The learning goals for today’s class are as follows:

• Use R-squared to assess predictive power of linear models
• Examine the residuals from linear models
• Think about outliers, leverage, and influence

To achieve these goals, we will spend today’s class analyzing the relationship between voting and ethnicity in recent general elections in Malaysia:

• 2013 General Election
• 2018 General Election

A codebook with descriptions of the variables and where they were circulated as part of the pre-class materials. The critical variables for the first half of today:

• DISTNO and DISTNAME: The constituency number and name.
• MALAY: Percent of voters in the constituency classified as ethnic Malay.
• CHINESE: Percent of voters in the constituency classified as ethnic Chinese.
• INDIAN: Percent of voters in the constituency classified as ethnic Indian.
• OTHER: Percent of voters in the constituency classified as “other” ethnicity (e.g. Eurasian, Thai).
• pctBN: Percent of the votes cast for ruling Barisan Nasional (BN) coalition out of the total number of valid votes cast in the constituency. The BN was in power from independence in 1957 until 2018.

Create a histogram. It appears to show a lot of variation in BN vote share across constituencies.

hist(ge13$pctBN, freq=FALSE, col="lightblue",
     main = "Histogram of BN Vote Share",
     xlab = "BN Vote Share (%)")

What do you expect to be the relationship between BN vote share and ethnicity? Why?

With the patterns you expect to see between BN vote share and ethnicity in mind, let’s now use the ethnic composition of constituencies to predict BN vote shares.

Let’s start with the percentages of ethnic Malay voters and ethnic Chinese voters. If you finish early try to do the same with the percentages of ethnic Indian and “Other” voters.

Estimate linear models:

lm13.BN.MALAY   <- lm(pctBN ~ MALAY, data=ge13)
lm13.BN.CHINESE <- lm(pctBN ~ CHINESE, data=ge13)

Now interpret the intercepts and slopes:

coef(lm13.BN.MALAY)   # display intercept & slope

## (Intercept)       MALAY 
##  26.9520696   0.3534643

coef(lm13.BN.CHINESE) # display intercept & slope

## (Intercept)     CHINESE 
##   60.873960   -0.412503

Now interpret the R-squares:

summary(lm13.BN.MALAY)$r.squared   # displays r-squared value

## [1] 0.4398053

summary(lm13.BN.CHINESE)$r.squared # displays r-squared value

## [1] 0.455778

Which model best accounts for the variation in BN vote share? Why?

What about the residual standard errors?

sd(lm13.BN.MALAY$residuals)   # displays residual standard errors

## [1] 9.812739

sd(lm13.BN.CHINESE$residuals) # displays residual standard errors

## [1] 9.671834

Which model best accounts for the variation in BN vote share? Why?

The violation of the straight enough condition is also clear if we add the line of best fit.

Let’s take a moment to pause, think, and talk about the models so far:

• What have we found?
  • Positive relationship between MALAY and pctBN
  • Negative relationship between CHINESE and pctBN
  • CHINESE more predictive of BN vote share than MALAY
  • Residuals suggest relationship not strictly linear
• Things to think about:
  • Did you expect CHINESE to be more predictive than MALAY?
  • What might explain the curve in the relationship?
    • What further might we do to find out?

For the second half of today’s activity, we will be using the dataset ge2018.csv to practice fitting and assessing linear models.

To provide some background on the 2018 general election:

• Fought largely between the Barisan Nasional (BN) coalition and the opposition Pakatan Harapan (PH) coalition
  • BN led by PM Najib Razak
  • PH led by former PM Mahathir Mohamad
• Election played out against backdrop of a major corruption scandal over embezzlement funds from Malaysia’s 1MDB sovereign wealth fund
  • PM Najib accused of taking USD 700 million - is still on trial

Despite extensive pre-electoral manipulation and against the expectations of observers, however…

• The PH coalition soundly defeated the BN
  • First non-BN/UMNO-led government since Malaysia’s independence
  • Transition raised hopes of a Malaysia Baru (a new Malaysia) and that the country could potentially move beyond the politics of race/ethnicity that had dominated since independence
• But do the results from 2018 suggest that ethnicity is no longer an issue?

Histograms indicate there more variation across constituencies in the proportion votes received by PH than in what proportion votes for BN. We can also confirm by calculating the standard deviation of vote shares.

sd(ge18$pctBN, na.rm = TRUE)

## [1] 15.18926

sd(ge18$pctPH, na.rm = TRUE)

## [1] 20.82529

Estimate linear models for PH Vote Share and all ethnicities:

lm18.PH.BUMI    <-lm(pctPH ~ BUMI,    data=ge18)
lm18.PH.CHINESE <-lm(pctPH ~ CHINESE, data=ge18)
lm18.PH.INDIAN  <-lm(pctPH ~ INDIAN,  data=ge18)
lm18.PH.OTHER   <-lm(pctPH ~ OTHER,   data=ge18)

Check the model output for BUMI in 2018:

coef(lm18.PH.BUMI)

## (Intercept)        BUMI 
##  93.3133200  -0.7146756

A quick subset will allow us to identify and investigate:

ge18[which(ge18$pctPH > 70 & ge18$BUMI > 90),]

##     DISTNO STATE DISTNAME VOTESBN VOTESPH VOTERS TURNOUT AREA BUMI CHINESE
## 189   P189 Sabah SEMPORNA    6135   26809  48248   34613 1177 98.1     1.6
##     INDIAN OTHER    pctBN    pctPH
## 189     NA   0.3 17.72455 77.45356

The constituency is named Semporna and is located in Sabah in East Malaysia.

For those who are familiar with Sabah, what makes Sabah different from other states in Malaysia? Alternatively, what happened in the state during the 2018 election that could help to explain why voters behaved differently than the model would predict?

First and foremost, ethnic identity is different and more complicated in Sabah.

• Much of Sabah’s population is legally classified as Bumiputera
• But this is not the same thing as saying that people in Sabah fit into the MCIO model of race/ethnicity that shapes identity and you have talked about in CSI.
  • Sabah only became part of Malaysia in 1963.
  • Prior to this, British colonial authorities did not use MCIO to divide and rule in Sabah as they had in Singapore and Malaya.
• The bumiputera population of Sabah is more diverse in terms of religion, language, and traditions than the bumiputera population in Malaya:
  • Not everyone who is Bumiputera in Sabah speaks the Malay language, or is Muslim.
  • Over 42 distinct ethno-linguistic groups and 200 sub-groups.

Second and as a result, the relationship between ethnic identity and politics is more complicated in Sabah than it is in peninsular Malaysia.

Other forms of localised identity also matter for national politics:

• Identity as a Sabahan — one who was born in or identifies with Sabah — is arguably more important that ethnic identity in many contexts.
  • Whether legally recognized as Bumiputera, Chinese, or Indian, you can (arguably) be a Sabahan.
• If you identify as a Sabahan, then it might also affect your politics, how you vote, and which parties or coalitions you support…

To this end, the 2018 General Election also saw the emergence of a new, multi-ethnic party in Sabah: The Sabah Heritage Party (Parti Warisan Sabah or Warisan).

• Warisan positioned itself in favor of greater regional autonomy for Sabah.
• It also allied itself with the opposition Pakatan Harapan (PH) coalition.
• It proved to be quite popular across the state among bumiputera and non-bumiputera voters…

The solid red points are from Sabah while the pink ones are not.

The observed PH vote share is shifted upwards from what the model would predict (i.e. leverage). Things are indeed different in Sabah. But does this influence the slope?

Comparing the two lines shows that excluding Sabah does not influence the line of best fit by that much. Sabah constituencies are still different, but not influential with regards to the slope between PH vote share and bumiputera voters.

To recap, the learning goals for today’s class were as follows:

• Use R-squared to assess predictive power of linear models
  • We used R-squared to compare the predictive power of models
• Examine the residuals from linear models
  • We used residual plots and the shape of the scatterplots to assess whether the relationship met the conditions for a linear model
  • We compared residual standard errors from a variety of models to discern which models accounted for more of the variability
• Think about outliers, leverage, and influence
  • We identified outliers from Sabah in the 2018 General election
  • We then showed that outliers may have high leverage, but they are not necessarily influential (i.e. they do not influence the slope by much)