My question is looking at the influence of simulation on student attitudes. My professor would like me to do regression analysis, but he says to do two regressions. I have my pre-test data and post-test data the only other information I have is student college. What I found in my class materials seems to indicate that I can complete a regression using the post-test as my dependent variable and the pre-test as my independent variable in SPSS. How would I do another regression? Should I work in the colleges as another dependent variable and if so, do I do them as a group or do I need to create a variable for each college?
Hello,
You can perform a second regression by using the post-test scores as your dependent variable and both the pre-test scores and college as independent variables. In this case, you would treat colleges as categorical variables, which means creating dummy variables for each college. This way, you can assess the influence of both the pre-test scores and college on the post-test scores.
Hope this helps.
I have some questions.
1) Was there some treatment (or intervention) between the baseline and followup scores? If so, did all subjects receive it, or only some of them? And if so to that, how were they allocated to intervention vs control?
2) How many colleges are there? If the number is fairly large, it may be preferable to estimate a multilevel model with subjects at level 1 clustered within colleges at level 2.
See the literature on Lord's paradox to help you to decide if ANCOVA is the appropriate approach. I assume that it would be incorrect to treat students as randomly allocated to colleges. This means that pre scores may be a collider. We actually had this situation in an example in my class last week. The initial problem is framed as a more complex procedure, but then it is shown that the problem would exist just with a normal regression. I'll attach my slide. In your example TeachV is your variable (whether treated as fixed or random as Bruce Weaver suggests may be appropriate) for college.
Bruce Weaver 's caution about the clustering of students within colleges is important.
The question is whether college is a nuisance factor or a variable of interest.
If variation between colleges is of no intrinsic interest and – importantly – it is unlikely to affect the effect of the intervention over and above the pretest scores of the students, then you could opt for regression in which you calculated robust variance estimates based on clustering within college.
You may, however, have grounds for testing the hypothesis that the relationship between the intervention and the outcome depends on the college: the intervention is more successful in some colleges than in others. For example, colleges with wealthy catchment areas might respond differently to colleges in more deprived areas. A healthy eating programme rolled out in the US, for example, would run into the problem that the areas in which poor people live are "food deserts" where only processed industrial foodstuffs are available in shops, while in rich areas you can buy actual foods.
If you have this kind of a hypothesis in which the intervention's effect depends upon context, and you have a reasonable number of colleges, then, as Bruce observed, you need a multilevel model.
A small caveat to Ronán Michael Conroy 's point. It is not just the colleges influencing the post scores, it can be some unmeasured variables influencing which college someone is at and the post scores. More information is necessary to advise, for example, if the ANCOVA (probably with college as a random variable) is appropriate.
You can do two regression if you have two groups (a group exposed to the intervention while the other will be a control group). However, regression is not the appropriate statistics to engage for an experimental study
Enoch Tomiwo Oladunmoye , I think I am misunderstanding the second part of your answer. Say I have an experiment with a control/placebo and an experimental group, and a single outcome. Assuming the assumptions are meant, wouldn't an OLS regression of the form:
y_i = B0 + B1 x_i + e_i
be appropriate, where B1 is the parameter of interest (i.e., a t-test)?
Daniel Wright, based on your explanation, the appropriate statistics is a one-way ANCOVA. It will take care of your pretest score as covariate (a continuous variable), your post test as dependent variable (a continuous variable), the treatment group (1 coded as experimental group, 2 coded as control group). This will produce what you need. If you are not satisfied and you want a statistics that can predict the DV. I will suggest you use a more advance statistics as such Suborn path analytical approach for experimental study using AMOS.
Enoch Tomiwo Oladunmoye , you said "regression is not the appropriate statistics to engage for an experimental study," and have not addressed my question where I asked if you thought a t-test is okay for some experiments. As far as what I said was appropriate, clearly I say in previous messages that not enough information is provided and this is an example where considering Lord's paradox is appropriate. Which of two common approaches described in Lord's paradox do you think is appropriate here?
- Badges
- Science topic
- Similar topics
- Geoscience
- Atmospheric Sciences
- Atmosphere