Which analysis is appropriate for our study design with multiple DV's at 3 levels?

Kyana · Apr 24, 2014

Our bachelor research is about distance perception in the peripersonal space. The experiment was already complex to design with the oculus rift, but the statistical analysis appears impossible :shakehead.
We are trying to fit the data we gathered till now into the right analysis but we are having a hard time

.

The experiment set up:

35 participants; stimuli seen through Oculus Rift (3D): approaching cones that disappeared at different distances (conditions: near (20 cm), mid (60 cm) and far (120 cm). Participants estimated the distance where the cones disappeared in the virtual 3D environment by moving a square backwards. The perceived threat for the three conditions was measured on a 7point-Likert scale (but it appears to be a 8point scale, since some judged it to be 0)

. Moreover, participants received a motor vibration cue on the left or right cheek depending on the heading direction of the cone. The instruction was to press a button as soon as possible after a motor vibration cue was perceived. Cue onset timing was either at a random interval (random cue condition), or depended on the time it took the cone from the distance it dissapeared to blow in the participants face (fixed cue condition). We also have STAI questionnaire scores for Trait and State anxiety levels (max score of 80 each). All variables were measured multiple times for all individuals that participated in all conditions.

So we have 1 IV (stimulus distance) with three levels (near, mid, far) and 3 DV’s: Distance Estimation, Threat and Response time (with 2 additional levels: fixed/random) all measured multiple times for each individual in all three distance conditions near, mid far.

The dataset contains a score (average) for every individual and includes:
1. Estimation in cm (3 levels): Distance_Estimation_Near, Distance_Estimation_Mid, Distance_Estimation_Far.
2. Perceived threat 0-7 scale (3 levels): Threat_Near, Threat_Mid, Threat_Far
3. Response speed in ms (2x3 levels): RS_Near_Fixed, RS_Mid_Fixed, RS_Far_Fixed and also RS_Near_Random, RS_Mid_Random and RS_Far_Random.
+ State and Trait anxiety score.

Our research goal is to examine whether cones/threat in the peripersonal space (near) would be more threatening and more underestimated in distance than outside the space more distant from the observers. Also if there are more estimation errors with high threat in the conditions. We want to find out if underestimation might be adaptive for a fast reaction, so we want to find out if response speed is also faster/slower for the distance estimation errors at different distance conditions (near/mid/far). We also included state and trait anxiety to find out of this had any influence on the estimation, threat and response speed. High state anxiety was thought to increase the distance underestimation, threat and a faster response speed in the peripersonal space condition.

I ran individual repeated measures ANOVA’s on distance estimation(3), threat(3) and response speed(3x2) in SPSS. However, we also need to find a way to include interactions into an analysis (such as threat on distance estimation, and possible modulation of distance estimation on response speed via threat or state/trait anxiety levels) to test or hypotheses.

Been through all three statistic's study books, but the appropriate analysis for this sort of design seems no-where to be found.

What analysis is possible or the most appropriate to test the hypotheses for this dataset?

Any advice is appreciated!

noetsi · Apr 24, 2014

Re: Which analysis is appropriate for our study design with multiple DV's at 3 levels

It is not clear to me what you hypothesis is (or are) specifically. You have to be very specific in that for anyone to help, particularly since no one here likely knows your research area (I got quickly lost....).

I suspect you have a series of hypothesis not just one and that you might be able to answer many of these with a single DV if you break down your question in narrow hypothesis. If you really want to estimate multiple DV at the same time you might consider structural equation models (if so remember you have categorical variables and they have to be handled differently, for example using weighted least squares as part of the analysis).

Kyana · Apr 24, 2014

Re: Which analysis is appropriate for our study design with multiple DV's at 3 levels

The dataset contains distance estimation deviation from actual difference in centimeters for every condition, threat measured as a 7 point scale for every condition, response time in ms for every condition split for fixed and random cue timing, so i don't have categorical data? The hypotheses are: the closer (near) the more threat. The closer, the more underestimations in distance, the more threat the more underestimations for near compared to far, and faster response participants also tend to underestimate the distance from the actual distance, possibly mediated by threat. And: high state enxiety enhances the effects we expect. But we found that distance estimation for near is the opposite, they overestimated (all but one) the near cone condition with 38 cm and far - 9 cm. I don't know how to compare all these different variables because the data is available for near mid and far. I can split the persons in groups that are fast for cones that approach close and a group that is slow?

Kyana · Apr 24, 2014

Re: Which analysis is appropriate for our study design with multiple DV's at 3 levels

Several studies found that persons tend to underestimate the distance for threat that approaches in the peripersonal space (distance from person to 60 cm). The aim was to find out why people perceive threat closer than the actual distance. We thought it was adaptive because the body has more time to act, (fight or flight) because the threat is actually farther away. This has never been studied before. Maybe it helps if I show how the dataset looks like? It is really confusing, our bachelor research professor does not know what analyses to do, she said we only had to do correlations. That is only descriptive i thought, not actually testing if the different conditions signifficantly differ.

noetsi · Apr 24, 2014

Re: Which analysis is appropriate for our study design with multiple DV's at 3 levels

Actually I was saying you do have categorical data for your dv and SEM assumes, in some versions, that you have interval data.

Which of these individual hypothesis involves more than one of your DV. That is when you test your hypothesis which do you have to test more than one DV at a time for.

Kyana · Apr 24, 2014

Re: Which analysis is appropriate for our study design with multiple DV's at 3 levels

View attachment 4406

Added a picture of data for some participants, I don't know how to use all these variables at the 3 levels in an analysis

Kyana · Apr 24, 2014

Re: Which analysis is appropriate for our study design with multiple DV's at 3 levels

View attachment 4410

I would love to have categorical variables, so these can be compared as groups in ANOVA or someting, but somehow we have three levels for all DV's and these are continuous, or am I seeing this wrong?

This is the best I can do to make the question less complicated (sorry):

We somehow wanted to find a relationship like this: cones that headed towards the person and come close would be found more threatening, and if they are more threatening the body would prepare for action to defend itself against harm. To actually find out if the person is more readily to act when -threat- comes at a close distance, we included response time. So threat > underestimation distance > more readily to act and thus faster response time to near distance approaching threat, which is what we found (even irrespective of whether the vibration cue corresponded to the time the cone would approach the person or not).

We have tested the main effects in repeated measures analysis.
Now we would like to somehow show that threat has an effect on distance estimation at these three different distance conditions: more threat, the stronger estimation errors at close distances. We also want to test the hypothesis that underestimation is higher in the near conditions when response time is lower (thus faster response), compared to high response time and compared to middle and far conditions. At last we want to relate these three variables in each condition to each other.

1. More proximal distance => more threat (supported by data)
2. More proximal distance => more underestimation (rejected -> it is the opposite direction!)
3. More proximal distance => faster response time (supported, but interaction for timingcue*distance)
4. More threat => more underestimations, especially at close distance (don’t know how to test this in these conditions)
5. More underestimations => faster response for close but not for far (maybe split in groups?)
6. Threat*response_time*distance => distance estimation at three distance levels
7. More proximal => more threat => more underestimation of distance => faster response.

How can we test this with a statistical analysis?

We can run an ANOVA (repeated measures) for main analysis but not to test for the interactions with other DV's.. Because we now have one IV (distance of cone at three levels) on 3 DV measurements for all three levels, and we also want to test the DV's with each other at all the levels. So then it is IV on DV on DV or something at 3 levels. Now I'm way to confused

I will try to figure this out again tomorrow, but we have been cracking our brains on this for over a month. We would be very happy (really) if someone can help us out.:yup:

Karabiner · Apr 25, 2014