Do we punish high income criminals too heavily?

Lott, John R., Jr.

I. INTRODUCTION A large literature exists on the purported "injustice" involved when wealthy criminals face a lower probability of conviction than do the less well-to-do for committing the same crimes.(1) Some critics have pointed to this as proof that the legal system is not efficient.(2) Others have argued that when the opportunity cost of imprisonment is accounted for, allowing accused individuals to buy legal services results in the probability of conviction varying inversely with the opportunity cost of the criminal and can produce the optimum level of deterrence.(3) Yet, even this does not adequately state the case for why individuals facing higher opportunity costs should face a lower probability of conviction, since that analysis focuses solely on the formal penalties imposed by the courts--time spent in prison, fines, and legal costs. I will show that for many of those convicted, especially those with the highest presentence incomes, the largest penalty takes the form of reduced legitimate earnings once they return to the labor force. Although the data only provides information for one year of postconviction earnings, just that single year's reduction in earnings often dramatically outweighs all other monetary penalties combined. Personal income may decline either because of a lost reputation for honesty and integrity or because of lost human capital while in prison. All else equal, if highly paid individuals face the greatest reduction in postconviction earnings, they should also face the lowest probability of conviction if all criminals are to face the same expected penalty from committing a crime. Obviously, to the extent that a criminal conviction depresses legitimate income for longer than a single year, my findings underestimate the size of the reputational penalty. The next section of the paper briefly reviews some arguments for why an individual's income is likely to decline with a criminal conviction. Section III provides empirical evidence on the size of the reduction in legitimate earnings due to conviction and discusses why these estimates probably underestimate the true reduction in earnings. Section IV examines how fines, restitution, and the income forgone while in prison vary with the characteristics of the criminal and then discusses how the likelihood of conviction must vary with income if total expected monetary penalties are to be equated across criminals. II. THE NATURE OF INDIVIDUAL REPUTATIONS Either violating an agreement with an employer or breaking the law is expected to reduce a person's earnings. When a wage premium is paid for honesty, the loss of one's reputation and the consequent threatened loss of quasi-rents serves as a deterrent to cheating on agreements. This explanation, in the context of individual reputations and the threatened loss of wage premiums in labor markets, has been discussed by economists since Adam Smith [(1776) 1976].(4) However, last-period problems exist in labor markets, greatly limiting the power of sunk investments to guarantee performance. Another hypothesis, advanced by Lott [1987a], Frank [1987], and Lott and Reed [1989], avoids the last-period problem by stressing the differing preferences workers have for honesty. Workers who value honesty will find it difficult to cheat on their employers because it lowers their own level of utility, and so the possibility exists that employers who can sort workers by of their utility functions can protect themselves from cheating.(5) More reliable individuals would receive higher wages simply because honest individuals are a scarce resource. A criminal conviction provides potential employers with information on a person's preferences. The more convictions, the more certain employers can be that the person was not falsely accused. The seriousness and type of crime also should provide information on the criminal's preferences. Once a person reveals that he does not have the desired preference for honesty, employers would no longer be inclined to offer him the premium wage.(6) People's incomes can also fall after a criminal conviction because of debarment: lawyers loose their licenses, executives in defense firms are forbidden to work in the defense industry, and stockbrokers are banned from working in the securities industry.(7) Not only do people forfeit existing licenses because of a conviction, but they will lose the opportunity to enter other occupations for which licensing is a prerequisite. Even an acquittal or a pardon does not preclude refusal or revocation of a license (Grant et al. [1970, 1007]). Finally, an individual's legitimate human capital is likely to depreciate while he is in prison. Presumably, those facing the largest losses in postconviction income will also tend to have the highest income before conviction. Debarment is not normally associated with low income individuals, and the premium paid for honesty seems probably to increase with the damage that a person can do to the firm--with the highest salaried individuals presumably being able to do the greatest damage. However, optimal penalty theory states that whenever two people are guilty of identical crimes, face the same probability of conviction, and have the same supply elasticities for offenses, they should be punished with the same total penalty. (Added complications will be discussed later, such as when criminals face systematically different costs of being punished.)(8) Since not only fines and the traditional opportunity costs of imprisonment but also the reduction in legitimate earnings impose a penalty on criminals, efficiency implies that the penalties from forgone postrelease earnings take the form of a lump-sum tax. If wealthy individuals suffer a larger reduction in income and greater forgone opportunity costs from imprisonment, optimal penalty theory predicts that they face a lower probability of punishment. This difference in probabilities is further increased if the fines are levied so that they also vary positively with income. Finally, these declines in legitimate earnings can arise because a convicted criminal with a large stock of wealth may find that his wage rate is so reduced that he prefers to consume leisure and live off his wealth. Presumably, even if Michael Milken loses most of his wealth, he will not consider many low wage jobs preferable to unemployment. The concern is then that such a decline would overestimate the penalty imposed upon the criminal to the extent that he prefers the same dollar loss from wage income (because he no longer has to work to obtain it) to that loss from a fine. However, it is by no means clear that completely leaving the labor force will occur that frequently since wealth is also likely to decline with income after conviction.(9) The large legal costs borne during trial are not the only reason that wealth should decline. For example, given how divorce law and practice divides property decidedly in favor of the spouse who seeks a divorce from a convicted felon, high divorce rates after conviction may substantially reduce personal wealth.(10) While this paper tests the relative importance of these different explanations to a limited extent, in the preceding discussion l have primarily sought to explain why income should fall after conviction. The remainder of the paper analyzes whether such a decline exists, its size, and its implications for optimal penalties. III. MEASURING THE REDUCTION IN POSTCONVICTION EARNINGS The total monetary penalty imposed by a criminal conviction consists of the reduction in legitimate income, the income lost while in prison, fines, legal costs, and the lost time resulting from the legal process leading up to conviction.(11) This paper tries to measure the first three types of penalties. However, if the cost of the first three penalties rises with income, leaving legal costs and the time costs of trial out of the discussion undoubtedly underestimates the absolute difference in penalties borne by high income individuals since they will likely spend more on their defense and have a higher opportunity cost of participating in the trial.(12) The higher the total penalty from conviction, the greater is the effort individuals make to defend themselves. Two categories of Federal crimes will be examined: embezzlement or fraud and larceny or theft. The appendix shows the Administrative Office of the U.S. Courts' definition from the Criminal Statistical Codes and Coding Procedures manual for the two categories of crime. Federal crimes in these two sections generally involve actions which directly affect government property or interfere with commerce or occur through the postal service. Crimes for embezzlement involve someone having lawful title to the property but using it in an unauthorized manner, while for fraud the crimes involve the use of false pretenses. The category larceny and theft generally refers to those actions where a person does not have title to an asset and carries it away with the intent of depriving the owner of its use. Table I provides some insight into the average characteristics of criminals who were at least twenty-two years of age: their prison sentence, criminal fines and restitution, and the size of the dollar loss associated with conviction in these two general categories of crime. The income figures show that the drop in legitimate income from the twelve months before sentencing to the last twelve months that a criminal was either on probation or parole is TABULAR DATA OMITTED roughly the same size as the total penalty from criminal fines, restitution, and lost income while in prison (assuming income remained at its presentence level). These income data include only legitimate sources of income and were obtained from the Administrative Office of the U.S. Courts data set entitled the Federal Probation and Parole Sentencing and Supervision Research File (FPSSIS). The Administrative Office relies on a presentence report drafted by the court and on the final report by the criminal's probation officer.(13) For embezzlement or fraud, the reduction in income between these two periods is large, equalling 74 percent of the sum of criminal fines, restitution, and the forgone income from imprisonment. Similarly, the reduction in income from a conviction for larceny or fraud equals 104 percent of the total penalty from the other sources. These numbers probably substantially underestimate the importance of this reduction in income since the reduction is likely to continue past the twelve month period observed.(14) A first analysis of the data is very consistent with the hypothesis that the criminals risking the largest reduction in legal income are also those who had the greatest legal income to begin with. Regressing the difference in the logarithmic pre- and postconviction incomes on the criminal's presentence income, preconviction income squared, and an intercept over the entire sample yields (1) [ln (ACTUAL POSTCONVICTION INCOME) - ln (ACTUAL PRECONVICTION INCOME)] = -.0000400 PRECONVICTION INCOME (8,564) + 3.81E-11 PRECONVICTION INCOME SQ. (6.198) - .20078 (1.533) F-statistic = 37.766 Adj-[R.sup.2] = .0504 N=1387, with the t-statistics shown in parentheses. The estimates indicate that the percentage reduction in legitimate income associated with conviction increases up to a preconviction income of $524,934. Criminals with the same mean preconviction income as larcenists and thieves ($10,312) can expect a reduction of $4705 (or 46 percent), while those with the same mean preconviction income as embezzlers and defrauders ($20,992) face a reduction of $13,451 (or 64 percent).(15) Table II provides TABULAR DATA OMITTED some evidence that this very simple result is not sensitive to different specifications of the endogenous variable (e.g., using only the logarithmic value of postconviction income) and different specifications of the income variables on the right-hand side of the regression (e.g., specifying preconviction income in logarithmic values). All eight specifications provide evidence that the absolute reduction in postconviction income increases with the level of preconviction income, and for each type of crime the different specifications predict reductions which differ by no more than two thousand dollars for preconviction incomes up to two standard deviations above the mean. The next step in measuring the reduction in income due to conviction is to analyze how the change in income also depends on the characteristics of the crime, the individual's past criminal record, his personal characteristics (e.g., race, marital status), and any changes that may have happened while in prison (e.g., whether he received an education while in prison). The data for these characteristics are also available from the FPSSIS data set. While I will provide additional estimates using the difference in actual incomes, one problem is that this approach does not account for what the individual's income would have been in the second period if he had refrained from committing a crime. Thus the true opportunity cost of committing the crime if real legitimate income would have changed over time is not measured. Presuming that the criminal's real legitimate income would have grown had he refrained from criminal behavior, using the drop between the criminals actual pre- and postconviction incomes underestimates the true penalty from conviction. One way to proxy for the changes that would have occurred in the criminal's legitimate income had he not been convicted is to observe how income changed for nonconvicted individuals with the same characteristics over the same time period. Income regressions were run for a sample from the general population using data from the Bureau of the Census's 1984 Survey of Income and Program Participate (SIPP).(16) Inserting the values for the criminals from the FPSSIS data set into the income regression for legitimate individuals predicted the income that these criminals would have had if they had been legitimate individuals with the same observed characteristics as the criminals.(17) Both the predicted honest income for the criminal's observed characteristics and their actual income are in real July 1984 dollars after deflating by the consumer price index. The goal is to compare the drop in the criminals income, not from what it was preconviction, but from what it would have been postconviction had not the individual been involved in a crime. To make this comparison, the endogenous variable is set up so that it measures the percentage change in a criminals actual legitimate earnings relative to the percentage change in legitimate earnings for a noncriminal who has the same observed characteristics as the criminal.(18) Using the endogenous variable which compares the percentage change in the income of noncriminals with the same observed characteristics as the criminals with the percentage change in criminals' legitimate income,(19) I estimated a regression model of the following form separately for embezzlement and fraud and for larceny and theft: (2) [ln (PREDICTED PRECONVICTION INCOME) - ln (PREDICTED POSTCONVICTION INCOME)] - [ln (ACTUAL PRECONVICTION INCOME) - ln (ACTUAL POSTCONVICTION INCOME)] = ACTUAL PRECONVICTION [INCOME.sub.i] [[Beta].sub.1] + ACTUAL PRECONVICTION [INCOME.sup.2.sub.i] [[Beta].sub.2] + PRISON [SENTENCE.sub.i] [[Beta].sub.3] + EDUCATED IN [PRISON.sub.i] [[Beta].sub.4] + DOLLAR [LOSS.sub.i] [[Beta].sub.5] + [BREACH.sub.i] [[Beta].sub.6] + PAST [CONVIC.sub.i] [[Beta].sub.7] + RACIAL [DUMMIES.sub.i] [[Beta].sub.8] + MARITAL STATUS [DUMMIES.sub.i] [[Beta].sub.9] + [WOMAN.sub.i] [[Beta].sub.10] + [AGE.sub.i] [[Beta].sub.11] + COURT DISTRICT [DUMMIES.sub.i] [[Beta].sub.12] + TYPE OF CRIME [DUMMIES.sub.i] [[Beta].sub.13] + [[Beta].sub.14] + [[Epsilon].sub.2i] where PRISON SENTENCE = the number of months that an individual was sentenced to prison by the court, EDUCATED IN PRISON = a dummy variable that equals 1 if the criminal was educated while in prison and 0 otherwise, DOLLAR LOSS = the estimated nominal dollar amount taken in the crime rounded to the nearest thousand dollars for values exceeding $1000, though this excludes any money recovered during the arrest(20) (unfortunately it is in nominal dollars since the date when the crime occurred was unavailable), BREACH = a dummy variable that equals 1 if the criminal committed a crime that involved a breach of trust and 0 otherwise (the FPSSIS data set defines a breach of trust to "indicate whether the offense involved a direct violation of the trust placed in the offender by virtue of his legitimate occupation (i.e., the offender used the trust he received by virtue of his employment to commit the offense)"), PAST CONVIC = a vector of dummy variables for whether the individual has been convicted once (Past Convic = 1), convicted twice (Past Convic = 2), or convicted three or more times (Past Convic [is greater than or equal to] 3), where the dummy variable equals 1 if the individual has that characteristic and 0 otherwise, RACIAL DUMMIES = a vector of dummy variables for whether a person is either an American Indian, Asian, Black, or Hispanic, where the appropriate dummy variables equals 1 if the individual has that characteristic and 0 otherwise, MARITAL STATUS DUMMIES = a vector of dummy variables for whether the person is either Divorced, Separated, Never Married, or Widowed, where the appropriate dummy variable equals 1 if the individual has that characteristic and 0 otherwise, WOMAN = a dummy variable that equals 1 if the criminal is a woman and 0 otherwise, AGE = the age of the criminal at time of sentencing, COURT DISTRICT DUMMIES = a vector of dummy variables for the eleven federal court districts and the District of Columbia,(21) and TYPE OF CRIME DUMMIES = a vector of dummy variables indicating the type of embezzlement or fraud and larceny or theft that the individual was convicted of (see the appendix for U.S. Code title and section). For larceny and theft the intercept represents the reduction in income for "Other--Misdemeanor" convictions (code 3800) and for embezzlement and fraud the intercept represents the reduction in income for "Other" fraud convictions (code 4999). The embezzlement or fraud cases contain thirty-two categories of crime, while larceny or theft cases contain seven categories. The specification assumes that a criminal's actual legitimate income would have grown at the same rate as the predicted noncriminal's income if the criminal had not been convicted. My primary focus is the effect that presentence income has on the decline in income due to a conviction. While the discussion in section II suggests that the absolute decline in income increases with Actual Preconviction Income, there seems to be no strong a priori reason to expect this decline to be either a falling, constant, or increasing percentage of presentence income. One suspects that the longer the length of the Prison Sentence, the greater the increase in the spread between a criminals actual pre- and postconviction incomes (the coefficient should thus be negative). First, the criminal can lose human capital while imprisoned (analogous to the reduction in income that women suffer when they leave the labor force to have children), and second, the length of the sentence can signal the seriousness of the crime. The signal could either provide additional information on the criminal's utility function or on the size of the sunk investment that he forfeited. However, it is quite possible that these two effects are not large. First, the average sentence lengths shown in Table I are relatively short and thus may not produce a very large loss of human capital. Second, the sentence length, while it is probably highly correlated with the total time that a person actually spends in prison, is not perfectly related to the actual time served (e.g., early release due to "good behavior"), and thus may provide downwardly biased estimates of the regression coefficients because of measurement error. If any formal education that a criminal receives in prison produces marketable human capital, the postconviction difference between the predicted and actual incomes should be smaller, and thus the coefficient should be positive. Especially in the cases of embezzlement and fraud, larger dollar losses (LOSS) should cause a greater reduction in income (a negative coefficient), since these would proxy for the amount of trust that had been placed in the individual and thus the premium that he had been receiving in his presentence income for that trust. This prediction is consistent with both the sunk investment hypothesis and the utility function type argument. Likewise, the reduction in income should be greater whenever the employee was able to commit the offense "by virtue of his legitimate occupation" (BREACH), since employers are bound to place a greater premium on honesty whenever the job affords opportunities for crime. If a larger number of past convictions causes less trust to be placed in the individual, the smaller is the reduction in income, and hence the coefficients for PAST CONVIC should be positive. Alternatively, if employers believe that one conviction alone provides little information on the likelihood of the worker committing another crime, a second conviction may result in a greater reduction in income than the first. Since I did not know how people learn from the past convictions of criminals, I first used dummy variables for the number of convictions so as not to impose a particular structure on the effect.(22) With regard to the other control variables, such as marital status, racial dummies, age, geographic location, and sex, it is not clear what type of systematic relationships, if any, should exist. These variables help answer the question of which characteristics are responsible for the decline in income due to conviction. Some groups may have a greater reputation for being honest, and thus possibly stand to lose more by committing a crime. The results for the two different sets of crime categories are shown in Tables III and IV, with Table III showing the case of embezzlement or fraud and Table IV larceny or theft. The first column in both tables shows the simple specification with the change in the (logarithmic) values of pre- and postconviction income. They provide strong evidence that incomes decline after conviction. Columns 2 through 6 show the change in the difference between the predicted and actual incomes in the two periods. This set of regressions provides some evidence on the robustness of the results through the successive removal of those variables that describe the charged crime, the past conduct of the criminal, the variables on race, marital status, age, and sex, and finally everything but the past income variables.(23) The most striking finding for both tables is the substantial and significant impact that presentence income has on the size and progressivity of the reduction in postsentence income. Using the second columns in both Tables III and IV, the net impact of presentence income in both regressions is to continually increase the percentage reduction in postsentence income up to a presentence level of $702,415 and $92,424 in Tables III and IV. Since these values are 11.4 and 5.5

standard deviations greater than their respective means, this effect is consistent across the sample. The net effect remains negative until $1,404,830 and $184,849.(24) The progressivity of the reduction in income can be seen from two hypothetical cases involving bank embezzlement and bank larceny. Assume that both cases involve a single, white, California male who has no previous criminal history. Using the coefficients from column 2 in Table III and assuming the sample mean for the bank embezzler's age, presentence income, and dollar amount taken,(25) this typical bank embezzler faces a reduction in income of $6,581 if his real income in the absence of a crime would have remained at his presentence level of $16,813, a reduction of 39 percent. Similarly, if he would have had a constant real income of $49,362 in both periods had he been law-abiding (which is one standard deviation above the mean presentence level) the loss would have been $40,484, or a decline of 82 percent. The reduction is so large for the higher income criminal that he actually ends up earning $1,353 less than the criminal with the mean presentence income. Possibly, the high income criminal faces a greater change in the types of jobs for which he is eligible, and thus might require a greater amount of retraining, or, alternatively, he may remain unemployed longer before he finds new employment.(26) By comparison, TABULAR DATA OMITTED TABULAR DATA OMITTED for this same difference in income, the 1988 Federal tax code produces only a 13 percent higher average tax rate. Using the sample mean age, presentence income, and dollar amount stolen for bank larcenists, along with the coefficient estimates from column 2 in Table IV,(27) the postconviction income declines by $3,408 if his real income in the absence of a crime would have remained at his presentence level of $8,254.65, a reduction of 41 percent. Similarly, someone with a constant real income of $14,986 (a one standard deviation above the mean presentence level) faces a loss of $10,107, a 67 percent reduction. Again, this progressivity greatly exceeds that in the Federal tax code which specifies only a 4 percent increase in the average tax rate between the two incomes. Longer prison sentences are consistently related to reduced postconviction earnings in the cases of larceny or theft and are significant at least at the .07 level for a single-tailed t-test. A one-month increase in sentence length causes a 5.5 to 32 percent greater reduction in postconviction income. This reduction could be due to either lost reputation being associated with the seriousness of the crime and hence the length of the sentence or because of lost human capital. However, the effects of imprisonment far exceed the relative depreciation rates for women who leave the labor force to raise a family, which suggests that some portion of the reduction is caused by something beyond human capital depreciation.(28) In addition, since debarment is a function of conviction and not the length of the prison sentence, debarment cannot explain the size of the prison sentence coefficient. The results for embezzlement or fraud are more mixed, with the coefficient being both negative and significant in only one of the five regressions at the .10 level for a single-tailed t-test. A partial explanation is that prison sentences are highly correlated with the size of the dollar loss and thus might not provide much additional information as to the seriousness of the crime.(29) Past convictions play a very important role in explaining how postconviction income changes for larceny and theft with employers of individuals who commit these crimes seemingly attaching more importance to an prospective employee's second and third convictions than to his first. After the third conviction, additional convictions seem to have relatively little effect on income. In the cases of embezzlement or fraud, the biggest reduction in income occurs after the first conviction. Embezzlers and defrauders thus do not seem to be as likely to be "given a second chance." A breach of trust generally has a much larger and a more significant impact for larceny or theft than it does for embezzlement or fraud. In fact, the coefficient is at least twice as large in the specifications for larceny or theft. Surprisingly, the dollar amount taken in embezzlements or frauds appears to have both a very insignificant and economically very small negative impact, while it is actually positive and significant for larceny or theft, though the coefficient is quite small.(30) There is very little difference in the results between the regressions that use the difference in the log of income between the two periods and those which rely on the logarithmic values of predicted and actual incomes in the two periods. While these results are interesting, there are at least two important reasons to believe that they underestimate the reduction in income from conviction.(31) According to Steve Reynolds at the Administrative Office of the Courts, the income estimates during the last twelve months of supervision are biased upward. If an individual was unable to find employment for the first six months of this period, it is standard practice to report the annual rate of the individual's last six-months income. In fact, any large improvements at the end of the twelve-month period are likely to be weighted heavily. A similar bias reportedly also exists in the other direction for the income estimates during the twelve months prior to sentencing. Income in the twelve months prior to sentencing is often already reduced by the individual having been arrested. If income tends to decline as sentencing approaches and the income earned closest to sentencing tends to be reported as the annual rate, the presentence income will be biased downward. The ideal measure of presentence income would be the income that the individual was earning before he was charged with the crime. Both effects thus work to narrow the difference between pre- and postconviction incomes. Discussion with probation officers also indicates that high income individuals take much longer to find new employment after a conviction than do low income individuals. However, as long as the criminal can find new employment by the time he reaches the end of supervision, he will be reported as having earned that income even over the period during which he was unemployed. The bias against finding a reduction in income would then be greatest for the highest income criminals.(32) An important question is whether substitution between legitimate and illegitimate income in the pre- and postconviction periods occurs with the result that total income remains relatively unchanged.(33) While the large drops in legitimate earnings undoubtedly suggest that individuals are supporting themselves through illegal means, two effects work to make criminals substitute out of illegal and into legitimate activities during the period that the criminals are on parole or probation. First, the expected punishment from crime is higher while criminals are on probation or parole than it was prior to the original sentencing. Not only is the probability of detection under supervised release higher, but the effective penalty the criminal faces for any new offense also increases because of the threat of serving out the remaining term of the prior conviction. Second, criminals will frequently no longer have the jobs which in some cases allowed them to commit their original crimes. There is at least then a presumption that during the period of supervised release both legitimate and illegal income will decline. While illegal income could increase between the pre- and postconviction periods as a result of the reduction in legitimate earnings, it seems unlikely that we will observe a large increase in illegal earnings while a criminal is on probation or parole or that it will offset the decline in legitimate income. If any substitution toward legitimate income does occur between periods, it biases the results against finding a reduction in income from conviction. IV. ESTIMATING THE TOTAL MONETARY PENALTY To compare the estimated reduction in income from conviction with the other penalties that criminals suffer as presentence income varies, I reestimated different regressions, generally using the specification shown in column 2 in Tables III and IV but consecutively replacing the endogenous variable with the length of the prison sentence, the level of the criminal fine, and the level of the criminal restitution. The only change in the exogenous variables is that the variables for the sentence length and whether one received a formal education while in prison are removed. The background of judges would probably be important in predicting the penalty, but controlling for the general characteristics of a region's judges through a court district dummy seemed a reasonable substitute.(34) It is interesting to note, however, that the sentencing commission guidelines set the length of sentences primarily as a function of the offense type, the dollar loss associated with the crime, and the defendant's past criminal history.(35) The results are provided in Table V. They suggest that while a higher income increases both fines and restitution (though the coefficients are not significant for larceny and theft),(36) a higher income also reduces the length of the prison sentence for both categories of crime (though in neither case is the effect significant). Income may be a more important determinant of monetary penalties in the cases of embezzlement and fraud simply because those individuals have higher incomes and judges tend to constrain dollar penalties by the defendant's ability to pay. While a breach of trust usually results in greater penalties, the dollar amount of the crime is only significant and positively related to the sentence length for embezzlement and fraud.(37) Finally, as one would expect, a past criminal history results in more severe current prison sentences, though the level of fines exhibits either no relationship or a negative one, possibly because a criminal record is related to low wealth. Examining the same four hypothetical bank embezzlers and larcenists as discussed above and using the regressions in Table V, we can predict the expected criminal fines and the forgone earnings from imprisonment. I should note that since I am using the length of prison sentences rather than the actual time served, the estimates of forgone income are biased upward. The estimates of the expected fine and restitution plus forgone earnings show that in all four categories of crime the combined penalty is less than the reduction in income from conviction. The criminal fines, criminal restitution, and the forgone income from incarceration all rise with income, despite the negative effect of income on prison sentence length in the first and fourth regressions in Table V, because the opportunity cost of time rises proportionally faster than the sentence length falls. As can easily be seen from Table VI, a bank embezzler with a mean income one standard deviation above the mean faces a total monetary penalty that is 4.94 times greater than that for an embezzler with a mean income, while the analogous bank larcenist faces a 2.1 to 1 ratio over a larcenist with the mean income. Therefore, if the low and high income criminals in both cases are to face the same expected penalties from conviction, the high income embezzler must face a probability of conviction that is only 20.7 percent of that of the mean income embezzler and the high income larcenist a probability that is only 47 percent of that of mean income larcenist. When the corresponding values for two standard deviations above the mean are used, the relative probability of convictions fall to only 11 and 30 percent. To the extent that the reduction in income due to conviction continues past the end of supervision, the preceding numbers can underestimate the true differences across criminals. For example, if the real reduction in earnings persisted for five years and the real interest rate was two percent, the present value of lost earnings for an average bank embezzler is $31,020. For embezzlers with incomes one and two standard deviations above the mean the present values are $190,818 and $364,028. The total penalty for an embezzler with a preconviction income one standard deviation above the mean would thus be 5.8 times greater than that for the mean income embezzler, and the penalty for the embezzler with the highest income would be 10.9 times greater. To equate the expected total monetary penalties across criminals, embezzlers whose incomes are one and two standard deviations above the mean should only be convicted 17 percent and 9 percent as frequently as embezzlers with the mean income. For bank larcenists, the comparable percentages are 37.6 and 22. TABULAR DATA OMITTED TABULAR DATA OMITTED Before determining how the expected penalty varies across criminals, additional information on how the probability of detection and conviction varies with the criminals preconviction incomes is required. Both probabilities should be lower for criminals with high preconviction incomes since higher penalties increase the return to defense expenditures (Lott [1987b]). Though these greater efforts to avoid conviction will work to narrow the difference in expected penalties between high and low income criminals, if greater efforts at avoiding conviction arise from increased legal expenditures, the difference in expected penalties will not be completely eliminated. The reader may refer to the last column in Table VI to decide whether it is plausible that these differences in effort can to a significant degree offset the differences in total monetary penalties. Unfortunately, data on the probabilities of detection and conviction are generally not available. One option is to use existing information for the probability of conviction conditional on being charged with a crime. Hagan [1974] finds at most a 16 percent difference in conviction rates across characteristics of groups and conviction rates for noncapital crimes in the twenty sociology papers that he surveys. If expected penalties are to be equal, Table VI implies that the probability of arrest must be extremely small for high income criminals. Fortunately, one exception where reasonable estimates of conviction conditional on committing the crime exist is for bank larcenies and robberies, where almost 95 percent end in conviction (Rhodes [1989]). Given that banks are likely to report virtually all of these crimes and that the photographic evidence used in these cases implies a relatively small probability that individuals will be falsely convicted, this conviction rate provides an unusually good estimate of the probability with which the crime is punished. Using the data shown in Table VI, it is extremely difficult to believe that the probability of detection of bank larceny will vary sufficiently so that expected penalties will be equal across all criminals. With 14.2 percent of the bank larceny sample having incomes between $14,986 and $21,717 and 4.8 percent having values greater than $21,717 and assuming that the probabilities of detection for these two groups are 47 and 30 percent, even if the probability of conviction for all remaining criminals were 100 percent, the probability of conviction for the whole group would be at most 89.1 percent. Even in that case, the expected penalties would be too low for those criminals whose preconviction income was below the mean income, if they are at the right level for those with the mean income and those at $14,986 and $21,717. To correctly answer whether the well-to-do criminals are in fact penalized "too heavily," it would be useful to examine efficiency explanations for why the current penalty structure exists. One plausible explanation involves the different costs of imposing penalties for the different types of criminals. While reputation is costly to produce, it may represent a much lower cost method of imposing penalties on certain people than does imprisonment. If so, it follows that a higher total penalty will be imposed upon high income individuals who have a great deal of reputation compared to low income individuals who have little or no reputation. Another possibility involves differences in supply elasticities between high and low income criminals, but it is difficult to have any strong beliefs on how these may differ a priori. A different position is taken by Andreoni [1989], who provides an efficiency explanation for imposing lower penalties on the wealthy than the poor by assuming that jurors value not convicting innocent individuals. Given the large differences in observed penalties, the obvious question for future papers is to explain why this pattern exists. Possible policy implications from these results depend on the source of the reduced earnings. If reductions in earnings result from market-imposed penalties, longer prison terms and/or greater criminal fines may be partially offset by less reliance on individual reputations to deter crime. With a greater reliance on criminal penalties, firms would reduce the wage premiums used to ensure behavior. If legally imposed criminal penalties and individuals' reputations are perfect substitutes in guaranteeing a persons behavior, higher criminal sanctions will be perfectly offset by reduced investments in reputation--leaving the total penalty unchanged. Higher income individuals then would not face total penalties different from those imposed on low income individuals. However, the legal system does not act as if these penalties are perfect substitutes, since larger reductions in legitimate income are associated with larger (not smaller) legally imposed penalties. Besides the possibility that there are differential costs to imposing these different penalties, a Klein and Leffler type framework also implies that there is no reason to presume perfect substitutability because the threatened loss of quasi-rents is useful in deterring shirking even when criminal penalties are irrelevant (Lott [1988]).(38) One last note, if the reduced earnings arise from debarment and other government restrictions, increased prison sentences or fines should not result in a smaller reduction in income upon conviction. V. CONCLUSION Conviction, by itself, produces a real penalty. The results indicate that the reduction in income from conviction is often a much larger share of the total penalty that criminals face than the combined total of fines and the forgone income costs of imprisonment. In view of the major reduction in earnings after a conviction, penalties appear less lenient than often believed.(39) Further, the reduction in income from conviction is extremely progressive. When added to the fact that fines and the forgone income costs of imprisonment tend to increase with preconviction income, it implies that conviction rates for individuals with relatively high incomes have to be dramatically lower than for low income individuals if expected total monetary penalties are to remain constant across individuals. Even a single standard deviation increase in income can result in a fivefold increase in the total monetary penalty. At least for bank larceny, where relatively reliable estimates exist for punishment conditional on the crime taking place, the evidence strongly indicates that even if a high income larcenist commits the same crime as a low income larcenist, he will face a much higher expected total monetary penalty. This discussion provides a possible explanation for why relatively few well-to-do individuals commit crimes: the total penalty that they bear is simply so much higher than it is for poor individuals.(40) This evidence provides a stronger case that if well-to-do criminals are not allowed to "buy justice" the expected penalties will be too large for the well-to-do, if they are at the correct level for the poor. Similarly, if the expected level of punishment equalled the total social cost for wealthy criminals and the probability of conviction is not appreciably higher for the poor, too much crime will be committed by the poor. Finally, given the size of these estimated reputational losses there is the concern that if arrest alone also produces a loss in reputation, economists may have misplaced their emphasis on the probability of conviction in their discussion of optimal penalties. 1. See Lott [1987b, 1313-1315] for a brief survey of this literature. For another recent perspective from an opposing point of view, see Kennedy [1988]. Work that I have done on this topic since this paper was originally written includes estimates of the total monetary penalties for drug convictions (Lott [1992]). The findings there are similar to the ones presented here with the reduction in postconviction income accounting for between 35 and 96 percent of the total pecuniary penalty borne by convicted criminals. One major advantage in examining the bank larceny data used here is that much better estimates exist for the probability of conviction conditional on the crime occurring. Little real confidence can be placed in government estimates of the rate at which drug smugglers or users are punished, and this makes it difficult for that data to answer the question addressed at the end of section IV: whether high income criminals face higher expected monetary penalties than do low income criminals. Karpoff and Lott [1990] provide estimates of the reputational loss that firms suffer from the first public announcement that they are being investigated for fraud. They estimate that the average drop in market value of firms suspected of frauds against private parties is $60.5 million, with at least $28.7 million being attributable to lost reputation. 2. Schulhofer [1988, 69]. 3. See Lott [1987b]. Easterbrook [1983] discusses the general tendency towards efficiency in the criminal justice system. 4. Becker and Stigler [1974], Barro [1973], and Klein and Leffler [1981] provide more recent examples of this argument. 5. A related case discussed in these other papers occurs in political markets. If voters can sort into office those politicians who value the same things as they do, this can solve the shirking and last-period problems. Politicians will not deviate from their constituents' interests even during the politicians' last terms in office (when they no longer face the threat of reelection) since doing so would lower their own level of utility. 6. The reduced legitimate earnings discussed here provide an additional type of private enforcement. See also Friedman [1979; 1984] and Benson [1990] for general theoretical discussions of the private enforcement of law. 7. The National Advisory Commission on Criminal Justice Standards and Goals [1973, 592] notes that, "every State and the Federal government make it difficult for persons convicted of a felony to obtain licenses to practice occupations regulated by the government. In many instances conviction of a felony is automatic grounds for denial of a license. In others, it is in practice impossible for a former offender to obtain a license." Lott [1992] discusses the difficulties that convicted individuals have in joining unions, securing public employment, and obtaining business licenses. Ex-convicts face many other forms of penalties such as being prevented from inheriting property, suffering partial or complete divestment of their assets, losing life and automobile insurance, and losing pension funds, and face the discontinuance of pension payments even if the individual is already retired (Grant et al. [1970, 1109-43]). Since the loss of inheritance and pension funds and divestment of assets undoubtedly impose a larger absolute penalty on the well-to-do than on the poor, the estimates presented here will underestimate how much penalties increase as a function of presentence income. Several Presidential Task Forces have emphasized the importance of these collateral penalties and expressed concern over how ignoring collateral penalties will create inequities in criminal penalties (President's Commission [1967, 88]). For the current list of states where conviction affects voting, parental rights, divorce, public employment, ability to serve as a juror, holding public office, and ownership of firearms, see Burton et al. [1987, 55], Benson [1984] and Jesilow [1982]. The difficulty in obtaining new employment after conviction can be quite striking. For some recent stories on the difficulties that white-collar criminals have in obtaining employment after conviction see the Wall Street Journal 4 May 1990, A 10; and 20 April 1990, p. 1). In one case, a debarred lawyer was reduced to selling newspapers out of the back of his stationwagon and working as a clerk in a liquor store, and in another a stockbroker decided it made more sense to start serving out his sentence while he waited for what turned out to be a successful appeal because he was unable to find employment. 8. See Friedman [1981] for a model of how these various concerns interact. 9. The last paragraph in footnote 4 provides some examples of the types of jobs lost after conviction. 10. I would like to thank an anonymous referee from this journal for bringing the issue in this paragraph to my attention. Burton et al. [1987, 55] provides the current rules pertaining to divorce. While no data are available, conversations with probation officers indicate that divorce is quite prevalent after conviction. 11. One part of the penalty that will not be measured is the value that the individual attaches to his forgone liberty while in prison. Lott [1987b] (particularly footnote 3) discusses why the opportunity costs of imprisonment are likely to vary positively with the wealth of the criminal. If this is true, it provides yet another reason to suspect this paper underestimates the absolute difference in penalties between lower and higher income individuals. Another penalty that will not be examined involves civil restitution. This penalty can be quite large. For example, in the case of insider trading, Pulver [1988] calculates that civil restitution usually amounts to two times the profit from the insider trading. The size of the multiple is closely related to the wealth of the defendant, with wealthier individuals facing a higher multiple than poorer ones (Pulver [1988, 28-31]). In general, I expect that civil restitution varies with income in a manner similar to the way criminal restitution varies with income. Data on criminal restitution will be examined more carefully in section IV. 12. This effect is enhanced by the fact that low income individuals may not have to pay for their own defense since they can be provided with a public defender. 13. The court's presentence report is usually based upon the individual's tax returns. The report filed by the probation officer relies upon a combination of the criminal's tax returns and a monthly report that he files with his probation office. Probation officers are involved in filing both the presentence and the final probation or parole reports. As I will discuss below, while the data are reported as an annual rate, if there has been a change in income over the preceding year, the reports have a tendency to list the most recent rate as the annual rate. This biases the findings against finding a reduction in income from conviction. The following types of income are included in the FPSSIS estimate: wages from any legitimate employment, disability income, welfare income, retirement income, income from legitimate investments, funds from academic scholarships, assistantships, or fellowships, and union unemployment benefits. The income measure used from the SIPP data set was the person's total income (PPINC). 14. Not surprisingly, those who are sentenced to prison face a larger percentage reduction in income than those who are simply convicted. For larceny and theft, the average income declines from $5,983.69 to $4,001.41 for those who are sentenced to prison (N=93) and from $11,263.62 to $8,876.70 for those who are not, a 33 versus 21 percent decline. For embezzlement and fraud, the average income declines from $21,564.03 to $12,524.67 for those who are sentenced to prison (N=153) and from $20,871.29 to $16,600.45 for those who are not, a 42 versus 21 percent decline. Another not too surprising relationship is the decline in the average income of criminals as the number of prior convictions increases. For embezzlement or fraud the presentence incomes are: no prior convictions, $23,384 (N=668, [Sigma]= $55,710); one prior conviction, $16,132 (N=86, [Sigma]= $16,560); two prior convictions, $13,761 (N=61, [Sigma]= $24,724); and three or more prior convictions, $7,817 (N=56, [Sigma]= $8,673). For larceny or theft the presentence incomes are: no prior convictions, $11,460 (N=328, [Sigma]= $22,355); one prior conviction, $11,565 (N=72, [Sigma]= $23,052); two prior convictions, $10,6741 (N=40, [Sigma]= $21,360); and three or more prior convictions, $3,976 (N=76, [Sigma]= $5,182). 15. The results shown in Tables III and IV below are consistent with these findings when [ln (ACTUAL POSTCONVICTION INCOME) -- ln (ACTUAL PRECONVICTION INCOME)] is replaced by ln (ACTUAL POSTCONVICTION INCOME) or the exogenous variables for preconviction income are replaced by their logarithmic values. 16. Separate income regressions were estimated for both men and women twenty-two years of age or older for each of the overlapping twelve-month periods during the thirty-six months in the sample yielding a total of fifty regressions (see footnote 17). They take the form: (A) ln [INCOME.sub.i] = [EXPERIENCE.sub.i] [[Alpha].sub.1] + [EXPERIENCE.sup.2.sub.i] [[Alpha].sub.2] + RACIAL [DUMMIES.sub.i] [[Alpha].sub.3] + EDUCATION [DUMMIES.sub.i] [[Alpha].sub.4] + MARTIAL STATUS [DUMMIES.sub.i] [[Alpha].sub.5] + COURT DISTRICT [DUMMIES.sub.i] [[Alpha].sub.6] + [[Alpha].sub.7] + [[Epsilon].sub.li] where EXPERIENCE = the individual's age minus the number of years of education minus six years, and EDUCATION DUMMIES = a vector of dummy variables for whether the individual has either elementary school (eight years), some high school (nine to eleven years), high school (twelve years), some college (thirteen to fifteen years), college (sixteen years), and postgraduate school (sixteen to eighteen years, this was truncated in SIPP at eighteen years). The appropriate dummy variable equals 1 if the individual has that characteristic and is zero otherwise. The intercept thus indicates those values for when the individual has less than eight years of education. See the list of variables after equation (2) for the definitions of the other variables. The primary constraint on which variables to include and which time period to select was the matching of the SIPP data with the FPSSIS data. The continuous variables for education had to be put into the form of dummy variables since the FPSSIS data do not have a continuous variable. If an individual's educational status or geographic residence changed during the course of the twelve months for which the income regression was being estimated, the dummy variables were weighted by the portion of the year for which the particular circumstances were true. (For instance, if the individual lived in the first court district for six months and the second court district for six months, the "dummy" variables for both districts would equal .5.) The adjusted [R.sup.2]s for these regressions averaged about .15 for the males and .21 for the females, and similarly the F-statistics range from the low 50s to the low 90s. The experience variable for the FPSSIS data set was constructed by assuming that the amount of schooling associated with elementary school was eight years; some high school, ten years; high school, twelve years; some college, fourteen years; college, sixteen years; and postgraduate school, eighteen years. 17. Income regressions were constructed for the twenty-five overlapping twelve-month periods during the thirty-six months beginning in July 1983 and ending in June 1986. Figure 1 helps to explain the general setup. The first twelve month income regression for the noncriminals is estimated starting at [t.sub.0], a second starting at [t.sub.1], a third starting at [t.sub.2], and so on up until [t.sub.24]. All the sentencing dates in the sample thus fall within the period from July 1984 to July 1985. The additional constraint was then that all criminals in the sample were released from supervision by July 1986. The data contains the entire set of criminals sentenced in the Federal system for violations of embezzlement, fraud, larceny, and theft. 18. While I am not making the assumption that the variables controlled for in predicting the income of the honest individuals affect the estimated legitimate earnings of those in the general population and the criminals in our sample equally, I will assume that they affect the growth rate similarly. Not making the first assumption recognizes that there may be unmeasured characteristics which help to explain why those in the criminal group became criminals and that these unmeasured differences may result in lower current wages for the prospective criminals. There is no obvious way to allow the growth rates to vary differently. In any case, the error introduced by not allowing different growth rates seems small relative to not recognizing that real income may be growing over time. 19. This specification thus measures the percentage change in ratios of predicted and actual pre- and postconviction incomes. The endogenous variable can also be rewritten as: [PREDICTED PRECONVICTION INCOME/PREDICTED POSTCONVICTION INCOME]/[ACTUAL PRECONVICTION INCOME/ACTUAL POSTCONVICTION INCOME]. 20. Without a measure of the Dollar Loss which controls for the level of funds recovered in the arrest it is difficult to draw any comparisons between the penalties imposed upon the criminal and the loss from the crime. Unfortunately, while no systematic data on this were available, Steve Reynolds at the Administrative Office of the Courts told me that in many cases this difference would be substantial. 21. The First U.S. District Court includes Maine, New Hampshire, Massachusetts, and Rhode Island; the second district includes New York, Connecticut, and Vermont, the third, Pennsylvania, New Jersey and Delaware; the fourth, Maryland, Virginia, West Virginia, North Carolina, and South Carolina; the fifth, Texas, Louisiana, and Mississippi; the sixth, Michigan, Ohio, Kentucky, and Tennessee; the seventh, Wisconsin and Illinois; the eighth, North and South Dakota, Nebraska, Minnesota, Iowa, Missouri, and Arkansas; the tenth, Utah, Wyoming, Colorado, Kansas, New Mexico, and Oklahoma; and the ninth, the remaining states in the west. The intercept indicates whether the individual lives in the eleventh court district, which includes Alabama, Florida, and Georgia. 22. Posner [1986, 214, 14] makes the interesting observation that "Another reason for heavier punishment of repeat offenders is that the stigma effect of criminal punishment may diminish with successive punishments." As we shall see, the evidence presented here generally supports Posner's conjecture. Interestingly, the same cannot be said for the data on drug offenses examined in Lott [1992]. Why this is true for the offenses examined here but not for those involving drug offenses is not immediately obvious. 23. While it is not reported in these tables, the same successive deleting of variables was tried using the endogenous variables in the first columns of Tables III and IV with virtually the same results. 24. The level of presentence income has an even larger negative net effect over a greater range of income in the first specifications in Tables III and IV. The reverse and direct regressions for the endogenous variables shown in those two tables were run solely on presentence income and presentence income squared as a simple test of measurement error for the income variables. In all cases the coefficients agreed in sign, and thus the maximum likelihood estimates were bounded. It is of some minor interest to note that when the regressions were specified so that the endogenous variables were replaced with the level of postconviction income, the statistical significance and relative impact of the presentence income variables were similar to those reported in the text. 25. The sample mean age of bank embezzlers is thirty-three, presentence income, $16,813, and the amount taken, $512,352. The standard deviation of presentence income was $32,549.09. Of the cases in this sample, 83 percent involved a breach of trust, so this example assumes that a breach of trust was involved. Bank embezzlement was the most frequently convicted crime in this category with 13 percent of the sample. Both bank embezzlement and bank larceny were chosen since they were the first crimes listed in the categories of embezzlement or fraud and larceny or theft. Using the above individual characteristics, the expected amount of prison time was calculated from the regression coefficients shown in Table V, yielding an expected time in prison of 1.26 months. California is within the jurisdiction of the Ninth Circuit Court. 26. A highly educated criminal could be at a disadvantage when competing for manual labor jobs with the poorer criminals who have performed such work in the past. Some employers might also view the educated criminal as overqualified and believe he would soon quit. Furthermore, the very fact that he is asking for such low-skilled employment may raise questions about why he has had difficulty finding employment similar to what he has had in the past. 27. The sample mean age for a bank larcenist is 31.3, presentence income is $8,254.65, and the amount taken $11,565.94. The standard deviation of presentence income was $6,730.93. Only 14 percent of these cases involved a breach of trust, so this example assumes that no breach of trust was involved. Using the above individual characteristics, the expected amount of prison time is calculated from the regression coefficients shown in Table V, yielding an expected time in prison of 1.89 months.