R. J. V AUG H AN B.Sc., M.Eng.Sc., M. E., F. S.S., Australia n Road Research Board Research Fellow, Universi ty of Adelaide G. G 0 U L B ERG H A.R.M .T.C. Cert.T. P.C., Research O fficer, Highways Department, South A ustralia A TWO-STAGE M ODEL FOR TRAFF IC A CCIDENT PREDICTION * (Pa per No . 723) The major street system of metropolitan Adelaide has been divided into some 250 road sections for the purpose of accident evaluation. The object of the study was to determine the worst sections. For each section the million vehicle miles (m.v.m.) was calculated and plotted against accident damage. A linear relationship was shown to exist between m.v.m. and accident damage; from this a weight was assigned to each section to make a comparison between sections possible. Sections with major geometric differences were not compared and the virtue of the system lies in that it can detect minor differences in road geometries to which the variation in accident damage is attributed. INTRODUCTI ON 1. To date little effective use has been made of the vast data on road accidents. The authors feel that the data can, and should be, used as a feedback mechanism in traffic engineering, to reduce accidents. The method outlined in this paper compares road sections (excluding intersections), to detect the minor geometric differences between them, to which accidents (excluding accidents at intersections) , may then be attributed. At present the only method to determine the effectiveness of a traffic engineering measure is to conduct a before and after study (Ref. 1). 2. It is assumed that all road sections with major geometric differences can be stratified (e.g. type 1 = hilly, type 2 = straight single carriageway, etc.) . In thi s paper, which deals only with metropolitan roads, the road sections are more or less homogeneous, so no stratification was made. After stratification, road sections cannot be compared until allowance is made of their relative accident exposure. As a measure of exposure the million vehicle miles (m.v.m.) travelled on the road section was used in this study. 3. For a given exposure, models for the expected distribution of fatal accidents, personal injury accidents, property damage accidents, total accidents and dollar damage were tested. 4. Tolerance limits can be constructed on any of these above factors, but here they are tested on weighted combination described in the next section . • ACKNOWL EDGEMENT - The authors would like to thank ti,e A .R.R.B. fo r their fin ancial assistance, the Comm issio ner fo r H ighways for his enco uragement, and Mr. W. N. Vena bles of the Stat istics Depa rtment, Adelaide University for invaluable statistical advice. Volume 5, Part 3, 1970 365 VAUGHAN, GOULBERGH - MODEL FOR TRAFFIC ACCIDENT PREDICTION DESCRIPTION OF DATA 5. The data used are those from a report by Goulbergh (Ref. 2) Road sections (excluding intersections) are listed with their corresponding m.v.m., number of fatal accidents, number of personal injury accidents, number of property damage accidents, and the total dollar damage of property. 6. Graphs of m.v.m. against number of personal injury accidents, and m.v.m. against dollar damage are given in Fig. 1 and 2. Graphs similar to Fig. 1 and 2 were made of the means and variances (for sections with similar m.v.m.) against m.v.m. 7. It was fo und that the mean number of personal injury accidents increased linearly with m.v.m. and the variance also increased linearly, or possibly quadratically, with m.v.m. The means and variances of dollar damage increased linearly with m.v.m. 8. To make a comparison between type of accidents possible, Goulbergh calculated the weighted damage for each section by allocating $A5000 to each fatal accident, $A1500 to each personal injury accident (which is consistent with English estimates of £ 2970 and £ 770 (Ref. 3)) and adding the dollar damage. USE OF TOLERANCE LIMITS 9. Tolerance limits are used to classify road sections into good, bad, and intermediate. The upper and lower tolerance limits determine the three regions of Fig. 2 in which I, II, III represent the above classifications. To classify a road section, on dollar damage alone, a point is plotted on the graph, with co-ordinates 30 '" . ........ . .. .. .. _........_"-'-" ..... . ., . ~ o ~~.:,. w o 10 MILLION VEHICLE MI LES Fig . 1 - 366 Each road section, Adelaide 1967, plotted with, number of personal injury accidents and m.v.m., as co-ord inates A.R .R.B. PROCEEDINGS VAUGHAN, GOUL BERGH - MODEL FOR TRAFFIC ACCIDENT PREDICTION 4IJ 11 30 If) a:: < -J -J 0 a ~ 20 0 If) a z < If) :::J 10 0 I t- 10 20 MILLION VEHICLE MILES 30 Fig . 2 - Ea ch road section, Adelaide 1967, plotted with, dollar damage and m.v. m., as co-o rdinates, together with the upper a nd lower 10 per cent tolera nce li mits the dollar damage and the m.v.m. for that section. Then the region in which the point lies determines its classification. THE POISSON MODEL 10. The first model investigated was the Poisson model of Norden et al (Ref. 4) used by Goulbergh in his report (Ref. 2). The above method compares each road section with sections on the same road and assumes that the number of accidents has a Poisson distribution . Approximate tolerance limits were given by using the percentage points of the normal distribution (Ref. 4). Am ± UeV Am where m is the m.v.m. of the road section, A is the overall accident rate for the road (total accidents/ total m.v.m.), and Ue is the lOOe per cent point of the normal distribution. 11. Faults in this method are that tolerance limits are inaccurate at low m.v.m. for here the normal approximation will not fit. Also the assumption of a Poisson distribution is invalid for dollar damage as the mean was found to be twice the variance. We tested the Poisson model to find it gave a very poor X2 value even for the number of accidents. 12. Goulbergh in using the Poisson model has obtained approximate tolerance limits which are of the correct form but err slightly in that the percentage tolerance to which they correspond is incorrect. Vo lu me 5, Pa rt 3, 197 0 367 VAUGHAN, GOULBERGH - MODEL FOR TRAFFI C ACCIDENT PREDICTION BASIC MODEL 13. Let us consider a given road section which has exposure x m.v.m. Let Y be a variate to describe the weighted damage, F be a variate to describe the number of fatal accidents, P be a variate to describe the number of personal injury accidents, D be a variate to describe the total amount of property damage, and C be a variate to describe the number of casualty accidents. 14. Ideally Y is calculated by the formula (1) Y = 5000 . F + 1500 P + D But since there were only 30 fatal accidents for the period, the fatal accidents were combined with personal injury accidents and called casualty accidents. The adjusted form of (1) becomes Y = 5000 . Y • C + 1500 P + (2) D where: number of fatal accidents number of casualty accidents The fitted models for C, P, and D are described in the next section . Y= DISTRIBUTION OF PERSONAL INJURY ACCIDENTS 15. The same basic model was used for P and C. It was assumed th at P, for a given x, had a negative binomial distribution with density : gp ( n) = PreP = n . x) . = (h +n- n !(h - 1) !(l 1) !(f3x)n + f3x)h +n (3) 11 = 0, 1,2, 3, .. . with mean hf3x and variance hf3x(1 + f3x ) where h, f3 are parameters and x is the m.v.m. for the road section. This distribution has the required property of a mean which increases linearly with m.v.m . and whose variance can increase linearly or qu adratically depending upon the parameters h, f3. 16. These parameters were estimated by the method of maximum likelihood (see Appendix A) and their values were as follows . 2 h X 17 f3 18.01 10.30 0.072963 Personal injury accidents 15.05 10.02 0.077697 Casualty accidents 46.56 5.870 All accidents 0.5089 TABLE I shows how, for personal injury accidents, the observed medians from the records compare with the expected m~dians calculated by the model. 17. The values for the X2 test of fit were very good for the number of personal injury accidents, and for the number of casualty accidents, but for the total number of accidents a very poor value resulted. The most likely reason for these values is that accidents with very light property damage are not reported , whilst nearly all casualty accidents are reported. 368 A.R.R.B. P ROCEED INGS VAUGHAN, GOULBERGH - MODEL FOR TRAFFIC ACCIDENT PREDICTION TABLE I CO MPAR ISO N OF OBSERVE D AN D EXPECTED M EDIANS Do llar Damage Personal Injury m.v. m. Obse rved Expected Observed Expected 1.0 1.0 1.0 167 1 492 3.0 2.0 2.0 1980 2038 5.0 I 7.0 3.0 3.0 3287 3568 5 .0 5.0 4750 5097 10.0 6.0 7.0 5780 7389 14.0 7 .0 10.0 10500 10445 INTERPRETATION 18. The negative binomial distribution is a member of the family of compound Poisson distributions. The theory of this family and some interpretation of the properties as they apply to traffic accidents, is given by Feller (Ref. 5). 19. To obtain om model we have assumed firstly that a particular person has an accident only as a result of a chance mishap; however some people are more accident-prone than others. For a person with accident proneness, A accidents per m.v.m. will have a mean number of accident AX in X m.v.m., and the probability of just n accidents is Pr(n accidents I AX) = e- J'"(,h)" n! n = 0,1,2, . . . (4) 20. To describe the population exposed to accidents, we give A a distribution, and we have assumed A has a gamma distribution with density e -;·/p;..n- l A> O ,Bhr(li) , with mean h,B (h, ,B are parameters to be estimated from the data). (5) 21. From equations 4, 5, we obtain the negative binomial distribution with density (3) , for all the accidents on the road section for the year. DISTRIBUTI O N O F DAMAG E 22. A model for the amount of damage was built as follows. It was assumed, on the basis of highway department experience, that the amount of damage per accident had a negati.ve exponential distribution. This observation led to the model: D = Dl + D2 + D3 + D4 + D5 + .. . + D k where the D b i = 1.2.3 ... , k are independent damage 'components' each with negative exponential density e-d/ aj ex, where ex is an unknown parameter and k has a Po isson distribution with density Volume 5, P a rt 3, 197 0 369' VAUGHAN, GOULBERGH - MODEL FOR TRAFFIC ACCIDENT PREDICTION e -eX(8xt k = 0, 1, 2, 3, .. . k! where 8 is a second unknown parameter and x is the m.v.m . for the road section. Then D will have density: go(d) = e- ef~(8x) + 8xe: d/a of{2' 8: d) ] (6) where t. is the dirac delta function, and <Xl OF1(a, z) = zn L -- n=O(a)nn! is the of! hypergeometric function , closely related to a Bessel Function. For a description of these functions see Sneddon (Ref. 6). A sketch of some uses of the density (6) is given by Fisher and Cornish (Ref. 7). 23. D then has mean ex8x and variance 2ex 2 8x, so that the mean and variance increase linearly with m.v.m . e, 24. The parameters ex were estimated by the method of maximum likelihood (see Appendix B) and the following values obtained e= r:t. The value of X 2! 7 1·540 = 504 was 26.4, which indicated that the model just fitted. 25 . An explanation of the mediocrity of this fit may be the fact that the total number of accidents does not follow the negative binomial distribution. Another explanation is the apparent overvaluation of light property damage accidents indicated by TABLE I. 26. One can roughly interpret a damage component as being due to a single a(;cident. Alternatively, since exact correspondence does not hold, we can think of it as an accident quantum, with a variable number of quanta to an accident. CONSTRUCTION OF TOLERAN CE LIMITS 27 . The upper 100 per cent tolerance limit for the weighted damage, y, (where y, is the point which is exceeded by E of the total probability) may be determined by the formula: YE = 5000 . }' . CE + 1500PE + de where c,' P., d, are the respective upper 100E per cent tolerance limits of the casualty, personal injury, and dollar damage distributions. They are given by the following equations : = 1- e Gp(p,) = 1 - e Go(d e) = 1 - e (7) G C(c e) (8) (9) where G is the relevant cumulative distribution given by: 370 A.R.R .B. PROCEEDINGS VAUGHAN, GOULBERGH - MODEL FOR TRAFFIC ACCIDENT PREDICTION c GC(C) = L: gC(n) n=O = L: gp(n) n=O (10) p Gp(p) GD(d) = 28. f gD(t)dt (11) (12) To solve (7) and (8) we cumulate the sum on the right of (10) and (11) until l-E is attained. 29. To solve (9) we use the expression for GD(d) given by Fisher and Cornish (Ref. 7): and Newton's method. 30. TABLE II shows the values of y., c., P. and d. for each m.v.m . in the case when E = 0 .1 and 0.9 . The d. curve obtained is that of Fig. 2 and each road section has been plotted to find approximately in which region it lies. Usually the ovcrall tcst cmploying the y. curve would be applied. Computer programmes have been developed to fit, test and plot these 31. tolerance limits, for the successful models, and are available on request. SUGGESTED ALLOCATION OF FUNDS 32. After determining the bad road section how should funds be allocated to each district and road type? We would first form the road sections into groups according to their road 33. type and district, e.g. group 1 = district 1, road type 1 group 2 = district 1, road type 2 group 3 = district 2, road type 1 Now let be the group number, j be the road section number, d ij be the vertical distance to the weighted damage upper tolerance limit if the road is bad, if good set d ij to zero, and (J' ij be the standard deviation determined by the type of road in group i and the m.v.m. of section j. Then we construct a measure for each group c.I = I d .. ---.!.1 j O'ij Then the fraction of funds allocated to a group is given by Volume 5, Part 3, 1970 371 V AUGH AN , GO ULBERGH - MODE L FOR TRAFFIC ACCIDE NT PREDICTIO TABLE II TOLERANCE LIMITS Upper 10% points m .v.m. Casualty Person a l Damage 372 Weighted Damage Injury 1.000 2 .000 2 .000 1950.245 5295.867 2 .000 3 .000 3 .000 3202.941 8221 .374 3.000 5 .000 4.000 4332 .851 11196.907 4 .000 6 .000 6 .000 5402.456 15439.322 5 .000 7 .000 7.000 6434.344 18144.021 6 .000 8 .000 8 .000 7439.765 20822.254 7.000 9 .000 9 .000 8425.289 23480.588 8.000 10.000 10.000 9395 .146 2 6 123.257 9.000 12.000 11 .000 10352 .253 28925 .986 10.000 13 .000 12.000 11298.721 31545.264 11 .000 14.000 13 .000 12236.133 34155.488 12.000 15.000 14.000 13165.717 36757.883 13 .000 16.000 16.000 14088.446 40853.423 14.000 17.000 17.000 i5005 .106 43442 .894 15.000 18 .000 18 .000 15916.343 46026 .943 16.000 19.000 19.000 16822.698 48606.108 17.000 21.000 20 .000 17724.626 51353.658 18.000 22.000 21.000 18622.515 53924.358 19.000 23.000 22 .000 19516.700 56491.355 20 .000 24 .000 23.000 20407.474 59054.939 21.000 25.000 24 .000 21295.089 61615.366 22 .000 26 .000 25 .000 22179.772 64172 .860 23.000 27 .000 26.000 23061.722 66727.620 24 .000 28 .000 27 .000 23941.115 69279.825 25 .000 30.000 28 .000 24818 .112 72002.443 26.000 31.000 30.000 25692.855 76049.998 27.000 32.000 31.000 26565.474 78595.428 28 .000 33 .000 32.000 27436.087 81138.852 29.000 34.000 33 .000 28304 .801 83680.377 30.000 35.000 34.000 29171.714 86220. 101 A.R.R.B. PROCEEDINGS VAUGHAN, GOULBERGH - MODEL FOR TRAFFIC ACCIDENT PREDICTION TABLE II (continued) lower 10% point s m .v. m . Cas ua lty Personal Damage Injury 1.000 0 .000 0 .000 Weighted Damage 0.000 0 .00 2 .000 0.000 0.000 0 .000 3 .000 0.000 0.000 574.386 574.386 4 .000 1.000 1.000 1031.712 2704 .523 5.000 1.000 1.000 1526.938 3199.7 49 0 .00 6 .000 2 .000 2.000 2048 .840 5394.462 7 .000 2.000 2 .000 2590.779 5936.401 8.000 2 .000 2.000 3148.477 6494 .099 9 .000 3 .000 3 .000 3718 .989 8737.422 10.000 3 .000 3 .000 4300.187 9318.620 11 .000 4 .000 4.0 00 48 90.473 11581.718 12.000 4.000 4 .000 5488. 613 12179.858 13.000 5 .000 5.000 6093 .628 14457.684 14.000 5 .000 5 .000 6704.728 15068.783 15.000 6.000 5 .000 7321.262 15858.129 16.000 6.000 6.000 7942 .689 17979.556 17.000 7.000 6.000 8568.552 18778.230 18.000 7 .000 7.000 9198 .461 20908 .138 19.000 8 .000 7 .000 9832.078 21714 .567 20 .000 8 .000 8 .000 10469.113 23851.602 21.000 9 .000 8.000 11109.310 24664. 6 10 22 .000 9.000 9.000 11752.444 26807.743 23 .000 9.000 9 .000 12398.314 27453 .613 24.000 10.000 10.000 13046.743 29774.853 25 .000 10.000 10.000 13697.571 30425.682 26.000 11 .000 11 .000 14350.655 32751.577 27.000 11 .000 11 .000 15005 .865 33406.786 28.000 12.000 11 .000 15663 .083 34236.815 29.000 12.000 12.000 16322.201 36395.934 30.000 13.000 12 .000 16983 .122 37229.666 Volume 5, Part 3, 1970 373 VAUGHAN , GOULBERGH - MODEL FOR TRAFFIC ACCIDENT PREDICTION CONCLUSION 34. The methods contained in this paper will enable a valid comparison of road sections over a wider range of exposure than was possible with the Poisson model ; particularly with respect to rural roads and intersection accidents to which the authors will next turn attention. 35. In comparing the two methods, it was found that they did not differ greatly indicating the robustness of use of the Poisson and normal distributions. The fact that the variance of the number of accidents increases almost 36. linearly with m.v.m . indicates that the increase is due mostly to cross factors (e.g. people crossing the road, side distractions, etc.) and not between vehicles themselves. This confirms a study by Raff (Ref. 8), in which he could find no relation between accidents and traffic volume on freeways, yet there was one on other roads. The method of this paper has shown that the original models for the ex37. planation of general accident behaviour of Greenwood and Yule (Ref. 9) is valid for road traffic accidents. The interpretation of these models is that for a given person, accidents will occur to him in a chance (Poisson) manner. For a given exposure, as each individual has a different accident proneness, the number of accidents occurring to a group of people will form a negative binomial distribution. Therefore each road section will have a different distribution of accidents according to its exposure and the type of drivers using the road section. APPENDIX A MAXIMUM LIKELIHOOD (M.l.) ESTIMATION OF hAND f3 38. Let (nb Xi) be the number of accidents and the m.v.m. for a road section i. From (3) the likelihood function is h L(h, f3:n j , x) = (h + n - l)!p Q"' j = l (h-l)!n j ! fI 1 p.= - - where I 1 + f3Xj Then the log likelihood is InL = c + m L {Iogrch + n j) -logrch) + h .10gP j + njlog QJ i= 1 and 374 A.R .R.B. PROCEEDINGS VAUGHAN, GOULBERGH - MODEL FOR TRAFFIC ACCIDENT PREDICTION Also E - a21nL) - .( afJ2 m hp2 x2 '\' II_t - 1= .L...1 -' Q i 2 E(- a2!n~) = ah -2 = m 11 f: PjX j = t12 = t21 afJoh - a21nL - j=1 1 " ,- I I I --- = j=1 k=1 (n + k)2 t22 We define the information matrix T = (til' t12) t 21, t22 and the score vector (:~) s = and the initial estimate vector where {30 and ho are initial estimates for {3 and h, in our least squares estimates. Then an iterative procedure, A A CX 1 = CXo + T-1s (13) is used to obtain a better approximation ~l to the m.l. estimates of {3, h . APPENDIX B MAXIMUM LIKELIHOOD ESTIMATION OF a AND 0 39. Let (Yb Xj) be the dollar damage and the m .v.m. for a road section i (i = 1, ... , n). From (6) and for convenience letting (3 = l / a, the likelihood function becomes L(O, fJ: Xj , Y;) = " fJex je- OX,- PY 'oF 1(2, (Jex jy;) IT i :::; 1 The log likelihood is InL = n log fJ + n log e + I" j [log Xj + log of 1(2, fJexjy;)] - ne x - nfJy = 1 Noting that Volume 5, Part 3, 1970 375 VAUGHAN , GOULBERGH - MODEL FOR TR AFFIC ACCIDE T PREDICTION we have n olnL I op ceri XiYi - Yi + I jP) = Sl i= l where and where or - i = ap l j I-'R - 2r. j PR - er·2 x.y. or. ---1 = oe 1j e - 2r.1 j e - I>Rr 1.- x 1.y.1 I I I I ? 40. Forming the information matrix T, the score and initial estimate vectors in Appendix A, we may use the same procedure as equation (13) to obtain m.l. estimates of () and f3 = 1/ 0:. a3 REFERENCES l. R esearch on road traffic, H.M.S. O. (London 1965). 2. Goulbergh, G ., Analysis of accident rates for metropolitan roads (for 1967) , S.A. Highw. Dept (Jan. 1969). 3. R esearch on road safety, H .M .S.O. (London 1963). 4. Norden, M., Orlansky, J . and Jacobs, H ., Application of statistical qualitycontrol techniques to analysis of highway accident data, H .R .B. Bull. 117 (1955). 5. Feller, W. , An introduction to probability theory and its applications, 1, 2nd ed. (1957) . 6. Sneddon, I. N., Special junctions of mathematical physics and chemistry, 2nd ed. , Oliver and Boyd (1961). 7. Fisher, R. A. and Cornish, E . A. , The percentile points of distributions having known cumulants, Techometrics, 2: 2, 209 (May 1960) . 376 A.R.R.B . PROCEE DINGS VAUGHAN , GOULBERGH - MODEL FOR TRAFFIC ACCIDENT PREDICTION DISCUSSIONS 8. Raff, M. S., Interstate highway-accident study, H.R.B. Bull. 74, 18 (1954). 9. Greenwood, M. and Yule, G. U. , An enquiry into the nature of frequency distributions with reference to the occurrence of accidents, J. Roy. Stat. Soc., 83,255 (1920). DISCUSSIONS D. J. B U C K LE Y, Ph.D., M.Eng .Sc., B.E., School of Traffic Engineerin g, University of New So uth Wales. 41. It is not immediately clear whether para. 33 refers to personal injury accidents or dollar-damage, but this discussion applies in both cases. 42. Am I correct in assuming that para. 33 defines a criterion which allocates funds so that (roughly speaking and subject to some mathematical reservations) roads having a high dollar-damage per million vehicle-miles will be improved first? This might be called a democratic criterion because it tends to produce an approximate equality of hazard to the individual driver. Opposed to this is the idea of using a minimization of dollar-damage (or accidents) criterion. This involves no mathematics and is equivalent to improving roads having large ordinates on Fig. 2 . Instead of reducing the variability of hazard it seeks to spend public money so as to reduce dollar-damage (or accidents). 43. I am sure that the authors could provide a better non-mathematical description of their criterion. Perhaps they might care to do so. L. A. FOLDVARY , B.E., B.S c. (Econ.), Ph .D., F.S.S .( Hun gary), Rese arch Officer, A.R.R.B. 44. The model for traffic accident prediction prese nted by the authors will be a useful addition to the repertoire of methods at hand , and the authors are commended for their contribution. 45. A little note about a misquotation of Raff in para. 36, i.e. " . .. he could find no relation between accidents and traffic volume on freeways ... " What Raff stated in his quoted paper is the opposite: that there is a direct relationship between the number of accidents and traffic volume on freeways , making accident rates per m.v.m. uniform throughout the volume range. D. C. AND REA SSE N D , B.Sc. , M.Tech ., M.H.F.S., Chief Engin eer, Traffic Commission of Victori a . 46. Have the authors considered investigating the use of traffic density (vehs/ ml) against the accidents per mile on the road section, i.e. using vehs/ ml instead of veh. miles? Work done by D. J. Delaney* concluded in one section that "the occurrence of accidents on a length of street, in a given period might be expressed as a function of the average density of vehicles on that street" and I wondered if the Adelaide data might confirm this? " Del aney, D . J., A pilot stlldy 0/ M elbo llrne's traffic problem , A.R.R.B . Proc. I : 1,388 (1962). Volume 5, Part 3, 1970 377 VA UGHAN, GOULBERGH - MODEL FOR TRAFFIC ACCID ENT PREDICTION A UTHOR'S CLOSURE 47 . I was also pleased to se that the authors have mad a positive point of omitting the accidents at intersections along the road sections they studied; whereas a lot of other 'practitioners' seem to still, incorrectly I feel, cheerfully ignore this aspect. Would the authors clarify which types of accidents remained in their road sections and whether any of these could be attributed to the nearby intersections? That is, has any arbitrary cut off point been used or has some other criterion been used to determine 'intersection accidents '? AU TH O RS' To J. CLOS URE F. M. B R Y ANT (See Introductory Remarks ) 48 . The models of Ashton referred to in Mr. Bryant's comments describe the idiosyncrasies of individual drivers and are only of interest in interpretation of the model. The extension of th e model presented to compare road sections over areas where the weather pattern is non-homogeneous would indeed be important. However, the authors would try to use weather as a measure of exposure together with m.v.m. as used in this paper. 49 . Mr. Goulbergh's paper (Ref. 2) is not difficult to obtain and has been distributed to the A.R.R.B ., State road authorities and libraries. Moreover, all the data has been plotted in Fig. 1 and 2 of our paper to avoid inclusion of very lengthy tables. 50. Mr. Bryant highlights the main point of the paper, namely, to determine those sections of the road system which require a detailed examination by a traffic engineer, who is informed automatically if the change is significant. To D. J. B U C K LEY 51. The authors propose that funds should be spent where they will be the most effective. The further the dollar damage ordinate is above the expected ordinate the more likely the section will respond to treatment. Funds could be allocated in proportion to this difference, if the expected variability about the expected number of accidents was constant. However, the variability increases with m.v.m . and to account for it we divide difference by the standard deviation and allocate funds in proportion. To L. A. F0 L D V A RY 52. The authors thank Dr. Foldvary for his comments. The authors have erred in para. 36 of pre-print, last sentence, 'accidents' should be 'accident rate'. To D. C. AND REA SSE N D 53 . Our paper would confirm that the accidents on a street would be proportion al to the vehs/ mile on that street. 54. All accidents within 30 ft of an intersection were considered to be intersection accidents. 378 A.R.R .B. PROCEEDINGS