3 Statistical Analysis II

Last week, we explored methods for comparing groups. This week, our focus shifts to examining relationships between variables, aiming to assess the extent to which two variables covary: how changes in one variable are associated with changes in another. For example, what is the link between greenhouse gas emissions (x) and global mean temperature (y)? How are earnings (y) related to the time spent in education (x)? To what extent are house prices or local health outcomes (y) associated with proximity to urban greenspace (x)? To answer these sorts of questions we need techniques for unpacking relationships between variables.

3.1 Lecture slides

You can download the slides of this week’s lecture here: [Link].

3.2 Reading list

Essential readings

  • Field, A. 2026. Discovering Statistics Using R and RStudio (2nd edition), Chapter 7: Correlation. [Link]
  • Field, A. 2026. Discovering Statistics Using R and RStudio (2nd edition), Chapter 8: The linear model (regression). [Link]

Suggested readings

  • Field, A. 2026. Discovering Statistics Using R and RStudio (2nd edition), Chapter 2: The SPINE of statistics. [Link]

3.3 Start of session

Before working through the tutorial, we will spend a few minutes looking at:

  • A regression prediction worked out by hand
  • The error “object ‘aged_50’ not found”

3.4 Election results in England and Wales

Today, we will investigate the political geography of England and Wales, focusing on the results of the July 2024 General Election, which was won by the Labour Party led by Keir Starmer. You will work with data extracted from two data sources: the constituency results from the election and socio-demographic information relating to age groups, economic status, and ethnic background from the 2021 Census, extracted using the Custom Dataset Tool. These datasets have been prepared and merged. Please download the file using the link below and save it in your project folder within the data directory.

File Type Link
England and Wales Parliamentary Constituencies GE2024 csv Download

To download a csv file that is hosted on GitHub, click on the Download raw file button on the top right of your screen and it should download directly to your computer.

To get started, open your GEOG0018 R Project and create a new script: File -> New File -> R Script. Save your script as w08-election-analysis.r.

We will start by loading the libraries that we will need:

R code
# load libraries
library(tidyverse)
library(janitor)
library(descr)
library(ggcorrplot)
library(easystats)

3.4.1 Data loading

Now that we are set up again, we can load the EW-GE2024-Constituency-Vars.csv file into R.

R code
# load election data
elec_vars <- read_csv("data/EW-GE2024-Constituency-Vars.csv")
Rows: 575 Columns: 28
── Column specification ────────────────────────────────────────────────────────
Delimiter: ","
chr  (6): constituency_code, constituency_name, region_name, winning_party, ...
dbl (22): eligible_voters, valid_votes, conservative_votes, labour_votes, li...

ℹ Use `spec()` to retrieve the full column specification for this data.
ℹ Specify the column types or set `show_col_types = FALSE` to quiet this message.

As always, we will start by inspecting the dataset to understand its structure and the variables it contains:

R code
# inspect
head(elec_vars)
# A tibble: 6 × 28
  constituency_code constituency_name  region_name winning_party eligible_voters
  <chr>             <chr>              <chr>       <chr>                   <dbl>
1 W07000081         Aberafan Maesteg   Wales       Lab                     72580
2 E14001063         Aldershot          South East  Lab                     78553
3 E14001064         Aldridge-Brownhil… West Midla… Con                     70268
4 E14001065         Altrincham and Sa… North West  Lab                     74025
5 W07000082         Alyn and Deeside   Wales       Lab                     75790
6 E14001066         Amber Valley       East Midla… Lab                     71546
# ℹ 23 more variables: valid_votes <dbl>, conservative_votes <dbl>,
#   labour_votes <dbl>, libdem_votes <dbl>, conservative_vote_share <dbl>,
#   labour_vote_share <dbl>, libdem_vote_share <dbl>,
#   aged_15_years_and_under <dbl>, aged_16_to_24_years <dbl>,
#   aged_25_to_34_years <dbl>, aged_35_to_49_years <dbl>,
#   aged_50_years_and_over <dbl>, eco_not_applicable <dbl>,
#   eco_active_employed <dbl>, eco_active_unemployed <dbl>, …
# inspect column names
names(elec_vars)
 [1] "constituency_code"       "constituency_name"      
 [3] "region_name"             "winning_party"          
 [5] "eligible_voters"         "valid_votes"            
 [7] "conservative_votes"      "labour_votes"           
 [9] "libdem_votes"            "conservative_vote_share"
[11] "labour_vote_share"       "libdem_vote_share"      
[13] "aged_15_years_and_under" "aged_16_to_24_years"    
[15] "aged_25_to_34_years"     "aged_35_to_49_years"    
[17] "aged_50_years_and_over"  "eco_not_applicable"     
[19] "eco_active_employed"     "eco_active_unemployed"  
[21] "eco_inactive"            "eth_asian"              
[23] "eth_black"               "eth_mixed"              
[25] "eth_white"               "eth_other"              
[27] "pop_50plus_40percent"    "pop_white_90percent"    

You can further inspect the dataset using the View() function.

3.4.2 Crosstabulation

In geography, we frequently work with categorical variables. One technique for assessing relationships between two categorical variables is crosstabulation. For example, if we hypothesise that constituencies with an older population are more likely to be won by the Conservative Party, we could crosstabulate the winning party against constituencies where more than 40% of the population is over 50 years old (pop_50plus_40percent).

R code
# crosstabulation
crosstab(elec_vars$winning_party, elec_vars$pop_50plus_40percent, prop.r = TRUE,
    plot = FALSE)
   Cell Contents 
|-------------------------|
|                   Count | 
|             Row Percent | 
|-------------------------|

================================================
                           elec_vars$pop_50plus_40percent
elec_vars$winning_party       no     yes   Total
------------------------------------------------
Con                          25      91     116 
                           21.6%   78.4%   20.2%
------------------------------------------------
Lab                         243     131     374 
                           65.0%   35.0%   65.0%
------------------------------------------------
LD                           24      42      66 
                           36.4%   63.6%   11.5%
------------------------------------------------
Other                         8      11      19 
                           42.1%   57.9%    3.3%
------------------------------------------------
Total                       300     275     575 
================================================

We observe that in the 116 constituencies won by the Conservatives, 78.4% have more than 40% of their population aged over 50. In contrast, only 35.0% of the 374 constituencies won by the Labour Party have a similar demographic. This suggests an association between a large (> 40%) proportion of individuals over 50 years old (x) and the election results (y).

Note that these data describe constituencies, not individual voters. We can say that constituencies with an older population tend to be won by the Conservatives, but not that older people vote for the Conservatives. Drawing conclusions about individuals from data about areas is called the ecological fallacy.

We can use a Chi-square test to assess whether this association is statistically significant:

R code
# chi-square
crosstab(elec_vars$winning_party, elec_vars$pop_50plus_40percent, expected = TRUE,
    plot = FALSE, chisq = TRUE)
   Cell Contents 
|-------------------------|
|                   Count | 
|         Expected Values | 
|-------------------------|

================================================
                           elec_vars$pop_50plus_40percent
elec_vars$winning_party       no     yes   Total
------------------------------------------------
Con                           25      91     116
                            60.5    55.5        
------------------------------------------------
Lab                          243     131     374
                           195.1   178.9        
------------------------------------------------
LD                            24      42      66
                            34.4    31.6        
------------------------------------------------
Other                          8      11      19
                             9.9     9.1        
------------------------------------------------
Total                        300     275     575
================================================

Statistics for All Table Factors

Pearson's Chi-squared test 
------------------------------------------------------------
Chi^2 = 75.53043      d.f. = 3      p = 2.79e-16 

        Minimum expected frequency: 9.086957 

The Chi-square test is a statistical method used to assess whether there is a significant association between two categorical variables. It compares the observed frequencies in each category to the expected frequencies, which are based on the assumption that there is no relationship between the variables (the null hypothesis). If the differences between observed and expected values are large enough, we reject the null hypothesis, indicating a significant association between the variables. Note that the test tells us whether an association exists, not how strong it is: for that, compare the percentages in the crosstab.

The \(p\)-value of the Chi-square test is < 0.001, so the association is statistically significant. The test assumes that the expected counts are large enough (conventionally at least 5 in every cell). The smallest expected count here is 9.1, so this assumption is met.

Crosstabulations are very useful if you have gathered your own survey data, where many variables will often be categorical.

3.4.3 Correlation

With continuous data, we can explore associations more deeply. For example, we could investigate the relationship between the proportion of individuals over 50 years old in each parliamentary constituency (aged_50_years_and_over) and the proportion of votes cast for the Conservative Party (conservative_vote_share). A scatterplot is an ideal starting point for comparing two continuous variables.

We can create a scatterplot as follows:

R code
# scatterplot
plot(elec_vars$aged_50_years_and_over, elec_vars$conservative_vote_share, xlab = "Proportion of population over 50 years old",
    ylab = "Proportion of votes for the Conservative party")
Figure 1: Quick scatterplot

The scatterplot supports our earlier hypothesis that the Conservative vote is higher in constituencies with a larger population of individuals over 50 years old. The next step is to quantify the strength and direction of this relationship. We can assess whether the variables are correlated using Pearson’s Correlation Coefficient, which ranges from -1 (perfect negative correlation) to 0 (no correlation) to +1 (perfect positive correlation).

Pearson’s Correlation Coefficient measures the linear relationship between two continuous variables. A coefficient close to 1 indicates a strong positive correlation, meaning that as one variable increases, the other tends to increase as well. Conversely, a coefficient close to -1 signifies a strong negative correlation, where an increase in one variable is associated with a decrease in the other. A coefficient near 0 suggests little to no linear relationship between the variables.

We can run a correlation as follows:

NotePredict first

Look at the scatterplot again. Do you expect the correlation coefficient to be positive or negative, and is it closer to 0 or to 1? Check your guess against the output.

R code
# correlation
cor.test(elec_vars$aged_50_years_and_over, elec_vars$conservative_vote_share, method = "pearson")

    Pearson's product-moment correlation

data:  elec_vars$aged_50_years_and_over and elec_vars$conservative_vote_share
t = 16.472, df = 573, p-value < 2.2e-16
alternative hypothesis: true correlation is not equal to 0
95 percent confidence interval:
 0.5086971 0.6199173
sample estimates:
      cor 
0.5668849 

The correlation coefficient is 0.57 (95% confidence interval: 0.51 to 0.62) with a \(p\)-value < 0.001. This indicates a modest (0.40 to 0.69 on the scale we use), statistically significant positive correlation: constituencies with a larger share of people over 50 tend to have a higher Conservative vote share.

The Pearson test is parametric and requires interval or ratio variables with a linear relationship. If the data are ordinal or skewed, Spearman’s rank correlation (method = 'spearman') is a better option, as it provides a non-parametric alternative.

It can be useful to look at a matrix of correlation coefficients between a selection of variables within your dataset. To generate this, we can use the cor() function and subsequently visualise the results:

R code
# select variables
elec_vars_sel <- elec_vars |>
    select("conservative_vote_share", "labour_vote_share", "libdem_vote_share", starts_with("aged"),
        starts_with("eco"), starts_with("eth"))

# correlation matrix
cor_mat <- cor(elec_vars_sel)

# correlation plot
ggcorrplot(cor_mat, outline.col = "#ffffff", tl.cex = 8, legend.title = "Correlation")
Figure 2: Correlation plot of selected variables.

3.4.4 Regression

Before running a regression, we need to know what type of variable we want to explain (the dependent variable, y) and what types of variables we use to explain it (the independent variables, x). The type of the dependent variable determines which type of regression we can use. The independent variables can be of any type, although categorical variables need to be converted into dummy variables. In this chapter we only use linear regression, but Table 1 also lists other types of regression. The methods from the previous chapter are summarised in the table in that chapter.

Table 1: Types of variables in regression models.
Type of variable Example As dependent variable (y) As independent variable (x)
Continuous Conservative vote share Linear regression Used as it is
Binary Whether the Conservatives won (yes or no) Logistic regression Dummy variable (0 or 1)
Count Number of crimes Poisson regression Used as it is, often transformed
Categorical with more than two categories Region Multinomial logistic regression Dummy variables, with one reference category
Ordinal Agreement on a five-point scale Ordinal logistic regression Treated as categorical, or as numeric if the steps between the categories can be assumed to be equal

3.4.4.1 Bivariate regression

A Pearson’s correlation describes the strength and direction of a linear relationship, but we can also quantify the change in the dependent variable (y) for a one-unit change in the independent variable (x) with an Ordinary Least Squares (OLS) regression.

Let us run a model with the proportion of the population that is over 50 years old as x and the Conservative vote share as y:

NotePredict first

Both variables are proportions. For each extra percentage point of people over 50, do you expect the Conservative vote share to rise by less than one percentage point, by about one, or by more than one? Check your guess against the slope.

R code
# regression
lm_voteshare <- lm(conservative_vote_share ~ aged_50_years_and_over, data = elec_vars)

# summary
summary(lm_voteshare)

Call:
lm(formula = conservative_vote_share ~ aged_50_years_and_over, 
    data = elec_vars)

Residuals:
     Min       1Q   Median       3Q      Max 
-0.26274 -0.05384  0.00359  0.05877  0.31701 

Coefficients:
                       Estimate Std. Error t value Pr(>|t|)    
(Intercept)            -0.03884    0.01772  -2.192   0.0288 *  
aged_50_years_and_over  0.74008    0.04493  16.472   <2e-16 ***
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Residual standard error: 0.08345 on 573 degrees of freedom
Multiple R-squared:  0.3214,    Adjusted R-squared:  0.3202 
F-statistic: 271.3 on 1 and 573 DF,  p-value: < 2.2e-16

The Estimates column shows the regression coefficients: -0.039 for the intercept (where the line crosses the y-axis when x = 0) and 0.740 for aged_50_years_and_over (the slope of the line, indicating how y changes on average as x increases by 1). Both variables are measured as proportions between 0 and 1, so an increase of 1 in x would mean going from no one to everyone being over 50. The intercept is not substantively meaningful here, as no constituency has a share of people over 50 anywhere near zero. A more useful reading is per percentage point: for every additional percentage point of people over 50 in a constituency, the Conservative vote share is on average 0.74 percentage points higher.

The \(R^2\) (R-squared) of a regression model shows how well the independent variables explain the variation in the dependent variable. It ranges from 0 to 1, where 0 means the model explains none of the variation, and 1 means it explains all of it. Our model suggests that 32% of variation in the dependent variable is explained by the model. The remaining 68% is due to factors not captured by the model or random error.

3.4.4.2 Multiple regression

Now, let us include additional predictors in our regression model to see if we can better explain the variation in the Conservative vote share. We will include eco_active_employed (the proportion of economically active employed individuals) and eth_white (the proportion of the population that identifies as white):

R code
# regression
lm_voteshare <- lm(conservative_vote_share ~ aged_50_years_and_over + eco_active_employed +
    eth_white, data = elec_vars)

# summary
summary(lm_voteshare)

Call:
lm(formula = conservative_vote_share ~ aged_50_years_and_over + 
    eco_active_employed + eth_white, data = elec_vars)

Residuals:
      Min        1Q    Median        3Q       Max 
-0.271771 -0.045878  0.006598  0.053589  0.218631 

Coefficients:
                       Estimate Std. Error t value Pr(>|t|)    
(Intercept)            -0.50862    0.04462 -11.399  < 2e-16 ***
aged_50_years_and_over  1.20724    0.06972  17.315  < 2e-16 ***
eco_active_employed     1.01807    0.08674  11.737  < 2e-16 ***
eth_white              -0.22281    0.03096  -7.197 1.95e-12 ***
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Residual standard error: 0.07408 on 571 degrees of freedom
Multiple R-squared:  0.4671,    Adjusted R-squared:  0.4643 
F-statistic: 166.8 on 3 and 571 DF,  p-value: < 2.2e-16

The model \(R^2\) has increased to 47%, and the adjusted \(R^2\), which accounts for the number of predictors in the model, to 46%. This indicates that adding both eco_active_employed and eth_white has improved our ability to predict Conservative vote share in 2024.

The coefficients in a multiple regression model represent the predicted change in the dependent variable (y) with a one-unit increase in each independent variable (x), while holding all other variables in the model constant. This makes the coefficients partial regression coefficients, meaning they show the association between each predictor and the dependent variable after accounting for the others.

This is why the coefficient for aged_50_years_and_over has changed from 0.74 to 1.2. The share of people over 50 is strongly related to the other predictors (its correlation with eth_white is 0.80), so its association with the Conservative vote share looks different once they are held constant.

NoteTry it yourself

Run the model again without eth_white, so with only aged_50_years_and_over and eco_active_employed. Do you expect \(R^2\) to go up or down? Do you expect the coefficient for aged_50_years_and_over to stay at 1.2 or to move back towards 0.74? Run it and compare.

This is a good start, but as geographers, we should also consider whether regional patterns influence voter share. We can use a boxplot to split out the vote shares by region to see whether there is any evidence of regional differentiation:

R code
# change region variable order
elec_vars <- elec_vars |>
    mutate(region_name = factor(region_name, levels = c("London", "South East", "South West",
        "North East", "North West", "East Midlands", "West Midlands", "East of England",
        "Yorkshire and The Humber", "Wales")))

# boxplot
boxplot(elec_vars$conservative_vote_share ~ elec_vars$region_name, xlab = "Region",
    ylab = "Conservative vote share")
Figure 3: Boxplot on the Conservative voter share by region.

The boxplots suggest some regional variation and so we will add the region_name variable to our regression model. To do this, we need to transform the categorical variable into dummy variables: variables that are coded 1 for the constituencies in a region and 0 for all other constituencies. With 10 regions, R creates nine dummy variables, because one region has to be left out as the reference category. In the code above we put London first in the order of the regions, so London is the reference category.

R will detect that you have a categorical variable if it is stored as a character or factor and will do the dummy generation for you. The coefficients for the included dummy variables indicate how much the predicted value of the dependent variable differs for constituencies in that region compared to the reference region, with the other variables in the model held constant.

R code
# regression
lm_voteshare <- lm(conservative_vote_share ~ aged_50_years_and_over + eco_active_employed +
    eth_white + region_name, data = elec_vars)

# summary
summary(lm_voteshare)

Call:
lm(formula = conservative_vote_share ~ aged_50_years_and_over + 
    eco_active_employed + eth_white + region_name, data = elec_vars)

Residuals:
      Min        1Q    Median        3Q       Max 
-0.212739 -0.040914 -0.000956  0.036534  0.251386 

Coefficients:
                                     Estimate Std. Error t value Pr(>|t|)    
(Intercept)                         -0.387395   0.049488  -7.828 2.48e-14 ***
aged_50_years_and_over               1.096071   0.065739  16.673  < 2e-16 ***
eco_active_employed                  0.747208   0.101295   7.377 5.86e-13 ***
eth_white                           -0.183626   0.037209  -4.935 1.06e-06 ***
region_nameSouth East                0.056109   0.013387   4.191 3.22e-05 ***
region_nameSouth West                0.013635   0.015542   0.877  0.38067    
region_nameNorth East               -0.020671   0.020219  -1.022  0.30705    
region_nameNorth West               -0.039032   0.015201  -2.568  0.01049 *  
region_nameEast Midlands             0.046112   0.015404   2.993  0.00288 ** 
region_nameWest Midlands             0.042555   0.014559   2.923  0.00361 ** 
region_nameEast of England           0.059563   0.014387   4.140 4.00e-05 ***
region_nameYorkshire and The Humber -0.002996   0.015842  -0.189  0.85006    
region_nameWales                    -0.053131   0.019101  -2.782  0.00559 ** 
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Residual standard error: 0.06529 on 562 degrees of freedom
Multiple R-squared:  0.5925,    Adjusted R-squared:  0.5838 
F-statistic:  68.1 on 12 and 562 DF,  p-value: < 2.2e-16

The coefficient for each region represents the average difference in the predicted Conservative vote share compared to the reference region, which is London in this case.

The summary() output gives a separate \(t\)-test for each region compared to London, but none of these tests tells us whether region matters overall. We can answer that question with an \(F\)-test, which compares the model with and without the region_name variable. This is the same logic as the ANOVA from last week:

ANOVA is a regression too

An ANOVA is a linear regression with only a categorical predictor. The coefficients are then the differences between each group mean and the mean of the reference group, and the F-statistic at the bottom of the summary() output is the F-statistic of the ANOVA. Our model also contains continuous predictors, but the idea is the same.

We use the drop1() function for this, which removes one variable at a time from the model and tests how much the fit changes. You may come across anova(model) used for the same purpose, but that function adds the variables in the order in which they appear in the formula, so only the row of the last variable gives the same result. Focus on the F value and Pr(>F) columns.

R code
# test each variable
drop1(lm_voteshare, test = "F")
Single term deletions

Model:
conservative_vote_share ~ aged_50_years_and_over + eco_active_employed + 
    eth_white + region_name
                       Df Sum of Sq    RSS     AIC F value    Pr(>F)    
<none>                              2.3958 -3125.4                      
aged_50_years_and_over  1   1.18511 3.5809 -2896.3 277.995 < 2.2e-16 ***
eco_active_employed     1   0.23197 2.6278 -3074.2  54.414 5.863e-13 ***
eth_white               1   0.10383 2.4997 -3103.0  24.355 1.057e-06 ***
region_name             9   0.73757 3.1334 -2989.0  19.224 < 2.2e-16 ***
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

The row for region_name has an F-statistic of 19.2 with a \(p\)-value < 0.001. We can conclude that region helps to explain the Conservative vote share, even after accounting for age, employment and ethnicity. This does not mean that every region differs from London: in the summary() output, for example, the South West does not.

3.4.4.3 Diagnostics

The validity of a regression relies on several assumptions being met. One of the most important assumptions is that the residuals (the errors of the model: the differences between observed and predicted values) should be independent and approximately normally distributed. The easystats package provides tools to quickly generate relevant figures and diagnostics to help you assess your model.

Check for normality of the residuals:

R code
# run diagnostics
performance::check_model(lm_voteshare, check = c("qq", "normality"))
Figure 4: Check for normality of residuals.

The dots in the left-hand panel follow the line closely, with only small deviations in the upper tail, and the distribution of the residuals in the right-hand panel is close to the normal curve. There is no serious cause for concern.

Another important assumption is homoscedasticity, which occurs when the residuals exhibit constant variance across all levels of the independent variable(s). In contrast, heteroscedasticity arises when the variance of the residuals is not constant. Linearity is another key assumption in linear regression: each extra unit of an independent variable is assumed to add the same amount to the dependent variable. If the relationship curves, the model misses part of it.

Check for linearity and homogeneity of variance (homoscedasticity):

R code
# run diagnostics
performance::check_model(lm_voteshare, check = c("linearity", "homogeneity"))
Figure 5: Check for linearity and homogeneity of variance.

In the linearity panel the reference line should be flat and horizontal. Here it is close to flat across the bulk of the data, and the bends at either end rest on very few constituencies (note the wide grey bands), so this is at most a mild deviation from linearity rather than a clear violation. In the homogeneity of variance panel the reference line is also roughly flat, so there is no clear sign of heteroscedasticity.

Multicollinearity is another potential issue. Multicollinearity arises when independent variables are highly correlated with each other. This situation can complicate regression analysis, as it makes it challenging to isolate the individual effects of each correlated variable on the dependent variable. Finally, a few influential observations can skew the results and affect the overall model fit.

Multicollinearity can be assessed using Variance Inflation Factors (VIF). VIF measures how much the variance of an estimated regression coefficient increases because your predictors are correlated. For each predictor it is calculated by regressing that predictor on all the other predictors and taking \(1 / (1 - R^2)\) of that regression. A VIF of 1 means that a predictor is not correlated with the others. As a rule of thumb, values above 5 indicate moderate and values above 10 indicate problematic multicollinearity.

We can check for influential observations and multicollinearity as follows:

R code
# run diagnostics
performance::check_model(lm_voteshare, check = c("vif", "outliers"))
Figure 6: Check for influential observations and multicollinearity.

All VIF values are below 10, and none of the constituencies fall outside the contour lines in the influential observations panel. The predictor eth_white has the highest VIF, slightly above 5, which is moderate and not a reason for concern here.

If you encounter multicollinearity issues, you have several options to address them. One straightforward approach is to remove one of the correlated predictors from your model. Alternatively, you could combine the correlated predictors into a single measure, such as an index. Converting a continuous variable into categories is not usually a good solution, as it discards information.

Common misconceptions about regression

  • “The variables need to be normally distributed.” The assumption concerns the residuals of the model, not the variables themselves. The variables can be skewed.
  • “A high \(R^2\) means that the model is a good model.” The \(R^2\) only tells us how much of the variation the model explains. It says nothing about whether the assumptions hold or whether the relationships are causal.
  • “A significant coefficient means that the effect is large.” A small \(p\)-value tells us that the coefficient is unlikely to be zero, not that it is large. Look at the size of the coefficient in the units of the variables.
  • “A significant regression shows that x causes y.” A regression describes associations. Whether an association is causal depends on the research design and on what is not in the model.
  • “Linear regression can only describe straight lines.” The model has to be linear in its coefficients. Transformations or squared terms of the independent variables can describe curved relationships.
  • “If the diagnostic plots look fine, the model is correct.” The plots check only some of the assumptions. They cannot tell us whether we have left out an important variable, or whether neighbouring constituencies resemble each other, which makes the observations not fully independent.

3.5 Homework task

This concludes this week’s tutorial. Now complete the following homework tasks making use of the elec_vars dataset:

  1. Analyse the association between the pop_white_90percent and winning_party variables.
  2. Build a multiple regression model with the labour_vote_share as dependent variable for constituencies in England.
NoteAssignment appendix

Paste the test outputs in the appendix of your assignment and include a few sentences interpreting the results.

3.6 Before you leave

This week, we focused on statistically assessing associations between different variables. Because this week’s material was a little technical, we will shift our focus away from R next week and turn our attention to QGIS. Time to sit down and process this week’s material before some more relaxed content next week!