Dear RG Members, I am conducting a research project that aims to assess the determinants of radicalization/extremism. The dependent variable is radicalization which is measured through 15 survey questions/items (5-point Likert scale). The survey has 24 explanatory variables each of which is measured through several questions/items on 5-point Likert scales. The respondents of the survey belong to three different regions of the province (North, Center, and South). The survey questions measure individuals' preception and I, therefore, term these as individual-level factors. However, radicalization usually takes place in specific settings and therefore contextual factors, such as regional differences in economic development, regional gross domestic products, health facilities, government capacity, etc., are also of key importance. Secondary data on these variables for just one point of time is available from the national statistical office, i.e. for the year 2021. I want to simultaneously analyze the individual and contextual factors to identify the important determinants of radicalization. After reading the relevant literature, I believe that my survey respondents are nested within regions which points to the hierarchal data structure. This took me to multilevel modeling. In this connection, I have the following questions.
- Choice of the Model: Which type/category of the multilevel model suits surveys data that is collected on a Likert scale? After reading a lot of literature for weeks, this confusion still remains.
- Dependent variable: As mentioned above, the dependent variable is measured using 15 questions. Is there any way in multilevel modeling that allows using all these 15 items as dependent variables without the need to combine them into a composite index? Ideally, I would like to obtain item-specific estimates as well as a summary measure of the overall impact of all items.
- Composite Index: In case it is mandatory to combine the items of the dependent variable into a composite index, which type of multilevel model would suit it?
I kindly request the forum members to please share their insights. I would be truly grateful.
If I understand correctly, you're looking at just 3 different regions, correct? With just 3 Level-2 units, it would make more sense for you to use conventional multiple regression with dummy (or other) coding of the regions rather than multilevel modeling. This is discussed in Cohen, Cohen, West, and Aiken's (2003) regression text (see their chapter on multilevel analysis).
If you wanted to examine multiple DVs, you could use path analysis. However, I wonder whether it is really necessary to consider each item separately as this would lead to a very large model and the need to use regression/path analytic techniques for ordinal outcomes.
I think this would be easier to address as a multilevel ordinal logistic regression than as a SEM model (as multilevel SEM is a pain and not necessarily required here as I'm not sure you have latent variables). However, these models can be difficult to fit. For best results I'd consider something like bmrs in R. This is a Bayesian approach and therefore very flexible, but its main selling point here would be getting convergence.
See this paper for an intro:
Article Ordinal Regression Models in Psychology: A Tutorial
You can run a 3 level MLM equation. Level 3 should be the region, while Level 2 should be the states. Remember, that before you begin you need to carryout normality tests, then conduct a PLUM regression. During the testing of the MLM equation you will check for evidence of clustering in the datasets, and also ensure the robustness of the selected model. When testing your data, kindly note that a maximum of 10 variables can be analyzed at a time, so that you won't encounter redundancy issues. Also, your dependent variable should be a composite of the indicators that you want to test. Lastly, kindly, ensure that you conduct validity and reliability testing before you run the MLM regression analysis, so that you don't go back and forth.
If you are really interested in contextual factors you have used an inappropriate design as you only have a sample of three - regions that is. See Analysing and Designing Extensive Studies in Which Place Matter
https://www.jstor.org/stable/20003374
and
Article Multilevel approaches to modelling contexuality: from nuisan...
If you put dummies in the model to represent regions you will not be able to include measured regional variables *GDP etc) as they will be exactly collinear. If you leave the dummies out, the model will treat the regional variables as if they were the much more numerous individual level variables giving far too narrow standard errors for the estimates.
Hello, I had a similar situation in my partly survey-based research. Thus, some thoughts on this topic:
1.If you have an ordinal dependent variable, it could be the best to estimate the model as an ordinal multilevel random intercept regression model (with a random effect for each level where unobserved heterogeneity is expected). In your case, individual level would be the first level and region level the second level. Alternatively, you could use a conditional or fixed effects ordinal model - in which the multilevel structure is represented by using cluster robust standard errors and/or incorporating the fixed effect at the region level (e.g. as dummies). You might want to think further about which of the three models fits best for your purposes. I usually estimate several models.
2. In principle, it is no problem to include ordinal variables in the regression model. The prerequisite would be that your sample is sufficiently large and has correspondingly acceptable degrees of freedom. However, 15 dependent variables already seem to me to be quite a lot. Perhaps you could think about whether it is really necessary to include all 15 - many variables are certainly correlated and simultaneous inclusion would decrease the estimate efficiency (multicolliniarity).
3. In principle it is possible. However, I cannot assess whether it is also theoretically sound.
Many greetings, Dmytro Katerusha
My guess is first define your dependent variable. Generate dependent variable using pca and sem. Once you do that you choose your regression model. By dividng the variable into two parts you can use logistics. If you divide it into more parts you can use ologit or multi logit. Once you fix your dependent variable then go for choosing independent variables .
Abouit levels you can use two-level ( multi level) modelling. Individuals are at first level and regions are at second level. See if it helps you.
- Badges
- Science topic
- Similar topics
- Counseling Psychology
- Confusion