed something that always needs careful thought. $\endgroup$ Feb 2, 2015 at 0:20

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.