Using MLwiN's Command Interface
In class we have illustrated the use of MLwiN using the Graphical User Interface (GUI). For serious work you should (1) log the results, and (2) use the Command Interface so you can store all your commands on a file and ensure that you can reproduce your results at a later time. The online help has a useful section on Using the command interface. Here are some hints.
Log File
The command logo filename logs the results on a file.
You can turn this feature on using logo 1 and off using logo 0.
You must use the last command to ensure that the log file is closed properly.
To log commands as well as results turn echo on using echo 1.
Note that if you read data and echo is on the entire dataset will be listed.
For best results set echo 0 before reading data and echo 1
afterwards.
Reading Data
Here's how to read the data from Snijders and Bosker (1999),
henceforth SB, that we have been using in class.
Make sure you use the correct local or network filename.
I also use put to generate a vector of ones and
name to assign variable names.
It is usually a good idea to compute means or other summary statistics to ensure you read the data OK.
echo 0 dinput c1-c4 d:\pop510\snijders4.dat echo 1 put 2287 1 c5 name c1 "school" c2 "pupil" c3 "iqverb" c4 "langpost" c5 "one"
Variance Components
Here's how to define the level id's and the response variable,
and how to fit a simple variance components model with constant
variation at levels 1 and 2.
All explanatory variables must be declared using expl
and are automatically added to the fixed part.
The random structure is declared using setv.
I use settings to ensure all is well.
iden 1 "pupil" 2 "school" response "langpost" explanatory 1 "one" setv 1 "one" setv 2 "one" settings
The feedback of the last command is shown below:
settings
EXPLanatory variables in one
FPARameters one
RESPonse variable in langpost
FSDErrors : uncorrected RSDErrors : uncorrected
MAXIterations 20 TOLErance 2 METHod is IGLS BATCh is OFF
IDENtifying codes : 1-pupil, 2-school
LEVEL 2 RPM
one
one 1
LEVEL 1 RPM(RESETTING OFF)
one
one 1
To fit a model using a macro file you must set batch mode on
using bath 1, otherwise you get only one iteration.
To list results use fixed and random.
batch 1 start fixed random
The output is shown below, with commands in bold.
batch 1 Batch mode is ON start Convergence achieved fixed PARAMETER ESTIMATE S. ERROR(U) PREV. ESTIMATE one 40.36 0.4263 40.36 random LEV. PARAMETER (NCONV) ESTIMATE S. ERROR(U) PREV. ESTIM CORR. ------------------------------------------------------------------------------- 2 one /one ( 1) 19.42 2.921 19.34 1 ------------------------------------------------------------------------------- 1 one /one ( 2) 64.57 1.967 64.58
These are the results in SB-Table 4.1.
Random Intercepts
To fit a random-intercept model we use expl 1 to add
an explanatory variable:
expl 1 "iqverb" start fixed random
This is the output, again with commands in bold:
expl 1 "iqverb" start Convergence achieved fixed PARAMETER ESTIMATE S. ERROR(U) PREV. ESTIMATE one 11.17 0.8788 11.16 iqverb 2.488 0.07005 2.488 random LEV. PARAMETER (NCONV) ESTIMATE S. ERROR(U) PREV. ESTIM CORR. ------------------------------------------------------------------------------- 2 one /one ( 1) 9.496 1.515 9.486 1 ------------------------------------------------------------------------------- 1 one /one ( 2) 42.23 1.286 42.23
This is the parallel lines model in SB-Table 4.2.
Random Slopes
To make the slope of verbal IQ random at the school level we
use setv specifying the level and variable name.
I also use like to print the likelihood
setv 2 "iqverb" start fixed random
Here's the output
setv 2 "iqverb" start Convergence achieved fixed PARAMETER ESTIMATE S. ERROR(U) PREV. ESTIMATE one 10.81 1.111 10.81 iqverb 2.526 0.08145 2.526 random LEV. PARAMETER (NCONV) ESTIMATE S. ERROR(U) PREV. ESTIM CORR. ------------------------------------------------------------------------------- 2 one /one ( 2) 65.86 19 65.82 1 2 iqverb /one ( 1) -3.63 1.352 -3.628 -0.976 2 iqverb /iqverb ( 1) 0.21 0.09914 0.2099 1 ------------------------------------------------------------------------------- 1 one /one ( 3) 41.48 1.291 41.48 likelihood 979232 spaces left on worksheet -2*log(lh) is 15230.8
The same results in the equation window look like this:
