R Notes for Topic 25: Inference for Two-Way Tables

Activity 25-2 b asks us to “use technology to compute the expected values, test-statistic, and p-value”. Doing the chi-square test is easy, because we can use chisq.test once the data are in a table. So, let's put the data in a table.

Happiness = data.frame(
  row.names = c("Very Happy", "Pretty Happy", "Not Too Happy"),
  Y1972=c(486,855,265), 
  Y1988=c(498,832,136), 
  Y2004=c(419,738,180))
Happiness

Okay, now let's get the important statistics.

chisq.test(Happiness)

But what about the expected values? That's a little bit harder. I'll also note that they don't ask you for them and don't provide them in the answer key. But, hey, they've asked, so let's at least think about it.

First, we'll make a vector of proportions of the three response values. We could use the values given in the table for the total number of responses for each value.

> ResponseProportions = c(1403/4409, 2425/4409, 581/4409)
> ResponseProportions
[1] 0.3182127 0.5500113 0.1317759

However, we are probably better off asking R to compute these sums for us.

RP = c(sum(Happiness[1,]), sum(Happiness[2,]), sum(Happiness[3,]))/sum(Happiness)
RP

Once the proportions are computed, we can build a new data frame.

ExpectedHappiness = data.frame(
  row.names = c("Very Happy", "Pretty Happy", "Not Too Happy"),
  Y1972 = RP*sum(Happiness$Y1972),
  Y1988 = RP*sum(Happiness$Y1988),
  Y2004 = RP*sum(Happiness$Y2004))
ExpectedHappiness

But that's a long way to compute something. R lets us write this much more concisely. The rowSums and colSums procedures let us compute the sums of rows and columns, respectively. We still want to convert the row sums to proportions by dividing by the total number of observational units. Then we build the expected values table by computing the outer product of these two vectors.

ExpectedHappiness = rowSums(Happiness) %o% colSums(Happiness)/sum(Happiness)

No matter how we compute the expected value table, we can see a simple comparison of the expected values to the actual values.

Happiness - ExpectedHappiness

And we can even see the various contributions to the chi-square test.

(Happiness - ExpectedHappiness)^2/ExpectedHappiness

We start by creating the table.

CardTips = data.frame(row.names=c("Tip","No Tip"),
  Joke = c(30,42),
  Ad = c(14,60),
  None = c(16,49))

Then we use the chi-square test.

chisq.test(CardTips)

To help ourselves think about the meaning, we build a table of expected values.

ExpectedTips = rowSums(CardTips) %o% colSums(CardTips)/sum(CardTips)

We then compare that table to our original table. (Do you notice anything interesting about the columns?)

CardTips - ExpectedTips

We then see what contributed to the high chi-square value.

(CardTips-ExpectedTips)^2/ExpectedTips

PoliticalHappiness = data.frame(
  row.names = c("Very Happy", "Pretty Happy", "Not Too Happy"),
  Liberal = c(90, 177, 51),
  Moderate = c(128, 291, 76),
  Conservative = c(187, 258, 48))
ExpectedPH = rowSums(PoliticalHappiness) %o% colSums(PoliticalHappiness)/sum(PoliticalHappiness)