Dear Stata Community,
I am currently working on a multilevel structural equation model to test a mediation (see attached figure). The data consists of lower level business units (BU_I, n = 165) which are nested within higher level business units (BU_II, n = 54) of a large company. All three variables (slack3, sotif, ebit) are measured at the (lower) BU_I level. In order to test for the indirect effect of slack3 on ebit via sotif, I would like to run a non-parametric pairs bootstrap. Van der Leeden, Meijer and Busing (2007) highlight the need to take into account the nested structure of data when bootstrapping multilevel models. With their advice in mind, I would like to run a bootstrap where in each simulation run:
1) At first higher level business units (BU_II) are resampled with replacement
2) lower level business units (BU_I) within these resampled higher level business units are then resampled
As far as I have understood the Stata tutorials on this issue, I can either resample higher level business units (BU_II) using the cluster() and idcluster() commands. I can also resample lower level business units (BU_I) using the strata() command. However, I cannot combine both commands.
Accordingly, I would like to put forwards the following two questions:
1) How can I implement a bootstrap procedure where data at both levels is resampled?
2) As I am not an expert on bootstrapping multilevel models and van der Leeden, Meijer and Busing (2007) are the only source that I have found on this topic, I am not sure whether it is actually necessary / desirable to resample data at both levels. What would be your suggestions and do you have source that I can cite to back this up?
Unfortunately, I cannot share a sample of my data, since it is confidential.
I greatly appreciate your help on this issue.
Best,
Christian
Source: Van der Leeden, Rien, Erik Meijer, and Frank MTA Busing. "Resampling multilevel models." Handbook of multilevel analysis. Springer New York, 2008. 401-433.
Code that I have used so far to resample at the BU_I (higher business unit level):
I am currently working on a multilevel structural equation model to test a mediation (see attached figure). The data consists of lower level business units (BU_I, n = 165) which are nested within higher level business units (BU_II, n = 54) of a large company. All three variables (slack3, sotif, ebit) are measured at the (lower) BU_I level. In order to test for the indirect effect of slack3 on ebit via sotif, I would like to run a non-parametric pairs bootstrap. Van der Leeden, Meijer and Busing (2007) highlight the need to take into account the nested structure of data when bootstrapping multilevel models. With their advice in mind, I would like to run a bootstrap where in each simulation run:
1) At first higher level business units (BU_II) are resampled with replacement
2) lower level business units (BU_I) within these resampled higher level business units are then resampled
As far as I have understood the Stata tutorials on this issue, I can either resample higher level business units (BU_II) using the cluster() and idcluster() commands. I can also resample lower level business units (BU_I) using the strata() command. However, I cannot combine both commands.
Accordingly, I would like to put forwards the following two questions:
1) How can I implement a bootstrap procedure where data at both levels is resampled?
2) As I am not an expert on bootstrapping multilevel models and van der Leeden, Meijer and Busing (2007) are the only source that I have found on this topic, I am not sure whether it is actually necessary / desirable to resample data at both levels. What would be your suggestions and do you have source that I can cite to back this up?
Unfortunately, I cannot share a sample of my data, since it is confidential.
I greatly appreciate your help on this issue.
Best,
Christian
Source: Van der Leeden, Rien, Erik Meijer, and Frank MTA Busing. "Resampling multilevel models." Handbook of multilevel analysis. Springer New York, 2008. 401-433.
Code that I have used so far to resample at the BU_I (higher business unit level):
Code:
bootstrap _b[sotif:slack3] _b[ebit:slack3] _b[ebit:sotif] (_b[ebit:sotif]*_b[sotif:slack3]) , cluster(bu) idcluster(nbu) seed(1234) reps(20): gsem (slack3 -> sotif, ) ( slack3 sotif -> ebit, )(M1[nbu] -> sotif, ) (M3[nbu] -> ebit, ), covstruct(_lexogenous, diagonal) latent(M1 M3 ) nocapslatent vce(robust)
estat bootstrap, percentile
Comment