I've come across some past questions asking about using a linear probability model in place of a probit model when the data generating function has uniformly distributed errors. However, I am uncertain about deriving the variance of the linear probability model. The problem is as follows:

enter image description here

The first part seems straightforward enough (shown below), though I think I may be missing something fundamental about how the e_i error term arises.

enter image description here

When it comes to calculating the e_i, we clearly must know the distribution first, and I am having a hard time figuring out how we can pin down how this term will be distributed. Any help in the right direction would be appreciated.

EDIT: I believe that I worked out the second part of the problem here:

enter image description here

I also seem to find that, with regards to part (c), White Standard errors will not yield a consistent estimate in this case though I am not sure why. Can anyone provide any intuition about this?

EDIT II: In fact, it seems White Standard Errors will work in this case. Is this because the e_i estimates are independent of one another?


1 Answer 1


A more analytic derivation for $(a)$ would be: $$Pr(d_i=1\mid x_i) = Pr(y_i>0\mid x_i) = Pr(x_i'\beta +u_i>0\mid x_i) = 1 - Pr(x_i'\beta +u_i<0\mid x_i) = 1- Pr(u_i<-x_i'\beta\mid x_i)$$ and since the error term $u_i$ is assumed uniform $U(-0.5,0.5)$, the probability above is the value of the cumulative density function of $U(-0.5,0.5)$ at $-x_i'\beta$, so we have:

$$Pr(d_i=1\mid x_i) = 1-\frac {-x_i'\beta -(-0.5)}{0.5-(-0.5)} = 1 + x_i'\beta -0.5 = x_i'\beta +0.5$$ indeed, assuming of course that $x_i'\beta$ stays inside the support of $u_i$, i.e. that $|x_i'\beta| \leq 0.5$.

For the second requirement, the derivation is straight forward since (verifying the expression)

$$E(d_i\mid x_i) = Pr(d_i = 1\mid x_i) = P_i$$ since $d_i$ is an indicator function. Then we can write $d_i = P_i + e_i$, for which by design, $E(e_i\mid x_i)=0$. Note that this also implies that $E(e_i)=0$, by the law of iterated expectations.

For $(b)$, since $e_i = d_i-P_i$, its conditional variance is

$$\operatorname{Var}(e_i\mid x_i) = E(e_i\mid x_i)^2 = E(d_i-P_i\mid x_i)^2 = E[d_i^2\mid x_i] - 2E[d_iP_i\mid x_i] + E[P_i^2\mid x_i]$$

For indicator functions we have $d_i^2 = d_i$ so

$$\operatorname{Var}(e_i\mid x_i) = E(d_i\mid x_i) -2E[d_iP_i\mid x_i] + E[P_i^2\mid x_i]$$ $$= P_i - 2E[(P_i+e_i)P_i\mid x_i] + P_i^2 $$ (the last transformation because $P_i$ is just a function of $x_i$, so it is "measurable" (deterministic) conditional on $x_i$, and then $$\operatorname{Var}(e_i\mid x_i) = P_i - 2P_i^2 -2 E[e_iP_i\mid x_i] + P_i^2$$

$$= P_i - P_i^2 - 2E[e_i(x_i'\beta +0.5)\mid x_i]$$

$$= P_i(1-P_i) - 2x_i'\beta E(e_i\mid x_i) - E[e_i\mid x_i] = P_i(1-P_i) -0 -0,\;\; i=1,...,n$$

I.e. what you too found is the variance of $e_i$ conditional on $x_i$ (which is what we want anyway). You wrote "$e_i$ takes the value XX with probability YY" -but these probabilities are conditional on the $x_i$ so implicitly you calculated the conditional variance.

For (c) and switching to matrix notation we have that

$$\sqrt {n}(\hat \beta_{OLS} - \beta) = (\frac 1n\mathbf P'\mathbf P)^{-1}\frac 1{\sqrt{n}} \mathbf P'\mathbf e$$

and its asymptotic variance-covariance matrix will be

$$\operatorname{AVar}(\sqrt {n}(\hat \beta_{OLS} - \beta) ) = (E[\mathbf P'\mathbf P])^{-1}E\big(\mathbf P'\mathbf e\mathbf e'\mathbf P\big)(E[\mathbf P'\mathbf P])^{-1}$$

Since the $e_i$'s are independent (because the $x_i$'s I presume are independent also, across $i$), and so they are not-autocorrelated, then

$$E\big(\mathbf P'\mathbf e\mathbf e'\mathbf P\big)$$ is consistently estimated by Whites' heteroskedasticity robust formula, for asymptotically correct standard errors.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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