A guide to interpreting regression tables

Semra Sevi
Université de Montréal
Department of Political Science
email: [email protected]
website: www.semrasevi.com

January 2021
1 Types of variables
• The type of model researchers use will depend on the variables they have. Take the time
to get to know the data and the research design. When you read the assigned articles,
ask yourself how are the dependent (DV) and independent variables (IV) measured.
The independent variable is the variable that influences the value of the dependent
variable (also known as the outcome of interest).

• Continuous variables: these are measured along a numerical scale. For example: age,
and income.

• Discrete variables are variables that take discrete values (0, 1, 2, 3, 4...). There are
two general types:

– Categorical or nominal variables. For example: religion (Protestant, Catholic,

Jewish, Muslim...), parties (Liberal, Conservative, NDP, Green, Bloc...), and oc-
cupation. When a variable takes two categories for example gender (male or
female) this variable is known as a ‘dummy’/dichotomous variable. When they
are an independent variable in a regression model, their interpretation is very
simple. As we move from one category to the next, the associated change in
the dependent variable is...? For example, “women are more likely to vote for
Left parties compared to men.” When there are multiple-category variables in the
models, we can use a similar comparison when we examine their effect. For exam-
ple, lets say we have three-category variable for religion (Catholic, Protestant and
Others). In this case you have to use a reference category. Suppose it is Others.
“Compared to Others, Catholics are 10 percent more likely to vote for the Liberal
Party of Canada, but Protestants are only 3 per cent more likely than Others to
do so.”
– Ordinal-level variables take on values that have a specific order. For example:
degree of satisfaction (very, somewhat, a little, not at all), and degree of interest
in politics (very high, somewhat high, somewhat low, very low).

1
2 Linear Regressions
Linear regressions are a method to test whether/to what extent whenever the IV increases
in value the DV systematically increases/decreases in value. By convention the DV is y and
the IV is x and the relationship between the two is represented by an equation.

Bivariate ordinary least squares (OLS) regression analysis:

• Bivariate means two variables. They can tell us something about the extent to which
two unobserved population variables, x and y, are related.
The equation for a bivariate linear regression can be simply expressed as:

y = β 0 + β 1 x1 + ε

Where :

y = Dependent variable
β0 = Intercept/Constant
β1 = Slope Coefficient
x1 = Independent variable
ε = Error term/epsilon

β0 is the average value of y when the independent variable equals 0. This is rarely meaning-
ful/interpretable so for our purposes it can be ignored.

β1 represents the coefficient for the slope which is the average change in y associated with a
unit change in x.

ε is the error term which represents the remaining variation in y that cannot be explained
with x.

Let’s look at this visually. Suppose you were interested in the relationship between student
grades (X) and earnings (Y ). Your hypothesis is that higher grades cause higher salaries.

Grades (X) Earnings (Y )

But do they really? Is it possible that some other variable is affecting both student grades
(X) and earnings (Y )? Could you think of a variable that affects both grades and earnings?

Grades (X) Earnings (Y )

2
Suppose you said ability. Students who are more competent tend to get better grades and
they also have higher salaries later in life. Perhaps grades does not matter. What matters
is their intrinsic ability. In regression, we call this an omitted variable bias because in the
earlier example we tried to explain a cause and effect without controlling for all potential
factors that could affect both student grades and earnings. This is why we have multiple
regressions.

Ability (Z)

Grades (X) Earnings (Y )

Multiple OLS regression analysis:

• Often times we need to include multiple variables to control for confounders or to

measure the impact of different IVs and which one has the strongest impact. Mul-
tiple regression provides estimates of the impact of each independent variable on the
dependent variable, accounting for the impact of the other variables in the model.
If there are two variables in the model, the coefficient on x1 (e.g. explanatory vari-
able) indicates its impact on y after controlling for x2 (e.g. additional explanatory
variable/confounding/control variable).

The equation for a multiple regression with two x’s looks like this:

y = β0 + β1 x1 + β2 x2 + ε

Same interpretation as the bivariate linear regression except that in a multiple regression,
β1 represents the coefficient for the slope which is the average change in y associated with a
unit change in x1 when all the other independent variables (x2 ) are held constant.

To get more information about regressions you can consult and play around in this app
created by the Centre for Computational and Quantitative Social Science and the Consortium
on Electorial Democracy at Western University.

3
3 Examples
1. DV & IV’s continuous

Political Ideology in Sweden. Ordinary least squares (OLS) regression models with
political ideology as the dependent variable.

(Model 1) (Model 2)
Age 0.011∗ 0.005
(0.005) (0.005)
Income 0.277∗∗∗
(0.069)
Political Information -0.013
(0.145)
Constant 4.309∗∗∗ 3.696∗∗∗
(0.246) (0.294)
Observations/N 921 921
R2 0.005 0.023
Standard errors in parentheses
∗
p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001

Source: Shane P. Singh’s Course on Statistics and Data Analysis at ICPSR

The first thing to do when reading a table is to look at the Observations (N) and
understand what they are. Here the N are the respondents in a survey.
The DV is left-right ideology (0 means extreme left and 10 means extreme right).
The independent variables are: Age (the age of the respondent in years), Income (in
categories from lowest income to highest income), Political information (number of
political knowledge questions answered correctly). The unit of analysis is respondents
in Sweden in this survey. For each respondent in the survey we have: ideology, age,
income and political information.
The numbers not in the parentheses is called the coefficient. The coefficient provides
the expected change in the dependent variable (party ideology) for one-unit increase
in the independent variable (which is age in model 1).
The numbers in parentheses is called the standard error (we know this because of the
note below the table), which is the estimate of the standard deviation of the coefficient.
A rule of thumb, if a coefficient is twice as large compared to its standard error, then
the coefficient is likely significant at the .05 level. For our purposes, you do not need
to interpret this. Instead, you can look at the p-values (more on this soon).
First thing you should do is figure out what the N represents, what the dependent
variable is and identify the important independent variables that you will base your
interpretation on. Then ask yourself if the coefficient for the IV(s) is positive or
negative. If it is positive, this means there is a positive relationship between the

4
IV and DV. As the IV increases, the DV increases. A negative relationship would
indicate that as the IV increases, the DV decreases. The size of the coefficient for each
independent variable gives you the size of the effect that variable is having on your
dependent variable.
Let’s try interpreting this table.
Here is the equation for model 1:

Ideology = β0 + β1 × Age + ε

From our table above, we get the following values for our coefficients:

Ideology = 4.309 + .011 × Age + ε

In model 1 (bivariate), we see that as your age increases by one year, your party ideology
goes .011 point more to the right. So we can say that a ten-year increase in age, is
associated with, on average, a 0.11 unit rightward shift in ideology (.011*10=0.11).
You can tell that this is statistically significant (at the .05 level) by looking at the
asterisks.
There are two basic hypotheses: the null and an alternative. The null hypothesis
means that x is not linearly associated with y which indicates that the slope is zero.
The alternative hypothesis means that x is linearly associated with y so the slope is
not zero.
So statistically significant means that the β1 which is the coefficient for the effect of
Age on political ideology in this model is statistically different from zero.
The legend tells you that this coefficient is significant at p < 0.05 level. What does
this mean? The ‘p-values’ indicate the statistical significance of the coefficient which
is based off the principle of random sampling. In this model, there is only a 5 in a 100
chance (or less) that there really is no relationship between political ideology and age.
If you want to read more about p-values and significance, you can read about it here.
Note that the it is possible to have a very significant p-value (***) for a very small
effect so the size of the p-value for a coefficient says nothing about the size of the effect
that variable is having on your dependent variable.
Is model 1 substantively significant? Recall that ideology ranges from 0-10 so aging
ten years only corresponds with an expected shift equal to 1 per cent of the scale.
Technically, the constant indicates the ideology of respondents aged zero which is found
to be center left. This obviously does not make any sense which is why we normally
do not interpret the constant. Sometimes, however, authors select variables in such a
way that the constant is interpretable. For example, if they shifted the age variable to
start at 18 instead of 0 then the constant would represent political ideology at age 18.
Now lets look at a multiple regression (model 2). The data tell us that those who are
older are more on the right. But you may have reason that this may not be because
of their age. It is rather that older people tend to be richer and it is the fact that they

5
are rich that ’causes’ them to be on the right, NOT the fact that they are old. So it
is important to control for income. This means we need to have multiple variables in
our model.
Here is the equation for model 2:

Ideology = β0 + β1 × Age + β2 × Income + β3 × Information + ε

From our table above, we get the following values for the coefficients:

Ideology = 3.696 + .005 × Age + .277 × Income + (−.013) × Information + ε

In model 2, with income and political information controlled, age no longer has a
significant impact on ideology. Perhaps the effect of age that we found in model 1 can
be explained by its relationship to income. Consider this as a speculation as the model
does not test for this relationship. However it is possible that in a bivariate model an
independent variable is a significant predictor of a dependent variable but is no longer
significant in a multiple regression. In a multiple regression, the significance levels
given for a given independent variable indicates whether that independent variable is
a significant predictor of the dependent variable, controlling for the other independent
variables.
R2 is low in both models. This is ok. It may mean that there are lots of other things
that contribute to political ideology that we did not add to the model. Or that the
relationship is not linear. But just so you know, R2 ranges from 0 to 1 and has a very
simple interpretation: It is the proportion of variance in y “explained” by x(s). We can
think of the model as defining a line in an n dimensional space where n is the number
of variables. R2 is a measure of how close the data is to this line. If all the points fall
exactly on the line this would indicate a perfect relationship between x and y. When
you add more independent variables to your model, R2 will always increase.

6
2. DV continuous & Main IV dichotomous

The gender gap in Canadian federal elections from 1921-2015. Ordinary least squares
(OLS) regression models with party vote share as the dependent variable.

(Model 1) (Model 2)
Woman -9.877∗∗∗ -0.450∗∗∗
(0.366) (0.135)
Vote share lag 0.276∗∗∗
(0.006)
Party performance 0.666∗∗∗
(0.004)
Incumbent Party 6.783∗∗∗
(0.148)
Distance from contention -0.015∗∗
(0.005)
Constant 27.682∗∗∗ 0.281
(0.139) (0.246)
Observations/N 23903 23903
R2 0.030 0.872
Standard errors in parentheses
∗
p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001

Source: Sevi, Semra, Vincent Arel-Bundock and André Blais. 2018. “Do Women Get Fewer Votes? No.” Canadian
Journal of Political Science. 52(1): 201-210

The equation for model 1 is:

Vote share = β0 + β1 × Woman + ε

From our table above, we get the following values for our coefficients:

Vote share = 27.682 + (−9.877) × Woman + ε

First thing is the N: they refer to all the candidates. In column 1, we have the per-
centage of the vote (DV) obtained by each candidate and whether the candidate is a
man or a woman (IV).
The woman variable is a dummy variable (1 = Woman 0 = Man).
Note that because gender is a dummy variable, the interpretation is slightly different.
It does not make sense to talk about the effect on vote share as woman increases or
decreases (woman is not measured as a continuous variable. See above.)
Model 1 (bivariate) shows that compared to men, women candidate vote share is on
average 10 percentage points lower. Another way to say this is when the candidate is

7
a woman, the vote share is reduced by 10 percentage points. This is a very big effect
and the estimate is statistically significant.
In model 2 (multiple), we have many controls (vote share lag, party performance,
incumbent party and distance from contention as we may expect these variables to
have an impact on both the dependant and the main independent variable). So the
interpretation is: everything else being equal, women candidates get 0.45 percentage
points fewer votes than men. But is this effect substantively important?
Here is the equation for model 2:

Vote share =β0 + β1 × Woman + β2 × Vote share lag

+β3 × Party performance + β4 × Incumbent Party
+β5 × Distance from contention + ε

From our table above, we get the following values for our coefficients:

Vote share =0.281 + (−450) × Woman + .276 × Vote share lag

+.666 × Party performance + 6.783 × Incumbent Party
+(−.015) × Distance from contention + ε

8
3. DV dichotomous

The impact of voting for Trump over Clinton in the 2016 election. Ordinary least
squares (OLS) and Logit regression models with voting for Trump as the dependent
variable.

(Model 1) (Model 2)
OLS Logit
Party 0.163∗∗∗ 1.020∗∗∗
(0.004) (0.048)
Female -0.066∗∗∗ -0.551∗∗∗
(0.018) (0.165)
Income -0.007 -0.108
(0.020) (0.179)
65 or over -0.065∗ -0.594∗
(0.029) (0.254)
Latinx -0.104∗∗∗ -0.805∗∗
(0.028) (0.246)
Black -0.132∗∗∗ -1.426∗∗∗
(0.028) (0.315)
Other -0.078∗ -0.569∗
(0.034) (0.278)
Education -0.059∗∗ -0.524∗∗
(0.019) (0.178)
Constant -0.047∗ -3.181∗∗∗
(0.024) (0.214)
R2 0.54
P seudo R2 0.48
Observations/N 1563 1563
Standard errors in parentheses
∗
p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001

Source: Dean Lacy’s Course on Maximum Likelihood Estimation for General Linear Models at ICPSR

The N is the the number of respondents in this survey.

In both the OLS and the logit model, the dependent variable is whether the respondent
voted for Trump. Because this DV takes only a 0 or 1 (0 = Vote for Clinton, 1 = Vote
for Trump), sometimes researchers use a Logit model instead of OLS (we will come
back to this).
The IV’s:
Party is the respondent’s party identification on a 7-point scale, from 1= strong Demo-
crat to 7=strong Republican.

9
Female is a binary variable (another way of saying this is a dummy variable) which is
0 if the respondent is male and 1 if the respondent is a female.
65 or over is also a dummy variable where 0 = under 65 and 1 = 65 and over.
Latinx, Black and Other are all dummy variables. The reference category (0) is White
people.
Education is a dummy variable. B.A. or higher = 1, everyone else = 0.
Note when the DV is dichotomous and its an OLS model, the interpretation is in terms
of probability.
Here is the equation for model 1:

Vote for Trump = β0 + β1 × Party + β2 × Female

+β3 × Income + β4 × 65 or over
+β5 × Latinx + β6 × Black
+β7 × Other + β8 × Education + ε

From our table above, we get the following values for our coefficients:

Vote for Trump = −0.047 + 0.163 × Party + (−0.066) × Female

+(−0.007) × Income + (−0.065) × 65 or over
+(−0.104) × Latinx + (−0.132) × Black
+(−0.078) × Other + (−0.059) × Education + ε

For example, in model 1, the probability of voting for Trump is 7 points lower among
women than among men. The probability of voting for Trump increases by 16 points
for every point shift towards Republicans...
If it’s a Logit model (model 2), we cannot interpret the Logit coefficients directly from
the table. The assessment of the relationship between our independent variable and
the dependent variable proceeds with a discussion of the significance and direction of
the model coefficients or odds ratio (for more information on the latter, consult this
site). Sometimes researchers will show marginal effects to get around this limitation.
Let’s interpret the logit model with significance and direction of the model coefficients:
Party is positive and significant which means that Republicans are more likely to vote
for Trump.
The coefficient for female is negative and significant which means women are less likely
to vote for Trump than men.
65 or over is negative and significant which means that voters that are 65 and over are
less likely to vote for Trump than voters below 65.
The dummy variables for race (Latinx, Black and Other) are all negative and significant
which means that compared to white people, they are less likely to vote for Trump.

10
This effect seems to be most pronounced among black voters, followed by latin and the
other category.
The coefficient on the education variable is negative and significant which means that
voters with a BA or higher are less likely to vote for Trump.
Note: While the coefficients in model 1 (OLS) and 2 (Logit) change, the statistical
significance for each of the independent variables in both models remain the same.
This suggests that we can simply use OLS for dichotomous DV variables to make the
interpretation easier.

11
4. Interactions: DV and IV dichotomous
OLS Interaction Between Partisanship and Political Knowledge on Pro-Choice Atti-
tudes

(Model 1) (Model 2)
Partisanship -0.270∗∗∗ -0.202∗∗∗
(0.016) (0.021)
Political Knowledge 0.137∗∗∗ 0.200∗∗∗
(0.016) (0.020)
Partisanship × Political Knowledge -0.152∗∗∗
(0.032)
Constant 0.642∗∗∗ 0.618∗∗∗
(0.012) (0.013)
Observations 3715 3715
R2 0.083 0.089
Standard errors in parentheses
∗
p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001

Source: ANES, 2012

The N in this dataset are respondents in the ANES in 2012.

DV: In favour of abortion = 1, Opposition of abortion = 0. So higher values = pro-
choice.
Partisanship (0 = Democrats, 1 = Republican). Independents are not included.
Political Knowledge (0 = Doesn’t know Speaker, 1 = Knows Speaker).
Let’s start with model 1.
Here is the equation for model 1:

Pro-choice = β0 + β1 × Partisanship + β2 × Political Knowledge + ε

From our table above, we get the following values for our coefficients:

Pro-choice = 0.642 + (−0.270) × Partisanship + 0.137 × Political Knowledge + ε

We see that the coefficient for Partisanship is negative and significant. This means that
the probability of Republicans supporting pro-choice is 27 percentage points compared
to Democrats. The coefficient for Political Knowledge shows that the probability of
respondents with political knowledge being pro-choice is 14 percentage points higher
than respondents without political knowledge.
For model 2, note there is an interaction so we cannot interpret the coefficients directly.
In general, an interaction effect exists when the impact of one independent variable
(x1 ) depends on the value of another independent variable (x2 ).

12
Here is the equation for model 2:

Pro-choice =β0 + β1 × Partisanship + β2 × Political Knowledge

+β12 × Partisanship × Political Knowledge + ε

From our table above, we get the following values for our coefficients:

Pro-choice =0.618 + (−0.202) × Partisanship + 0.200 × Political Knowledge

+(−0.152) × Partisanship × Political Knowledge + ε

The Partisanship coefficient can be interpreted as the effect of Partisanship when Po-
litical Knowledge = 0.
The Political Knowledge coefficient can be interpreted as the effect of Political Knowl-
edge when Partisanship variable = 0 (among Democrats).
The intercept (0.618) is the probability of being pro-choice among low knowledge
Democrats.
The interaction shows the moderation effect of Political Knowledge on the relationship
between Partisanship and pro-choice. But in order to understand the interaction, we
need to do some calculations.

In what follows, Party is Partisanship and Knowledge is Political Knowledge. This is

simply abbreviated to fit the equation.

Here is the equation for model 2 again:

Pro-choice = β0 + β1 × Party + β2 × Knowledge + β12 × Party × Knowledge + ε

Let’s look at this in more detail (these are the models predicted y values):

y (the probability of being pro-choice) for low knowledge (knowledge = 0) Democrats

(Party = 0) is:

0.618 = 0.618 + (−0.202) × 0 + 0.200 × 0 + (−0.152) × 0 × 0

y (the probability of being pro-choice) for high knowledge (knowledge = 1) Democrats

(Party = 0) is:

0.818 = 0.618 + (−0.202) × 0 + 0.200 × 1 + (−0.152) × 0 × 1

y (the probability of being pro-choice) for low knowledge (knowledge = 0) Republicans

(Party = 1) is:

0.416 = 0.618 + (−0.202) × 1 + 0.200 × 0 + (−0.152) × 1 × 0

13
y (the probability of being pro-choice) for high knowledge (knowledge = 1) Republicans
(Party = 1) is:

0.464 = 0.618 + (−0.202) × 1 + 0.200 × 1 + (−0.152) × 1 × 1

As you can see, there is a big effect for Democrats but not really a difference for
Republicans. This means that knowledge has a big effect among Democrats and
only a small impact among Republicans. Among Democrats, knowledge increases the
probability of being pro-choice by 20 percentage points (from 0.618 to .818) while it
increases that probability by only 5 percentage points among Republicans (from 0.416
to 0.464). The impact of knowledge varies across parties. Knowledge really makes a
difference only among Democrats.
The calculation we did above could also be presented in a figure.
Here is an example:

Figure 1: Marginal Effect Plot By Partisanship

.9
Predicted probability for pro-choice

.8

.7

.6

.5

Low Political Knowledge

.4 High Political Knowledge

Democrat Republican

14