0:00
Hi. Anderson Smith.
And we're talking about threats to internal validity.
We've talked about maturation, and history,
and regression and now we're going to talk about attrition.
The fact that over time,
people drop out of experiments and that
dropping out of experiments means that what you're measuring
at the end of the experiment is not the
same as you are measuring at the beginning of the experiment.
This is called attrition error and the error occurs if inferences are made on the basis
of those participants that have participated only from the start to
the end because they may be different because of attrition.
So attrition is a potential problem when
more than one measurement is made at more than one point in time.
That is I have measurement of 0.1,
measurement of 0.2 and I assume that those are the same when in fact they're
different threat to internal validity of those comparisons at those times,
could be because the people at 0.2 are different from the people at 0.1.
Now, people who drop out are probably different from people who stay in the experiment,
and that really is the nature of the attrition error.
So, let me give you an example.
The percentage of group members having quit smoking
a post-test after an intervention is found
to be much higher in a group having received
a quit-smoking training program than in the control a group.
So, if you quit smoking
after a training intervention than a group that doesn't have a training intervention.
But, the experimental group at
the end of the experiment only has 60% of the original sample.
So 40% of the people have dropped out, and who dropped out?
Probably the people that didn't like the intervention,
that didn't want to stop smoking.
So people who stayed in and got the whole intervention,
they showed a decrease from group that had no intervention,
but it was different from the original group.
So if the attrition is related to the feature of the study,
then the nature of the independent variable,
the instrumentation, dropping out leads to relevant bias between the groups.
And we need to worry about that when we are looking at
differences in a sample at the end of
an experiment than at the beginning of the experiment.
So why do we have attrition?
Well, there's death.
Research that I do in aging,
if I'm comparing a group of 70-year-olds to 50-year-olds I get the difference,
is that because that a group of 50-year-olds are not representative of
the same population as 50-year-olds because people have died.
There're more people dead at age 70 who were originally
age 50 than people that are age 50, people are age 70.
So that can be attrition just to the death.
Like I said, people who study health,
people who study aging studies longitudinally, this is a real problem.
It could also be boredom or fatigue.
I simply don't like that experiment.
I don't want to participate any more, I'm dropping out.
And you drop out in the middle of an experiment,
then that means that the beginning and the end conditions are not the same.
This happens if the experimental group has
a more rigorous task than
people would drop out more in the experimental group than in their control group.
It's simply that because I'm getting fatigued.
I don't like doing this very rigorous task or it could be a very boring task, same thing.
Then more people would drop out of experimental groups,
so the group is not the same when the task is over.
There's also this problems with the experiment rationale.
For example, I don't want to stop smoking
and I'm getting out of this stupid experiment or
my memory is terrible and I'm tired of just
showing these people that my memory is terrible so I'm going to drop out.
So now, the group at the end of
the experiment is very different from the group at the beginning of the experiment.
So, here we have a population and we have a sample of that population.
We have a population and we take a sample of it.
And then in my experiment, I'm going to divide that sample up into old and young,
men and women and I do that.
But the problem is when I do that and I'm now dividing it by age and gender,
all of a sudden now my original sample population are very different from each other.
Here's what I assume I would get equal number of people in each one of the groups,
but in fact when I look at it,
I get very different because they're simply
fewer old males than
young males and there are fewer old females than young females because of death.
But in fact, females live longer than males significantly.
I'm involved in a public service nonprofit retirement facility.
We have 105 people living there.
We have six males and 99 females.
So these differences in attrition due to mortality, due to age.
So those are not quite the same comparisons because the conclusions about age that
I make and the conclusions about gender that I make are really
compromised in the experiment.
So, here's a possible experiment.
I'm going to look at whether exercise is an anti-aging variable.
And the answer is yes.
It's the only anti-aging intervention I need to
make when I'm measuring memory and cognition.
So we know that aerobic exercise improves memory and cognition in older adults.
And you can start at any time. You start at age 60 and prove it.
So it is truly an anti-aging intervention even though I
know no single drug that does that by the way,
that improves cognition in older adults.
So I'm going to have experiment that has no exercise.
Exercise young and old,
and I'm a measure memory and cognition.
If we did the experiment above,
which groups should show the greatest attrition?
Well, it's going to be both age and the intervention
itself because we have more attrition and people have to do
the exercise and have more attrition in the old age group.
So already both of my independent variables are
compromised due to differential attrition.
So what do you do about differential attrition?
Well, we can maintain a good tracking system for
all our participants and make sure we understand at least who the stayers are
and who the people that leave are and we can follow up with
telephone and postcard reminders to tell
people that their second task is coming up, please show up.
So we can actually remind people.
That will increase participation and reduce attrition.
We can look to see at the end of the experiment whether there're
differences between the people who stayed and the people that
dropped out and let's at least give us measurements
of what the attrition is done to our independent variable.
If attrition is not large,
we can simply drop score in one of the group that chose the less attrition to
match the score of a dropper in the group that has the less attrition.
So we can simply match the experiments based on who the droppers are,
but that's only can be done if the attrition is fairly small.
But the important thing is to look to see if
the droppers had extreme scores because if they had extreme scores,
then we would expect differences simply due to regression to the mean.
And that also means it's less likely,
if they had extreme scores,
that there's not going to be in effect of dropping out
because that extreme score is going to affect the mean and the variance.
So differential attrition, let's do an experiment.
Here we have scores on the first measurement 10, 8.5,
8.0, 7.0, 6.0, 5.8 and 4.0 and the mean is 7.04.
That's the mean of those scores.
If two people drop out and those people that drop out have extreme scores,
then the mean goes down to 6.16.
If these people drop out,
people that don't have extreme scores,
then the mean doesn't change.
So who drops out,
the nature of the people the dropout really can affect the distribution,
affect the variance and that means affect the level of
dependent variable that's going to look at whether
the independent variable is actually being changed or not.
So what can we do about differential attrition?
All these things are talked about but we could
also look at statistical techniques that actually account for missing data.
We have missing data that we statistically going to
adjust their scores based on those missing data.
It's debated because of the fact I just mentioned about
extreme scores dropping out can cause more of an effect than non-extreme scores.
But there're also statistical techniques for missing data that are available. Thank you.