Vous êtes sur la page 1sur 14

Applied Economics, 2002, 34, 2353 2365

An ordered response model of test cricket performance


R O B ER T D B R O O K S , R O BE R T W F A F F * { and DAVID SOKULSKY School of Economics and Finance, RMIT, GPO Box 2476V, Melbourne, Victoria 3001, Australia. { Department of Accounting and Finance, Faculty of Business and Economics, PO Box 11E, Monash University, Victoria 3800, Australia

The paper analyses the prediction of test cricket outcomes using an ordered response model. The results, based on data over the period 1994 to 1999, suggest that the ordered categorized production outcome of test cricket (win, draw, loss) can be explained by simple measures of the batting and bowling labour inputs. For example, across all countries the model correctly predicts 71% of test cricket outcomes. Further, it is found that losses are correctly predicted most often at 81% of the sample but that the model faces its biggest challenge predicting test match draws only getting 57% of these cases correct. Also analysed are the circumstances in which the model produces incorrect predictions and it is found that the most common events are unsuccessful last innings runs chases; successful last innings runs chases and rain-a ected matches. An analysis of failed model predictions in terms of country factors suggests that (relative to all other countries) Pakistan has a higher tendency to be involved in such matches, whereas Sri Lanka has a higher tendency to be involved in matches that are `predictable. A `style analysis using this model suggests that ve test cricket styles are evident. Style I is that of `Bowling and Batting Performance and describes Pakistan, the West Indies and (perhaps to a lesser extent) Zimbabwe. Style II is that of `Batting Performance and describes England, New Zealand and (perhaps to a lesser extent) India. Style III is `Bowling Performance and describes Australia. Style IV is `Bowling Performance/Batting Strike Rate and describes South Africa. Finally, Style V is `Bowling Performance and Strike Rate and describes Sri Lanka. Finally, the model is used to analyse which country can claim to be the world champions of test cricket over the sample period. In an initial analysis based on average performance over this period, South Africa has the best claim. However, in a `heavyweight title contest between South Africa and Australia, Australia has the superior claim.

I. INTRODUCTION The growth in professionalism in sport since the 1970s has been a worldwide trend, spread across many sports. As a result, there is now a body of literature which applies the techniques of statistical, mathematical, psychological and economic analysis to a number of facets of professional sport. Examples of the statistical approach include Clarke
* Corresponding author. E-mail: robert.fa @buseco.monash.edu.au

and Norman (1995); Crowe and Middeldorp (1996); Chedzoy (1997); Holder and Nevill (1997); and Magnus and Klaassen (1999a and 1999b). Clarke and Norman (1995) and Holder and Nevill (1997) examined the home advantage in English soccer and in international tennis and golf tournaments, respectively. Magnus and Klaassen (1999a and 1999b) examined four years of Wimbledon data to assess whether there is an advantage to the server

Applied Economics ISSN 00036846 print/ISSN 14664283 online # 2002 Taylor & Francis Ltd http://www.tandf.co.uk/journals DOI: 10.1080/0003684021014808 5

2353

2354
of using new balls and whether there is an advantage of serving rst in tennis. Crowe and Middeldorp (1996) and Chedzoy (1997) investigate the impact of umpiring decisions in test match cricket. Specically, Crowe and Middeldorp (1996) conduct a comparison of leg before wicket rates between Australians and their visiting teams played in Australia over the period 19771994, while Chedzoy (1997) analyses the e ect of umpiring errors. One-day cricket has also received some attention in the literature and represents a good illustration of the mathematical approaches to the analysis of sport. For example, Clarke (1988) employs a dynamic programming approach to analyse optimal scoring rates and Johnson et al. (1993) using the same methodology, assess player performance. In a more recent paper, Duckworth and Lewis (1998) presented a method for resetting the target for interrupted one-day matches a method that currently is being applied to all one-day internationals. Other aspects of cricket have also been investigated in the literature. For example, Clarke and Norman (1998) apply a dynamic programming approach to the issue of protecting the weaker batsman. From a psychology point of view, several papers have been written on cricket. Thelwell and Maynard (1998) investigated the anxiety and performance relationship; Craven (1998) analysed the psychology of the leg-beforewicket rule; and Totterdell (1999) studied the mood and performance relationship. Finally, with regard to the economics approach, in early work, Neale (1964) and Demmert (1973), address the basic issue of whether sporting teams can be viewed as `rms. A natural extension of this work was to apply a production function methodology to professional sports and this was performed in the case of professional basketball (Zak et al., 1979; Scott et al., 1985), professional baseball (Scully, 1974; Zech, 1981) and professional hockey (Jones and Walsh, 1987). In the context of cricket, two of the main papers apply a production function approach to rst class cricket. These are Schoeld (1988) in his study of English county cricket, and Bairam et al. (1990) in their comparative study of the Australian and New Zealand domestic cricket competitions. In their modelling of rst class cricket both Schoeld (1988) and Bairam et al. (1990) seek to explain team success (the production output) as a function of batting and bowling (the production inputs). First class cricket is organized into regular seasons and success is measured either as a percentage of total wins or as a percentage of total points for a given season. In this context the success variable is continuously observed and can be modelled in a conventional regression-based approach. There are problems in generalizing the approach of Schoeld (1988) and Bairam et al. (1990) to test cricket, the peak form of the game. This is due to the fact that test cricket is not organized into regular seasons, but instead is

R. D. Brooks et al.
organized into tours of one country by another country. While some of these tours are on a regular calendar cycle, for example, the Ashes series between Australia and England, other series do not t a regular calendar pattern. This implies that it is not possible to dene a season over which countries play the same number of tests and the same set of countries over a given time interval. This implies that the Schoeld (1988) and Bairam et al. (1990) measures of success will not work well in this context. It also implies that it will be di cult to identify the world champion of test cricket at any point. This has been problematic in the cricket literature since Australias defeat of the West Indies in 1995. Since that time, many countries have had competing claims to be the test world champion. For any given test, there are three possible discrete outcomes, namely a win, a draw and a loss. These outcomes are also ordered, in that a win is superior to a draw, which in turn is superior to a loss. This suggests a possible strategy of applying an ordered response model to the outcome of individual tests as a function of both batting and bowling measures. The fact that a draw is a possible and regularly occurring outcome makes this a non-trivial exercise as it means that a team with superior batting and bowling performance does not necessarily win. The plan of this paper is as follows. Section II outlines the modelling framework in terms of an ordered response model. Section III presents an empirical analysis of test cricket starting from the series in which Australia defeated the West Indies in 1995. This series represents a natural breakpoint because it was the rst time the West Indies had lost a test series since 1980. Thus, from this time onwards, other countries could credibly stake a claim that they were the test world champions. All nine test-playing nations (Australia, England, India, New Zealand, Pakistan, South Africa, Sri Lanka, the West Indies and Zimbabwe) are analysed separately. Section IV utilizes the results of the ordered response model to compare the performance of teams and to draw out information regarding the leading test cricket-playing nations claims to be the world champions. Section V contains some concluding remarks.

I I . M O D E L L I N G F R A M EW O R K Ordered response model The modelling framework employs an ordered response model. The model explains a categorical data outcome (win, draw, loss) where those categories possess a natural ordering. To motivate the ordered response model, consider the latent variable model,
0 y i x i "i

Ordered response model of test cricket performance


where is an unobservabl e latent variable that measures team performance, xi is a vector of explanatory variables, in this case, batting and bowling measures, is a vector of unknown parameters and "i is a random disturbance term. If the distribution of "i is chosen to be normal then we ultimately produce an ordered probit model. This model for the latent variable, yi , generates the observed variable yi . There are three possible observations for yi generated in the following manner, yi Win; yi Draw; yi Loss; if if if yi > W L < y i < W y i < L
yi

2355
win, draw or loss for any given cricket test match. The tted value for any given cricket test is therefore the outcome (win, draw or loss) with the maximum probability. Denition of the independent variables The exact denition and consequent measurement of the independent variables in Equation 8 is not unique however. Perhaps the most important issue here is whether these variables should be measured concurrently with the test outcome indicator variable (that is, taking batting and bowling `performance and `strike rate measures from the current test match only) or whether they should be measured on some sort of historical basis up to but not including any given test match. It seems appropriate that the guiding principle for making this fundamental choice should quite simply be what is the primary aim of the research? If the central research objective is to produce a model of forecasting test cricket outcomes then the appropriate choice for measuring the independent variables would be the `historical base. Alternatively, if the central research objective is one of developing a model that simply seeks to explain current performance then the `concurrent measure of the independent variables is most appropriate. Given the predominantly exploratory nature of the current paper, it seems legitimate to select the explanatory model focus as a starting point. That is, knowing how well current performance can be explained logically warrants investigation prior to the question of forecasting. Moreover, the explanatory model provides a useful research framework for exploring worthwhile subsidiary issues such as the `styles of test cricket di erent countries employ and the determination of which test team has the 1 most legitimate claim to being world champion.

2 3 4

where W and L are unknown parameters to be estimated subject to the ordering rule that W > L . In this context the probabilities for observing the three possible categories of yi are,
0 Pryi Lossjx; ; F L xi 0 0 Pryi Draw jx; ; F W xi F L xi 0 Pryi Winjx; ; 1 F W xi

5 6 7

where, F is the cumulative normal distribution function. To estimate the model we need to select the variables that make up xi . In this context, Schoeld (1988) and Bairam, Howells and Turner (1990) are followed and two batting measures used, runs per wicket (RPW) and runs per over (RPO), and two bowling measures, opposition runs per wicket (ORPW) and opposition runs per over (ORPO). Bairam et al. (1990) describe each variable as follows (a) RPW `batting average; (b) RPO `attacking batting; (c) ORPW `bowling average; and (d) ORPO `defensive bowling. We choose to give them more generic descriptions which will aid a `style analysis of test cricket discussed later in the paper. Specically, RPW and ORPW are classied as `performance measures RPW measures `batting performance and ORPW measures `bowling performance. Likewise RPO and ORPO are classied as `strike rate measures RPO measures `batting strike rate and ORPO measures `bowling strike rate. Therefore the model to be estimated is, yi 1 RPWi 2 RPOi 3 ORPWi 4 ORPOi "i 8 Given the estimated parameters ( ; ) and the observed variables (xi ) one can then calculate the probability of a

III. EMPIRICAL ANALYSIS This section reports the empirical results for each of the nine test playing nations. The results are initially reported sequentially for each of the test cricket playing nations. Then, a comparative discussion is presented in which several major themes are identied. 2 All data is sourced from the Cricket Info website. It should be noted that there are three tests that are omitted from the analysis due to the fact that only a single side batted because of bad weather and/or pitch problems.3

The forecasting objective is of course a very interesting and worthwhile one. The data collection and modelling process is a little more involved than that required in the current paper and is the focus of ongoing research by the authors. 2 The current web address for this site is: http://www-aus.cricket.org/link_to_database/ARCHIVE/. 3 These three tests are India versus New Zealand 2nd Test 1995/6; India versus Sri Lanka 2nd Test 1997/8 and England versus West Indies 1st Test 1997/8.

2356
Table 1. Estimation of ordered response model for test cricket outcomes Batting coe cients Test cricket playing nation Australia England India New Zealand Pakistan South Africa Sri Lanka West Indies Zimbabwe `Performance 1 0.096 (0.000) 0.059 (0.000) 0.025 (0.030) 0.048 (0.042) 0.097 (0.000) 0.001 (0.904) 0.025 (0.031) 0.061 (0.001) 0.120 (0.063) `Strike Rate 2 70.369 (0.601) 70.313 (0.488) 0.591 (0.348) 0.178 (0.787) 1.332 (0.014) 0.178 (0.773) 70.489 (0.303) 1.124 (0.341) Bowling coe cients `Performance 3 70.139 (0.000) 70.020 (0.091) 70.008 (0.380) 70.001 (0.930) 70.064 (0.000) 70.090 (0.000) 70.051 (0.021) 70.042 (0.001) 70.111 (0.045) `Strike Rate 4 70.434 (0.552) 70.546 (0.223) 70.660 (0.108) 70.957 (0.102) 0.570 (0.337) 71.705 (0.012) 0.375 (0.447) 70.120 (0.911)

R. D. Brooks et al.
Hypothesis tests H10 : 240 0.985 (0.611) 1.548 (0.461) 3.032 (0.220) 8.325 (0.016) 6.325 (0.042) 1.310 (0.519) 1.080 (0.583) H20 : 1 3 3.417 (0.065) 5.190 (0.023) 2.081 (0.149) 4.107 (0.043) 3.057 (0.090) 15.394 (0.000) 2.150 (0.143) 1.073 (0.300) 0.031 (0.861)

The table reports the results of estimating the following ordered response model for the nine test playing nations: yi 1 RPWi 2 RPOi 3 ORPWi 4 ORPOi "i

where the dependent variable yi , takes on three ordered categories of test cricket match outcomes: (1) a loss; (2) a draw and (3) a win; and the independent variables comprise two batting measures runs per wicket (RPW) and runs per over (RPO), and two bowling measures opposition runs per wicket (ORPW) and opposition runs per over (ORPO). The table also reports the outcome of tests of the hypotheses, H10 : 2 4 = 0 and H20 : 1 3 . P-values are reported in parentheses throughout.

Australia Data on all of Australias tests from the 1994/5 defeat of the West Indies in the Carribean through to the one-o test against Zimbabwe in October 1999 were used. Over this period of time Australia played 50 tests, producing 26 wins, 10 draws and 14 losses. The results for estimating the ordered response model for Australia are reported in Table 1, and show that the `strike rate measures of RPO and ORPO are insignicant. They also show that improved batting performance (a higher RPW) and improved bowling performance (a lower ORPW) both increase the probability of winning. For Australia there is evidence that bowling performance is more important as the coe cient on bowling is signicantly greater than the coe cient on batting. Table 2 reports the `predictive accuracy of the model for 4 Australia. The model does very well at predicting the wins and losses. We are able to predict 25 of the 26 wins and 10 of the 12 losses. However, the model does not do as well at predicting draws and only predicts four of the 10 draws. An examination of the cases where the estimated model fails to predict the result illustrates the factors not captured
4

in the current modelling. In the case of Australian test matches over the sample period the model fails to predict the result of rain-a ected tests; tests in which there were last innings batting collapses; tests in which the team batting last saves the match and a test in which a record individual innings was made. Specically, rst there were three rain-a ected tests (the rst Australia v England test in 1998/9 and the second and third Australia v Sri Lanka tests in September 1999). Second, there were three last innings batting collapses (Australia v Pakistan 3rd test 1995/ 6, Australia v England 6th test 1997, and Australia v England 5th test 1998/9). Third, there were two tests where the team batting last saved the match (Australia v New Zealand 3rd test 1997/8 and Australia v South Africa 3rd test 1997/8). Finally, there was the test in which Mark Taylor compiled his Australian Test record score of 334 not out (Australia v Pakistan 2nd test in October 1998).

England Data on all of Englands tests from the 1995 series against the West Indies to the 1999/2000 series against South

The issue here is one of `in-sample prediction as opposed to out-of-sample forecasting the latter, as discussed earlier, being beyond the scope of the current paper.

Ordered response model of test cricket performance


Table 2. Comparison of actual test cricket outcomes and those predicted from estimated ordered response models Test cricket playing nation Australia (50 tests) England (49 tests) India (31 tests) New Zealand (33 tests) Pakistan (37 tests) South Africa (42 tests) Sri Lanka (34 tests) West Indies (41 tests) Zimbabwe (25 tests) Aggregates (342 tests) Number of actual occurrences 26 10 14 12 18 19 7 14 10 9 12 12 15 12 10 20 14 8 11 12 11 13 10 18 2 10 13 115 112 115 Number of predicted occurrences Wins 25 4 0 4 6 0 1 1 0 5 2 0 10 3 0 20 1 4 9 1 0 11 3 0 1 1 0 86 22 4 Draws 1 4 2 7 8 1 6 10 4 3 8 2 5 7 3 0 8 0 2 8 4 2 3 1 1 8 1 27 64 18 Losses 0 2 12 1 4 18 0 3 6 1 2 10 0 2 7 0 5 4 0 3 7 0 4 17 0 1 12 2 26 93

2357

Outcome Win Draw Loss Win Draw Loss Win Draw Loss Win Draw Loss Win Draw Loss Win Draw Loss Win Draw Loss Win Draw Loss Win Draw Loss Win Draw Loss

No. correct/ No. of tests 41/50 (82.0%) 30/49 (61.2%) 17/31 (54.8%) 23/33 (69.7%) 24/37 (64.9%) 32/42 (76.2%) 24/34 (70.6%) 31/41 (75.6%) 21/25 (84.0%) 86/115 (74.8%) 64/112 (57.1%) 93/115 (80.9%) 243/342 (71.1%)

The table reports a comparison of the actual and predicted outcomes for each test cricket-playing nation. The three outcomes are a test match win, a test match draw and a test match loss. Predicted outcomes are produced from the estimated ordered response model: 8 yi 1 RPWi 2 RPOi 3 ORPWi 4 ORPOi "i where the dependent variable yi , takes on three ordered categories of test cricket match outcomes: (1) a loss; (2) a draw and (3) a win; and the independent variables comprise two batting measures runs per wicket (RPW) and runs per over (RPO), and two bowling measures opposition runs per wicket (ORPW) and opposition runs per over (ORPO). The predicted value for any given cricket test match is the outcome with the maximum probability according to the estimated equation.

Africa were used. Over this period of time England played 49 tests producing 12 wins, 18 draws and 19 losses. The results for estimating the ordered response model for England are reported in Table 1. The results show that the strike rate measures (RPO and ORPO) are insignicant. They also show that improved batting and improved bowling increase the probability of winning. However, in contrast to the outcome for the Australian model, the bowling measure is signicant only at the 10% level. The

evidence suggests that the batting measure has a more signicant impact on the probability of winning for England. Table 2 reports the predictive accuracy of the model for England. The model does very well at predicting Englands losses where it predicts 18 of the 19 cases. However, the model does not do as well at predicting wins and draws and frequently wrongly chooses between these outcomes. Overall the model is only able to predict four of Englands 12 wins and eight of its 18 draws.

2358
An examination of the cases where the model fails to predict the result illustrates the factors not captured in the current modelling. In the case of England test matches over our sample period, the most common circumstances in which the model fails to predict the result are rain-a ected tests; tests in which both sides make very large rst innings totals; tests in which successful last innings run chases occurred and tests in which unsuccessful last innings run chases occurred. Specically, and rst, the model fails to predict the outcome of three rain-a ected tests (3rd test against South Africa 1995/6; 2nd test against Australia 1997 and 3rd test against New Zealand 1999). Second, there are three tests in which both sides made very large rst innings totals (6th test against the West Indies 1995; 3rd test against India 1996 and 2nd test against Pakistan 1996). Third, there were ve cases in which successful last innings chases occurred (4th test against the West Indies 1995; 3rd test against New Zealand 1996/7; 2nd and 3rd tests against the West Indies 1997/8; and 1st test against New Zealand 1999). Fourth, there were four cases in which failed last innings chases occurred (2nd test against the West Indies 1995; 6th test against Australia 1997; 5th test against South Africa 1998 and 4th test against Australia 1998/9). Fifth, one test draw (5th test against the West Indies 1997/8) could not be predicted because there was insu cient last innings time to either score the runs or to take the wickets. Sixth, there was a drawn test (1st test against New Zealand 1996/7) in which New Zealand completed a match-saving and record 10th wicket partnership. Finally, there were two drawn matches where both teams achieved the same run total (1st test against Zimbabwe 1995/6 and 3rd test against South Africa 1998). India Data on all Indias test matches from the 1995/6 series against New Zealand to the Asian Test Cricket Championship of March 1999 were used. Over this period of time India played 31 tests producing seven wins, 14 draws and 10 losses. Di culty was experienced estimating the complete specication of the ordered response model of Equation 8 for India. The problem was attributed to the joint inclusion of the two batting measures (RPW and RPO). Accordingly, it was chosen to estimate a modied version of the model for India in which the batting strike rate variable (RPO) was omitted. The results of this estimated model are reported in Table 1. The table reveals that only the batting measure (RPW) is signicant for India. Despite this, as reected by the outcome of testing H20 , the impact of the bowling and batting measures are not signicantly di erent from each other in the case of India. Table 2 reports the predictive accuracy of the model for India. As can be seen from the table, the model does rea-

R. D. Brooks et al.
sonably well at predicting draws in that it is able to predict 10 of Indias 14 draws. The model is not as strong at predicting losses in that it only identies six of its 10 losses correctly. However, the model does very poorly at predicting Indian wins in test match cricket only one of its seven victories are correctly predicted. Once more insight can be gained into the modelling process by investigating the 14 cases in which the model fails to correctly predict the actual outcome. First, the model fails to predict the result of two rain-a ected matches (4th and 5th tests against the West Indies 1996/ 7). Second, there were three tests in which successful last innings chases occurred (1st test against New Zealand 1995/6; the one-o test match against Australia 1996/7; and 2nd test against New Zealand 1998/9). Third, there were seven tests in which unsuccessful last innings runs chases occurred (1st and 3rd tests against South Africa 1996/7; 3rd test against the West Indies 1996/7; 1st test against Australia 1997/8; one-o test against Zimbabwe 1998/9; 2nd test against Pakistan 1998/9 and India versus Pakistan in the Asian Test Cricket Championship). Fourth, there was one drawn test (2nd test against Sri Lanka 1997/ 8) involving two large rst innings totals. Finally, there was a drawn test match between India and Sri Lanka (1st test 1997/8) in which Sri Lanka scored a world record six declared for 952 when Jayasuriya made 340 and Mahanama made 225. New Zealand Data on all of New Zealands tests from the 1995/6 series against India in India through to the 1999 series against England in England were used. Over this period of time New Zealand played 33 tests producing nine wins, 12 draws and 12 losses. The results for estimating the ordered response model for New Zealand are reported in Table 1. The table shows that the strike rate measures of RPO and ORPO are insignicant. In contrast to several other countries the bowling measure (ORPW) for New Zealand is insignicant while the batting measure (RPW) is signicant. Given this result it is not surprising to nd that the outcome of testing H20 shows batting to be signicantly more important than bowling for New Zealand. Table 2 reports the predictive accuracy of the model for New Zealand. Here it can be seen that the model does best at predicting losses where it predicts 10 of the 12 cases. Furthermore, the model does a reasonable job at predicting draws and wins it is able to predict eight of the 12 draws and ve of the nine wins. An examination of the 10 cases in which the model produces incorrect predictions again illustrates the factors not captured in the current modelling. First, the model fails to predict the results of two rain-a ected tests (New Zealand v India 3rd test 1995/6 and New Zealand v England 3rd test 1999). Second, there are three unsuccess-

Ordered response model of test cricket performance


ful last innings run chases which the model does not predict (New Zealand v Pakistan 1st test 1996/7, New Zealand v Sri Lanka 2nd test 1996/7 and New Zealand v England 4th test 1999). Third, there are two tests in which successful last innings chases occur (New Zealand v England 3rd test 1996/7 and New Zealand v India 2nd test 1998/9). Fourth, there are two instances where the team batting last had a large total to chase and was not bowled out (New Zealand v Zimbabwe 2nd test 1995/6 and New Zealand v Zimbabwe 1st test 1997/8). Finally, there was a test in which Sherwin Campbell scored a large double century (New Zealand v West Indies 1st test 1995/6). Pakistan In the case of Pakistan, data on all of its tests beginning with the series against Sri Lanka in 1995/6 through to the Asian Test Cricket Championship were used. Over this period of time Pakistan played 37 test matches producing 15 wins, 12 draws and 10 losses. Similar to the situation of India, there was di culty in estimating the ordered response model, although on this occasion it was due to problems caused by including both of the bowling measures (ORPW and ORPO). Accordingly, a modied version of the model was estimated for Pakistan in which the bowling strike rate variable (ORPO) was excluded. The results for this estimated model are reported in Table 1. The table reveals that in the case of Pakistan, the batting strike rate measure is insignicant. Furthermore it is seen that both improved batting and improved bowling increase the probability of Pakistan winning test matches. Based on the outcome of testing H20 (signicant at the 10% level), there is some weak evidence that batting is more important than bowling. Once again, Table 2 reports the predictive accuracy of the model for Pakistan. It can be seen from the table that the model does a reasonable job of predicting all three types of outcome for Pakistan. Specically, the model is able to correctly predict 10 of Pakistans 15 test wins, seven of its 12 draws and seven of its 10 losses. An examination of the 13 cases in which the model provides incorrect predictions for Pakistan reveals the following. First, it is found that the model fails to predict the result of one rain-a ected match (2nd test against South Africa 1997/8) and one fog-a ected match (2nd test against Zimbabwe 1998/9). Second, there are six cases in which the model could not correctly predict unsuccessful last innings runs chases (2nd test against Sri Lanka 1995/6; 3rd test against Australia 1995/6; 1st test against New Zealand 1996/7; 3rd test against South Africa in Pakistan 1997/8; 2nd test against South Africa in South Africa 1997/8; and 2nd test against India 1998/9). Third, there was one drawn match against Zimbabwe (1st test 1997/8) in which there was insu cient time to score the runs or to take the wickets. Fourth, there was a single

2359
test match in which Pakistan defeated India (1st test 1998/9) by the very narrow margin of 12 runs. Fifth, there was a Pakistan victory over Zimbabwe (2nd test 1997/8) chasing a score of 192 in which they lost 7 wickets. Sixth, there was the test match in which Wasim Akram scored 257 not out against Zimbabwe which helped move Pakistans rst innings score from 6/183 to 553 all out (1st test 1996/7). Finally, we again have the test match between Pakistan and Australia in which Mark Taylor scored 334 not out.

South Africa For South Africa the data sample comprises all its tests beginning with the one-o match against Zimbabwe in October 1995 through to the series against England in 1999/2000. Over this period of time South Africa played 42 tests producing 20 wins, 14 draws and eight losses. Table 1 shows the results of estimating the ordered response model for South Africa. Unlike other countries, the South African estimated model shows that while the conventional batting performance measure (RPW) is not signicant, the batting strike rate measure (RPO) is signicant. However, consistent with several other countries the conventional bowling performance measure (ORPW) is signicant, while the bowling strike rate measure (ORPO) is not signicant. Once again, Table 2 reports the predictive accuracy of the model this time for South Africa. It can be seen from the table that the model does extremely well at predicting South Africas test match wins, correctly predicting all 20 cases. However, the model does struggle to correctly predict matches that South Africa draws (loses), achieving an eight (four) out of 14 (eight) success rate. An examination of the 10 cases in which the model provides incorrect predictions for South Africa reveals the following. First, the model fails to predict the results of two rain-a ected tests (2nd test against Pakistan in Pakistan 1997/8 and 1st test against Pakistan in South Africa 1997/8). Second, there are three failed South African run chases that the model does not correctly predict (1st test against India 1996/7; 2nd test against Pakistan in South Africa 1997/8; and 5th test against England 1998). Third, there was a successful Australian run chase (2nd test 1996/7) that was not predicted. Fourth, there were two occasions in which there was a lack of time to score the runs or to take the wickets in the nal innings (4th test against England 1995/6 and 1st test against Australia 1997/8). Fifth, there was the 1st test match between South Africa and Pakistan in 1996/7 in which both sides made very large rst innings totals. Finally, there is the 1st test match against New Zealand in 1998/9 in which Darrell Cullinan scored an unbeaten 275 and the game was drawn.

2360
Sri Lanka Data on all of Sri Lankas tests from the Sri Lanka v Pakistan series in Pakistan in 1995/6 to the Sri Lanka v Zimbabwe series in 1999/2000 were used. Over this period of time Sri Lanka played 34 tests producing 11 wins, 12 draws and 11 losses. The results for estimating the ordered response model for Sri Lanka are reported in Table 1. Unlike other countries the bowling strike rate measure (ORPO) is signicant for Sri Lanka. Consistent with the majority of other countries the batting and bowling performance measures (RPW and ORPW, respectively) are signicant for Sri Lanka. However, the results suggest a similar impact of the batting and bowling measures. Table 2 reports the predictive accuracy of the model for Sri Lanka. The model does best at predicting wins and is able to predict nine of the 11 cases. The model also does a reasonable job at predicting draws and losses in that it predicts eight of the 12 draws and seven of the 11 losses. An examination of the 10 cases in which the model provides incorrect predictions for Sri Lanka reveals the following. First, the model fails to predict the result of one raina ected test (Australia v Sri Lanka 3rd test 1999/2000). Second, there are four failed last innings run chases that are not predicted (Pakistan v Sri Lanka 2nd and 3rd tests 1995/6, New Zealand v Sri Lanka 2nd test 1996/7 and South Africa v Sri Lanka 1st test 1997/8). Third, there are three successful last innings run chases involving Sri Lanka that are not predicted (West Indies v Sri Lanka 1st test 1996/7, South Africa v Sri Lanka 2nd test 1997/8 and Australia v Sri Lanka 1st test 1999/2000). Finally, there was the test match in which Ganguly and Tendulkar shared a 250 run 4th wicket partnership (India v Sri Lanka 3rd test 1997/8) and the drawn test against Pakistan as part of the Asian Test Cricket Championship in March 1999.

R. D. Brooks et al.
Once again in Table 2 the predictive accuracy of the model is reported this time for the West Indies. As revealed in the table, this model does very well at predicting wins (11 out of 13 cases) and losses (17 out of 18 cases). In contrast, the model does quite poorly at predicting draws, with only a three out of 10 strike rate. An examination of the 10 cases in which the model provides incorrect predictions for the West Indies reveals the following. First, the model fails to predict the results of two rain-a ected tests (4th and 5th tests against India 1996/7). Second, there were two failed last innings runs chases that the model failed to correctly predict (2nd test against England 1995 and 1st test against India 1996/7). Third, there were two successful last innings runs chases that the model was unable to predict (1st test against Australia 1994/5 and 2nd test against England 1997/8). Fourth, there was a single test against England (5th test 1997/8) in which insu cient time remained in the last innings to make the runs or to take the wickets. Finally, there were three tests in which both sides made large rst innings totals (6th test against England 1995; 2nd test against New Zealand 1995/6 and 1st test against India 1996/7).

Zimbabwe Data on all of Zimbabwes tests from the one-o match against South Africa in October 1995 up to the second test of the series against Sri Lanka played in Zimbabwe in 1999 were used Over this period of time Zimbabwe played 25 tests, producing two wins, 10 draws and 13 losses. The results for estimating the ordered response model for Zimbabwe are reported in Table 1. They show that the strike rate measures of RPO and ORPO are insignicant. They also show that improved batting (a higher RPW) and improved bowling (a lower ORPW) both increase the probability of winning. In addition, as reected by the outcome of testing H20 , results show that batting and bowling performance have an equal impact on Zimbabwes probability of winning. Table 2 reports the predictive accuracy of the model for Zimbabwe. The model does very well at predicting the draws and losses. Specically, eight of the 10 draws and 12 of the 13 losses are able to be predicted. Further, it is found that the model predicts one of Zimbabwes two wins. An examination of the four cases in which the model provides incorrect predictions for Zimbabwe reveals the following. First, in the 1st test against India 1998/9, the model fails to predict a last innings batting collapse. Second, in the 1st test against Pakistan 1997/8 the team batting last saved the match. Third, in the 2nd test against Pakistan of the same series a team successfully chased in the last innings. Finally, there was the test match in which Wasim Akram scored 257 not out against Zimbabwe (1st test 1996/7).

West Indies The data used here covers all tests played by the West Indies starting from the 1994/5 series against Australia through to the 1998/9 series against Australia. Over this period of time the West Indies were involved in 41 test matches which produced 13 wins, 10 draws and 18 losses. The outcome of estimating the ordered response model for the West Indies is shown in Table 1. There it can be seen that neither of the strike rate measures (RPO and ORPO) are signicant. However, it can be seen that RPW and ORPW are signicant, indicating that improved batting and bowling performance increase the probability of winning. Indeed, as reected by the test of H20 , it is found that these two measures are statistically of equal importance in inuencing the probability of winning test matches.

Ordered response model of test cricket performance


Discussion The purpose of this section is to make some comparative comments and conclusions based on the sections above. First, on the basis of the totality of results reported in Table 2, some comment can be made on the relative success of the ordered response models across countries in terms of the degree to which successful predictions are produced. To assist this discussion we refer to the nal column of Table 2 where an indication of the `success ratio of predictions is listed for each country. As can be seen from the table the highest level of correct predictions is achieved for Zimbabwe at 84% (21/25). This is closely followed by the Australian model in which an 82% `prediction rate is achieved. In contrast, the Indian model at 55% (17/31) produces the poorest prediction rate. However, even this rate is reasonable when compared to na ve benchmarks. For example, even if one chose the most frequent Indian result (a draw) as the prediction for every game, this would have been correct only 45% of the time (14/31). Hence, it can be concluded that the simple ordered response model

2361
performs very well indeed across all models 71% (243/ 342) of test outcomes are correctly predicted. A second issue also relates to the totality of the results reported in Table 2, but from a di erent perspective. Specically, how do the results stack up if one aggregates across countries and assesses the consolidated ability to correctly predict wins, draws and losses? To help answer this question we refer to the nal row of Table 2 which presents these aggregated results. From the table it can be seen that the models generally predict test losses best in the sample at a rate of 81% (93/115). Test match wins are also reasonably well predicted at a rate of 75% (86/115). However, it is the outcome of a test match draw that presents the model with its biggest challenge only 57% (64/ 112) of cases can be correctly predicted. A third issue relates to bringing together common themes in the circumstances in which the ordered response model fails to correctly predict test match outcomes. To assist this discussion, a summary of the circumstances identied earlier is contained in Table 3. From the table it can be seen that weather-a ected matches (in all cases but one,

Table 3. Summary of circumstances in which ordered response models fail to correctly predict test outcomes Ordered response model for nation Circumstances 1. Weather-a ected matches* 2. Unsuccessful last innings runs chase 3. Successful last innings runs chase 4. Large rst innings totals by both sides 5. Insu cient time for a win/lose result 6. Large individual score 7. Last innings batting collapse 8. Last batting team saves match 9. Runs scored level but match drawn 10. Miscellaneous Total cases Aus 3 1 3 2 9 Eng 3 4 5 3 1 1 2 19 Ind 2 7 3 1 1 14 NZ 2 3 2 2 1 10 Pak 2 6 1 2 2 13 SAf 2 3 1 1 2 1 10 SL 1 4 3 2 10 WI 2 2 2 3 1 10 Zim 1 1 1 1 4 Total 17 29 17 8 7 7 4 4 2 4 99

* One fog-a ected match, all others were rain-a ected. Table 4. Summary of opposing teams involved in non-weather a ected matches in which ordered response models fail to correctly predict test outcome Ordered response model for nation Opponent 1. Australia 2. England 3. India 4. New Zealand 5. Pakistan 6. South Africa 7. Sri Lanka 8. West Indies 9. Zimbabwe Weather-a ected Total Aus 2 1 2 1 3 9 Eng 2 1 3 1 2 6 1 3 19 Ind 2 2 2 2 2 1 1 2 14 NZ 2 1 1 1 1 2 2 10 Pak 2 2 1 2 1 3 2 13 SAf 2 2 1 1 2 2 10 SL 1 1 1 3 2 1 1 10 WI 1 4 2 1 2 10 Zim 1 3 4 Total 10 10 9 10 14 9 4 9 7 17 99

2362
due to rain) account for the second highest number at 17 out of 99 cases. However, the more interesting circumstances involve the remaining 82 instances in which it seems that an important feature is a major or perhaps dramatic occurrence in terms of cricketing performance, unrelated to mother nature. The most dominant of these cases is the situation in which there has been an unsuccessful last innings runs chase this circumstance occurred in 29 cases which represents 35% of instances other than those explained by the weather. As revealed in the table other major circumstances coinciding with failed model predictions were (a) successful last innings runs chases; (b) large rst innings totals made by both sides; (c) insu cient time for a win/lose result; and (d) tests in which one batsman had compiled a very large individual score. A fourth issue that follows on from the previous one, but looks at the failed predictions from a di erent angle, is whether there is any discernible `country e ect in evidence. To assist this analysis we refer to Table 4 which summarizes the distribution of opposing teams involved in (non-weather a ected) cricket test matches in which the ordered response model failed to correctly predict the outcome. Of most interest in this table are the totals produced in the nal column. Here it can be seen that Pakistan has the highest incidence of cases (14) which represents approximately 38% of its tests. The next highest rate is around 30% for India, New Zealand and Zimbabwe. This suggests that tests in which these countries are the opposition will tend to pose problems for the model. In contrast, Sri Lanka has the lowest rate at around 12% (four cases out of 34), suggesting that when this nation is the opposing team the model is likely to perform relatively well. Put another way, Pakistan has a greater tendency to be involved in test matches that are somewhat `unpredictable whereas Sri Lanka has a greater tendency to be involved in test matches that are somewhat `predictable. A nal issue worthy of discussion here is that of test cricketing `style or cricketing `philosophy that emanates from the analysis. That is, based on the estimated ordered response models, to what extent can each cricket playing nation be classied according to a particular style in terms of bowling/batting and `performance/`strike rate? To help answer this question Table 5 is referred to. This
Table 5. `Style analysis of test cricket playing nations based on the outcome of estimated ordered response models `Performance Bowling Style I: Pak/WI/Zim Style III: Australia Style IV: South Africa Style V: Sri Lanka Style I: Pak/WI/Zim Stile II: Eng/NZ/India `Strike Rate

R. D. Brooks et al.
table succinctly converts the key ndings of the models reported in Table 2 into a `2 2 matrix in which along one dimension `performance versus `strike rate is measured and along the other dimension bowling versus batting is measured. To help identify the di erent `styles one can begin by making two broader observations. First, it is apparent that (not surprisingly) the overwhelmingly popular direction is one of performance (either batting or bowling). Only two countries seem to vary slightly from this Sri Lanka (bowling strike rate) and South Africa (batting strike rate). Notably, both of these two nations have a (batting) performance focus as well. The second general comment is that neither batting nor bowling seem to be a dominant focus across our nine-country sample. Table 5 suggests that ve `styles of playing test match cricket have emerged over the period of our analysis. Style I is that of `Bowling and Batting Performance and it describes Pakistan, the West Indies and (perhaps to a lesser extent) Zimbabwe. Style II is that of `Batting Performance and it describes England, New Zealand and (perhaps to a lesser extent) India. Style III is `Bowling Performance and it describes Australia. Style IV is `Bowling Performance/ Batting Strike Rate and it describes South Africa. Finally, Style V is `Bowling Performance and Strike Rate and it describes Sri Lanka.

I V . TH E W O R L D C H A M P I O N O F T E S T CRICKET Rankings based on the probabilities of winning and losing The proposed modelling strategy enables provision of some information on the claims to be world test cricket champion over the period of the study, based on probability information produced by the models. Specically, this works as follows. For each of the nine test cricket-playing nations we are able to calculate their average batting strike rate and performance (based on RPO and RPW) and their average bowling strike rate and performance (based on ORPO and ORPW) over the sample period. Given Equation 1, one can then obtain an average measure of their overall performance which feeds into Equations 5 to 7. From this we obtain each teams probabilities of winning, drawing and losing test matches based on their average performance over the sample period. Accordingly, the proposed means of assessing the world test cricket champion nation is to identify the team with the highest probability of winning and/or the lowest probability of losing. With this goal in mind, consider Table 6, which reports the above-mentioned probability estimates for each of the nine test-playing countries. Initially, the results suggest that, based on the average playing performance over the period, South Africa has claim to the world champion title as it has the highest probability of winning and also the

Batting

Style V: Srk Lanka Style IV: South Africa

Ordered response model of test cricket performance


lowest probability of losing test matches. The results then suggest that Australia follows at number 2 and Pakistan at number 3. Based on the probability estimates displayed in Table 6, the rankings for places 4 through 8 are unclear. For example, if the teams are ranked by the higher probabilities of winning then the list would be the West Indies at 4, New Zealand at 5, India at 6, England at 7 and Sri Lanka at 8. In contrast, if the teams were ranked by the lower probabilities of losing then the list would be Sri Lanka at 4, India at 5, New Zealand at 6, England at 7 and the West Indies at 8. If some combination of the winning and losing probabilities were used [for example, P(win)P(loss)] then yet another ranking would be possible. What is reasonably clear however, is that the ninth ranked test-playing nation is Zimbabwe due to having both the lowest probability of winning and the highest probability of losing.

2363
in 1997/8. Accordingly, the probabilities of winning, drawing and losing for Australia and South Africa were recalculated based on their average performance in these six tests. The outcome of this analysis is reported in Table 7. The results shown in Table 7 provide a very di erent picture to the overall average performance probabilities. Specically, in these six tests Australias probability of winning is 0.736 while South Africas probability of winning is only 0.053. Therefore, in a `heavy-weight title contest Australia would appear to have the legitimate claim to be the test cricket world champion.

Test cricket `style re-visited It is worth making a comment regarding the interplay between the rankings and the `styles of test cricket identied earlier. Recall that Australia was classied as a `Bowling Performance orientated team and clearly, given Australias claim to being world champion of test match cricket, this has been a very successful strategy over the sample period. Of course this is not to say that all other nations must/should adopt this strategy to be successful indeed it doesnt even suggest that such a strategy will be as successful for Australia in the future. The situation is much too complex to be so simplistically applied for example, team personnel and their relative skills are an important factor that may change relatively quickly due to retirements, dropping of form and injuries. This will be true not only for future Australian teams but also

A `heavy-weight title analysis There is, however, an issue with using average strike rate and performance data to produce the ranking list: these averaged data ignore the `heavy-weight title factor involved in clashes between the top nations. This is best illustrated by considering the two test series between South Africa and Australia over the sample period. Despite South Africas higher ranking on average performance, Australia have won both of these series. Australia defeated South Africa 21 in South Africa in 1996/7 and 10 in Australia

Table 6. Ordered response model probabilities of winning, drawing and losing test match cricket: all nations Test cricket-playing nation Australia England India New Zealand Pakistan South Africa Sri Lanka West Indies Zimbabwe Probability of winning (Ranking) 0.417 0.164 0.182 0.205 0.351 0.492 0.135 0.248 0.002 (2) (7) (6) (5) (3) (1) (8) (4) (9) Probability of drawing 0.467 0.492 0.540 0.495 0.513 0.418 0.646 0.366 0.538 Probability of losing (Ranking) 0.116 0.345 0.277 0.300 0.136 0.090 0.219 0.385 0.460 (2) (7) (5) (6) (3) (1) (4) (8) (9)

The table reports the results of the probabilities of winning, drawing and losing test cricket matches. These probabilities are produced from the estimated ordered response model based on the average team batting and bowling performance over the sample period.

Table 7. Ordered response model probabilities of winning, drawing and losing test match cricket: Australia versus South Africa Test cricket-playing nation Australia South Africa Probability of winning 0.736 0.053 Probability of drawing 0.243 0.345 Probability of losing 0.021 0.602

The table reports the results of the probabilities of winning, drawing and losing test cricket matches between Australia and South Africa. These probabilities are produced from the estimated ordered response model based on the average team batting and bowling performance over the sample period when these two nations played each other.

2364
future teams that it plays. However, what can be said is that it is likely Australia currently believe they have a winning formula and it is also likely that they feel it best, at least in the near future, to try and perpetuate the successful style. What may be concluded about the South African `style of `Bowling Performance and Batting Strike Rate? Again, given South Africas very successful period of test match cricket, one would have to say that its style has served it very well. However, when one compares its style to that of Australias, it may be concluded that, perhaps, South Africa would be better served by switching its batting strategy away from a `strike rate to a performance-base d approach. This may or may not have ramications for how well South Africa (and indeed national teams generally) manages the continual transition between test match cricket and one-day internationals. Indeed, an interesting research question beyond the scope of the current paper is whether the di erence in the Australian and South African test match styles is related to the `two-team policy implemented by Australian selectors in recent years. The fact that a `strike rate measure is important in explaining South African test outcomes provides some plausibility to this speculation. Further, this could be linked to the importance of the bowling `strike rate measure for Sri Lanka given that it undoubtedly rode on the glory of its one-day cricket World Cup victory in 1996 for some time thereafter. This very interesting question we leave to future research endeavours. One nal comment relates to the plight of Zimbabwe. Over the sample period its `style is one of `Bowling and Batting Performance a general style that also seems to have been adopted by Pakistan and the West Indies. While this style seems an eminently sensible one (Pakistan have achieved a 3rd ranking using it) and we would not suggest that Zimbabwe necessarily try to change from it, this illustrates that simply having a particular `style does not guarantee success.

R. D. Brooks et al.
richer specication is needed to better model all of these inuences. An analysis of failed model predictions in terms of country factors suggests that (relative to all other countries) Pakistan has a higher tendency to be involved in such matches, whereas Sri Lanka has a higher tendency to be involved in matches that are `predictable. Furthermore, a `style analysis using the model suggests that ve test cricket styles are evident. Style I is that of `Bowling and Batting Performance and describes Pakistan, the West Indies and (perhaps to a lesser extent) Zimbabwe. Style II is that of `Batting Performance and describes England, New Zealand and (perhaps to a lesser extent) India. Style III is `Bowling Performance and describes Australia. Style IV is `Bowling Performance/ Batting Strike Rate and describes South Africa. Finally, Style V is `Bowling Performance and Strike Rate and describes Sri Lanka. The model is also used to provide evidence on which test cricket-playing nation has a superior claim to being current world champion. Based on an initial analysis of average performance since 1994/1995, the model identies South Africa as the number one nation just ahead of Australia. However, this ignores the `heavy-weight title nature of clashes between the top teams. Consequently, when this feature is incorporated into the analysis of the two top ranked nations, then Australia prevails as the champion test cricket playing country. By producing a general model a large range of factors have been ignored which may be important in some contexts. While not exhaustive, a good list of such factors would include: the identity of the opponent; the location of tests (home versus away; Australia v sub-continent v England v South Africa); other input measures such as elding, captaincy, inclusion of allrounders; which captain wins the toss; team balance (e.g. balance of left- and righthanded batsmen); rst innings versus second innings performance; who bats rst; types of wickets (e.g. a turning pitch, a greentop or a at track and so on); the potential interaction between test and one-day cricket performance; and the potential impact of weather and other playing condition variables. The simple model employed in the current paper is very encouraging and suggests that exploring more complete model specications that take on board at least some of the above listed factors is a worthwhile research exercise. A start on this work has begun and it is hoped that a future models will, in particular, improve on the existing one by: better predicting test match draws and better coping with the most common circumstances that are associated with the failed model predictions. Only time will tell whether more complex models do better than the basic one employed currently for the moment it provides a very useful preliminary insight into modelling test match cricket outcomes.

V. C ONCLUSION This paper has explored the modelling of test cricket performance from the perspective of an ordered response model. From an economic perspective modelling test cricket is interesting because the production output is an ordered categorical variable (win, draw, loss) which is produced using specialized labour inputs, namely, batting and bowling. Results suggest that a couple of simple batting and bowling measures are able to predict the majority of test results. Interestingly, the model does poorly in predicting matches (a) impacted by weather; (b) that involve unsuccessful last innings runs chases; (c) that involve successful last innings run chases; and (d) with large rst innings scores made by both teams. Clearly, a

Ordered response model of test cricket performance


A C K N O W LE D G E M E N TS The authors wish to thank Craig Brooks for his helpful assistance with the data collection. Part of this research was conducted while the second author was employed at RMIT University. R E F E R E N C ES
Bairam, E. I., Howells, J. M. and Turner, G. M. (1990) Production functions in cricket: the Australian and New Zealand Experience, Applied Economics, 22, 8719. Chedzoy, O. (1997) The e ect of umpiring errors in cricket, The Statistician, 46, 52940. Clarke, S. R. (1988) Dynamic programming in one-day cricket optimal scoring rates, Journal of the Operational Research Society, 39, 3317. Clarke, S. R. and Norman, J. M. (1995) Home ground advantage of individual clubs in English soccer, The Statistician, 44, 50921. Clarke, S. R. and Norman, J. M. (1998) Dynamic programming in cricket: protecting the weaker batsman, Asia-Pacic Journal of Operational Research, 15, 93108. Craven, B. J. (1998) A psychophysical study of leg-before-wicket judgments in cricket, British Journal of Psychology, 89, 555 78. Crowe, S. and Middeldorp, J. (1996) A comparison of leg before wicket rates between Australians and their visiting teams for test cricket series played in Australia, 197794, The Statistician, 45, 25562. Demmert, H. G. (1973) The Economics of Professional Sports Teams, Heath, Lexington, MA. Duckworth, F. C. and Lewis, A. J. (1998) A fair method for resetting the target in interrupted one-day cricket matches, Journal of Operational Research Society, 49, 2207. Holder, R. and Nevill, A. (1997) Modelling performance at international tennis and golf tournaments: is there a home advantage? The Statistician, 46, 5519.

2365
Johnston, M. I., Clarke, S. R. and Noble, D. H. (1993) Assessing player performance in one day cricket using dynamic programming, Asia-Pacic Journal of Operational Research, 10, 4555. Jones, J. C. H. and Walsh, W. D. (1987) The World Hockey Association and player exploitation in the National Hockey League, Quarterly Review of Economics and Business, 27, 87 101. Magnus, J. and Klaassen, F. (1999a) The e ect of new balls in tennis: four years at Wimbledon, The Statistician, 48, 23946. Magnus, J. and Klaassen, F. (1999b) On the advantage of serving rst in a tennis set: four years at Wimbledon, The Statistician, 48, 24756. Neale, W. C. (1964) The peculiar economics of professional sports, Quarterly Journal of Economics, 78, 114. Schoeld J. A. (1988) Production functions in the sports industry: an empirical analysis of professional cricket, Applied Economics, 20, 17793. Scott, F. A., Long, J. E. and Somppi, K. (1985) Salary vs. marginal revenue product under monopsony and competition: the case of professional basketball, Atlantic Economic Journal, 13, 509. Scully, G. W. (1974) Pay and performance in Major League Baseball, American Economic Review, 64, 91530. Thelwell, R. C. and Maynard, I. W. (1998) Anxiety-performance relationships in cricketers: testing the zone of optimal functioning hypothesis, Perceptual and Motor Skills, 87, 67589. Totterdell, P. (1999) Mood scores: mood and performance in professional cricketers, British Journal of Psychology, 90, 31732. Zak, T. A., Huang, C. J. and Siegfried, J. J. (1979) Production e ciency: the case of professional basketball, Journal of Business, 52, 37992. Zech, C. E. (1981) An empirical estimation of a production function: the case of Major League Baseball, American Economist, 25, 1923.

Vous aimerez peut-être aussi