[讨论]Modeling count outcomes and robust standard errors interpretation

1311

收藏 2014-01-21

I’m running a model where 14-day diaries (L1) are nested within 72 subjects (L2). I’m using HLM 6.0 and the outcome is a count variable (number of drinks) and at Level 1 I have weekend (dichotomous), positive affect (continuous), and negative affect (continuous). All level-1 variables are measured at the daily level. At level 2, the intercept contains parent education (less than some college vs. some college or more), race (white vs. non-white), gender, full-time college (yes/no), and the average positive and negative affect over the 14-days. Interactions between level-1 and level-2 variables include weekend(L1)*full-time college(L2), positive affect(L1)*full-time college(L2), and negative affect(L1)*full-time college(L2).

The method of estimation reported in HLM output is restricted PQL. I specified the distribution of the outcome variable as Poisson (constant exposure).

HLM output gives me two types of results: unit-specific model and population-specific model. Then for each, I also have results for robust standard errors. I’m interpreting the population-specific model. When looking at the difference in standard errors between the robust and non-robust standard errors, they are different (in most cases going from significant to non-significant results).

Looking thru previous threads, someone suggested Mass and Hox’s 2004 paper on robustness issues in multilevel analysis. Do the suggestions from Mass and Hox also apply for models with count and dichotomous type of outcomes? Or is that unrelated to the use of robust standard errors? I’m also getting a little confused because my method of estimation is not ML but restricted PQL. Does that play any role in which standard errors I need to report (robust vs. non-robust)?

扫码加我拉你入群

请注明：姓名-公司-职位

以便审核进群资格，未注明则拒绝

全部回复

Lisrelchen

2014-1-21 10:46:50

The Impact of Small Cluster Size on Multilevel Models: A Monte Carlo Examination of Two-Level Models with Binary and Continuous Predictors

Bethany A. Bell; Grant B. Morgan; Jeffrey D. Kromrey; John M. Ferron

Abstract
Recent methodological research has addressed the important issue of sample size at each level when estimating multilevel models. Although several design factors have been investigated in these studies, differences between continuous and binary predictor
variables have not been scrutinized (previous findings are based on models with continuous predictor variables). To help address this gap in the literature, this Monte Carlo study focused on the consequences of level-2 sparseness on the estimation of fixed
and random effects coefficients in terms of model convergence and both point and interval parameter estimates. The 5,760 conditions simulated in the Monte Carlo study varied in terms of level-1 sample size, number of level-2 units, proportion of singletons (level-2 units with one observation), type of predictor, collinearity, intraclass correlation, and model complexity.

http://www.amstat.org/sections/srms/proceedings/y2010/Files/308112_60089.pdf

扫码加我拉你入群

请注明：姓名-公司-职位

以便审核进群资格，未注明则拒绝

Lisrelchen

2014-1-21 10:51:41

Sufficient Sample Sizes for Multilevel Modeling

Cora J. M. Maas and Joop J. Hox

Sufficient Sample Sizes Joop Hox 2005.pdf
大小:(66.5 KB)
马上下载

扫码加我拉你入群

请注明：姓名-公司-职位

以便审核进群资格，未注明则拒绝

扫码加我 拉你入群

扫码加我 拉你入群

扫码加我 拉你入群

分享

扫码加好友，拉您进群

扫码加我拉你入群

扫码加我拉你入群

扫码加我拉你入群