Examining the research underlying the sentencing guidelines concept in Denver, Colorado: A partial replication of a reform effort

Examining the research underlying the sentencing guidelines concept in Denver, Colorado: A partial replication of a reform effort

Journal of Criminul Jusricc. Vol. 9. pp. 51-Q lx~47-2352/x1101051-1?$0z wo Copyright ‘5: IYXI Pcrgamon Press Ltd (IYXI) Pergamon Press. Prmted i...

924KB Sizes 0 Downloads 13 Views

Journal

of Criminul

Jusricc.

Vol. 9. pp. 51-Q

lx~47-2352/x1101051-1?$0z wo Copyright ‘5: IYXI Pcrgamon Press Ltd

(IYXI)

Pergamon Press. Prmted in U.S.A.

EXAMINING THE RESEARCH UNDERLYING THE SENTENCING GUIDELINES CONCEPT IN DENVER, COLORADO: A PARTIAL REPLICATION OF A REFORM EFFORT

JOHN D. HEWITT

Department

of Criminal Justice and Corrections Ball State University Muncie. Indiana 47306

BERT LITTLE Department of Anthropology University of Texas-Austin Austin. Texas 78712

ABSTRACT This paper examines the empirical basis for the criminal sentencing guidelines developed in Denver. Colorado. Unlike many other sentencing reforms, such guidelines have generally been developed out of an empirical analysis of past sentencing decisions, which identifies those variables most predictive of sentence. Empirical arguments, as a part of a reform effort, are often more persuasive than nonempirical arguments. However, when the analysis is inadequate or faulty. the resultant reform effort map be called into question. Presented in this paper is a reanalysis of the original data used to develop the Denver guidelines. Questions are raised regarding implications of extensive missing observations across cases and the resulting shrinkage of cases available for multivariate analyses. The original data is reanalyzed in both its original .form and in a more complete form b_v estimating the missing data through a complex regression technique. Our analyses suggest that there are serious methodological weaknesses in the original study. The implications of these weaknesses are discussed.

INTRODUCTION Sentencing guidelines for the structuring of judicial decision making are increasingly

being considered. implemented, and evaluated in a variety of jurisdictions. For example. single jurisdiction guidelines have been developed and implemented to some

52

JOHN D. HEWITT

degree in Denver, Newark, Chicago. Phoenix, and Philadelphia (cf. Bohnstedt. 1979: Wilkins et al., 1978: and Calpin et al.. 1978). In addition. a major research grant. funded by the National Institute of Law Enforcement and Criminal Justice. has been undertaken to test the feasibility of developing multijurisdictional sentencing guidelines (LEAA, 1978). It is the purpose of the research discussed here to examine. through a partial replication, the empirical creation of guidelines in Denver, Colorado. In the publication, Sentencing Guidelines: Structuring Judicial Discretion (Wilkins et al., 1978), the procedures used in the construction of judicial sentencing guidelines in Denver are presented. The research (data collection and analysis) was conducted by a group from the Criminal Justice Research Center (CJRC) at Albany, New York. The research strategy, briefly. was to collect a random sample of cases of criminal convictions. From this random data set the immediate goal was to extract those variables that best predicted sentence outcome (incarceration versus nonincarceration). The ultimate goal was to create a grid or model using variables identified as aiding in the prediction of past sentencing outcomes. While these variables are, essentially, descriptive of past decisions. the theory behind guidelines suggests that the judges within the jurisdiction would, in the end. prescribe which variables would eventually go into the grid. The guideline or grid functions by computing offense and offenders’ scores on a scale derived from those variables selected and plotting them on the grid. The grid then specifies the type and length of sentence. The CJRC study generated a random sample of 200 cases drawn during the period from November 1975 to February 1976. A total of 206 items of information variables covering a diverse range of attributes (in the statistical sense) was collected from the Denver Criminal Court’s Docket Files (Wilkins et al., 1978: 41-44). As in the development of federal parole guidelines (Gottfredson et al.. 1978), a large initial set of variables was identified primarily to exhaust all measurable items which might have possible

and BERT

LITTLE

impact on the sentence decision. Inasmuch as judges (and parole board members) tend to believe that they incorporate a large amount of information into their decision making. an empirical demonstration suggesting that. in.fact. only a limited number of items of information can adequately predict sentence (or parole) decision would be important.’ Thus one aim of the CJRC research was to show that a limited number of variables. easily gathered and quantifiable. could be arranged into a grid structure and could provide judges with adequate information for making their decisions. It became evident early in the Denver project that not all 206 items of information were generally available to the judges at the time of sentencing. The missing data for the 200 cases examined ranged from 0.0 percent to over 90.0 percent. It is possible to argue that much information across variables is redundant and that judges may react to nonspecific. generalized information rather than to a specific item. thus eliminating any real substantive problem with extensive missing data. On the other hand. the widespread missing data noted earlier may pose significant empirical problems for subsequent analysis of the data set. Some of these problems will be discussed shortly. To the extent that CJRC was interested in predicting sentence outcomes. the identification of independent variables and their respective weights led the CJRC group to choose multiple linear regression as its mode of analysis. Multiple regression allows the user to calculate a linear combination of scores for the independent variables that best predict the dependent variable. As a result of this process the researcher is able to identify those variables. or combination of variables. in a given regression model. that best predict sentence. A special form of multiple regression. stepwise multiple rewas used by the CJRC group gression. (Wilkins et al.. 1Y7X: 10). While one problem is the issue of randomness. a second problem. more limiting. arises in the treatment of cases with missing data in a regression analysis. The CJRC group selected to use a listwise

Examining

the Research

Underlying

the Sentencing

deletion method of dealing with missing values. Listwise deletion excludes from the analysis any case which contains missing data on any variable included in the regression equation. Consequently. case attrition is much higher when there are even a few missing values on a wide number of variables across cases. The results of the CJRC regression analysis are presented in two tables in the CJRC publication (Wilkins et al., 1978). The regression analysis for these tables was performed on a subset of the original 200 cases. This was not a random sample drawn from the original sample. Instead, it was formed by the apparently arbitrary selection of only those cases containing information on the one variable. “statutory class of offense at time of conviction.” Information on this variable was available in only 120 of the 200 cases. The 16 independent variables. selected from the original 206 items and used in the CJRC analysis, are presented in Table 1. In the first stepwise regression equation computed by CJRC. 14 variables were included. which collectively accounted for 53 percent of the variation in the sentence variable. A total of six independent variables was considered to be statistically significant (P < 0.01). These variables were: number of offenses for which the offender was convicted. number of previous incarcerations, seriousness of offense at conviction. use of weapon in offense, legal status of offender at time of offense, and length of offender’s employment prior to offense (Wilkins et al.. 1978: 13). The second equation included two additional variables, number of previous arrests and age at first arrest, but did not greatly increase the amount of explained variation. The second equation explained 57 percent of the total variance. increasing the amount of variance explained, by only 4 percent. Four additional variables. however. were found to be statistically significant. In order. they were: number of probation revocations. number of previous arrests. number of previous convictions. and age at first arrest.

Guidelines

Concept

.

in Denver.

Colorado

53

TABLE 1

VARIABLES NAMES OF ITEMS INCLUDED IN REGRESSIONANALYSIS

CJRC

Variable Name Disposition offense

after

sentence

for

present

Statutory class at conviction/deferred cution present offense-first charge Offender’s

job status

prose-

at time of offense

Number of times offender cerated after conviction

previously

incar-

Offender’s relationship with criminal system at time of offense

justice

Offender’s convictions Harm

total

inflicted:

number

present

of

offense

Offender’s

age at first conviction

Offender’s

use of drugs

Highest

school

grade

Number of offenses convicted Weapon

usage:

Probation

attained of which

present

revocation

offender

was

history

total number

Offender’s

age at first arrest

Offender’s

use of alcohol

revocation

by offender

offense

Offender’s

Parole

previous

of previous

arrests

history

The CJRC analysis raises, perhaps. more questions than it answers in terms of the influence of specific variables on sentencing outcome. First. there would appear to be some question about the actual number of cases on which the regressions were computed. In the tables presented in the CJRC document. it is indicated that 120 cases were

JOHN D. HEWITT and BERT LITTLE

54

used. However, inasmuch as this set of 120 cases was created as a separate set of cases. it became the base, or maximum number of cases, for the regression analysis. CJRC clearly states that it used a listwise deletion technique for handling missing data in the regression analysis (Wilkins et al., 1978: 11). To the extent that there were missing data on even a few of the 16 independent variables spread across a number of cases, the listwise technique would eliminate a potentially large number of the cases in the analysis. A second question concerned the possibility of high intercorrelations among a number of the independent variables (i.e.. number of previous arrests, number of previous convictions, and number of previous incarcerations) entered in the equations. A final question raised has to do with the dependent variable, “sentencing decision,” used to develop the guidelines. CJRC states that, instead of viewing sentence in terms of an “in-out” (incarcerated-not incarcer-

ated) type of decision, it was treated as an interval variable. “All incarcerative sentences were assigned a value according to the number of years to be served. All nonincarcerative sentences were assigned a value of zero” (Wilkins et al., 1978:84). Two possible constructs of the dependent variable were identified in analytical material provided by CJRC. The first construct of the variable, “max-one.” is based on the maximum number of years collapsed into 11 categories. The second construct, “maxtwo,” is further collapsed. resulting in only 5 categories. It is noted that neither of these constructs defines sentence strictly in an interval measure.’

THE REANALYSIS Using the original data as collected in the Denver project, the first step was to compute stepwise regressions with listwise dele-

TABLE 2 STEPWISE REGRESSION WITH “MAX-ONE” AS DEPENDENT VARIABLE* (N = 70)

Variable

Seriousness of the offense at conviction Legal status of offender at time of offense Number of prior incarcerations (juvenile and adult) Injury to victim Narcotics abuse Weapon usage Employment history Age at first conviction (juvenile and adult) Education level l

R’

R’ Change

B

0.154

0.154

0.950

0.313

0.299

0.145

0.423

0.312

0.347 0.389 0.424 0.437 0.439

0.048 0.042 0.035 0.013 0.002

0.471 -0.114 0.138 0.485 0.147

0.225 -0.311 0.202 0.171 0.049

0.440 0.440

0.001 0.000

0.107 0.323

0.028 0.019

“Max-one” is an ordinal measure of maximum number of years sentenced.

RZ = 0.440 Adjusted R2 = 0.356 F = 5.24

df = 9160 sig = P < 0.01

.

Beta

Examining the Research Underlying the Sentencing Guidelines Concept in Denver. Colorado

TABLE

STEPWISE

REGRESSION

WITH

Number of prior incarcerations (juvenile and adult) Seriousness of the offense at conviction Legal status of offender at time of offense Injury to victim Narcotics abuse Weapon usage Employment history Age at first conviction (juvenile and adult)

3

“MAX-TWO”

AS DEPENDENT

VARIABLE*

(N

B

=

70)

R’

R’ Change

0.154

0.154

0.260

0.261

0.255

0.101

0.458

0.318

0.325 0.364 0.389 0.397 0.399

0.070 0.039 0.025 0.008 0.002

0.166 -0.495 0.560 0.178 0.713

0.258 -0.285 0.173 0.132 0.050

0.400

0.001

0.646

0.036

Variable

55

Beta

*“Max-two” is a collapsed ordinal measure of maximum number of years sentenced. R’ = 0.400 Adjusted R’ = 0.321 F = 5.08 df = S/61 sig = P < 0.01

tion of cases with missing observations, as CJRC reports having done. The sentence variable (operationalized first as a dichotomous “incarceration-not incarceration” decision and then as ordinal length of sentence) was regressed on the set of independent variables identified in the CJRC equations. The regressions were performed on the “nonrandom” sample of 120 cases. As noted earlier, the 120-case data set had been arbitrarily selected because there were no missing observations on statutory class of offense at sentencing. A number of the other independent variables had varying frequencies of missing observations ranging from 1.O percent to 40.0percent. The listwise deletion technique employed in the CJRC study, and thus in this replication. greatly reduced the sample size. resulting in the regression statistics’ being based upon only 70 cases. The results of regressing the second construct of sentence length (max-one) on nine independent variables are presented in Table 2. After examining the distributions. correla-

tions, and previous regressions, these variables were entered in the equation. The R’ of 0.440 in Table 2 is statistically significant. However, because of the relatively small number of cases. the adjusted R’ is probably a more correct measure of the amount of variance explained. The adjusted R’ of 0.356 is less significant than the unadjusted R’, yet still meaningful. Only four of the nine independent variables entered in the equation were deemed significant at the 0.01 level of inclusion. These variables were, in order: seriousness of offense at conviction, legal status of offender at time of offense, number of previous incarcerations, and age at first conviction. The first two variables alone account for well over half of the variance explained by this model. In Table 3. the results of regressing max-two (a collapsed version of max-one) on the same set of independent variables are presented. Using the adjusted R’ again, it is noted that these variables explain only 32.1 percent of the variance. However, in this

56

JOHN

D. HEWITT

equation six, rather than four, independent variables are significant at the P < 0.01 level, with employment history and injury to victim being added. It is important to note that not only is the order of variables entered into the equation different in the second treatment, but the unstandardized regression coefficients are very different as well. This also suggests that the relative stability of the coefficients in these equations across the 120-case and 200-case data samples was questionable. An examination of the regression results reveals that only three of the nine variables used in the Denver guidelines were shown to be significant in replication equations. A fourth guideline variable (injury to victim) entered in the equations was included in the stepwise regression but explained no significant amount of variance. Finally, two variables not included in the CJRC guidelines that entered in the replicative equations were found to be significant. The variables were “age at first conviction” and “employment history.” One of the problems inherent in the CJRC analysis was that there were simply too few cases for the multiple regression analysis employed, considering the number of variables, to produce reliable regression coefficients (Harris, 1975:231). For example. while CJRC implies that its regressions were based on 120 cases, this appears questionable, given the listwise deletions that left only 70 cases in the replicative study. A related problem involves the caseto-variable ratio. While no absolute ratio is set by statisticians, a general rule of thumb suggests that there should be somewhere between 30 and 50 cases for each variable entered into a multiple regression equation (Harris, 1975). In a footnote to the CJRC document it is admitted that “the pilot study samples did not meet the rule of thumb ratio of cases-to-variables” (Wilkins et al., 1978:83). To meet this general rule, at least 270 cases would have been required for the regression equations with only Y variables. and CJRC’s equations with 16 variables would have required a conservative minimum of 480 cases.

and BERT

LITTLE

Perhaps a more critical problem with the data involves the use of a number of highly skewed variables in the analysis. The most highly skewed variable encountered was “number of offenses of which offender was convicted.” Slightly over 94 percent of the offenders in the 120-case data set were convicted of a single offense. In addition, eight other dichotomous variables were split beyond the 40-60 rule of thumb. with three falling outside an SO-20 split. The inclusion of highly skewed variables produces very erroneous results, such as distorted R’ statistics. This will also affect the correlation for the regression equation. The F statistic will. in addition, be misleadingly higher for the skewed variables (Hertel. 1976:459-74). Thus the skewed variables will enter the equation. deleting the normally distributed variables. A further problem identified with the data provided by CJRC concerns the representativeness of the “nonrandom” sample of 120 cases. An examination of the correlation matrices. the R”s in the various equations. and the frequency distributions strongly suggests that the 120-case data set was not representative of the 200 cases. It is obvious that the difficulties with extensive missing observations in the data set collected for the CJRC pilot study severely limit the reliability of any replication attempts. The shrinkage in number of cases resulted from the use of the listwise deletion technique for handling missing observations. The listwise technique is one of the three alternative methods for multiple regression in the SPSS software package used by the CJRC team in its analysis. The two other methods of dealing with missing observations. pairwise deletion and inclusion of missing values. are generally regarded as less reliable treatments (Hertel. 1976:460-62). It is therefore understandable that CJRC selected the listwise procedure from these alternatives. However. given the limitations of listwise deletion. a more rigorous method for estimating the missing observations was undertaken in the replicative study.

Examining

REANALYSIS

the Research

Underlying

the Sentencing

WITH ESTIMATED DATA

The estimation of missing data has been advocated in recent publications for a variety of reasons (Dixon and Brown. 1979). The important advantage of estimating the missing data is the increase of data sample wholeness. With a more complete data sample. the statistical analyses are enhanced whereas. with the crude solution of case deletion. sample size becomes a critical issue. With the estimation of missing data, the unnecessary discarding of the available data is avoided. Thus the assumptions made on the basis of statistical analyses are based on a larger sample. Consequently, interpretations of the statistics are made on a more tenable basis. However, it should be noted that there are some disadvantages to the estimation of missing data. For example, highly skewed variables will cause exaggerated estimates and misleading statistics and will affect the overall estimates if used as predictor variables. More specific problems are encountered in the methodological assumptions underlying the methods of estimation. Briefly. there are five methods of estimating missing data available in the BMDP software package (Hertel, 1976). The estimations may be made on the basis of group means, or four variations of regression. By far. the regression techniques are the most reliable (Dixon and Brown, 1979), and best suited for the data in this study. The multiple regression technique “step”” was used in the replication study to estimate the data because there were linear relationships among the variables and clusters of correlations. The missing observations are estimated by using all variables which meet the F-to-enter criterion. A stepwise regression sequence is then used to select among the variables to be used as predictor variables. Thus a stepwise regression is performed. and each variable within each case containing a missing observation is dependent. By employing a very useful option of the BMDP software. the estimated values may replace the missing values and be written in a data file for use in further statistical

Guidelines

Concept

in Denver.

Colorado

57

analysis. Using this option, it was possible for the replication study to perform analyses on a more complete data set. The replication analyses made on the estimated data set and the subsequent interpretations are statistically more tenable than those made on the original 200-case data set. The first step in the replicative analysis with the estimated data involved stepwise regression. with the sentence variable defined as the maximum number of months to be served. The independent variables used in the CJRC regressions were entered in the equation (see Table 4). A total of six variables met the F-to-enter inclusion level of 2.67. The six variables (number of previous incarcerations, weapon usage, narcotics abuse. injury to victim, offense severity, and parole revocation history) accounted for 39 percent of the variance. In examining both the change in R'and the beta coefficients it can be seen that the variable, “number of prior incarcerations,” is by far the most influential variable entered, accounting for almost one-third of the explained variance. With the sentence decision treated as a dichotomous “in-out” decision (see Table 5), slightly different results were obtained. Again, six variables entered the equation, but only slightly over 35 percent of the variance was explained. The “number of prior convictions” variable alone accounted for 26 percent of the variance. Offense severity, weapons usage, and number of prior incarcerations were the only variables consistently significant in all equations. Two additional variables that were identified through earlier regressions and an examination of the correlations matrix were added to the equation. These variables were “offender liberty status between arrest and sentencing” (a measure of being released on bail) and “number of juvenile arrests.” Both of these variables met the F-to-enter criterion and were entered in the equation (see Table 6). These two variables force “offender liberty status at time of arrest” and “number of prior incarcerations” out of the equation. Also, “offender liberty status between arrest and sentencing” enters the equation first and alone accounts for 27

-

-

- ._,

-. -

-

*This equation included all independent Multiple R = 0.624 Multiple RZ = 0.389 St. error of estimate = 39.463 F = 20.506 df = 6,193 P = < 0.001

Number of prior incarcerations Weapon usage Narcotics abuse Injury to victim Offense severity Parole revocation history

Variable

,_. -

variables

-.

(except

-

“number

0.223 0.284 0.330 0.360 0.379 0.389

,_, -

_ -. ,_,

of offenses convicted”)

0.223 0.061 0.046 0.030 0.019 0.010

.,

R2 Change

._ . -

from the original

tables.

_ ,_,

CJRC

11.322 31.436 21.889 -24.386 - 3.138 9.388

B

.”

STEPWISE REGRESSIONWITH MAXIMUM NUMBER OF MONTHS TO BE SERVED AS DEPENDENT VARIABLE (200 CASES-ESTIMATED DATA)*

TABLE 4

_

0.341 0.283 0.205 -0.180 -0.155 0.129

Beta

TABLE 5 STEPWISEREGRESSIONWITH SENTENCEAS“IN-OUT” DECISION AS DEPENDENT VARIABLE (200 CASES-ESTIMATED

Variable

Number of prior convictions Offense severity Probation revocation history Weapon usage Offender liberty status at time of arrest Number of prior incarcerations *This equation included all independent CJRC Tables. Multiple R = 0.596 Multiple RZ =0.355 St. error of estimate =0.380 F = 17.699 df = 6,193 P = < 0.001

DATA)*

R’Change

B

Beta

0.260 0.294 0.325 ‘0.336

0.260 0.034 0.031 0.011

-0.057 0.026 -0.166 -0.138

-0.297 0.140 -0.154 -0.132

0.346 0.355

0.010 0.009

-0.119 -0.043

-0.106 -0.140

R2

variables (except “number of offenses convicted”)

from the original

TABLE 6 REVISEDSTEPWISEREGRESSIONWITH SENTENCEAS “IN-OU? ESTIMATEDDATA)*

Variable

Offender liberty status between arrest and sentencing Number of prior convictions Number of juvenile arrests Probation revocation history Offense severity

R’

R’Change

B

Beta

0.273 0.395 0.424 0.433 0.442

0.273 0.122 0.029 0.009 0.009

-0.396 -0.052 -0.044 -0.130 0.018

-0.328 -0.269 -0.169 -0.120 0.096

*Equation added “offender liberty status between arrest and sentencing” variables in equation for Table 5. Multiple R = 0.665 Multiple R’ = 0.442 St. error of estimate = 0.352 F = 30.708 df = 5.194 P < 0.001

DECISION (200 CASES-

&nd number of juvenile arrests” to

60

JOHN

D. HEWITT

percent of the variance. In addition, the multiple R2 for the equation increases from 0.355 to 0.442 with the addition of these variables. Finally, offense severity, while entering in this last equation, adds relatively little to the explained variance. It should be noted that the two variables were also added to the equation with sentence length as the regressor, but that neither reached the significance level for inclusion.

CONCLUSIONS While it is difficult to reconstruct all of the empirical procedures used in a project done by others, it is even more difficult to reconstruct the thinking, motives, and unwritten goals that also evolved in a project. It appears clear that, in the development of the Denver guidelines, two primary goals guided the project: first, the identification of a limited number of variables that would clearly predict past sentencing decisions. By so demonstrating, it could be shown to the judges that they were, in fact, not using all of the information usually requested and presented to them in the Presentence Information Report. Rather, accurate knowledge of between only 6 and 12 items would be sufficient to make decisions similar to those they had already been making (Wilkins et al., 1978:26). A second goal implied in the CJRC report was to provide only descriptive, not prescriptive, models to the judges so that the judges themselves could examine, evaluate, and possibly reconsider the kind of information that they had been using: “They summarize expected sentences in a given jurisdiction on the basis of recent practice, and they indicate the relative weights given to what apparently are the most important factors considered. They do not tell what either the sentences or the criteria oughr to be” (Wilkins et al., 1978:31-32). But while these models were forwarded as only a descriptive tool, the report concluded: “It is, finally, our view that once the judges of a given jurisdiction are accurately informed as

and BERT

LITTLE

to what they have been doing in the past. then they can more clearly focus on what they should do in the future” (Wilkins et al.. 1978:32). The first goal of the CJRC project was obviously accomplished and empirically supported by the current replication. The accomplishment of the second goal is very much more open to question. It is important that, when one is telling a group of judges that the following x number of variables have been empirically identified as the variables which best predict their past sentence decisions and that the relative weights of the variables show their relative importance, such statements truly be empirically accurate and supportable. Admittedly, the replicative study found some empirical support for the selection of some of the variables used in Denver sentencing guidelines. However. it would be misleading to suggest that this support confirms that those variables were adequate predictors of past sentencing practices in the Denver courts. The data base used by the CJRC group for its analysis was simply inadequate. There were too few cases and too many variables for reliable multivariate analyses. There were also extensive and biasing missing observations across cases. Further, many of the variables were either poorly structured or redundant, or both. Multicollinearity was present among critical independent variables, and many of the variables had such highly skewed distributions that they had to be excluded from the analysis. The second phase of the replicative study was able to enhance the existing data by a complex multivariate regression program for estimating the missing data. This eliminated those problems imposed by missing observations and small numbers of cases. The variables included in the guidelines are not statistically strong enough to explain the variance in the sentences imposed. In addition. the single best predictor of the “in-out” decision. “offender liberty status between arrest and sentencing,” was not included in the guidelines. The exclusion of

Examining

the Research

Underlying

the Sentencing

this variable from sentencing guidelines is not only understandable. but obviously appropriate. The point. however. is that this variable has’ apparently been an important one in the past. one which judges subjectively or unconsciously used. The guidelines, however, ignore it. In conclusion. it would appear that in developing sentencing guidelines for the Denver courts. relatively sophisticated research informed the judges that a limited number of variables seemed to be consistent indicators of past sentencing patterns (i.e.. items found in the CJRC study. the present replication study, as well as most other studies of sentencing). Unfortunately. the overall variance explained by these variables is rather small. The guideline variables as implemented by the Denver Court are shown in Appendix A. As can be seen. these are not a one-to-one-map of either the first or second sets of regression equations. It is unfortunate that the specific process of selecting those variables eventually used in the guidelines is left unclear in the CJRC report. If the initial research was used merely to indicate to the judges that a small number of variables is sufficient to look at in sentence decisions, then the CJRC activity was a useful exercise. To the extent that the judges were influenced in their decision of selecting the final variables to be included in the guidelines by the rather questionable data presented in the CJRC report. then the final variables may have been inappropriately biased. Inasmuch as it is not possible to know what the judges were thinking at the time or how influenced they were by the CJRC research. one can only speculate on the actual relationship between the research activity and the ultimate policy-making activity. Perhaps in the final analysis. courts considering the use of research in developing policy changes should monitor research but not allow it to dictate or even heavily influence the decisions in a policy-making context. In such contexts. research. with its many pitfalls. should only guide. not decide.’

Guidelines

Concept

in Dcnvcr.

APPENDIX

Colorado

61

A

Seriousness of offense Weapon usage Injury to victim Number of prior convictions Number of prior incarcerations Legal status of oflcndcr at time of offense Prior probation revocations Prior parole revocations Employment status Length of employment School status Years of school

ACKNOWLEDGMENTS We wish to thank Todd R. Clear of Rutgers University. Duane F. Alwin of the University of Michigan. and Peter B. Hoffman of the U.S. Parole Commission for their helpful comments and suggestions on earlier drafts of this paper.

NOTES



For a more complete discussion of the process and impact of reducing information items in parole and sentencing decisions see Gottfredson et al.. 1978.

’ “Max-one” used the following year categories: I .2. 3. 3% to 3. 5.7.8.9.9!4 to 10. 15 to 30. “Max-two” collapsed them further into: I. 1%. 5. 7 to 10. 15 to 30. A discriminant analysis was also performed. using the same independent variables with sentence dichotomized as “in-out” as the grouping variable. The results were essentially similar to those presented in the regression statistics in Table 6 (‘.offender liberty status between arrest and sentencing” was again the single most discriminating variable). In addition. these variables correctly classified the grouping variable in 82 percent of the cases. It has recently been learned that an evaluation of the impact of the Denver sentencing guidelines. conducted by the National Center for State Courts. concludes that the guidelines. as implemented. have had oo statistically significant effect on reducing sentence discrepancies (Todd R. Clear, personal conversation). Whether these findings suggest that the principles of guidelines are not reflective of practice. whether the guidelines were implemented on the basis of defective data and variable selection. or whether this merely Indicates that without an effective enforcement mechanism judges will tend to ignore the guidelines. the actual potential impact of guidelines for sentencing is as vet unresolved.

62

JOHN

D. HEWITT

and BERT

Harris.

REFERENCES

LITTLE

R.J. (1975). A primer of multi\~uriare York: Academic Press.

stuns-

rics. New

Bohnstedt, lines

Justice

M. (1979).

source

Parole

and sentencing

Sacramento.

book.

CA:

guide-

American

Institute.

Calpin. J .C. ; Kress. J .M. ; Chandler. M.A. : Margarita, M.; Mitchell-Herzfeld. S.; Gelman. A.M.: and Broderick, B.A. (1978). The analytical basis for the formulation of sentencing policy. Albany, NY: Criminal Justice Research Center. Dixon, biomedical

University

W.J.,

and

Brown,

computer

of California

M.B.

programs,

(1979).

P-series.

BMDP

Berkeley:

Press.

Gottfredson. D.M.: Wilkins. L.T.: and Hoffman. P.B. (1978). Guidelines for parole and sentencing. Lexington, MA: D.C. Heath and Company.

Hertel. B.R. (1976). Minimizing error variance introduced by missing data routines in survey Sociological methods und research analysis. 4:459-74.

Administratton Assistance Law Enforcement (LEAA) ( 1978). Multijurisdictional sentencingguidelines program test design. Washington. DC: U.S. Government Printing Office. Wilkins, L.T.: Kress. J.M.: Gottfredson. D.M.: Calpin. J.C.: and Gelman. A.M. (1978). Sentencmg guidelines: structuring iudicial discretion. Washington. DC: U.S. Government Printing Office.