Can I use SEM with small sample size and indirect measures?

10 October 2018 2 7K Report

To all

I am new to SEM and would welcome all answers, suggestions.

First, I created a simple questionnaire to survey famers' attitude toward climate change and their adoption of climate resilient coffee farming practices. The survey was built to capture more of a CONSUMAT model, ie certainty, satisfaction etc. The answers to the survey were supposed to be parameterized for an ABM.

However, during the field work, I realized the farmers consistently refuse to answer these questions, leaving an incomplete survey with only information on: impact of climate change on crop (ranking), observation of climate-related events (count), sources of climate-related techniques (count), level of climate information seeking (active or passive), observation of environmental impact of coffee (count) and different climate adaptation techniques applied (count), along with some data of the household and the farms.

Upon reading many other papers in the field, I created a SEM based on the Theory of Planned Behavior. Based on the salvaged survey, I only have indirect measures of Attitude (sources of climate-related techniques, impact of coffee), Norms (sources of climate-related techniques, level of climate information seeking), Perceived Behavior Control (number of people in the HH - proxy for human resources), Intention is actually measured by different climate adaptation techniques applied.

I ran a CFA on this model and the standardized result is attached.

My questions are whether SEM makes sense:

1 Can I do SEM on only 30 observations?

2 Can I run SEM based only on indirect measures of latent vars instead of direct measures of latent vars (due to the construction of my survey questions) ? Can I have only 1-2 items per latent var?

3 If SEM does not make sense, can you please suggest other methods and some readings?

Thanks a lot for your help

Alejandro Ros-Gálvez

My answer to questions 1 and 2: that's possible. However, the use of single items in latent variables has some has some drawbacks Read this:

Bergkvist, L., & Rossiter, J. R. (2007). The predictive validity of multiple-item versus single-item measures of the same constructs. Journal of marketing research, 44(2), 175-184.

Holger Steinmetz

Hi,

the conditions of you study are really not "ideal". Actually, I personally would start all over again. I know this is not really constructive, but please bear with me. I see the following issues.

1) The fact that many farmers did not response very likely imposes the problem of nonrandom missing if downstream TPB-variables (e.g., attitude) caused skipping the study. This is known was endogeneous selection (conditioning on a dependent variable). I recommend to find out why the reactions were so negative and avoid these conditions.

2) I actually do not really like the labels "indirect" vs. "direct" measures as the indirect variables are beliefs thought of causing the direct "versions". For instance, beliefs about consequences of a behavior are supposed to influence the attitude towards the behavior.

And to answer your question: Of course, you can estimate an effect of beliefs on intention without having the mediator (e.g., attitude) in the model. Every direct effect can finally be separated by any yet not modeled mediator, going down to an infinite specificiation of more tiny mediators (in the extrem quantum processes in the brain). This does not mean that you cannot estimate the overall effect.

3) I have some problems to recognize the TPB concepts in your descriptions (with regard to the different "souces"). Be sure to adhere to Ajzens proposal how to set up a TPB questionnaire:

Ajzen, I. (2001). Constructing a TPB questionnaire: Conceptual and methodological considerations.

This is a report to be gained from his website.

4) Unfortunately, N=30 is really low. You can get something in low-N situations with the swain-correction, but this is really extremely low. Beyond the low power to really test the model, the parameter estimates will be strongly biased.

I always can recommend to strongly pretest any questionnaire before going into the field, starting with cognitive probing about the validity of the measues up to a full-blown run using the complete questionnaire with a limited number of particpants from the target population.

I know repeating a full study sounds awkward (we had to do this several times) but at the end you will get all your problems solved.

HTH

Holger (@HolgerSteinmetz)

Could you recommend any books about Number Theory?

What's the best way to measure growth rates in House sparrow chicks from day 2 to day 10?

After a reference from Int J Body Compos Res, can any body help?

What techniques can I use to determine the miscibility of polymer blends?

Arena simulation - how do i model the following model in ARENA (see description) (new to ARENA)?

How to repeat an operation on an atom selection for multiple times in PyMol?

Is there a powerful system for the security of the systems distributed on IoT systems?

New Journal, "Integrative Psychological and Behavioral Science" -- do you not know, and have you not seen, this done before?

How Relativistic effects is one reason for lanthanide contraction?

Shall I skip the measurement model in SmartPLS?