E 31501/4150 Properties of OLS estimators (Monte Carlo Analysis)

Transcription

1 E 31501/4150 Properties of OLS estimators (Monte Carlo Analysis) Ragnar Nymoen 10 February 2011

2 Repeated sampling Section of the HGL book is called Repeated sampling The point is that by drawing repeated samples of size 40 from a large population of households, they repeat the food expenditure regression several times (they do 10 repetitions). The average of the estimates for β 2 (the 10 b 2 values) is 78.8 The interpretation is that if the regression model is correctly specified, then 78.8 is a better estimate than a single one.

3 Monte Carlo analysis I Repeated sampling is costly in practice, but we can do something similar at a low cost: We can use the computer to generate a large number of data sets (call it M), estimate the regression model on all M data sets and average all of the M estimates of β 2. Since the data sets are independent (the computer takes care of that), the average M j=1 b 2 = b 2j M. will come close to the true E(b 2 ) when M is large. This is with reference to the Law of Large Numbers

4 Monte Carlo analysis II The procedure in an example of Monte Carlo analysis. In the lecture we have shown that if the classical assumptions hold, the OLS estimator b 2 is unbiased: E(b 2 ) = β 2 We can use Monte Carlo analysis to check this result. To do that, we instruct the computer to generate data in accordance with the regression model with classical assumptions. We say that the model is the data generating process, and we expect to find that: b 2 M β 2

5 Monte Carlo analysis III We illustrate Monte Carlo methodology in two steps. A manual Monte Carlo, where M is very small. But it still illustrates the idea Only in class last week: The approach is to let Stata or PcGive generate say 4 artificial data sets and average the estimation results for b 2. Already this suggests that unbiasedness does indeed hold (if the assumptions of the regression model holds). Automatic Monte Carlo when M can be as large as we want The following

6 Monte Carlo 1: Deterministic X y i = 10 β1 + 2 β2 x i + ε i ε i N(0, ) homoskedasticity and no disturbance correlation x i : Deterministic N = 100 (maximum sample) but we show graphs for recursive estimation with N = 5, 6,..., 100. Which will then illustrate whether the consistency result is of practical interest or only has theoretical interest-

7 Monte Carlo 2: Stochastic X y i = 10 β1 + 2 β2 x i + ε i ε i N(0, ) homoskedasticity and no disturbance-correlation x i : Stochastic N = 100 (maximum sample)

8 We can compare also numerical results from the two Monte Carlos. The estimate of the true expectation of b 2 : M j=1 E(b 2 ) = b 2j M Results for deterministic and stochastic (model 2) regression model: b 2 bias Monte Carlo Monte Carlo

9 The variance of the OLS estimators in the two experiments (N = 100). Var(b2 ) Var(b 2 ) Stata Monte Carlo Monte Carlo Var(b2 ) is the Monte Carlo estimate of Var(b 2 ). Var(b 2 ) Stata is the estimate we get from Stata, PcGive or another programme

10 In our third experiment we keep x i stochastic and N = 100, but we increase σ 2 from 0.1 to 1, and we reduce sx 2. Call the resulting Monte Carlo 3: bias Var(b2 ) Monte Carlo Monte Carlo Var(b2 ) Stata would again give the correct estimate of the standard error of b 2.

11 Mis-specification I Next consider departures from the classical assumptions. Let us check what happens if The Monte Carlo 4 is cov(ε i, ε j ) 0 y i = 10 β1 + 2 β2 x i + ε i ε i = 0.6ε i 1 + ε i ε i N(0, ) disturbances are correlated x i : Stochastic N = 100 (maximum sample)

12 Mis-specification II In other respects we have the same situation as in Monte Carlo 2. bias Var(b2 ) Var(b 2 ) Stata Monte Carlo Monte Carlo This shows that when ε i and ε i 1 are positively correlated, we have E(b 2 ) = β 2 as in the classical case. We also have consistency. But note that Var(b 2 ) is underestimated! To fix that, can modify the OLS estimator. However we have to leave that for later

13 Mis-specification III Heteroskedasticity leads to similar (limited) problem with the OLS estimator. This shows that correctness of specification is important.