Utilisateur
när exogeneity assumption fails, vilket behövs for ols to work. det vetyder att vi inte skulle kunna med hjälp av våra regressors, ta reda på om u är large eller small
var(X)
= 0
om det varit X,aY hade man kunnat flytta ut a och haft a•cov(X,Y)
a^2 • var(X)
if exogen. assu. does not hold, we cant say that there is correlation (cov) between u and x, since there are more assumtions that couldve broken the assumption
se have omitted imoortant variables that have effect on Y and correlates with x1. the effect end up in u and breaks the exog. assu.
by adding it as a control variable to the model
long model controls for X2
short model have omitted variablds in u.
B1 for X1 will not be the same between the long and short model
= B1^ + ( côv(x1,x2) / vâr(x1) ) • B2^
u would contain the total effetct of X2. dvs u = B2x2
over/underestimate with the correct sign
estimating the wrong sign
exog. assu.
large sample
det finns fortfarande kvar trots att vi fått bort estimation error
we have omitted imortant controll variables that are correlated with the variable of interest and that have its own effect (B2) on the outcome variable.
det är två raka linjer med dots, en linje vid 0 och en vid värdet 1 på x-axeln
B1^= B1 + ( E[Y| dummy=1] - E[Y| dummy=0] ) • B2^
för att the two groups (0,1) differ along other dimensions other than just the variable of interest, which can happen if there is a confounding variable
the difference between the true value and the measurement
när vi har ex. ability som vi inte kan samla in data på, men så har vi IQ som reklekterar lite av det vi vill ha reda på. då kan vi använda IQ som en proxy variable och run a regression med den för att fånga lite av ability. så en proxy measures the unobserved ability, but also randomness. we therefore say that we have replaced the regressor with a messurement (proxy)
med en *
E[u| x1*] = 0
u = -B1W+u
det säger att measurenent error exists
vi ser att om W≠0 så har vi ett problem att u inte är lika med varandra och vi har ett endogeneity problem, då u kan be predicted by loooking at x1*. då exogen. assu. ≠0
the edtimated model
1) E[W]=0, so on average x1* messures x1 correctly
2) W is independent of u and x1, so there is no systematic mismeasurement
3) var(W) > 0 which rules out the case where W is always 0. this tells us that measurement error exists.
causal is the long model while the estimated model is short since it includes the prixy variable without accounting and showing that there is a measuremnt error
E[u| x*, w] = 0
E[u| x1+W, W] = 0
we know due to the 2) assumption that W is useless and independent fron x1 and u, so
E[u|x1]=0 so the exogeneity assumption for the two regressors E[u|x1,w]=0 is SATISFIED
B1^ = B1 + ( cov(x1*,w) / var(x1*) ) • B2
where B2 = -B1
= ( var(w) / var(x1) + var(w) ) • B1
det är en scaling factor som minskar the causal effect.
är ett värde mellan 0-1 och kan inte göra så att B1 byter sign.
the effect that we are estimating is less strong than the true causal effect
bias toward zero since we are scaling and moving towards zero
then we will estimate a B^ that is bised toward zero. (underestimate eith the correct sign)