Lec 19 and 20

Question 1

plenty of times, context known, looks unimportant?

Answer 1

exploring data

Question 2

can afford detail but looks vital?

Answer 2

results in paper

Question 3

appeal from afar but have close up detail

Answer 3

poster at conference

Question 4

must sell idea in seconds

Answer 4

presentation or talk at conference

Question 5

counts in categories?

Answer 5

table or barchart which is barplot in R - these show frequency more clearly than pie chart

Question 6

why not to use pie chart

Answer 6

Pie charts are usually a poor way of representing frequency (count) data because the human eye is far better at comparing lengths than angles

Question 7

Does a paper graph need colour and title?

Answer 7

no

Question 8

stacking of groups,. helpful or no?

Answer 8

This might be useful if you are particularly interested in the sums across the primary x-axis categories, but it isn’t easy to compare the breakdown by the second categorical variable. For this, it is easier to have unstacked bars, side by side.

Question 9

when to use colour?

Answer 9

for talk or poster

Question 10

what is a mosiac plot for?

Answer 10

Mosaic plot for exploration of complex contingency tables

Question 11

how many colours does default R palette have?

Answer 11

8

Question 12

how to overcome overlap?

Answer 12

use open symbols and jitter

Question 13

What chart is used to represent distribution in diff groups

Answer 13

stripchart and histogram

Question 14

when to represent anaylsis and when to represent the data themselves?

Answer 14

analysis for talk/poster
data themselves for paper (although many articles present analysis summaries, so best to use both so reader can decide if right analysis has been used)

Question 15

repeated measures - what to plot and what not to plt

Answer 15

dont plot the simple treatment medians (or means) as it conceals the paired design. Instead, Plot the differences between treatments (but only really viable for two treatments)
for two or more treatments, remove subject differences first

Question 16

plotting correlation or regression

Answer 16

y ~ x as it matches the analysis

Question 17

regression line?

Answer 17

Ordinary Least Squares regression

When you are predicting y from x

Question 18

reduced major axis reg line?

Answer 18

only when no prediction involved or causation implied

for correlation or allometry

Question 19

relationship between many variables

Answer 19

pairs () uses whole dataset

Question 20

shingling

Answer 20

break down the second continuous variable into categories. By default, the ‘shingles’ overlap by 50%

Question 21

how to make 3d plot easier to read

Answer 21

add drop lines or Adding the plane describing the regression of z on y and x

Question 22

How is analysis represented? and how is data represented?

Answer 22

anaylsis represented using means and CI

data represented using boxplots

Question 23

what is data mining

Answer 23

True exploration when you don’t know what patterns might exist
May have a large number of candidate predictors and no strong theory to predict which should be more important
And as the number of variables goes up, the number of possible interactions goes up geometrically
** Before ‘data mining’ check for missing values and outliers and ‘clean up’ the data

Question 24

what does vertical spacing in CART represent

Answer 24

variance explained

Question 25

when pruning a tree in CART do errors go up or down with no of predictors?

Answer 25

error goes down with no of predictors

Question 26

what is Cross-validation?

Answer 26

For any classification method, your ‘best model’ will always perform well on the sample you developed the model for So a risk of over-fitting So, build model using one set of data and test it on another (=cross-validation) partitioning the original sample into a training set to train the model, and a test set to evaluate it

Question 27

pos and negsof CART

Answer 27

``` It’s crude: doesn’t make use of continuous information; simply splits such variables as ‘high’ or ‘low’ But it’s robust Variables can have any distribution And powerful Examines all interactions And gives results in an easy-to-use form A decision tree ```

Question 28

two classes of clusters - definitions

Answer 28

Supervised’ learning – know the true identity of some clusters and use these to develop a predictive model for data where you don’t know group membership Unsupervised’ learning – don’t know what’s ‘right’ or ‘wrong’, so try and find natural clustering patterns in the data (which can then be used in future prediction)

Question 29

examples of supervised and unsupervised

Answer 29

supervised -Discriminant function analysis Logistic regression (two groups) Multinomial logistic regression (>2 groups and no order) Support Vector Machines Neural networks Genetic algorithms unsupervised - k-means clustering Specify number of clusters to find Allocates data to clusters to minimise within-cluster sums-of-squares

Question 30

when to use CART vs clustering?

Answer 30

Use robust methods like CART to explore predictive relationships Use clustering to expose unexpected ‘structure’ (groupings) in your data

Question 31

what is multinomial logistic regression

Answer 31

Multinomial logistic regression (>2 groups and no order)

Question 32

plotting raw data?

Answer 32

stripchart or dotplot

Question 33

a histrogram represents..

Answer 33

partially summarised data

Question 34

plotting continuous data

Answer 34

stripchart, sctterplot, histogram (to explore whole distributions)

Question 35

notched boxplot

Answer 35

If the notches of 2 boxes don’t overlap, medians are likely to differ

Question 36

how much of the data is represented in a boxplot? and what is the interquartile range?

Answer 36

A boxplot consists starts at the 25th percentile and ends at the 75th percentile, so this box contains 50% of the data. The interquartile range is the difference between the 25th and 75th percentiles, which are also referred to as the lower and upper quartiles.

Question 37

ellipses package

Answer 37

Simplify correlations to ellipses with hue & intensity indicating direction & strength

Question 38

which test is this?..the slopes are assumed to be the same (parallel) for each level of the categorical factor..

Answer 38

ANCOVA - can represent in sep panel or Or in same plot, with separate regression lines USE SHINGLING to help visualisation

Question 39

what graph summaries the data?

Answer 39

boxplot

Question 40

what is inferential statisitcs?

Answer 40

testing hypothesis, makes inferences about populations using data drawn from the population.

Question 41

what is a conditioning plot and when is it used?

Answer 41

for investigating the relationship between multiple variables , is a plot of y against x conditioned upon (or broken down by) a third variable (or more if you have them). If the third variable is categorical (a factor with discrete levels/groups), then you get a plot of y against x for each level of the third variable, each in a separate panel If the third variable is continuous, the third variable is broken into similarly-sized groups, somewhat overlapping, and you get a plot of y against x for each of those.

Question 42

what is k-clustering?

Answer 42

allocates data to clusters to minimise within-cluster sum of squares

Question 43

how does hierarchical or agglomerative clustering work?

Answer 43

need a measure of distance between points, join the clostest two points, when two or more are joined they are treated as an item with a single 'location' ... once all joined up, then create a pattern of linkage tree

Question 44

when to use CART vs hierarichcal/agglomerative?

Answer 44

CART for exploring predictive relationships, agglo/hierrchical to expose unexpected structures.