Confounding Variables

 

Let us come focus now on a concept that we have seen a few times before: confounding variables.  A confounding variable is an extraneous variable (i.e., a variable that is not a focus of the study) that is statistically related to (or correlated with) the independent variable.  This means that as the independent variable changes, the confounding variable changes along with it.  The result is that subjects in one condition are different in some unintended way from subjects in the other condition.  This is bad because the point of an experiment is to create a situation in which the only difference between conditions is a difference in the independent variable.  This is what allows us to conclude that the manipulation is the cause of differences in the dependent variable.  But if there is some other variable that is changes along with the independent variable, then this confounding variable could be the cause of any difference.

 

Imagine a randomized experiment concerning the effect of noise on concentration.  There are 50 subjects in the quiet condition and 50 in the noisy condition.  The ideal version of this experiment is represented by the table below.  The independent variable is labeled “IV” and extraneous variables are labeled “EV.”  Note that the only variable that differs systematically across the two conditions is the independent variable itself (which is highlighted in the table).  There is little noise in the quiet condition and lots of noise in the noisy condition.  The rest of the variables—IQ, room temperature, etc.—are the same across the two conditions.  They have been controlled.  So if there is a difference in the concentration levels of subjects in the quiet and noisy conditions, it must be caused by the independent variable.

 

 

Variables

Quiet Condition

Noisy Condition

Noise Level (IV)

Low

High

IQ (EV)

Average

Average

Room Temperature (EV)

68 Degrees

68 Degrees

Sex of Subjects (EV)

60% F

60% F

Task Difficulty (EV)

Moderate

Moderate

Time of Day (EV)

All Different Times bet. 9 – 5

All Different Times bet. 9 – 5

Etc. (EV)

Same as Noisy Cond.

Same as Quiet Cond.

Etc. (EV)

Same as Noisy Cond.

Same as Quiet Cond.

 

An Ideal Experiment

        

Now imagine a less than ideal version of this experiment, with some other variables that differ systematically across conditions.  These are confounding variables (all highlighted).  Now if there is a difference in the concentration levels of subjects in the quiet and noisy conditions, it could be caused by the independent variable … but it could also be caused by any of the confounding variables.  If subjects in the quiet condition have greater concentration levels, is it because it was quiet, because the temperature was not too hot, or because they were tested in the morning?  There is no way to tell.  Obviously, this is less than ideal.

 

 

Variables

Quiet Condition

Noisy Condition

Noise Level (IV)

Low

High

IQ (EV)

Average

Average

Room Temperature (EV)

68 Degrees

82 Degrees

Sex of Subjects (EV)

60% F

60% F

Task Difficulty (EV)

Moderate

Moderate

Time of Day (EV)

Morning

Afternoon

Etc. (EV)

Same as Noisy Cond.

Same as Quiet Cond.

Etc. (EV)

Same as Noisy Cond.

Same as Quiet Cond.

        

A Non-Ideal Experiment

 

How Do You Control Confounding Variables?

 

There are many ways to control extraneous variables so that they do not become confounding variables.  Essentially all person variables can be controlled by random assignment.  If you randomly assign subjects to conditions, then on average they will be equally intelligent, equally outgoing, equally motivated, and so on.  Again, random assignment does not guarantee that this is true of every extraneous variable in every experiment.  But when you use large enough samples (we will talk more about what is “large enough” later), it works extremely well and you will never be second-guessed for using it.

 

One way to control situation variables or task variables is simply to hold them constant.  In the noise-concentration experiment, for example, you would simply hold the room temperature constant by setting the thermostat and testing everyone in the same room.  You would, of course, hold task difficulty constant by giving subjects in both conditions the same task.  In general, instructions that are given to subjects are usually written down or recorded and presented in exactly the same way for every subject.

 

Sometimes it is difficult or impossible to hold a situation or task variable constant.  In such cases, random assignment comes to the rescue again.  Imagine that you cannot test all subjects in a two-group experiment in the same room; the schedule is such that on Tuesday you have to use Room 222 and on Wednesday you have to use Room 247.  Note that if the Tuesday subjects are randomly assigned to conditions, then half will be in Condition A and the rest in Condition B.  The same is true of the Wednesday subjects.  This means that both conditions will have roughly the same percentage of subjects tested in Room 222 and 247.  Note also what would happen if you did not use random assignment and instead tested the Tuesday subjects in Condition A and the Wednesday subjects in Condition B.  Now you have two confounding variables.  Subjects in Conditions A and B were tested on different days of the week and in different rooms.  Any difference between conditions could be due to the independent variable, the day of the week, or the room.

 

Experiments vs. Correlational Studies Again

 

Remember that experiments include the manipulation of an independent variable.  This is really what makes an experiment an experiment.  But good experiments also include some attempt to control extraneous variables so that they do not become confounding variables—usually through random assignment, holding extraneous variables constant, and other means.  An ideal experiment ends up looking like the first table above.

 

Remember also that correlational studies do not include the manipulation of an independent variable.  This means no random assignment and usually no other means of controlling extraneous variables so that they do not become confounding variables.  The table below represents a correlational study in which one simply asks people whether or not they got an allowance and measures their financial responsibility.  Note that there now all sorts of confounding variables (which are often called third variables in the context of correlational studies).  Again, this is why you would not want to interpret a statistical relationship between getting an allowance and financial responsibility as evidence that former caused the later.  The cause could be any of a large number of confounding variables.   

 

 

Variables

Got an Allowance

Did Not Get an Allowance

Allowance (Quasi-IV)

Yes

No

IQ (EV)

Higher on Average

Lower on Average

Parents’ Financial Resp. (EV)

Higher on Average

Lower on Average

Socioeconomic Status (EV)

Higher on Average

Lower on Average

General Responsibility (EV)

Higher on Average

Lower on Average

Family Income (EV)

Higher on Average

Lower on Average

Etc. (EV)

???

???

Etc. (EV)

???

???

        

A Correlational Study