AFL Drafts From 1999-2009
Is there an advantage in having early picks?
Oliver Fitzpatrick s3686342
Last updated: 20 October, 2017
Data
- The data was collected from footywire.com - using their list of drafts each year to make my own list of data
- The years I chose to include were from 1999-2009 - going any more recently would limit the effectiveness because many of the players drafted that recently are still playing and so their games played will not be as accurate
- Information collected was: Draft Round, Pick Number, Player, Team, Games Played, Average Disposals, Brownlow Votes, All Australian Selections, Draft Year, Draft Position (in groups of 20), and which draft they were taken in
AFL_Drafts <- readxl::read_excel("C:/Users/sarah/Desktop/Oliver/AFL_Drafts(1999-2009).xlsx",
col_types = c("text", "numeric", "text",
"text", "text", "numeric", "numeric",
"text", "numeric", "numeric", "numeric",
"text", "text", "text"))
View(AFL_Drafts)
Data Cont.
- The main variables to be used in the analysis are:
- Draft Position: this groups picks into groups of 20 - so the first 20 picks are in a group called Top 20, the next are 20-40 and so on until >120. This factor was ordered.
- Games Played: This is the number of games the player has played at AFL level since being drafted correct at the start of the 2017 AFL season
- AA: this is whether the player has ever been selected in an All Australian team or not
- Brownlow: This is the total number of Brownlow votes that a player has received in their career (up to 2016)
AFL_Drafts$Draft_Position <- AFL_Drafts$Draft_Position %>% factor(levels = c('Top 20','20-40','40-60','60-80','80-100','100-120','>120'),ordered = TRUE)
AFL_Drafts$AA <- AFL_Drafts$AA %>% factor(levels = c('Yes','No'),ordered = TRUE)
Descriptive Statistics and Visualisation
- First, we will look at the distribution of games played by all people who are drafted to get an idea of how many games would be expected for an average player and to note the shape of the distribution
AFL_Drafts$Games_Played %>% hist(xlab = "Number of Games Played",ylab = "Frequency", main = "Distribution of Games Played By AFL Footballers",col = "dodgerblue")
- We can see the data is heavily positively skewed, with the vast majority of players playing less than 50 games.
Summary Stats Cont: Games Played by Draft Position
- Lets look at the statistics of games played when grouped by where they were taken in the draft:
AFL_Drafts %>% group_by(Draft_Position) %>% summarise(Q1 = quantile(Games_Played,probs = .25,na.rm = TRUE)%>%round(1),
Median = median(Games_Played, na.rm = TRUE)%>%round(1),
Q3 = quantile(Games_Played,probs = .75,na.rm = TRUE)%>%round(1),
Max = max(Games_Played,na.rm = TRUE),
Mean = mean(Games_Played, na.rm = TRUE)%>%round(1),
SD = sd(Games_Played, na.rm = TRUE)%>%round(1),
n = n(),
Missing = sum(is.na(Games_Played))) -> table1
knitr::kable(table1)
| Top 20 |
48.0 |
125.0 |
197.5 |
353 |
128.3 |
88.9 |
220 |
0 |
| 20-40 |
12.8 |
62.5 |
139.5 |
319 |
82.6 |
80.1 |
220 |
0 |
| 40-60 |
4.0 |
27.0 |
111.0 |
332 |
63.9 |
75.8 |
220 |
6 |
| 60-80 |
0.0 |
13.0 |
59.8 |
251 |
43.2 |
63.0 |
220 |
66 |
| 80-100 |
0.0 |
2.0 |
61.0 |
275 |
41.6 |
65.3 |
220 |
35 |
| 100-120 |
0.0 |
1.0 |
45.5 |
289 |
38.4 |
66.9 |
220 |
16 |
| >120 |
0.0 |
3.0 |
31.0 |
290 |
27.5 |
49.6 |
477 |
78 |
- It is clear that players taken in the first 20 picks average far more games than players taken anywhere else
Summary Stats Cont: Boxplots
AFL_Drafts %>% boxplot(Games_Played ~ Draft_Position, data = ., main = "Box Plot of Games Played by Draft Position", xlab = "Draft Position", ylab = "Games Played", col = "firebrick1")
- The Boxplots help us notice the obvious difference between the top picks and later picks - The trend seems to be that the later the pick, the less games the player will play, although it does apppear that the groups 60-80, 80-100 and 100-120 are somewhat similar - Further analysis will need to be done to see whether there is a statistically significant difference between the groups
Summary Stats Cont: Histograms
AFL_Drafts %>% histogram(~Games_Played | Draft_Position,data=.,main = "Distribution of Games Played by Draft Position", xlab = "Games Played", ylab = "Frequency", layout=c(4,2))
- The histograms are interesting, they appear to add weight to the idea that earlier picks play more games - the distribution amongst the top 20 group is almost uniform, but as the picks get later, the distribution becomes more and more positively skewed - This tells us that more players taken with late picks play only a few games than players taken with early picks
Summary Stats Cont: Brownlow Votes
- Another way to compare players’ value is by looking at how many Brownlow votes they received in their career:
AFL_Drafts %>% group_by(Draft_Position) %>% summarise(
Max = max(Brownlow,na.rm = TRUE),
Mean = mean(Brownlow, na.rm = TRUE)%>%round(1),
SD = sd(Brownlow, na.rm = TRUE)%>% round(1),
n = n(),
Missing = sum(is.na(Brownlow))) -> table2
knitr::kable(table2)
| Top 20 |
210 |
23.5 |
38.0 |
220 |
3 |
| 20-40 |
220 |
10.3 |
29.1 |
220 |
8 |
| 40-60 |
186 |
5.3 |
17.2 |
220 |
23 |
| 60-80 |
35 |
1.7 |
5.6 |
220 |
40 |
| 80-100 |
72 |
2.6 |
7.9 |
220 |
35 |
| 100-120 |
146 |
3.2 |
15.0 |
220 |
25 |
| >120 |
66 |
1.7 |
6.5 |
477 |
68 |
- Earlier picks appear to receive more Brownlow votes
Summary Stats Cont: All Australian
- The other measure used is whether a player has ever been selected in an All Australian team:
table(AFL_Drafts$AA,AFL_Drafts$Draft_Position) %>% prop.table(margin = 2)%>% round(3)
##
## Top 20 20-40 40-60 60-80 80-100 100-120 >120
## Yes 0.227 0.095 0.045 0.014 0.014 0.014 0.013
## No 0.773 0.905 0.955 0.986 0.986 0.986 0.987
AAfreq <- table(AFL_Drafts$AA,AFL_Drafts$Draft_Position) %>% prop.table(margin = 2)%>% round(3)
barplot(AAfreq, main = "Proportion of All Australians From Each Draft Position", ylab = "Proportion", xlab = "Draft Position", legend.text = TRUE)
- Earlier picks are also far more likely to make an All Australian team - over 20% of top 20 picks will be picked at some point in their career, and this fades to around 1% as the picks get later
Hypothesis Testing
- To test whether there is a significant difference in the means of games played between the groups, we will firstly look at the 95% confidence intervals of each group
AFL_Drafts %>% group_by(Draft_Position) %>% summarise(Mean = round(mean(Games_Played, na.rm = TRUE),2),
SD = round(sd(Games_Played, na.rm = TRUE),3),
n = n(),
tcrit = round(qt(p = 0.975, df = n - 1),3),
SE = round(SD/sqrt(n),3),
`95% CI Lower Bound` = round(Mean - tcrit * SE,2),
`95% CI Upper Bound` = round(Mean + tcrit * SE,2))
p1 <- ggplot(data = AFL_Drafts, aes(x = Draft_Position, y = Games_Played))
p1 + geom_dotplot(binaxis = "y", stackdir = "center", dotsize = 1/2, alpha = .25,binwidth = 10) +
stat_summary(fun.y = "mean", geom = "point", colour = "red") +
stat_summary(fun.data = "mean_cl_normal", colour = "red",
geom = "errorbar", width = .2)
- The confidence intervals show us that Top 20 picks are significantly more likely to play more games than players taken at any other time in the draft - The confidence levels do appear to decrease as the picks get later in the draft, but there is some overlap
Hypothesis Testing Cont: Anova
- To compare the means of each group, I used Anova testing to determine whether the differences are significant or not
fit <- aov(Games_Played ~ Draft_Position, data = AFL_Drafts)
TukeyHSD(fit)
## Tukey multiple comparisons of means
## 95% family-wise confidence level
##
## Fit: aov(formula = Games_Played ~ Draft_Position, data = AFL_Drafts)
##
## $Draft_Position
## diff lwr upr p adj
## 20-40-Top 20 -45.668182 -65.18472 -26.1516460 0.0000000
## 40-60-Top 20 -64.426805 -84.07966 -44.7739476 0.0000000
## 60-80-Top 20 -85.075974 -106.58215 -63.5697990 0.0000000
## 80-100-Top 20 -86.671867 -107.09063 -66.2531065 0.0000000
## 100-120-Top 20 -89.924332 -109.81986 -70.0287985 0.0000000
## >120-Top 20 -100.765995 -117.95484 -83.5771528 0.0000000
## 40-60-20-40 -18.758624 -38.41148 0.8942342 0.0726378
## 60-80-20-40 -39.407792 -60.91397 -17.9016172 0.0000015
## 80-100-20-40 -41.003686 -61.42245 -20.5849247 0.0000001
## 100-120-20-40 -44.256150 -64.15168 -24.3606167 0.0000000
## >120-20-40 -55.097813 -72.28665 -37.9089710 0.0000000
## 60-80-40-60 -20.649169 -42.27913 0.9807925 0.0725470
## 80-100-40-60 -22.245062 -42.79416 -1.6959635 0.0239661
## 100-120-40-60 -25.497526 -45.52680 -5.4682504 0.0033553
## >120-40-60 -36.339189 -53.68266 -18.9957197 0.0000000
## 80-100-60-80 -1.595893 -23.92404 20.7322546 0.9999927
## 100-120-60-80 -4.848358 -26.69905 17.0023319 0.9948498
## >120-60-80 -15.690021 -35.10848 3.7284416 0.2053812
## 100-120-80-100 -3.252464 -24.03377 17.5288461 0.9992733
## >120-80-100 -14.094127 -32.30091 4.1126546 0.2516957
## >120-100-120 -10.841663 -28.45965 6.7763206 0.5368725
- The anova testing’s p-values show that all groups are significantly different except for: 20-40&40-60, 40-60&60-80, 60-80&80-100, 60-80&100-120, 60-80&>120, 80-100&100-120, 80-100&>120 and 100-120&>120.
- This tells us that a top 20 pick is significantly more valuable than any other pick. A pick in the 20-40 range is not significantly different to one in the 40-60 range, but is signifcantly better than all other lower picks. Similarly, a pick in the 40-60 group is not better than a 60-80 pick but better than other lower picks
- There is not a statistically significant difference in picks taken later than 60
Hypothesis Testing Cont: Correlation
- Now, let’s have a look at the correlation between simply draft number and games played
bivariate<-as.matrix(dplyr::select(AFL_Drafts, Draft_Number,Games_Played)) #Create a matrix of the variables to be correlated
rcorr(bivariate, type = "pearson")
## Draft_Number Games_Played
## Draft_Number 1.00 -0.39
## Games_Played -0.39 1.00
##
## n
## Draft_Number Games_Played
## Draft_Number 1797 1596
## Games_Played 1596 1596
##
## P
## Draft_Number Games_Played
## Draft_Number 0
## Games_Played 0
CIr(-0.39, n = 1596, level = .95)
## [1] -0.4308229 -0.3475842
- A Pearson’s correlation was calculated to measure the strength of the linear relationship between Draft Number and Games Played. The negative correlation was statistically significant, r=-0.39, p<.001, 95% CI [-0.431,-0.348]