Hello Emad,

don't you have a supervisor? Because at the end, s/he (and the comittee) will decide whether this is ok. Overall, any non-significant effect can be a contribution if the hypothesis' underlying rationale was solid or you tested conventional wisdom (and can no inform readers that these beliefs were wrong) and if the study was solid.

In contrast, a non-significant effect is unfortunate when

1) the hypothesis was not plausible in the first way (that is, you have to build up arguments for a implausable hypothesis and then you show that it is not true)

2) there are problems with the study (low N/power, bad measures, or other failures).

Best,

Holger

Dear Emad,

If you get unexpected results, it is important to analyze the reasons behind them and justify your findings. If possible, collect more data or refine your hypotheses. It is normal to iterate between the data and hypotheses during this process. Often, the reasons behind unexpected results are hidden, but many are likely due to poor sampling or a small sample size. If you have multiple moderators, it is advisable to test them individually. Additionally, you may consider using the PLScons algorithm to compare differences in R-squared values.

Thanks Holger Steinmetz and Salah Ahmed for your valuable comments and input.

My sample size was 155 as per the inverse square root method proposed (Kock & Hadaya, 2018). I've collected 201 usable responses. yet it seems not enough to detect the moderation effect.

A sample size of N=155 is considered small and may not be sufficient to rigorously test the moderation effects. Increasing the sample size would likely lead to more reliable results.

Is there any way to collect a larger sample using a better sampling technique?

Hi Salah,

I would not see it so pessimistic. The power of detecting a moderator effect depends on

1) The strength of the effect

2) The amount of measurment error (reliability and validity) of the predictor and the moderator

3) The N

That is, N = 155 may be totally fine when there are clean measures and a substantial effect. Or turned around: When we can assume clean measures (which can be supported by diverse analyses) then with the given sample size you would detect a substantial effect. Otherwise it is negligible. This could form the basis for a discussion.

Besides that: Emad, you throw away much data. I would suggest to apply some missing data techniques (multiple imputation for FIML).

HTH

Holger

Thanks again Holger Steinmetz and Salah Ahmed for your comments.

In fact, I had collected a total of 276 responses. However, after data cleaning procedure (treating missing data, assessment of outliers, data normality, non-response bias) I ended up with 201 valid responses.

emad

Hi Emad,

as I said, if "treating missing data" means elimination of cases, you loose much statistical power and perhaps worse, bias your results.

Multiple imutation and/or FIML (e.g., implemented in teh R-lavaan package) is much better and keeps the N up (lavaan.org)

The same goes for outliers. Don't remove them. Either use robust estimators or apply winsorizing (changing the outlying value to the 95% percentile of the sample).

And what do you men with "normality". If you use OLS, the only real issue is homoscedasticity (conditional variances of the residuals must be equal). And if not, again use robust SE.

Best,

Holger