Can causality be inferred in a study with an experience followed by two sets of measures

Question

I came across this study as part of a mock exam paper and was confused to say the least.

Context:

The study investigates cognitive and behavioural factor related to the experience of anxiety in MRI scanners.

Participants completed the following questionnaires 5 mins after the scan:

a measure of the frequency of anxiety experienced in the scanner
a measure of the frequency of physical symptoms of panic
a measure of the frequency of coping strategy employment
a measure of the frequency of claustrophobic-related thoughts

Participants were also asked to fill in and submit general measures at home after the scan:

level of claustrophobic fears regarding restriction and suffocation
depression
anxiety
health anxiety

The return rate was 75%, i.e. only 97 of the original 130 participants completed the second batch of questionnaires.

Overall, the level of anxiety in the scanner was RELATED TO:

level of claustrophobic fears regarding restriction and suffocation
frequency of coping strategy to manage scanner anxiety
no. symptoms of panic during the scan
no. claustrophobia-related thoughts about the scan.

No stats are given.

Questions:

What type of design is this?
Is it possible to understand the causes of anxiety experienced in MRI using this type of design?

Initial Thoughts:

In my mind, causality can not be inferred because there is no comparison with a control group, and correlation would probably have been used rather than AVOVA. Is this the answer to the questions above or have I missed the point?

score 1 · Answer 1 · answered Feb 12 '11 at 23:40

I think you're on the right track. Before drawing any conclusions about causality, I'd want to know...

...to what degree the 97 were representative of the 130, and the 130, to the population of interest.

...the magnitude of any relationships found as well as their statistical sig. (Whether correlation or anova was used is not, I think, directly relevant to your question.)

...the psychometric properties (validity and reliability indicators) of all measures used (perhaps one would want to adjust for reliability attenuation, for instance).

...what the subjects' characteristics were at baseline, i.e., before undergoing the scanner experience.

...as you've said, how a control group would compare.

The bottom line is, this is a correlational/observational study and, certainly as reported, says little about causal relationships.

Thanks, that's great. The paper also stated that the first set of post-scan questionnaires (5 mins after) for the 97 particpants who did return the second set of questionnaires were compared to those of the dropouts. No mention was made of any indication of representativeness. The paper then asks WHY they did this- I'm assuming that is to check that scores weren't significantly higher for the dropouts, as particularly high anxiety could be the reason for dropping out? Wow, talk about a trick question! — , Feb 13 '11 at 19:01

score 1 · Answer 2 · answered Feb 13 '11 at 05:34

Design:

The study is an observational design. Some may call it a correlational design, but I'm not a fan of such terminology because it can encourage the false assumption that the ability to draw a causal inference is related to the statistical test used.

Following up on this point, ANOVA does not permit causal inference any more than correlation. The issue is whether a variable was experimentally manipulated. It's just that most of the time variables manipulated are categorical and not numeric, so we end up using ANOVAs to analyse their effect.

As a general rule validating a causal proposition is easier where you experimentally manipulate the variable of interest. However, I would argue that correlations drawn from an observational design are still informative, particularly if they are combined with solid theoretical knowledge about the phenomena.

Causal Inference:

Many of the things related to anxiety experienced in the scanner are themselves related to the concept of anxiety. For example, panic, claustrophobia related thoughts and so on. Thus, showing that they are correlated, to my mind does not suggest causality. Rather, I would say that they are syndrome of co-occurring mental states that presumably have some common cause. I suppose this would be an example of me using prior knowledge of the phenomena and combining this with the correlational evidence.

The study is also limited to examining what causes individual differences in MRI anxiety. It does not examine how MRI implementation could be altered to make it more or less anxiety evoking.

However, I suppose you could use the reported levels of anxiety in the scanner as evidence that the scanner causes anxiety. Presumably we have an implicit control group in our minds that says most of the time people are not anxious about being in a scanner (particularly if they are not actually in a scanner).

I don't think the return rate is a major issue, although it might be worth considering whether this is biasing your results.

Some of the other variables are presumably getting at the question of whether anxiety experienced in a scanner is related (or even caused) by a general disposition of the person.

The fact that the general measures of anxiety, depression, and so on are measured after the MRI study is potentially a confound, although a researcher would probably argue that such measures are relatively stable over time.

score 1 · Answer 3 · edited Apr 13 '17 at 12:44

Here is another question from this websites that says among other things the following:

What can a statistical model say about causation?

This led to his motto:

NO CAUSATION WITHOUT MANIPULATION

which emphasized the importance of restrictions around experiments that consider causation. Andrew Gelman makes a similar point:

"To find out what happens when you change something, it is necessary to change it."...There are things you learn from perturbing a system that you'll never find out from any amount of passive observation.

check it out: Statistics and causal inference?