An Emprcal Analss of the Determnants of Gult Plea Dscount Jose Pna-Sánchez PhD student n Socal Statstcs at the Unverst of Manchester
Eecutve Summar In ths report, I assess the applcaton of the 2007 Sentencng Gudelnes Councl gudelne, Reductons for a Gult Plea, emprcall usng data collected on the Sentencng Councl s Crown Court Sentencng Surve n the ear 2011. I begn b usng an eplorator analss to observe the relatonshp between the level of dscount appled and the stage at whch the gult plea was entered. I then consder the possble mpact of other factors taken nto account when sentencng on the reducton appled for a gult plea. For ths, I specf dfferent models for dscrete data to regress the level of dscount on a broad set of eplanator varables. Results pont towards a substantal degree of agreement between the recommendatons provded n the 2007 gudelne and the actual level of dscount receved b offenders who plead gult. In partcular, the stage based approach recommended n the 2007 gudelne was found to be the major factor determnng the level of dscount appled. However, the results also show there to be a number of departures from the gudelne, such as: a a hgh proporton of cases where the reducton gven was hgher than the mamum recommended level of 33%, wth these anomales concentrated n specfc Courts; b the presence of partcular aggravatng factors, on average, leadng to lower levels of dscount after controllng for the stage when the gult plea was entered; and c the presence of the mtgatng factor remorse, on average, havng a postve sgnfcant effect on the level of dscount. 1
1. Introducton In 2007 the Sentencng Gudelnes Councl, ssued a revsed gudelne for judges to determne the approprate reducton to appl n cases where a gult plea was entered. Ths bult on the then current gudelne whch was ssued n 2004. The new 2007 gudelne mantaned the consderaton that the level of reducton appled to an offender s sentence when the plead gult should contnue to be based on pragmatc justfcaton. In partcular, t descrbes gult plea dscounts as recognton that, b avodng the full court process, a gult plea saves courts from the cost of conductng further admnstratve work and sometmes, the cost of a tral, but also spares vctms and wtnesses from the stress of attendng or gvng evdence at the tral. The man novelt of the 2007 revson was the hardenng of the process used n the 2004 gudelne to determne the approprate level of gult plea dscount. B law, when determnng the reducton to appl to the sentence of an offender who has pleaded gult to an offence, the court must take nto account: 1 the stage n the proceedngs for the offence at whch the offender ndcated hs ntenton to plead gult; and 2 the crcumstances n whch ths ndcaton was gven. The sstem proposed n the revsed 2007 gudelne establshes a clearer process for takng account of the legslaton, whch, n practce, should reduce nconsstenc n the level of dscount appled across dfferent courts and judges. To address pont 1, the gudelne recommends that the level of reducton should be gauged on a sldng scale rangng from a recommended one thrd where the gult plea s entered at the frst reasonable opportunt n relaton to the offence beng sentenced, reducng to a recommended one quarter where the plea s entered after a tral date has been sent, down to a recommended one tenth where the gult plea s entered at the door of the court or after the tral has begun Roberts and Bradford, 2013. Fgure 1: The sldng scale of dscounts defned n the 2007 gudelne Source: Reducton n Sentence for a gult plea, p. 6. To address pont 2 whch specfes that the crcumstances under whch the ndcaton was gven should be taken nto account, the 2007 gudelne dentfes a lst of cases n whch the dscount ndcated b the sldng scale could be modfed see Secton E of the 2007 gudelne. Ths ncludes a recommendaton that n cases where there s overwhelmng evdence pontng at the culpablt of the offender, the mamum amount of dscount should be lmted to a mamum of 20%, whlst later pleas whch mght have otherwse attracted a reducton of 25% or 10% should receve a lower reducton. Another central topc covered b the 2007 gudelne s the clarfcaton that the reducton prncple s derved from the need for the effectve admnstraton of justce and therefore, should not be seen as an aspect of mtgaton. In partcular, the revsed 2
gudelne emphasses that remorse and assstance provded to prosecutng authortes are separate ssues from whether a gult plea was entered and makes clear that the approach to calculatng the reducton should not take account of the latter crcumstances. That s, when decdng on the most approprate length of sentence, sentencers should address separatel the ssue of remorse, and an other mtgatng features, before calculatng the reducton for a gult plea. Ths process was establshed to prevent sentencers from double countng for an mtgatng factors when dervng the fnal sentence. In ths paper I analse emprcall the data from the Crown Court Sentencng Surve CCSS on for sentences passed n 2011 where the offender entered a gult plea. The am s twofold: to test whether dscounts for a gult plea are consstent wth the recommendatons provded n the 2007 gudelne; and to dentf whether other factors taken nto account when sentencng have a sgnfcant effect on the level of dscount receved. In the followng secton I provde an overvew of the data used, coverng ts strengths and weaknesses. The results of m analss are then presented n Secton 3, whch s dvded nto three subsectons: the frst showng results from an ntal eplorator analss; and the second and thrd showng the use of an ordered and two bnar logt model, respectvel. Secton 4 concludes wth a summar of the man fndngs. 2. Data The Crown Court Sentencng Surve CCSS s a census surve of all cases sentenced at the Crown Court n England and Wales. It records the ke factual elements of each case that were taken nto account b the judge to determne an approprate sentence for the offender. For eample, nformaton s recorded on the serousness of the offence, the aggravatng and mtgatng factors present, and the number of recent, relevant prevous convctons of the offender. In cases where the offender entered a gult plea, the surve also collects nformaton on the stage of the proceedngs at whch the plea was entered, whether t was entered at frst reasonable opportunt, and the dscount gven. Before ths surve, there was no comprehensve source of data on the determnants of sentences that would allow ths knd of eploraton of sentencng decsons. Ths dataset allows us to eplore gult plea reducton n much more detal than was prevousl possble. As Roberts and Bradford 2013 note, before the ncepton of the CCSS, the onl prevous nsght nto the magntude of pleas dscounts has come from the sentencng statstcs publshed b the Mnstr of Justce. However, these statstcs are ver lmted as the do not allow dfferences to be dentfed between the set of offenders who enter a gult plea to those who do not. For eample, no nformaton on aggravatng or mtgatng factors s collected, and nformaton s onl avalable on whether a gult plea was entered or not, not the level of dscount receved or stage of the plea. On the other hand, lke most other surve data, the CCSS s prone to data qualt ssues that should be kept n mnd. The ke concerns that affect the analss are ponted out below. Frst of all, rather than capturng the eact dscount level receved for a gult plea, the surve uses a rather coarse unt of measure to capture the dscount level no dscount gven, 1% to 10%, 11%,to 20%, 21% to 32%, a thrd, 3
or more than thrd. The use of ths dscrete metrc represents a loss of nformaton snce the varable derved from ths queston s not contnuous but ordnal. In addton, and n spte of the mprovement n the amount of nformaton avalable, there are stll some relevant areas n the stud of gult plea reductons whch are not covered b the CCSS. For eample, there are no questons capturng the presence of overwhelmng evdence or ndcatng whether the case ncluded a Newton hearng 1, both of whch mght legtmatel nfluence the dscount level wthout gong aganst the recommendatons of the gudelne. Although the surve s a census, t s not a mandator requrement and therefore suffers from a sgnfcant problem of mssng data. In 2011, the average response rate of the CCSS was 61%. Ths s partcularl problematc as there s no other source of data that wll allow us to rule out the possblt that ths mssng data s non-gnorable 2. In partcular, t could be argued that judges who thnk postvel about the sentencng gudelnes mght be the ones wth a hgher response rate. In the analss I also dscarded cases for whch the secton on gult plea dscounts on the form had not been full completed. Where an of the questons on stage, dscount, or frst opportunt were unanswered, the record was removed. These restrctons reduced the sample sze to a total of 40,783 sentences mposed at the Crown Court n 2011. 3. Results 3.1. Eplorator Analss Table 1 shows the dstrbuton of reductons receved accordng to the dscrete categores used on the surve forms. The majort of cases, 64.3% receved a dscount of eactl 33%. A small proporton of sentences receved dscounts outsde of the mnmum 10% and mamum 33% specfed b the sldng scale approach of the 2007 gudelne. In total, 1.6% of receved a dscount of 0% and 6.8% were gven a dscount of more than 33%. Cases where no dscount was receved mght be eplaned where there was overwhelmng evdence pontng to the gult of the offender and the plea was entered at the last possble opportunt. However, the revewed gudelnes do not consder stuatons where the offender mght get more than the mamum 33% dscount. Table 1. Percentage of Cases b Categor of Dscount 0% 1-10% 11-20% 21-32% 33% >33% 1.6% 8.2% 7.6% 11.6% 64.3% 6.8% Lookng at the stage of the proceedngs at whch the gult plea entered, I observe a more unform dstrbuton of cases. Table 2 shows that all categores have 10% cases or more, wth the man concentraton of cases, 45%, ndcated to have entered ther plea at the plea and case management hearng PCMH. 1 These are a small number of cases when the defendant ma, on the da of tral, enter a plea to a lesser charge, and when ths occurs, he or she ma stll beneft from the mamum recommended reducton as ths pont s consdered the frst opportunt to enter a plea to ths specfc allegaton. 2 See Rubn 1987 for a classfcaton of the mplcatons and possble adjustments for the dfferent mssng data mechansms. 4
Table 2. Percentage of Cases b Stage of Plea At Magstrates Pror to PCMH At PCMH After PCMH At Tral 15% 11.1% 45% 1.10% 18.8% In order to see how these dfferent stages match the dscounts that we have observed, n Table 3, I provde a contngenc table wth these two varables. Although the categores avalable on the surve form for both of these varables dffer to the ones defned n the 2007 gudelne, the varables collected stll provde a good pro for the stages and dscounts specfcall referenced n the sldng scale shown n Fgure 1. As epected, Table 3 shows a strong assocaton between earler pleas and hgher levels of dscount. When pleadng gult at the tral, the majort of sentencers 45.1% appled a dscount of 1-10%. For pleas entered after the PCMH but before tral, dscounts are manl concentrated n the three categores 11-20%, 21%-32% and 33%, wth the bggest proporton of cases 36.8% recevng reductons n the range 21-32%. The other three stages, whch represent the earlest ponts of the court proceedngs, have ver smlar compostons. In each of these categores, the vast majort of cases about 80% obtan reductons of 33%. However, dfferences can be seen between these stages when lookng at the proporton that receve dscounts of 21-32% or >33%. The former dscount categor s nversel related wth earler stages, whle the latter s drectl related. In partcular, we can now see that the probablt of obtanng anomalous cases of >33% s hgher when the gult plea s entered at the magstrates court than n an other stage. Table 3. Contngenc Table of Dscounts vs Stages At Magstrates Pror to PCMH At PCMH After PCMH At Tral 0%.5% 0%.4% 1.1%.4% 1-10%.7% 0%.3% 4.7% 45.1% 11-20% 1.7%.4% 1.7% 22.1% 27.2% 21-32% 3.7% 6.3% 13.1% 36.8% 9.8% 33% 77.2% 81.8% 81.6% 33.5% 12.4% >33% 16.2% 11.5% 3.0% 1.8%.1% In order to eamne whether these patterns are general across the full sample or whether the are nfluenced b partcular courts, I now look at the dstrbutons of these two ke varables level of dscount and stage b court. For reasons of space, ths analss s shown graphcall. Append I ncludes hstograms of the number of cases fallng nto each of the dscount and stage categores for each of the 76 Crown Court centres covered b the CCSS Fgure A1 and A2, respectvel. A thrd chart showng the proporton of all gult plea sentences that receved a dscount of >33% for each court centre s also provded Fgure A3. From the two frst hstograms we can observe a strong varablt n terms of the number of volume of sentences b court centres. Court centres such as Dorchester, Taunton, or Salsbur, sentenced ver few cases, 75, 71, and 32 cases respectvel, whlst court centres such as Lverpool returned forms for 2,249 sentences. We can also see substantal varablt amongst court centres n the proporton of cases fallng nto each of the dscount and stage categores. For eample, Mold Crown Court onl recorded a few cases 26 n total, 4% of the gult pleas entered n ths court of gult 5
pleas entered on the da of the tral, whlst Sheffeld Crown Court shows a smlar proporton for pleas entered on the da of the tral to pleas entered at the earl stage of the PCMH, whch as we saw n Table 2, s the stage where most of the pleas are entered. In order to vsualze dfferences between court centres n relatve terms, we can observe the proporton of dscounts receved that were larger than 33% for each court centre. Ths s shown n Fgure A3. Here we can see that the proporton of anomalous >33% reductons s hgher n Nottngham, Derb, or Truro 19.4%, 17.3%, and 14.1%, respectvel than n other court centres. In contrast, for court centres such as Swndon, Stoke-on-Trent, or Wnchester, these cases are eceptonal 2.2%, 2.1%, and 1.5%, respectvel. In summar, n ths eplorator analss I have detected both epected and unepected patterns. On one hand, I have found that the recommended sldng scale lnkng reductons to the stage at whch the plea was entered s largel compled to, wth pleas entered at the earler stages beng more lkel to receve a hgher level of reducton. On the other hand, there stll est a far number of anomales that fall outsde of the gudelne approach, ncludng >33% dscounts whch are not contemplated b the gudelnes. There s also evdence to suggest that the approach taken dffers amongst court centres 3. I now proceed to etend the analss usng nferental statstcs. In partcular, I use an ordered and a multnomal logt model to regress the categor of dscount on a set of eplanator varables drawn from the CCSS dataset. These two models wll be used to test whether gult plea dscounts are unquel determned b the gudelne s sldng scale or whether there are other varables that also have a sgnfcant effect. The aggravatng and mtgatng factors avalable on the surve form dffer accordng the tpe of offence beng sentenced. To ensure that ever record n the sample contans nformaton on the same set of eplanator varables, I have restrcted the analss to sentences passed for offences covered b the offence tpe of assault and other publc order offences. Ths restrcton gves us a sample sze of 9,187 forms for sentences where the offender pleaded gult, and the followng lst of eplanator varables: stage of gult plea at magstrates, pror to PCMH, PCMH, after PCMH, and da of tral used as the reference case, tpe of assault GBH, GBH wth ntent, common assault, affra, and ABH used as the reference case, sentence length, gender, prevous convctons, a group of aggravatng and mtgatng factors remorse, vulnerable, publc worker, sustaned, and drugs, and the dfferent Crown Court centre at whch the sentence was passed wth Alesbur as the reference case. 3.2. The ordered logt model The frst regresson model I present s an ordered logt model where the set of eplanator varables s regressed on the varable capturng levels of dscount. Ths model assumes that the response varable can be nterpreted as an ordnal varable 4. That s, the dfferent levels of dscounts can be ranked from low to hgh, wthout necessarl knowng what the dstances between adjacent categores are. 3 Smlar fndngs on lack of consstenc n the applcaton of gult pleas across Crown Court centres were found n Robertshaw and Mlne 1992. Here, the authors looked at changes n whether a custodal sentence s passed or not n 60 Crown Courts n 1987 and 1988. The found that on average those who pleaded gult had lower custod rates, but that ths was not the case n ever Court. 4 See Append II 6
Results from ths model are ncluded n Table 4 below, wth the ecepton of the coeffcents for the dfferent Crown Courts, whch for reasons of space have been relegated to Append III. I fnd that, as epected, the strongest predctors of sentence dscounts are the dfferent stages where the GP was ntroduced, and whether ths was done at the frst opportunt. Furthermore, most of the addtonal varables ncluded n the model were not found statstcall sgnfcant. In partcular, t s nterestng to note the lack of sgnfcance of sentence length n logs and prevous convctons. These two varables could be understood as proes for severt of the offence and dangerousness of the offender. So, the fact that the are not assocated wth levels of dscount shows evdence on lack of double mtgaton, but t also dsproves the focal hpothess. Ball 2006 used ths hpothess to eplan how dspartes n plea barganng mght be due to judges havng ncomplete nformaton about defendants and ther cases and, thus, rel on a perceptual shorthand to whch the appl ther own bases and nterject stereotpes regardng the dangerousness of a partcular offender. However, other factors have an effect on GP reductons, whch contradcts the gudelnes recommendatons, and shows evdence of both double aggravaton and double mtgaton. The aggravatng factors of perpetratng an assault sustaned n tme and beng under the nfluence of drugs reduce the probablt of obtanng a hgher dscount; whle the mtgatng factor of remorse, and n partcular the nteracton of remorse and enterng the GP at the magstrates court ncrease the probablt of obtanng a hgher dscount. In addton cases of affra and GBH wth ntent also have a postve effect on levels of dscount. It s possble that the sgnfcant negatve effect of the aggravatng factors of a sustaned assault and beng under the nfluence mght be related to the problem of omtted relevant varables. The model does not nclude a varable to capture the effect of reduced dscounts for reasons of overwhelmng evdence. Cases where there s overwhelmng evdence often occur when the prosecuton s able to provde DNA or CCTV evdence aganst the offender. It could be argued that sustaned assaults or those perpetrated under the nfluence of drugs or alcohol mght be more susceptble to beng captured on CCTV. If so, the legtmate negatve nfluence of the presence of overwhelmng evdence would be wrongl pcked up b varables whch ndcate the presence of these two aggravatng factors, therefore negatvel basng ther estmates n the model. The postve effect of remorse represents a general departure from the approach recommended b the gudelne. Regardless of the stage at whch the plea was entered, results ndcate that the presence of remorse, on average, has the effect of nflatng the dscount level receved. Ths departure from the gudelnes mght represent an element of double countng for ths mtgatng factor n cases when the judge has alread accounted for remorse at an earler stage. In addton, b ncludng nteracton effects between remorse and the dfferent stages of plea, I fnd that ths effect s especall pronounced for cases where the plea was entered at the magstrates court. The postve effects for GBH wth ntent and affra are more dffcult to understand. It mght be argued that offences that are generall consdered to be less serous compared to other offences are lkel to show a better agreement wth the gudelnes, whle the most serous cases are subject to deeper and more subjectve 7
delberatons from the sentencer 5 so ma be more lkel to depart from the gudelne. Ths could eplan the departure that we see for cases of GBH wth ntent, snce these are more serous offences than the reference case of ABH, however, ths hpothess does not eplan the postve effect that we see for cases of affra. Table 4. Ordered Logt Model Varable Coeffcent Standard Error Magstrates 4.8.14 Pror to PCMH 5.03.17 PCMH 3.9.08 After PCMH 1.85.09 Frst op. 1.37.06 Log-length -.004.01 Female.07.08 GBH.05.1 Intent.26.06 Common -.002.08 Affra.31.07 Prevous conv. -.04.03 Remorse.27.09 RemMagst..64.18 RemPCMH.05.12 RemProrPCMH.1.23 RemAftPCMH.08.14 Carer -.11.1 Gang.01.06 Vulnerable.09.08 Pub. Worker -.09.1 Sustaned -.12.06 Drugs -.21.05 Thresholds 0% / 1-10% -2.25.29 1-10% / 11-20%.49.28 11-20% / 21-32% 1.81.28 21-32% / 33% 3.21.28 33% / >33% 8.54.30 In bold results whch are statstcall sgnfcant for a 95% confdence level. We have seen that there are a number of factors that have an effect on the dscount level receved for a gult plea outsde of the stage of plea and whether the plea was entered at the frst reasonable opportunt. Ths could show evdence that n some was, current sentencng practce does not full compl wth the 2007 gudelne. However, to assess the magntude of ths level of departure, t mportant to look at the effect sze of the varables found problematc, and compare ths to the 5 Ths s known as the lberaton hpothess, see Ball 2006. 8
effect sze of those varables that we epect should determne the dfferent levels of dscount receved. I llustrate ths comparson graphcall n Fgures 3 and 4. In these fgures I plot the nverse probablt of recevng a partcular level of dscount broken down b the stage at whch the plea was entered Fgure 3. and when the effect of remorse and ts nteractons wth the stage are ncluded Fgure 4. 6 Both n Fgures 3 and 4, the black lne represents the probablt of makng a transton across adjacent levels of dscount for an average offender. For eample, we can see that the probablt of gong from a 0% dscount to at least a 1-10% s.98. As the level of dscount ncreases, the probablt of movng to the net adjacent level of dscount steadl descends, wth the probablt of movng from 21-32% to 33% beng.63. After ths pont, the probablt of obtanng an even hgher dscount >33% descends drastcall to.04. In Fgure 3, the other lnes show these probabltes accordng to the stage at whch the plea was entered. Ths shows that enterng a plea at the magstrates court, pror to the PCMH, or at the PCMH, ncreases the probablt of obtanng a hgher than average dscount, whlst pleadng gult after the PCMH 7 substantall reduces the probablt of obtanng a hgher level of dscount. Fgure 4 shows the same probabltes after ncludng the estmated effect of the mtgatng factor remorse. Comparng ths to Fgure 3 gves us a sense of how the probablt of recevng each of the levels of dscount changes when remorse s present as mtgaton. We can see that, compared to Fgure 3, the dfferent probabltes have been shfted up. In partcular, we see that the red lne, whch depcts cases where the plea was entered at the magstrates court, shows more of an upwards shft than the other lnes. Fgure 3. Inverse Cumulatve Probabltes of Dscounts b Stage 6 Snce the coeffcents from Table 4 use a logt scale, we need to transform them to calculate the probabltes used n Fgures 3 and 4. I do that b takng the eponental of the coeffcents for the dfferent thresholds, stage of plea, and remorse, over one plus that same eponental. 7 Probabltes for gult pleas entered the da of the tral have not been ncluded because, for reasons of multcollneart, the had to be left out of the model. 9
Fgure 4. Inverse Cumulatve Probabltes of Dscounts b Stage and Remorse In summar, although there s an observed effect for remorse, t s mportant to see ths effect n relatve terms. When comparng how much the curves shft when remorse s added to the model wth the shft assocated wth addng the dfferent stages of plea o the model, we can clearl see the effect of the former s mnmal. Moreover, ths concluson could be generalsed to the rest of the varables whch I have found have an unepected effect on the level of gult plea dscount GBH wth ntent, affra, sustaned attack, and under the nfluence, snce none of ther coeffcents n the model s larger than the one found for the effect of remorse. 3.3. Bnar logt models One of the ke assumptons underlng ordnal logt regresson s that the relatonshp between each par of outcome groups s the same. In other words, ordnal logt regresson assumes that the coeffcents that descrbe the relatonshp between, sa, the lowest versus all hgher categores of the response varable are the same as those that descrbe the relatonshp between the net lowest categor and all hgher categores, etc. In other words, the probablt of movng up a dscount categor s the same, regardless of whch dscount categor ou start n. Ths s called the proportonal odds assumpton or the parallel regresson assumpton. Because the relatonshp between all pars of groups s the same, there s onl one set of coeffcents. If ths was not the case, we would need dfferent sets of coeffcents n the model to descrbe the relatonshp between each par of outcome groups. I specf two bnar logt model, the frst analsng the probablt of obtanng 0% dscount aganst all the other levels of dscount, and the second lookng at the probablt of >33% dscounts aganst all the other levels. These two models allow us lookng nto anomalous cases at the etreme of the gult plea dscount levels n more detals, and eplorng partall the possblt of unparallel effects. Results from the frst model are shown n Table 5 below. None of the prevous nteracton effects between remorse and stages of gult plea were found sgnfcant so the were removed from both models. Almost an stage where gult plea was entered granted a lower probablt of obtanng a 0% dscount than enterng the plea at the da of the tral. The onl ecepton s pror to PCMH, whch s estmated wth a 10
much greater standard error than an other coeffcent, hence, renderng the estmate statstcall non-sgnfcant. Ths problem s probabl derved from a reduced number of cases pleadng gult at that stage and gettng a 0% dscount. In addton frst opportunt s not sgnfcant ether. Therefore the tmng of the gult plea s not as mportant to determne whether offenders get 0% dscounts as t was before when I was consderng all the levels of dscount. On the other hand, remorse has a sgnfcant negatve effect, stronger than before. Ths represents that the double mtgaton from remorse s more lkel to occur n transtons from 0% to other levels of dscount, than from one partcular level to the net. Other varables whch are now found sgnfcant are common assault, GBH, and prevous convctons, wth a postve effect, and sentence length wth a negatve effect. Table 5. Bnar Logt Model 0% vs Other Dscounts Varable Coeffcent Standard Error Constant -3.39.31 Magstrates -1.80.41 Pror to PCMH -16.26 414 PCMH -2.26.27 After PCMH -1.27.29 Frst op. -.09.24 Log-length -.19.04 Female.43.25 GBH 1.42.39 Intent.45.27 Common.72.25 Affra.42.24 Prevous conv..31.12 Remorse -.91.21 Carer -.04.33 Gang -.33.26 Vulnerable -.46.40 Pub. Worker -.13.47 Sustaned.10.25 Drugs.17.20 Results whch are statstcall sgnfcant at the 95% confdence level are shown n bold. Results from the bnar logt model lookng at reductons of more than 33% are shown n Table 6 below. Here, the sldng scale s perfectl represented, as the probablt of gettng more than a 33% dscount ncreases for ever earler stage. The double mtgaton effect from remorse s also present, and the negatve effect of drugs whch we saw n the ordered logt model s now twce as bg, whch ndcates that ths erratc result s specall concentrated n the hghest levels of dscount. In addton, just lke n the prevous bnar logt model, sentence length has a negatve and GBH a postve effect. These two effects mght be ndcatng that eceptonal dscounts greater than 33% are gven n stuatons where the offence was categorsed too severel. In such stuatons judges mght be usng gult plea reductons as a tool to 11
rebalance the approprate sentence length wthout havng to change the tpe of offence from GBH to ABG, for eample. Table 6. Bnar Logt Model >33% vs Other Dscounts Varable Coeffcent Standard Error Constant -4.72.27 Magstrates 2.81.24 Pror to PCMH 2.45.25 PCMH.99.24 After PCMH.45.32 Frst op..38.11 Log-length -.11.02 Female -.15.19 GBH 1.06.22 Intent.08.15 Common -.02.18 Affra -.01.15 Prevous conv..07.07 Remorse.54.11 Carer.15.20 Gang -.08.16 Vulnerable -.16.20 Pub. Worker -.21.26 Sustaned -.07.14 Drugs -.40.12 Results whch are statstcall sgnfcant at the 95% confdence level are shown n bold. 4. Dscusson In ths report I have emprcall assessed the applcaton of the 2007 gult plea gudelne. I have run an eplorator analss usng gult pleas recorded b the CCSS n 2011, and focusng on the relaton between levels of dscount and the stages when the gult pleas were entered. In a second part, I have used dfferent models for dscrete data to regress levels of dscount on a broad set of eplanator varables offences of assault. Results pont at a substantal degree of agreement of gult plea sentences wth the recommendatons ncluded n the 2007 gudelne. In partcular, the sldng scale recommended n the 2007 gult plea gudelnes was found to be the major factor determnng levels of dscount. However, I have also obtaned results that show departures from the gudelnes. In the eplorator analss I found that 6.8% of the gult pleas were resolved wth reductons hgher than 33%. Ths s problematc snce the gudelne does not recommend levels of reducton beond the 33%. In addton, I have found that the orgn of such anomalous dscounts vares substantall across Courts, and t s especall concentrated n the Courts of Nottngham and Derb. 12
From the models that I used I obtaned addtonal evdence on departures from the gudelnes. In the ordered logt regresson model I found that after controllng for the stages when the gult plea was entered, tpes of offences such as GBH wth ntent or affra were assocated wth hgher dscounts, whle aggravatng factors such as sustaned assault or under the nfluence of drugs were assocated wth lower dscounts. I also found that the mtgatng factor of remorse has a postve and sgnfcant effect, showng evdence on double mtgaton. Furthermore, ths effect s twce as strong when the gult plea was entered at the Magstrates. Fnall, from the bnar logt models that was used, I showed that the sldng scale s not beng appled n the lowest level of dscount 0%, and that the reduced dscounts from drugs seem to appl onl to cases that got more than a thrd dscount. 13
References Ball, J. 2006. Is It a Prosecutor s World? Determnants of Count Barganng Decsons. Journal of Contemporar Crmnal Justce, Vol. 22, pp. 241-260. Roberts, J. 2013, Complng wth Sentencng Gudelnes: Latest Fndngs from the Crown Court Sentencng Surve. In Ashworth, A. and Roberts, J. eds. Structured Sentencng n England and Wales: Perspectves on the Defntve Gudelnes. Oford: Oford Unverst Press. Robertshaw, P. and Mlne, A. 1992 The Gult Plea Dscount: Rule of Law or Role of Chance? The Howard Journal, Vol. 31, No 1, pp. 53-75. Scott Long, J. 1997, Regresson Models for Categorcal and Lmted Dependent Varables. SAGE: London. Sentencng Councl 2012 Crown Court Sentencng Surve. London: Sentencng Councl of England and Wales. Sentencng Gudelnes Councl 2004 Reducton n Sentence for a Gult Plea. Gudelne. Sentencng Gudelnes Secretarat. Sentencng Gudelnes Councl 2007 Reducton n Sentence for a Gult Plea. Defntve Gudelne. Sentencng Gudelnes Secretarat.
Append I. Eplorator Analss b Courts Fgure A1. Hstogram of Gult Plea Dscounts b Court 15
Fgure A2. Hstogram of Gult Plea Stages b Court 16
Fgure A3. Hstogram of the Percentage of >33% Dscounts over Total Dscounts b Court 17
Append II. Specfcaton of the Ordered and Multnomal Logt Models Frst I present the process used to specf the ordered logt model. For that I draw from Scott Long 1997 8. The ordnal regresson model can be derved from a measurement model n whch a latent varable Y rangng from - to s mapped to an observed varable Y. The varable Y s thought of as provdng ncomplete nformaton about an underlng Y accordng to the measurement equaton: m f m1 m for m = 1 to J The s are known as thresholds. The etreme categores 1 and J are defned b open-ended ntervals wth and. So, n m case, the gult pleas 0 reductons captured b the CCSS form, that s the observed s related to accordng to the measurement model: 1 0%... f 1. 2 111%... f.. 1 2.. 3 11 20%... f.. 2 3.. 4 21 32%... f.. 3 4. 5 33.3%... f.. 4 5. 6 33.3%... f.. 5.. The structural model s specfed as follows, j where and β are vectors representng the eplanator varables and ther respectve coeffcents, and ε s a vector of resduals. However, snce I do not observe I cannot estmate equaton 2 drectl. The estmaton process uses mamum lkelhood under the assumpton that the error term follows a logstc dstrbuton wth a mean of 0 and varance π 2 /3. So, under ths assumpton, the cumulatve dstrbuton functon s epressed as, 1 2 ep 3 1 ep Once the dstrbuton of the errors s specfed the probabltes of observng values of gven can be computed as the area of the dstrbuton of resduals between two thresholds equaton 1. I llustrate ths process usng the eample of dscounts of 1-11%, that s, when = 2. Followng the measurement model defned n equaton 1, I observe = 2 when falls between 1 and 2. Ths can be more formall epressed as, Pr 2 Pr 1 2 substtutng from equaton 1 and rearrangng I obtan, Pr 2 Pr 2 Pr 1 8 Chapter 5 n Scott-Long 1997 covers ordered logt and probt models. Multnomal logt and probt models are ncluded n Chapter 6.
19 ths s equvalent to the condtonal cumulatve dstrbuton functon of resduals equaton 3 delmted b the two thresholds, 2 Pr 1 2 Ths process can be generalsed for the rest of the levels of dscount. So, for m model wth 6 observed outcomes the fnal ordered model could be specfed b the followng sstem of equatons, 1 6 Pr 5 Pr 4 Pr 3 Pr 2 Pr 1 Pr 5 4 5 3 4 2 3 1 2 1 4 The bnar logt model The bnar logt model can be understood as a smplfed veron of the ordered logt where onl two levels and one threshold are consdered. Lke n the ordnal logt model, the bnar logt models use a latent varable to map the observed values 0,1. So, we have the followng smplfed verson of equaton 1 for the two logt models that I run n ths paper,.... 0. 0%... 1 f other f.... 0. 33%... 1 f other f
Append III. Court Effects from the Ordered Logt Model Varable Coeffcent Standard Error Bradford.41 0.31 Hull.91 0.37 Brmngham.17 0.29 Bournemouth -.04 0.39 Dorchester.93 0.55 Brstol 1.21 0.32 Burnle.5 0.33 Cambrdge.91 0.43 Cardff 1.15 0.31 Carlsle.65 0.34 Central -.19 0.45 Chelmsford.01 0.33 Chester.93 0.36 Chchester.78 0.4 Coventr.64 0.36 Crodon.62 0.43 Derb 1.33 0.3 Doncaster -.06 0.39 Wolverhampton.45 0.3 Durham.73 0.33 Eeter.56 0.37 Gloucester.77 0.4 Great Grmsb.36 0.32 Ipswch 1.47 0.37 Kngston.62 0.39 Blackfrars -.39 0.46 Leeds.19 0.3 Lecester.98 0.3 Lewes.83 0.33 Lncoln.88 0.35 Lverpool.12 0.28 Madstone -.005 0.34 Manchester Cr..5 0.32 Manchester M..92 0.29 Merthr Tdfl.67 0.32 Mold.43 0.31 Newcastle.45 0.28 Inner London 1.52 0.43 Northampton.89 0.63 Norwch.86 0.34 Nottngham 1.17 0.3 Oford.46 0.38 Plmouth.01 0.34 20
Portsmouth 1.1 0.39 Preston.58 0.3 Readng.39 0.35 St. Albans.88 0.38 Sheffeld.37 0.29 Shrewsbur.08 0.4 Snaresbrook.5 0.3 Southampton 1.19 0.41 Stafford.25 0.32 Stoke-on-Trent.43 0.33 Swansea.71 0.3 Swndon.53 0.48 Taunton.51 0.55 Teessde 1.21 0.37 Basldon -.26 0.34 Warwck.57 0.35 Wnchester.08 0.41 Worcester.18 0.34 York.008 0.32 Harrow.77 0.37 Wood Green.97 0.33 Bolton.99 0.31 Southwark 1.25 0.42 Woolwch 1.21 0.41 Peterborough.09 0.35 Guldford.35 0.38 Isleworth.79 0.32 Luton.54 0.34 Truro -.1 0.4 Newport.31 0.5 Canterbur.35 0.38 Salsbur 1.07 1.32 In bold results whch are statstcall sgnfcant for a 95% confdence level. 21