R Commands for Introduction to Statistical Modeling

D ANIEL K APLAN O CTOBER 5, 2008

This sheet is intended to help you remember R com- > mod = lm( width ~ length + sex, data=kids)
mands and some of the ways they are used. It’s as- Simple Descriptive Statistics or
sumed that you already understand the statistics and For describing data one variable at a time. > mod = with(kids, lm( width ~ length + sex))
purpose of the commands. Relevant operators: mean, sd, median, IQR, or even
> marks the command you type. summary, quantile, table, prop.table. > mod = lm(kids$width~kids$length+kids$sex)
+ marks the second line, if any, of the command. There are two basic styles when selecting a variable Display the coefficients:
from a data frame: using with or using the $ reference > mod
Installing R syntax: Coefficients:
Download and execute the “binary” file appropri- > with( kids, mean(width)) (Intercept) length sexG
ate for your operating system: Windows, Mac OS X, [1] 8.992308 3.641 0.221 -0.233
others. > mean( kids$width ) • R-squared
Starting R [1] 8.992308 > r.squared(mod)
Datasets and convenience functions for the Intro- Either way is fine. I encourage the $ method. [1] 0.45954
duction to Statistical Modeling course are contained in • Quantitative Variables • Regression table including standard errors:
> summary(mod)
the “workspace” file ISM.Rdata. Double-click on that > sd( kids$width )
Coefficients:
file to start a new session of R. [1] 0.5095843 Estimate Std. Error t value Pr(>|t|)
Functions defined in ISM.Rdata are: resample, > median( kids$width ) (Intercept) 3.6412 1.2506 2.91 0.0061
shuffle, r.squared, do. [1] 9 length 0.2210 0.0497 4.45 8e-05
> IQR( kids$width ) sexG -0.2325 0.1293 -1.80 0.0806
Reading in Spreadsheet/Tabular Data
[1] 0.7 • ANOVA table
A data table (called a “data frame” in R) is orga-
> quantile( kids$width, 0.60 ) > anova(mod)
nized into cases and variables. Analysis of Variance Table
60%
• Data from the ISM course
9.08
Relevant operators: ISMdata. This takes a file name Response: width
> summary( kids$width )
(in quotes) and returns a data frame Df Sum Sq Mean Sq F value Pr(>F)
Min 1st Qu. Median Mean 3rd Qu. Max
> kids = ISMdata("kidsfeet.csv") (Intercept) 1 3154 3154 21287.88 < 2e-16
7.90 8.65 9.00 8.99 9.35 9.80
length 1 4 4 27.38 7.4e-06
> runners = ISMdata("ten-mile-race.csv") • Categorical Variables sex 1 0.48 0.48 3.23 0.08
• Your own data Count the number of cases at each level: Residuals 36 5 0.15
Store your data in a spreadsheet in CSV format. > table( kids$sex )
Resampling
There should be a header row. After that, each row is B G
Relevant operators: resample, shuffle, do
one case, each column is one variable. 20 19
• Bootstrapping a Standar Error
Convert the count to a proportion. The standard error reflects variability due to sam-
> prop.table(table( kids$sex )) pling.
B G
> with( kids, mean(width) )
0.5128205 0.4871795
[1] 8.9923
Linear Modeling > trials = do(500)*
Constructing linear models. + with( resample(kids), mean(width) )
Relevant operators: lm, r.squared, summary, anova > sd(trials)
Relevant operators: read.csv
Fit a model. All of these three styles are equivalent, [1] 0.076145
> mydata = read.csv("fish.csv")
but I recommend the first one: For a model:
 > campaign.spending
> lm( width~length+sex, data=kids) Causal Network with 4 vars:
(Intercept) length sexG 9.5
==============================
3.641 0.221 -0.233 9.0

width
popularity is exogenous
> trials = do(500)* 8.5
polls <== popularity
+ lm( width~length+sex, data=resample(kids)) 8.0
spending <== polls
> sd(trials) B G
vote <== popularity & spending
(Intercept) length sexG
• Observational Study
0.939525 0.036822 0.120385 • Histograms
> histogram( ~ age, data=runners) > run.sim(campaign.spending, 4)
• Hypothesis Testing popularity polls spending vote
Don’t forget the leading tilde (~).
Implement the null hypothesis that sex is not re- 1 44.141 46.875 59.484 42.780
lated to width, but length might be: 25 2 25.577 30.991 46.322 18.655

Percent of Total
20
3 47.376 46.616 45.275 52.207
> trials = do(500)* 15
4 48.393 50.003 56.839 54.389
10
+ lm( width~length+shuffle(sex), data=kids) 5

> mean(trials) 0
• Experimental Study
(Intercept) length shuffle(sex)G
20 40 60 80 Two types of experiments are supported.
age
2.8575581 0.2480365 0.0051818 Create the experimental intervention of the desired
> sd(trials) Extra: Try also length:
(Intercept) length shuffle(sex)G > histogram( ~ age | sex, data=runners) > intervene = rep( c(0,100), length.out = 5)
0.2228825 0.0085814 0.1326187 • Bar Charts 1) Impose the intervention directly.
> barchart( table(kids$sex), horizontal=FALSE) > run.sim( campaign.spending, 4,
Graphics + spending=intervene)
20

• Scatter Plot popularity polls spending vote

15
1 49.931 50.071 0 48.009
Freq

10
> xyplot( width ~ length, data=kids) 2 66.392 69.752 100 74.800
5
3 73.079 71.553 0 58.396
0
B G 4 51.752 53.136 100 57.155
9.5
2) Add the intervention on top of the “natural” val-
width

9.0

Simulations ues:
8.5

8.0 Simulations generate data from a hypothetical > run.sim( campaign.spending, 4,

22 23 24 25 26 27 causal network. + spending=intervene, inject=TRUE)
length
Relevant function: run.sim popularity polls spending vote
Available hypothetical causal networks in- 1 54.129 52.972 59.853 67.638
Try: xyplot(width~length|sex,data=kids)
clude: campaign.spending, jock, university.test, 2 70.974 65.608 126.752 85.580
• Box and Whiskers Plot 3 75.549 71.873 34.934 67.062
heights, electro, aspirin, salaries, etc.
To see the variables, type the name of the hypothet- 4 50.391 48.954 160.389 74.577
> bwplot( width ~ sex, data=kids)
ical causal network