How ProPublica Analyzed Pardon Data
Click here to download a PDF of this methodology.
Collecting the data
ProPublica's project on presidential pardons relied on data about individuals who were denied and granted pardons during the George W. Bush administration. As a matter of practice, his advisers said, President Bush relied almost exclusively on recommendations from the Office of the Pardon Attorney inside the Justice Department. The data thus provide an opportunity to assess the office's impact on final pardon decisions.
Through a Freedom of Information Act request, ProPublica obtained the names of petitioners who were denied pardons during Bush's two terms -- 1,729 individuals. The names of 189 petitioners who received pardons came from the pardon office website. ProPublica pulled a random sample of 500 names from the combined list. Because some could not be found, the final sample numbered 494. For each of these individuals, the White House followed the pardon office's recommendation to grant or deny a pardon.
For all those in the sample, ProPublica staff and freelance researchers searched public records or made phone contacts to gather demographic data such as age, race, gender and marital status. Researchers gathered sentencing information from federal court records and FOIA requests, and searched other records to collect data on bankruptcies, liens, financial judgments and any additional criminal history.
Although we attempted to achieve complete background checks, it is possible that some petitioners had committed crimes that were too old to be in electronic databases, had records expunged or committed crimes in areas where there is no public access to criminal records.
In response to a separate Freedom of Information Act request, the Justice Department released correspondence from 196 members of Congress to the Office of the Pardon Attorney during Bush's two terms. This allowed reporters to compare pardon outcomes with congressional contacts and political contributions data.
The federal pardons process
According to the Office of the Pardon Attorney, pardons are granted "on the basis of the petitioner's demonstrated good conduct for a substantial period of time after conviction and service of sentence."
A petitioner must wait five years after conviction or release from confinement (whichever is later) before applying. First among factors the office considers is "post-conviction conduct, character and reputation," which is defined as the "ability to lead a responsible and productive life for a significant period after conviction or release."
The pardon office website says examinations focus on "financial and employment stability, responsibility toward family, reputation in the community, participation in community service, charitable or other meritorious activities and, if applicable, military record."
Interviews with current and former Justice Department lawyers established that important factors include whether the applicant was married and financially stable. As a proxy for financial stability, we used bankruptcies and liens. Other factors the office considers important include the type of offense and how recently it occurred, both of which could be determined from court and prison records.
The office also weighs factors that are difficult to quantify. These include the extent to which a petitioner has accepted responsibility for his or her criminal conduct and whether a pardon is needed, either for a job or another reason. A complete description of pardon standards can be found at the Office of the Pardon Attorney's website.
Our statistical analysis
We conducted a binary logistic regression. The dependent variable was whether the petitioner received a pardon. We tested all other available variables against the outcome. In the end, variables that we included in the analysis were:
- race;
- number of years from sentencing to petition;
- offense;
- sentence;
- gender;
- marital status;
- whether the petitioner had a bankruptcy;
- whether the petitioner had a lien or judgment filed against him or her;
- whether the crime was committed while the petitioner was in the military;
- whether an elected official sent a letter to the pardon attorney on the petitioner's behalf.
In raw percentages, our analysis found that 12 percent of white petitioners and 10 percent of Hispanics were pardoned. No African-Americans in our sample received a pardon.
Other variables that showed a positive relationship with being pardoned were:
- Time since sentence — As the length of time between petition and sentence increased, so did likelihood of being pardoned.
- Sentence — Those who received probation instead of prison time were likelier to be pardoned.
- Marital status — Married petitioners where were likelier to be pardoned.
- Congressional letter — Petitioners with a letter from a member of Congress to the pardon office were likelier to be pardoned.
But accounting for the effects of these variables did not eliminate the strong influence of race on getting a pardon. After testing all available variables, we found that whites were still nearly four times as likely to be pardoned as minorities overall.
We were unable to determine the race for 20 people in our sample. None of the individuals in that subgroup was granted a pardon, so we created "what if" variables — "what if white" and "what if black" — to test the impact of race. Neither test significantly changed the finding that whites had far greater odds of getting a pardon.
Hispanic petitioners are classified as white in most federal court and prison records. To identify Hispanics in our sample, we relied on Hispanic surname or race as designated in other public records.
Several experts reviewed our methodology and findings: George Woodworth, emeritus professor of statistics and actuarial science at the University of Iowa; Jack Glaser, associate professor of public policy at the University of California, Berkeley; Mary Rose, associate professor at the University of Texas at Austin School of Law; and Richard Rosenfeld, professor of criminology at the University of Missouri-St. Louis.
Following are descriptive statistics and the result of our logistic regression:
Descriptives
Marital Status
Total | Granted | % Granted | |
---|---|---|---|
Married | 329 | 39 | 12% |
Not married | 149 | 8 | 5% |
Sentence
Total | Granted | % Granted | |
---|---|---|---|
Prison | 288 | 21 | 7% |
Probation only | 186 | 23 | 12% |
Subsequent crime found for petitioner
Total | Granted | % Granted | |
---|---|---|---|
Subsequent crime | 119 | 7 | 6% |
No subsequent crime found | 367 | 40 | 11% |
Correspondence sent on petitioner's behalf
Total | Granted | % Granted | |
---|---|---|---|
Correspondence | 44 | 9 | 20% |
No correspondence | 450 | 38 | 8% |
Offense
Total | Granted | % Granted | |
---|---|---|---|
Financial crime | 99 | 10 | 10% |
Theft/larceny | 35 | 7 | 20% |
Other | 71 | 5 | 7% |
Fraud | 135 | 8 | 6% |
Drug-related | 114 | 11 | 10% |
Weapons-related | 39 | 2 | 5% |
Violent crime | 20 | 1 | 5% |
Individuals may have committed multiple crimes or have crime
Financial problems
Total | Granted | % Granted | |
---|---|---|---|
Bankruptcy | 92 | 3 | 3% |
Lien or judgment | 180 | 8 | 4% |
Years since sentencing
Total | Granted | % Granted | |
---|---|---|---|
20+ | 91 | 23 | 25% |
10 to 19 | 191 | 19 | 10% |
Less 10 | 200 | 4 | 2% |
Unknown* | 12 | 1 | 8% |
* Small sample, interpret with caution.
Regression variables
Nagelkerke pseudo R square: .29
Hosmer and Lemeshow Goodness of fit: p=.42
Variable | B | S.E. | Wald | Sig. | Exp(B) | Reference category |
---|---|---|---|---|---|---|
Non-Hispanic White | 1.31 | 0.58 | 5.18 | 0.02 | 3.71 | All minorities |
Probation only | 0.81 | 0.38 | 4.51 | 0.03 | 2.25 | |
Military-related crime | 1.05 | 0.77 | 1.85 | 0.17 | 2.84 | |
Female | 0.77 | 0.52 | 2.24 | 0.13 | 2.17 | |
No subsequent crimes found | 0.49 | 0.51 | 0.92 | 0.34 | 1.63 | |
Correspondence written on petitioner's behalf | 1.14 | 0.48 | 5.74 | 0.02 | 3.12 | |
Married | 0.73 | 0.46 | 2.51 | 0.11 | 2.08 | |
No bankruptcy found | 1.06 | 0.67 | 2.52 | 0.11 | 2.89 | |
No lien or judgement found | 0.88 | 0.43 | 4.11 | 0.04 | 2.4 | |
Crimes | ||||||
Financial crime | -0.19 | 0.52 | 0.14 | 0.71 | 0.82 | |
Drug-related | -0.14 | 0.58 | 0.06 | 0.81 | 0.87 | |
Gambling/racketeering | -0.91 | 1.17 | 0.61 | 0.44 | 0.4 | |
Weapons-related | -1.34 | 0.89 | 2.28 | 0.13 | 0.26 | |
Theft/larceny | -0.04 | 0.66 | 0 | 0.96 | 0.96 | |
Fraud | -0.91 | 0.53 | 2.95 | 0.09 | 0.4 | |
Other | -0.83 | 0.64 | 1.69 | 0.19 | 0.43 | Violent crimes/threat |
20 years or more since sentencing | 1.76 | 0.37 | 22.15 | 0 | 5.82 | |
Constant | -6.55 | 1.1 | 35.47 | 0 | 0 |
Shading indicates significant at p<.05
Explanation of variables
Subsequent crime: Using public records, internal documents, interviews and news stories, we determined whether applicants broke the law — state or federal — again after the crime for which they sought a pardon. Those with a subsequent crime were less likely to be granted a pardon.
Bankruptcies/liens: Using public records, internal documents, interviews and news stories, we determined whether petitioners had filed for bankruptcy or had a lien filed against them. These applicants were less likely to be granted a pardon.
Letter/correspondence from elected official: We obtained congressional correspondence to/from the pardon office through a FOIA request. In some cases, members of Congress wrote letters supporting the petitioner's pardon. In other cases, the member forwarded the petitioner's application. We found that petitioners for whom correspondence was submitted were more likely to be pardoned.
Hispanic: Federal Bureau of Prisons records classify anyone who is Hispanic as white. To identify Hispanics in our sample, we used other public records and interviews. We also classified individuals with Hispanic surnames as Hispanic. Although not statistically significant in the regression, these individuals had a higher rate of pardons than other minorities.
Years from sentence to petition: In most cases in our sample, we know the sentencing date. We did not have offense dates. The number of years between the sentence and the pardon petition was used to determine how long ago the crime occurred, which officials say is one factor that is considered in a pardon application. As time from the sentencing increased, so did the likelihood of a pardon.
Age at petition: This is the estimated age of the petitioner at the time he or she applied for a pardon. Because some petitions take several years for review, it may not be the age of the petitioner at the time a pardon decision was made.
Sentence: Many petitioners received both prison and probation. We found those that received only probation had a higher rate of pardons. We also tested whether the length of the sentence — both prison and probation — had an effect on the grant rate, and did not find a significant pattern.
![Presidential Pardons](https://webarchive.library.unt.edu/web/20141030154547im_/http://www.propublica.org/images/uploads/series/gt_presidential_seal_pen_70x70_111201.jpg)
Presidential Pardons: Shades of Mercy
White criminals seeking presidential pardons are nearly four times as likely to succeed as people of color, a ProPublica examination has found.
Latest Stories in this Project
- Three Things Obama's New Clemency Initiative Doesn't Do
- President Obama Tells Clarence Aaron He Can Finally Go Home
- The Sweeping Presidential Power to Help Prisoners That Holder Didn’t Mention
- Clarence Aaron Still Waiting for Clemency, Months after Report Found Pardon Atty Misrepresented Case
- Despite New Pardons, Obama’s Clemency Rate is Still Lowest in Recent History
7 comments
A. Carruthers
Dec. 4, 2011, 12:37 p.m.
“Sentence: Many petitioners received both prison and probation.”
Isn’t incarceration and probation for a single conviction mutually exclusive by definition?
Are you confusing probation with parole? That would seem pretty basic. Or are you referring to different sentences for multiple counts or multiple offenses?
You say you conducted only a bi-variate analysis, but the text indicates a multi-variate regression model.
Also, what about multi-collinearity? I’d say that race might just be (hugely) correlated with marital status, getting a letter from your member of Congress, prior and subsequent crime (by type)—-even bankruptcy (e.g., foreclosures).
What’s your R-squared value?
George Gooding
Dec. 4, 2011, 5:16 p.m.
“But accounting for the effects of these variables did not eliminate the strong influence of race on getting a pardon. After testing all available variables, we found that whites were still nearly four times as likely to be pardoned as minorities overall.”
How can that be the case when Hispanics and Asians ended up with nearly the same pardon rate as whites? Where’s the statistical information pertaining to whether or not race was still a factor after everything else was controlled for? This whole report seems very vague on this point - which is the most important point for the entire angle you’re going with.
Also, your “Contrasting Colors, Contrasting Results” page is laughable. In every case you cite blacks against whites, there’s a glaring difference between them that makes it obvious why the black person was not given a pardon. In most of the “comparisons”, the situations aren’t even remotely comparable.
Take the last column from here:
http://www.propublica.org/special/contrasting-colors-contrasting-results
You compare a black man who did not disclose a subsequent crime he was charged with, with random white people who committed completely different crimes - and did not leave something out of their application. When the policy is that you automatically don’t get pardoned if you file an incomplete application, I can’t see how these cases are being compared - at all.
With lacking statistics for the most important point, and contrived comparisons, it seems like there’s very little to this whole analysis. How about numbers on how many of the petitioners were black? How about cross-tabs with race and the other factors combined? Wouldn’t that be a bit stronger in showing this supposed finding?
Michael Greene
Dec. 5, 2011, 1:23 p.m.
This is a very good study, with a few statistical limitations.
1. Small sample pardon component of the sample. The most serious limitation is the small sample size of the pardon component of the sample. Looking at the descriptives, a change up or down of 1 or 2 in the number of granteds might change a logistic regression coefficient (B or exp(B), the odds ratio). The sample is overly dominated with denied pardons. A better sampling strategy would have been to select all 189 pardoned cases and then randomly select from the denied cases. This is in the nature of a case-control study and can be analyzed with logistic regression.
2. Too many categories for offense with a small sample size. Another example of how the small sample of pardoned individuals affects the analysis is the absence of statistically significant results for offense. It is broken down into too many categories for the associated sample sizes.
3. Additional descriptive for other variables should be included. It would have been nice to see descriptive for the other variables used in the logistic regression such as gender, military related crime, race,
4. Treatment of missing data. In addition to race, there seems to have been missing data for some other variables. The discussion of “what if white” testing for race was a good idea, but not as useful as multiple imputation of missing data for race or other variables. The methodology does not state what was done for missing data in other variables, e.g. from the bivariate tables above: 16 missing marital status (494 - 329-149=16), 8 missing subsequent crimes, etc.
Gary Griffin
Dec. 6, 2011, 1:45 a.m.
Was there any data on Jewish pardons? Jewish people are said to have the wealthiest demographic and give reportedly 60 percent of the funding to the DNC and another 15 to 20 percent to the RNC. Jews are disproportionately represented in Congress. They roam the corridors of power on K-street as can be seen in the AIPAC parade every year. It would be interesting to see the numbers on the powerful people that we aren’t supposed to talk about.
Also, if Jews are disproportionally represented that could also skew the numbers for gentile whites inadvertantly. Like for example if you take out the Jews at Harvard the white population is cut in half.
If you consider the sizable wealth and influence of the tribe it would not be surprising to see the disproportionately represented in both Presidential pardons and commutations.
Rosemary Peternel
Dec. 6, 2011, 11:52 a.m.
Has there been a similar analysis of the patterns of racial discrimination and Congressional letters of support and the incidence of commutation of federal sentences?
Tom Reiland
Dec. 6, 2011, 4:06 p.m.
It needs to be made clearer in the discussion whether we are discussing probability or odds. They are NOT the same.
Logistic regression deals with odds ratios p/(1-p) where p is the probability.
The coefficient for “non-Hispanic White” reported above is 1.31. This means that the odds ratio for non-Hispanic Whites is e^1.31 = 3.71 times higher than for other racial classes (all other variables unchanged). I assume this is where the “nearly four” originates in the story’s headlines.
This need not mean that the probabilities differ by very much.
For example, if the odds ratio for event A is 24 (that is, 24 to 1), and the odds ratio for event B is 96 (that is 96 to 1), the odds ratio for event B is 4 times that of event A.
However,
p(event A) = 24/25 = .96
p(event B) = 96/97 = .9897.
Fenit Nirappil
Dec. 12, 2011, 12:15 a.m.
I’m curious to hear more about how the race of applicants were determined. This alludes to public records and interviews-but I’m guessing not all 500 applicants were reached and it seems like there are issues with public records and identifying race. For example, how the prison records indicate Hispanic as white. How did the analysis take people of mixed race into account?
Commenting on this story is closed.