FAQs Research Topics

Below are some common queries that we have received on our Telegram Group (link: https://t.me/joinchat/yfYAwnrvpaliMDI1). We update this page regularly.

SEM and Psychometrics

Question: The negatively worded items were reverse scored, however all the reverse scored items loaded on to a different factor.

This is usually seen. You have two options: either work with two factors or drop the negatively worded items

Question: How should we calculate sample size? Is the thumb rule - 10 per item always valid?

See chapter 6 of DeVellis book. You will also get references from there.

The ratio of 1:10 is advisable but not it does not apply in a linear manner. A sample size of 250-300 should be good even if you have more items in your scale.

A sample size of 400 or above is good even if you have large item sets and the formula of 1:5 or 1:10 is violated.

Reference: DeVellis, R. F. (2016). Scale Development: Theory and Application. Sage.

What are various scale options for Likert scale?

Refer to pages 139-146 of DeVillis book. It explains various scale options for Likert scale.

Reference: DeVellis, R. F. (2016). Scale Development: Theory and Application. Sage.

Question: How can we statistically decide between two variables that which one should be dependent and which one should be independent?

This is not possible by looking at statistics. It has to be determined by theory.

Question: When can we call a construct as perceived? For instance some papers use the construct Enjoyment but others regard it as Perceived Enjoyment! Does it depend on how we are manipulating the variable?

Usually when you are doing an individual level analysis, you call it as perceived enjoyment. This is perception and differs from person to person. If you do group level or organisational level analysis (when you aggregate individual responses to get to a group or organisational score), you call it as enjoyment. Because now this is the aggregated score of all members. You will need to do multilevel analysis here.

Question: Difference between SEM and Regression Testing

SEM is testing relationships between latent constructs. This is done in a SEM software (e.g., AMOS).

Regression testing is done between observed, imputed or averaged constructs and a regression software (e.g., SPSS) is used.

SEM = CFA + Regression

Question: In case where any item is showing somewhat poor loading lets say between .45 to .6 and we want to keep that item…. what should be reported in such cases?

Do nothing, just let it remain. It is not too bad

Question: Should I use EFA or CFA for dropping an item?

You should not reduce items on your own. EFA is not needed unless you are developing your own scale or have some issues with the existing measure.

Best way is to do CFA (3 or 7 or 15 items) and then prune items that have poor loadings. Ideally, you must retain as many items as possible. Do not drop items of existing measures without any reason.

Question: What is the difference between AMOS and PLS-SEM?

Amos uses covariance-based SEM (CB-SEM) whereas Partial Least Squares (PLS) method for performing SEM.

If multivariate normality assumption is violated or you have small sample size or the purpose of your study is to predict than confirmation of theory, then you may go for PLS-SEM.

Amos uses the covariance-based method whereas Smart PLS uses the partial least squares approach. The algorithms used in both the software are different. However, some researchers see both these methods of model testing as complementary to each other.

Both can be used for testing parallel mediation. PLS SEM may not give you model fit indices except for R-square.

Here’s a playlist on Partial Least Squares: Structural Equation Modelling by Prof. Arun, a faculty at S K Somaiya College, Mumbai and an SkillsEdge community member. Hope you may find it to be useful. These sets of videos will help researchers understand the nuances of Structural Equation Modelling using a partial least squares approach.

https://youtube.com/playlist?list=PLGBKkGD8cqiKF27-_UnVbYfPzOYs10zQQ

Question: How to calculate MSV (Maximum shared variance) and ASV (Average shared variance)?

Please refer https://www.analysisinn.com/post/how-to-calculate-average-variance-extracted-and-composite-reliability/

Question: I have one latent variable which has 3 sub-dimensions. Each sub-dimension contains 5 items. So, overall there are 15 items. Second latent construct contains 5 sub-dimensions, each sub-dimension containing 5 items, i.e., overall 5×5=25 items. My 3rd latent construct doesn’t have sub-dimensions, but it has 12 items. In this case how to perform CFA?

You can do hierarchical CFA (second-order CFA) for constructs 1 and 2. Do first -order CFA for construct 3.

Another way is to do parceling of items for dimensions of constructs 1 and 2. Then do first order CFA using parcels for constructs 1 and 2, and normal first-order CFA for construct 3.

Question: What are control variables? Don’t we use covariate to include controlled variables in our equations?

Control variables are nothing but independent variables that also impact your DVs. So, it is fine to have the other IV as a covariate. The only limitation is that PROCESS will give indirect effect and other outputs only for the variable that you declare as IV, and not for the covariate. To compute for the covariate, you should declare it as IV and the IV as a covariate in the next run.

Question: Suppose the strength of correlation shows very low in the final data received but literature shows the higher strength of the variables…Which test to perform in this case?

Suggestions:

  1. Report the correlations as it is. It is not necessary that the findings should be in line with the literature. However, this low correlation could be because of measurement error ( Random or Systematic error).
  2. Most likely the respondents have messed up. If possible take a fresh sample.
  3. Also check the control variables and boundary condition taken by the literature and compare it with your study.

Mediation and Moderation

Question: Is it possible to test a continuous moderator with a dummy independent variable and a continuous dependent.

The first example discussed in Hayes (2013) textbook, chapter 7 is of this kind.

Reference:

Hayes, A. F. (2013). Introduction to Mediation, Moderation, and Conditional Process Analysis: A Regression-Based Approach. Guilford Press, New York.

Question: Suppose the direct relationship between IV and DV is insignificant. It is significant through mediation. In this case, can we report that the mediation is present? Or is this called mechanism?

Yes, you can use the term and can report the mediation. See the section 6.1 of Hayes (2013). Presence of the relationship between IV and DV is not necessary for testing of mediation.

Reference:

Hayes, A. F. (2013). Introduction to Mediation, Moderation, and Conditional Process Analysis: A Regression-Based Approach. Guilford Press, New York.

Model Building

Question: How advisable is it to fit a model according to data? Which is better - to draw inferences from data science or should we always follow the literature?

Whatever results one presents should have a strong theoretical backing and should be generalisable. If one just goes by what the data is telling, then it may not be very interesting.

Question: In addition there is a problem of harking (Hypothesising After the Results Are Known).

Refer Kerr, N. L. (1998). HARKing: Hypothesizing after the results are known. Personality and social psychology review, 2(3), 196-217.

While developing our research models, talking to practitioners is very important. This is what makes our research models relevant and practical.

Question: What is Moderated-Mediation?

Moderated-mediation happens only when any (one or more than one) of the paths to or from the mediator is moderated. (If you are talking about moderation of the mediation, then it will be conditional process analysis or moderated-mediation)

Question: If I am using a model like Hayes Model Number 5 where there is a moderator included in the model.. what should be the procedure? Should i just test Common Method Bias (CMB) on mediation sub-model or should I include moderator also?

If you are using a moderator too, I suggest that you create a measurement model with all constructs (including the moderator) along with the CMB factor. Then, impute the latent factor scores from this measurement model and try running the model using PROCESS.

Question: Can we use Hayes process Macro for two independent (metric) variables with 2 mediating and 1 dependent variable?

Yes, it should be possible. Declare one as IV and the other variable as a covariate in PROCESS.

Normality Checks

Question: If kurtosis is more than 2 for few items what should be the course of action for those items?

It is fine. You don’t need to drop the items. But, you should definitely report bootstrapping results to show you have taken care of non-normality in the data

don’t we use covariate to include controlled variables in our equations?

Control variables are nothing but independent variables that also impact your DVs. So, it is fine to have the other IV as a covariate. The only limitation is that PROCESS will give indirect effect and other outputs only for the variable that you declare as IV, and not for the covariate. To compute for the covariate, you should declare it as IV and the IV as a covariate in the next run.

Research Gap

Question: Suppose I find a gap that I believe could be worked upon but i could not find any research evidence that backs the existence of such a gap. In that case, Will that gap be considered valid or not?

You must situate the gap using literature. If you don’t cite literature, it will not be considered valid and can be questioned

Scale For Innovative Behavior

Refer: https://www.tandfonline.com/doi/pdf/10.1080/09585192.2013.870311

Question: Should I used a “Not Applicable Option” in my survey?

Does not apply option should usually be used during pilot testing. This will help you understand whether the items in your scale are making sense to the respondents. In actual data collection, it is recommended not to use the does not apply option

Theory Building

Question: How many theories should be applied to explain our research model? Is one sufficient?

You should try to limit yourself to 1-2 per paper. Managing multiple theories may become difficult and make the writing less coherent and fragmented.

Using Referencing Software

Question: In Mendeley, is it necessary to save the research paper as per the format, only then will it reflect correctly?

Yes. You have to enter all details about the paper (correct format of author names, title, journal, volume, issue and page nod.) correctly in Mendeley. This is the first step. Only then will the citation be proper.

One suggestion for quick and efficient entry in Mendeley. In case DOI is available, go to “Add Entry Manually “ (Just below Files, there is a dropdown option). A new window opens. Go at the bottom, and enter DOI details. Save. All fields will get filled automatically and accurately

Qualitative Research

Question: What is the right number of citations in qualitative research?

Will depend on the kind of paper. If you are doing empirical paper, then anywhere between 20-40 should be good. If it is a review paper, then you will have to show that you have read almost everything that is relevant. Number of papers will be high for a reviewer paper. Usually, there is not upper limit. Exercise your judgment.

2nd suggestion: Check papers published in the journal you are targeting, to get a rough range for the references and citations in your paper.

Question: Does an interview published in a newspaper qualify for qualitative study?

We can do sentiment analysis of text. The method used for most of the research papers with studies of printed text or blogs.

Question: What comes first? Research Objectives or Research Question?

Objectives come before question. Objectives would define what you want to do and is broad/abstract. Research questions are more directed and specific.

Question: Please suggest some good strategies for writing papers.

Take reviewer comments very seriously.

  1. The Article-Journal fit is critical. It is important to understand the focus of a journal. Many papers are desk rejected because they are sent to “wrong” journals. For example, AMJ prefers empirical papers, but AMR prefers theoretical papers, MIT Sloan Mgmt Review is more practitioner-oriented, and so on. So, it is essential to understand what the editors are looking forward to.
  2. The literature gap is over-emphasized. Convince the editors (of a good journal) why your idea is worth it. What is the compelling reason to research a specific topic?
  3. Important – avoid over-citation. If 10 out of 30 pages of a paper are of citations, then what is your idea? Cite sufficiently but not overtly.
  4. Don’t get disheartened by terse feedback. It is a reality check. Better than getting a desk-reject after a wait of 30-40 days.
  5. Avoid getting emotionally attached to bad ideas. (Easier said than done, but be prepared to move on)