MDE and cluster specification for power calculations using -sampsi- and -sampclus- in Stata 12

[DB Sobol]

Hi Statalist,

I am running a power calculation for a two-sample comparison of means in a multilevel (5 cluster) data set. I have two questions about this:

1. Is it possible to specify the minimum detectible effect (MDE) desired, using sampsi and sampclus? If not, is there a default MDE that the commands assume? For example, I know that the clustersampsi command allows for the specification of detectable difference, but as my data is quasi-experimental, this command seems inappropriate.

2. Per literature in the field of this study (Community Driven Development), I am specifying the Intraclass Correlation Coefficient (ICC) at the level where treatment is assigned. The data is organized in 5 nested clusters, from highest to lowest. For simplicity I will refer to them as: L5 L4 L3 L2 L1. Treatment and control assignment was made at the L4 level. I ran a mixed effects regression using xtmixed followed by estat icc to identify the ICC for the L4 level. Then, following the sampsi command, I run sampclus and specify the number of clusters and the icc as follows:

sampclus, numclus(number of L4 clusters) rho(ICC of L4)

This may be a silly question, but must the numclus() specify the number of clusters at the level which the ICC is measured (L4)? If I want to identify the number of lower-level clusters needed to reach different power thresholds, should I be using the ICC calculated for that lower level cluster?

Thank you in advance!
Danielle

[DB Sobol]

Hi Statalist,

Re-posting this in hopes that someone will be willing to help me out. Thanks!!

I am running a power calculation for a two-sample comparison of means in a multilevel (5 cluster) data set. I have two questions about this:

1. Is it possible to specify the minimum detectible effect (MDE) desired, using sampsi and sampclus? If not, is there a default MDE that the commands assume? For example, I know that the clustersampsi command allows for the specification of detectable difference, but as my data is quasi-experimental, this command seems inappropriate.

2. Per literature in the field of this study (Community Driven Development), I am specifying the Intraclass Correlation Coefficient (ICC) at the level where treatment is assigned. The data is organized in 5 nested clusters, from highest to lowest. For simplicity I will refer to them as: L5 L4 L3 L2 L1. Treatment and control assignment was made at the L4 level. I ran a mixed effects regression using xtmixed followed by estat icc to identify the ICC for the L4 level. Then, following the sampsi command, I run sampclus and specify the number of clusters and the icc as follows:

sampclus, numclus(number of L4 clusters) rho(ICC of L4)

This may be a silly question, but must the numclus() specify the number of clusters at the level which the ICC is measured (L4)? If I want to identify the number of lower-level clusters needed to reach different power thresholds, should I be using the ICC calculated for that lower level cluster?

Thank you in advance!
Danielle