In case of multilevel data, it is obvious to use for example - mixed -, but using - mixed - it is still possible to use - vce(cluster clustervar) - to cluster standard errors at the higher level: is this in some way useful (since - mixed - already takes the multilvel structure into account)? Why (not)?
Would an ordinary linear regression - regress - with standard errors clustered at the higher level be a valid alternative? I'm just wondering about the differences between these alternatives and reasons to prefer one or the other ...
I have a small example below:
Code:
webuse nlswork eststo clear * (1) Multilevel regression mixed ln_w grade age ttl_exp tenure || id: eststo M_lev * (2) Multilevel regression with clustered standard errors mixed ln_w grade age ttl_exp tenure || id:, vce(cluster id) eststo M_lev_clus * (3) Single level regression with clustered standard errors reg ln_w grade age ttl_exp tenure, vce(cluster id) eststo S_lev_clus * (4) Single level regression (just for comparison) reg ln_w grade age ttl_exp tenure eststo S_lev esttab, mtit
Code:
---------------------------------------------------------------------------- (1) (2) (3) (4) M_lev M_lev_clus S_lev_clus S_lev ---------------------------------------------------------------------------- main grade 0.0741*** 0.0741*** 0.0744*** 0.0744*** (41.23) (35.87) (34.43) (71.46) age -0.00449*** -0.00449*** -0.00526*** -0.00526*** (-6.78) (-5.00) (-5.57) (-9.70) ttl_exp 0.0306*** 0.0306*** 0.0296*** 0.0296*** (26.92) (18.05) (16.08) (30.94) tenure 0.0136*** 0.0136*** 0.0195*** 0.0195*** (15.96) (10.18) (11.96) (22.72) _cons 0.635*** 0.635*** 0.652*** 0.652*** (23.20) (19.37) (19.30) (37.12) ---------------------------------------------------------------------------- lns1_1_1 _cons -1.319*** -1.319*** (-96.79) (-80.67) ---------------------------------------------------------------------------- lnsig_e _cons -1.216*** -1.216*** (-262.93) (-98.91) ---------------------------------------------------------------------------- N 28099 28099 28099 28099 ---------------------------------------------------------------------------- t statistics in parentheses * p<0.05, ** p<0.01, *** p<0.001
Mike
0 Response to Multilevel regression and/or clustering standard errors
Post a Comment