The concept of bivariate normal distibutions is very familiar to even beginning statistics students. Scatter plots and Pearson corelation are tools for examing bivariate normal distributions. Less familiar for some students might be using bivariate response variables in multivariate analyses. In the case of bivariate probit analysis we have two binary response variables that vary jointly. We want to esitmate the coefficients needed to account for this joint distribution.
As you would expect the likelihood function for bivariate probit is more complex than when there is only one esponse variable,
Example 1
use http://www.gseis.ucla.edu/courses/data/schvote, clear probit priv years ptax inc Probit estimates Number of obs = 80 LR chi2(3) = 1.14 Prob > chi2 = 0.7680 Log likelihood = -29.572798 Pseudo R2 = 0.0189 ------------------------------------------------------------------------------ priv | Coef. Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------- years | -.0092045 .023364 -0.39 0.694 -.0549971 .0365882 ptax | -.1427311 .6937362 -0.21 0.837 -1.502429 1.216967 inc | .4313241 .5792655 0.74 0.457 -.7040154 1.566664 _cons | -4.40218 4.938369 -0.89 0.373 -14.08121 5.276846 ------------------------------------------------------------------------------ probit vote years ptax inc Probit estimates Number of obs = 80 LR chi2(3) = 13.62 Prob > chi2 = 0.0035 Log likelihood = -45.576114 Pseudo R2 = 0.1300 ------------------------------------------------------------------------------ vote | Coef. Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------- years | -.0080101 .015893 -0.50 0.614 -.0391598 .0231395 ptax | -2.013629 .7192403 -2.80 0.005 -3.423314 -.6039441 inc | 1.582937 .5671639 2.79 0.005 .4713161 2.694558 _cons | -1.353637 4.411823 -0.31 0.759 -10.00065 7.293378 ------------------------------------------------------------------------------ biprobit priv vote years ptax inc Bivariate probit regression Number of obs = 80 Wald chi2(6) = 11.91 Log likelihood = -74.171253 Prob > chi2 = 0.0640 ------------------------------------------------------------------------------ | Coef. Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------- priv | years | -.0146627 .0264275 -0.55 0.579 -.0664596 .0371342 ptax | -.0923143 .6922562 -0.13 0.894 -1.449112 1.264483 inc | .3644544 .5588324 0.65 0.514 -.7308371 1.459746 _cons | -4.040363 4.872994 -0.83 0.407 -13.59126 5.510529 -------------+---------------------------------------------------------------- vote | years | -.008866 .0159739 -0.56 0.579 -.0401742 .0224422 ptax | -2.054462 .7310168 -2.81 0.005 -3.487229 -.6216959 inc | 1.574388 .5638432 2.79 0.005 .469276 2.679501 _cons | -.9732729 4.487075 -0.22 0.828 -9.767779 7.821233 -------------+---------------------------------------------------------------- /athrho | -.3425239 .2536544 -1.35 0.177 -.8396774 .1546297 -------------+---------------------------------------------------------------- rho | -.3297287 .2260769 -.6856382 .1534089 ------------------------------------------------------------------------------ Likelihood ratio test of rho=0: chi2(1) = 1.95532 Prob > chi2 = 0.1620 test years ( 1) [priv]years = 0.0 ( 2) [vote]years = 0.0 chi2( 2) = 0.69 Prob > chi2 = 0.7079 test ptax ( 1) [priv]ptax = 0.0 ( 2) [vote]ptax = 0.0 chi2( 2) = 8.15 Prob > chi2 = 0.0170 test inc ( 1) [priv]inc = 0.0 ( 2) [vote]inc = 0.0 chi2( 2) = 8.86 Prob > chi2 = 0.0119 mfx compute Marginal effects after biprobit y = Pr(priv=1,vote=1) (predict) = .05187385 ------------------------------------------------------------------------------ variable | dy/dx Std. Err. z P>|z| [ 95% C.I. ] X ---------+-------------------------------------------------------------------- years | -.0019032 .00259 -0.73 0.463 -.006986 .003179 8.77500 ptax | -.110602 .08091 -1.37 0.172 -.269192 .047988 6.93727 inc | .1141173 .06655 1.71 0.086 -.016321 .244556 9.96772 ------------------------------------------------------------------------------Example 2
use http://www.gseis.ucla.edu/courses/data/ms00, clear describe Contains data from http://www.gseis.ucla.edu/courses/data/ms00.dta obs: 200 vars: 7 8 Feb 2001 11:23 size: 6,400 (99.2% of memory free) ------------------------------------------------------------------------------- storage display value variable name type format label variable label ------------------------------------------------------------------------------- id float %9.0g female float %9.0g fl honors float %9.0g enrolled in honors read float %9.0g reading test write float %9.0g writing test nss float %9.0g national science scholar mma float %9.0g mooberry math award ------------------------------------------------------------------------------- summarize Variable | Obs Mean Std. Dev. Min Max -------------+----------------------------------------------------- id | 200 100.5 57.87918 1 200 female | 200 .545 .4992205 0 1 honors | 200 .525 .5006277 0 1 read | 200 52.23 10.25294 28 76 write | 200 52.775 9.478586 31 67 nss | 200 .165 .372112 0 1 mma | 200 .115 .3198225 0 1 tab1 female honors -> tabulation of female female | Freq. Percent Cum. ------------+----------------------------------- male | 91 45.50 45.50 female | 109 54.50 100.00 ------------+----------------------------------- Total | 200 100.00 -> tabulation of honors enrolled in | honors | Freq. Percent Cum. ------------+----------------------------------- 0 | 95 47.50 47.50 1 | 105 52.50 100.00 ------------+----------------------------------- Total | 200 100.00 tabulate nss mma national | science | mooberry math award scholar | 0 1 | Total -----------+----------------------+---------- 0 | 155 12 | 167 1 | 22 11 | 33 -----------+----------------------+---------- Total | 177 23 | 200 biprobit nss mma read write honors female Bivariate probit regression Number of obs = 200 Wald chi2(8) = 56.75 Log likelihood = -105.31311 Prob > chi2 = 0.0000 ------------------------------------------------------------------------------ | Coef. Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------- nss | read | .0493064 .0156631 3.15 0.002 .0186073 .0800055 write | .0679724 .0208136 3.27 0.001 .0271786 .1087663 honors | -.5517298 .2878803 -1.92 0.055 -1.115965 .0125051 female | -.9337984 .2767985 -3.37 0.001 -1.476313 -.3912834 _cons | -6.748716 1.144578 -5.90 0.000 -8.992047 -4.505385 -------------+---------------------------------------------------------------- mma | read | .0525179 .0197319 2.66 0.008 .0138441 .0911917 write | .1091292 .0364161 3.00 0.003 .037755 .1805034 honors | .8246593 .4328874 1.91 0.057 -.0237844 1.673103 female | -.1103348 .31435 -0.35 0.726 -.7264495 .5057799 _cons | -11.20763 2.338403 -4.79 0.000 -15.79081 -6.624443 -------------+---------------------------------------------------------------- /athrho | .3552813 .2288336 1.55 0.121 -.0932244 .8037869 -------------+---------------------------------------------------------------- rho | .3410509 .2022167 -.0929552 .6661485 ------------------------------------------------------------------------------ Likelihood ratio test of rho=0: chi2(1) = 2.52696 Prob > chi2 = 0.1119 test honors ( 1) [nss]honors = 0.0 ( 2) [mma]honors = 0.0 chi2( 2) = 8.31 Prob > chi2 = 0.0157 test female ( 1) [nss]female = 0.0 ( 2) [mma]female = 0.0 chi2( 2) = 11.41 Prob > chi2 = 0.0033 mfx compute Marginal effects after biprobit y = Pr(nss=1,mma=1) (predict) = .00303069 ------------------------------------------------------------------------------ variable | dy/dx Std. Err. z P>|z| [ 95% C.I. ] X ---------+-------------------------------------------------------------------- read | .0005465 .00051 1.07 0.286 -.000457 .00155 52.2300 write | .0010095 .00083 1.22 0.224 -.000616 .002635 52.7750 honors*| .0034291 .00419 0.82 0.413 -.004778 .011636 .525000 female*| -.0043393 .00507 -0.86 0.392 -.01427 .005591 .545000 ------------------------------------------------------------------------------ (*) dy/dx is for discrete change of dummy variable from 0 to 1The ancillary parameter rho measures the correlation of the residuals from the two models. As it turns out, the two equations were not strongly associated, rho = .34, which was not significant (chi-square = 2.53, df = 1, p =.11)
Seemingly Unrelated Bivariate Probit Example
It is also possible to run biprobit as a seemlying unrelated bivariate probit in which each of the equations has different predictors. The equations are not independent since they are computed on the same set of subjects.
biprobit (nss = female write)(mma = read write) Seemingly unrelated bivariate probit Number of obs = 200 Wald chi2(4) = 50.04 Log likelihood = -113.97205 Prob > chi2 = 0.0000 ------------------------------------------------------------------------------ | Coef. Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------- nss | female | -1.016887 .2584353 -3.93 0.000 -1.523411 -.5103629 write | .0860565 .0175865 4.89 0.000 .0515876 .1205253 _cons | -5.273082 .9810768 -5.37 0.000 -7.195957 -3.350206 -------------+---------------------------------------------------------------- mma | read | .057744 .0199113 2.90 0.004 .0187187 .0967694 write | .1078431 .0345801 3.12 0.002 .0400674 .1756189 _cons | -10.86132 2.238518 -4.85 0.000 -15.24873 -6.473901 -------------+---------------------------------------------------------------- /athrho | .2028541 .2053828 0.99 0.323 -.1996887 .6053969 -------------+---------------------------------------------------------------- rho | .2001167 .1971579 -.1970762 .5408787 ------------------------------------------------------------------------------ Likelihood ratio test of rho=0: chi2(1) = .986217 Prob > chi2 = 0.3207 test write ( 1) [nss]write = 0.0 ( 2) [mma]write = 0.0 chi2( 2) = 32.42 Prob > chi2 = 0.0000 test read ( 1) [mma]read = 0.0 chi2( 1) = 8.41 Prob > chi2 = 0.0037 display "chi-square approximation = " 2.90^2 chi-square approximation = 8.41Again it turns out that these two equations are not stongly correlated, rho = .2, which is not statistically significant (chi-squar1 = .99, df = 1, p = .32).
Instrumental Variable Example
biprobit (nss = female mma)(mma = female read write), nolog Seemingly unrelated bivariate probit Number of obs = 200 Wald chi2(5) = 90.74 Log likelihood = -118.5046 Prob > chi2 = 0.0000 ------------------------------------------------------------------------------ | Coef. Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------- nss | female | -.580865 .2214549 -2.62 0.009 -1.014909 -.1468214 mma | 2.115765 .3094269 6.84 0.000 1.509299 2.72223 _cons | -.9611232 .152064 -6.32 0.000 -1.259163 -.6630833 -------------+---------------------------------------------------------------- mma | female | -.222301 .3007045 -0.74 0.460 -.811671 .367069 read | .0675854 .0196951 3.43 0.001 .0289836 .1061871 write | .1076171 .0333168 3.23 0.001 .0423173 .1729168 _cons | -11.19188 2.087091 -5.36 0.000 -15.2825 -7.101252 -------------+---------------------------------------------------------------- /athrho | -1.301447 .7019475 -1.85 0.064 -2.677239 .0743445 -------------+---------------------------------------------------------------- rho | -.8620953 .1802543 -.9905906 .0742078 ------------------------------------------------------------------------------ Likelihood-ratio test of rho=0: chi2(1) = 10.6787 Prob > chi2 = 0.0011
Categorical Data Analysis Course
Phil Ender