You need more than 1 in each level 2 unit as otherwise level 2 and level 1 variance will be confounded.

In then depends on the target of inference - if you want to say something about the level 2 variance you need lots of level 2 units. If you want to say something about a specific level 2 entity you need lots of level 1 units - eg more than 25 as a minimum.

For general orientation on this - see

http://www.bristol.ac.uk/cmm/software/mlwin/mlwin-resources.html#datadesign

You can use the following software to calculate the power

calculations

http://www.bristol.ac.uk/cmm/software/mlpowsim/

the manual is a very good primer 

and

Optimal Design for Longitudinal and Multilevel Research Documentation for the “Optimal Design” Software

http://people.cehd.tamu.edu/~okwok/epsy652R/OD/od-manual-20080312-v176.pdf

In dyadic research the sample size at level 1 is groups of 2. That is fine, but in this case I would not try to estimate slope variation, the level 1 sample size is way too low for that!

Joop Hox

Thanks alot Joop Hox for your answer.

Can you please give abit more elaboration or suggest some reading regarding this

avid Kenny and Deborah Kashi have published a lot on dyadic modeling and the restrictions that follow from this very small 'group size', look at Kenny's homepage at davidakenny.net

Hello,

I have a similar issue (in terms of sample size).

I have collected data from 250 students enrolled in 4 highschools and I thought I can run a Hierarchical Linear Multilevel analysis with student characteristics at Level 1 and School at level 2. Predictors are at Level 1.

Initially, I was planning to focus on between groups comparisons, but as far as I can tell, that is not possible with my data. More recently, I have read in several papers that for this type of analysis, it is recommended to have a large number of units at Level 2 (at least 50), which is not my case.

Given this, what type of statistical analysis is recommended?

Thank you 

Fixed effects single level model - dummies for each school as predictors

the downside is that you cannot include level 2 predictors

https://www.researchgate.net/publication/233756428_Explaining_Fixed_Effects_Random_Effects_Modeling_of_Time-Series_Cross-Sectional_and_Panel_Data

Article Explaining Fixed Effects: Random Effects Modeling of Time-Se...

Hello,

I have done a multilevel SEM, focusing only on the Level 1 mediation. I have dataset from name-generating questionnaire, so I have asked individuals to rate their network partners, which produced a clustered dataset, each individual could have rated up to 5 network partners. I have 163 individuals that together rated 569 members. In an attempt to keep as much data as possible, I have included in the analysis those that rated only 1 person. I estimate random intercept, but the slopes are fixed to be equal across groups, since the number of "groups" is substantial, and the slopes are calculated using information from 569 relationships, this should be OK sample. However the reviewer thought having up to 5 individuals within group is biasing my results. I would agree if I wanted to estimate random slopes. If you know some good articles that address this issue please let me know.

Thank you,

AM