Are Stand Your Ground Laws in Florida Racist?: Some regression evidence
I have previously posted here a long discussion on the simple averages regarding the characteristics of those involved in "Stand Your Ground" cases in Florida. The data from the Tampa Bay Tribune is available here (also here). Yet, there are real limits to using just simple means because that approach assumes that the cases involving blacks and whites are the same. The Tribune has collected a lot of information on everything from the race and gender of the person shot and the shooter to the following questions:
Everything else equal, in cases with only one person killed, killing a black rather than a white increases the defendant's odds of being convicted doubles, though the result is not statistically significant. If you also include multiple murder cases, killing a black increases the chances of conviction even more.
Regression looking at the odds of someone being convicted of murder for those who have killed one person.
xi: logit convicted VictimHispanic VictimWhite VictimBlack VictimMale DefendantHispanic DefendantWhite DefendantBlack DefendantMale DidVictimInitiateConfrontation WastheVictimArmed WasVictimCommittingCrime DidDefendantPursueVictim CouldDefendantRetreat WasDefendantonHisProperty DidSomeoneWitnessAttack WasTherePhysicalEvidence othermurdered casetype_2-casetype_25 year_2006-year_2012 if pending=="Decided" & MurderVictim2sRace =="NA", or robust
Did the victim initiate the confrontation?
Was the victim armed?
Was the victim committing a crime that led to the confrontation?
Did the defendant pursue the victim?
Could the defendant have retreated to avoid the conflict?
Was the defendant on his or her property?
Did someone witness the attack?
Was there physical evidence?
Case type
Alleged Home Invasion
Alleged sexual assault
Argument over love interest
Argument turned violent
Attempted car theft
Attempted home invasion
Attempted robbery
Burglary
Citizen enforcing the law
Dispute over money/property
Domestic argument
Domestic dispute
Drug deal gone bad
Fight at bar/party
Home invasion
Neighborhood dispute
Retaliation
Road Rage
Robbery
Roommate Dispute
Teenage bullying
Trespassing
Unknown
Unprovoked attack
Case yearBefore I lose people by reporting the regressions below, let me provide a brief verbal discussion. There is a simple problem with comparing the mean conviction rates as I have done earlier. Just because two people are charged with murder doesn't mean the two cases are identical. Using the Tribune data, blacks killed in these confrontations were 13 percentage points more likely to be armed than the whites who were killed, thus making it more plausible that their killers reasonably believed that they had little choice but to kill their attacker. By a 43 to 16 percent margin, the blacks killed were also more often committing a crime. Further, there were also more cases with a witness around when a black was killed (69 to 62 percent).
Everything else equal, in cases with only one person killed, killing a black rather than a white increases the defendant's odds of being convicted doubles, though the result is not statistically significant. If you also include multiple murder cases, killing a black increases the chances of conviction even more.
Regression looking at the odds of someone being convicted of murder for those who have killed one person.
xi: logit convicted VictimHispanic VictimWhite VictimBlack VictimMale DefendantHispanic DefendantWhite DefendantBlack DefendantMale DidVictimInitiateConfrontation WastheVictimArmed WasVictimCommittingCrime DidDefendantPursueVictim CouldDefendantRetreat WasDefendantonHisProperty DidSomeoneWitnessAttack WasTherePhysicalEvidence othermurdered casetype_2-casetype_25 year_2006-year_2012 if pending=="Decided" & MurderVictim2sRace =="NA", or robust
Logistic regression Number of obs = 66
Wald chi2(29) = .
Prob > chi2= .
Log pseudolikelihood = -20.7842 Pseudo R2 =0.5408
-------------------------------------------------------
| Robust
convicted | Odds Ratio Std. Err. z P>|z|
-------------+-----------------------------------------
VictimHisp~c | .0009022 .002332 -2.71 0.007
VictimWhite | .4247123 .9166847 -0.40 0.692
VictimBlack | 1.174415 4.496167 0.04 0.967
DefendantW~e | 34.6601 89.92937 1.37 0.172
DefendantB~k | 4.915077 12.41981 0.63 0.529
DefendantM~e | .340511 .5529446 -0.66 0.507
DidVictimI~n | .0137108 .0348234 -1.69 0.091
WastheVict~d | .0721759 .2389135 -0.79 0.427
WasVictimC~e | 3.043378 12.33578 0.27 0.784
DidDefenda~m | 1.635232 3.278233 0.25 0.806
CouldDefen~t | 1.475613 2.438766 0.24 0.814
WasDefenda~y | 4.778653 5.087016 1.47 0.142
DidSomeone~k | 22.62614 40.55751 1.74 0.082
WasTherePh~e | .2503723 .2216539 -1.56 0.118
casetype_3 | 7.49e+08 1.75e+09 8.77 0.000
casetype_4 | 8.24e+08 2.21e+09 7.63 0.000
casetype_8 | 1.74e+10 3.81e+10 10.76 0.000
casetype_9 | 2.10e+09 5.48e+09 8.22 0.000
casetype_10 | 1.60e+09 2.58e+09 13.13 0.000
casetype_12 | 1.84e+09 5.58e+09 7.04 0.000
casetype_13 | 1.08e+12 3.20e+12 9.37 0.000
casetype_14 | 1.46e+09 4.26e+09 7.24 0.000
casetype_15 | 4.84e+08 1.89e+09 5.12 0.000
casetype_17 | 1.11e+08 3.26e+08 6.34 0.000
casetype_25 | 2.62e+08 . . .
year_2006 | .6355305 1.720501 -0.17 0.867
year_2007 | .0931599 .4768633 -0.46 0.643
year_2008 | .0573326 .2092783 -0.78 0.434
year_2009 | 1.008875 2.457142 0.00 0.997
year_2010 | 63.62403 200.2368 1.32 0.187
------------------------------------------------------
Note: 1 failure and 0 successes completely determined.
. test VictimWhite=VictimBlack
( 1) VictimWhite - VictimBlack = 0
chi2( 1) = 0.04
Prob > chi2 = 0.8386
. test VictimHispanic=VictimBlack
( 1) VictimHispanic - VictimBlack = 0
chi2( 1) = 2.55
Prob > chi2 = 0.1100
. test VictimHispanic=VictimWhite
( 1) VictimHispanic - VictimWhite = 0
chi2( 1) = 4.91
Prob > chi2 = 0.0267
. test DefendantWhite=DefendantBlack
( 1) DefendantWhite - DefendantBlack = 0
chi2( 1) = 0.23
Prob > chi2 = 0.6316
. test VictimWhite=VictimBlack
( 1) VictimWhite - VictimBlack = 0
chi2( 1) = 0.04
Prob > chi2 = 0.8386
. test VictimHispanic=VictimBlack
( 1) VictimHispanic - VictimBlack = 0
chi2( 1) = 2.55
Prob > chi2 = 0.1100
. test VictimHispanic=VictimWhite
( 1) VictimHispanic - VictimWhite = 0
chi2( 1) = 4.91
Prob > chi2 = 0.0267
. test DefendantWhite=DefendantBlack
( 1) DefendantWhite - DefendantBlack = 0
chi2( 1) = 0.23
Prob > chi2 = 0.6316
Regression looking at the odds of someone being convicted of murder for those who have killed one or more people.
. xi: logit convicted VictimHispanic VictimWhite VictimBlack VictimMale DefendantHispanic DefendantWhite DefendantBlack DefendantMale DidVictimInitiateConfrontation WastheVictimArmed WasVictimCommittingCrime DidDefendantPursueVictim CouldDefendantRetreat WasDefendantonHisProperty DidSomeoneWitnessAttack WasTherePhysicalEvidence othermurdered casetype_2-casetype_25 year_2006-year_2012 if pending=="Decided", or robust
Logistic regression Number of obs = 78
Wald chi2(32) = .
Prob > chi2= .
Log pseudolikelihood = -22.785937 Pseudo R2 =0.5735
-------------------------------------------------------
| Robust
convicted | Odds Ratio Std. Err. z P>|z|
-------------+-----------------------------------------
VictimHisp~c | .0000949 .0003103 -2.83 0.005
VictimWhite | .238639 .4879525 -0.70 0.483
VictimBlack | 3.390464 9.382387 0.44 0.659
DefendantH~c | 5.55e-13 1.35e-12 -11.61 0.000
DefendantW~e | 7.55e-11 2.26e-10 -7.78 0.000
DefendantB~k | 1.91e-12 . . .
DefendantM~e | .2819811 .5277879 -0.68 0.499
DidVictimI~n | .0078562 .0144318 -2.64 0.008
WastheVict~d | .0895871 .2060086 -1.05 0.294
WasVictimC~e | 2.951656 9.628308 0.33 0.740
DidDefenda~m | 1.935009 3.692359 0.35 0.729
CouldDefen~t | 1.207219 1.75638 0.13 0.897
WasDefenda~y | 3.68262 2.776331 1.73 0.084
DidSomeone~k | 34.60143 52.71921 2.33 0.020
WasTherePh~e | .236634 .2656798 -1.28 0.199
othermurde~d | 54.95588 119.1862 1.85 0.065
casetype_3 | 240.5917 643.6653 2.05 0.040
casetype_4 | 71.61738 152.6067 2.00 0.045
casetype_8 | 4369.197 16026.35 2.29 0.022
casetype_9 | 1132.737 3854.253 2.07 0.039
casetype_10 | 183.0676 402.9866 2.37 0.018
casetype_12 | 468.6694 1215.575 2.37 0.018
casetype_13 | 553160.6 2506482 2.92 0.004
casetype_14 | 1170.289 3029.217 2.73 0.006
casetype_15 | 84.6564 416.3267 0.90 0.367
casetype_17 | 24.15446 60.33759 1.27 0.202
casetype_25 | 37.81938 87.88588 1.56 0.118
year_2006 | .1661872 .3844092 -0.78 0.438
year_2007 | .0113472 .0417041 -1.22 0.223
year_2008 | .0095219 .0326906 -1.36 0.175
year_2009 | .3936484 .9631961 -0.38 0.703
year_2010 | 44.73127 123.881 1.37 0.170
year_2011 | .0005799 .001551 -2.79 0.005
-----------------------------------------------------
Note: 0 failures and 1 success completely determined.
. test VictimWhite=VictimBlack
( 1) VictimWhite - VictimBlack = 0
chi2( 1) = 0.57
Prob > chi2 = 0.4505
. test VictimHispanic=VictimBlack
( 1) VictimHispanic - VictimBlack = 0
chi2( 1) = 6.13
Prob > chi2 = 0.0133
. test VictimHispanic=VictimWhite
( 1) VictimHispanic - VictimWhite = 0
chi2( 1) = 6.41
Prob > chi2 = 0.0113
. test DefendantWhite=DefendantBlack
( 1) DefendantWhite - DefendantBlack = 0
chi2( 1) = 1.51
Prob > chi2 = 0.2198
. test DefendantWhite=DefendantHispanic
( 1) - DefendantHispanic + DefendantWhite = 0
chi2( 1) = 6.22
Prob > chi2 = 0.0127
. test DefendantBlack=DefendantHispanic
( 1) - DefendantHispanic + DefendantBlack = 0
chi2( 1) = 0.26
Prob > chi2 = 0.6105
I have tried other specifications, but there is no evidence that black and white defendants or black and white victims are treated differently. For example, here is the simplest specification with just the victim's race and gender and defendant's race and gender as well as the number of people murdered.
. xi: logit convicted VictimHispanic VictimWhite VictimBlack VictimMale DefendantHispanic DefendantWhite DefendantBlack DefendantMale othermurdered if pending=="Decided", or robust
Iteration 0: log pseudolikelihood = -71.958988
Iteration 1: log pseudolikelihood = -65.974128
Iteration 2: log pseudolikelihood = -65.940597
Iteration 3: log pseudolikelihood = -65.940546
Iteration 4: log pseudolikelihood = -65.940546
Logistic regression Number of obs = 111
Wald chi2(9) = 8.59
Prob > chi2 = 0.4762
Log pseudolikelihood = -65.940546 Pseudo R2 = 0.0836
----------------------------------------------------
| Robust
convicted | Odds Ratio Std. Err. z P>|z|
-------------+----------------------------------------
VictimHisp~c | .3791321 .580078 -0.63 0.526
VictimWhite | 1.009958 1.446379 0.01 0.994
VictimBlack | .4897925 .7041905 -0.50 0.620
VictimMale | .1134137 .1374925 -1.80 0.073
DefendantH~c | 1.415624 2.026775 0.24 0.808
DefendantW~e | 1.587212 1.956111 0.37 0.708
DefendantB~k | 2.120857 2.711162 0.59 0.556
DefendantM~e | .7934841 .4977044 -0.37 0.712
othermurde~d | 6.797993 7.429584 1.75 0.079
-----------------------------------------------------
. test VictimWhite=VictimBlack
( 1) VictimWhite - VictimBlack = 0
chi2( 1) = 1.93
Prob > chi2 = 0.1650
. test VictimHispanic=VictimBlack
( 1) VictimHispanic - VictimBlack = 0
chi2( 1) = 0.10
Prob > chi2 = 0.7566
. test VictimHispanic=VictimWhite
( 1) VictimHispanic - VictimWhite = 0
chi2( 1) = 1.41
Prob > chi2 = 0.2353
. test DefendantWhite=DefendantBlack
( 1) DefendantWhite - DefendantBlack = 0
chi2( 1) = 0.28
Prob > chi2 = 0.5939
. test DefendantWhite=DefendantHispanic
( 1) - DefendantHispanic + DefendantWhite = 0
chi2( 1) = 0.02
Prob > chi2 = 0.8918
. test DefendantBlack=DefendantHispanic
( 1) - DefendantHispanic + DefendantBlack = 0
chi2( 1) = 0.23
Prob > chi2 = 0.6324
I suspect that there are real biases in how this data is collected. An obvious example is how the Tampa Bay Tribune classified the Zimmerman case.
For example, many would strongly disagree with the newspaper's contention that Martin did not initiate the confrontation, that Zimmerman was pursuing Martin at the time of their confrontation, and that Zimmerman could have retreated to avoid the conflict. The point here is that even using the data with the obvious liberal bias in terms of how this data was entered, the results do not support the claims of bias against blacks.
I have tried other specifications, but there is no evidence that black and white defendants or black and white victims are treated differently. For example, here is the simplest specification with just the victim's race and gender and defendant's race and gender as well as the number of people murdered.
. xi: logit convicted VictimHispanic VictimWhite VictimBlack VictimMale DefendantHispanic DefendantWhite DefendantBlack DefendantMale othermurdered if pending=="Decided", or robust
Iteration 0: log pseudolikelihood = -71.958988
Iteration 1: log pseudolikelihood = -65.974128
Iteration 2: log pseudolikelihood = -65.940597
Iteration 3: log pseudolikelihood = -65.940546
Iteration 4: log pseudolikelihood = -65.940546
Logistic regression Number of obs = 111
Wald chi2(9) = 8.59
Prob > chi2 = 0.4762
Log pseudolikelihood = -65.940546 Pseudo R2 = 0.0836
----------------------------------------------------
| Robust
convicted | Odds Ratio Std. Err. z P>|z|
-------------+----------------------------------------
VictimHisp~c | .3791321 .580078 -0.63 0.526
VictimWhite | 1.009958 1.446379 0.01 0.994
VictimBlack | .4897925 .7041905 -0.50 0.620
VictimMale | .1134137 .1374925 -1.80 0.073
DefendantH~c | 1.415624 2.026775 0.24 0.808
DefendantW~e | 1.587212 1.956111 0.37 0.708
DefendantB~k | 2.120857 2.711162 0.59 0.556
DefendantM~e | .7934841 .4977044 -0.37 0.712
othermurde~d | 6.797993 7.429584 1.75 0.079
-----------------------------------------------------
. test VictimWhite=VictimBlack
( 1) VictimWhite - VictimBlack = 0
chi2( 1) = 1.93
Prob > chi2 = 0.1650
. test VictimHispanic=VictimBlack
( 1) VictimHispanic - VictimBlack = 0
chi2( 1) = 0.10
Prob > chi2 = 0.7566
. test VictimHispanic=VictimWhite
( 1) VictimHispanic - VictimWhite = 0
chi2( 1) = 1.41
Prob > chi2 = 0.2353
. test DefendantWhite=DefendantBlack
( 1) DefendantWhite - DefendantBlack = 0
chi2( 1) = 0.28
Prob > chi2 = 0.5939
. test DefendantWhite=DefendantHispanic
( 1) - DefendantHispanic + DefendantWhite = 0
chi2( 1) = 0.02
Prob > chi2 = 0.8918
. test DefendantBlack=DefendantHispanic
( 1) - DefendantHispanic + DefendantBlack = 0
chi2( 1) = 0.23
Prob > chi2 = 0.6324
I suspect that there are real biases in how this data is collected. An obvious example is how the Tampa Bay Tribune classified the Zimmerman case.
For example, many would strongly disagree with the newspaper's contention that Martin did not initiate the confrontation, that Zimmerman was pursuing Martin at the time of their confrontation, and that Zimmerman could have retreated to avoid the conflict. The point here is that even using the data with the obvious liberal bias in terms of how this data was entered, the results do not support the claims of bias against blacks.
Labels: CastleLaw, george zimmerman
posted by John Lott at 2:23 PM
3 Comments:
So, if the victim is black rather than white, the shooter has double or more chance of conviction. Wondering if that differs whether the shooter was balck or white?
Seriously? 66 observations with 30 variables to explain the data? Did you even bother reading a book on statistics before posting this? Anyone who reads this should be aware that there is no statistical foundation for the claims in this post.
(I have little doubt that the blog author will delete this comment, but that is conservative bias for you. Seriously, take a statistics class.)
Dear MaverickNH:
The regression simultaneously accounts for both the race of the person shot and the person doing the shooting, so the answer to your question is no.
Dear Stephanie:
There were 112 observations. Since this is a logit regression, the observations that are perfectly predicted are dropped from the regression. I showed three different specifications. The last specification accounted for nine variables and you can see the number of observations shown there. If anything the probability for conviction is greater when killing a white person. I also tried other specifications, but the results are the same. The ones reported seemed like the most obvious ones to report. The point is that whether you use all the variables provided by the Tampa Bay Tribune or just the race and gender info, the estimates show that those who kill whites have a higher probability of conviction. The results are not statistically significant, but the coefficients are the opposite of what people are claiming. Given that so many are claiming the opposite is "true," I thought that these results were of interest. That was my point, not that the results were significant. If you want statistically significant results (and some are significant), look at the comparison between Hispanics and others. The sample size wasn't so small that it prevented those results from being significant.
I have taught econometrics and statistics at places such as the University of Chicago. Please tell me exactly what I said that was wrong or misleading.
Post a Comment
<< Home