Meta-analysis of 9 parallel-group and cross over studies

15 May 2018

Paper: Risk for nocturnal hypoglycemia: A meta-analysis
Program: PaulBrownPhD/metaanalysis


  • this work is from 2007
  • primary analysis: number of nocturnal hypos (NHE), Novomix 30 v BHI 30
  • 9 studies: parallel group (7) and cross-over designs (2)
  • heterogeneity: I2, radial plot
  • use inverse variance method
  • model:the number of NHEs will be analysed as using a negative binomial regression model. The model will include the number of episodes as the dependent variable, the log-transformed exposure time as an offset variable, baseline HbA1c as a continuous covariate, treatment and type of diabetes as fixed effects

GENMOD: parallel group

NLMIXED: crossover

Log window

NLMIXED alert: "proc nlmixed reports incorrect statistics for all values associated with the NEGBIN distribution. To circumvent this problem, program the NEGBIN distribution using the GENERAL likelihood specification." (

GENMOD with subjects as fixed (for initial estimates)

NLMIXED has no CLASS statement, create indicator variables:

Specify log-likelihood in NLMIXED to get around using NEGBIN:


  • analyse studies individually and combine in meta-analysis in a second step (inverse variance method)
  • PARMS: get initial estimates from GENMOD
  • BOUNDS: use to specify offset variables
  • no CLASS statement, create indicator variables
  • use GENERAL and specify LL, instead of NEGBIN
  • no LSMEANS statement, use ESTIMATE

Notable sources:

Creating a suite of macros for meta-analysis in SAS: A case study in collaboration