External Globals:

The global variables are called ``options'' in Amelia for Windows and appear on the options screen (without the leading underscore; i.e., _AMempri in Amelia for Gauss is AMempri in Amelia for Windows). In Amelia for Windows, a few of the less useful options on this list are unavailable.

_AMburn

number of iterations to burn in IP, SIP, or VIP (default=1000)

_AMchk

0 not to check inputs; 1 (default) to check inputs

_AMdrTol

0.00001 (default) tolerance for EM convergence

_AMembrn

0 (default) to use convergence criteria for EM; 1 to do _AMburn number of iterations of EM

_AMempri

$-2$ (default) not to use prior; if equal to or greater than zero, implement ridge prior where the value of _AMempri determines the degrees of freedom on which the ridge prior is based (the higher the number, the more heavily weighted is the prior; use this with data with few observations relative to variables, lots of missingness, or high correlations); if equal to $-3$, the user sets all hyperparameters values (_AMtaupr, _AMmupr, _AMsigpr, and _AMmpr) to define various types of informative or diffuse priors. Prior can be used for any of the algorithms. See Section 7.1.

_AMemt

1 implement $t$-distributed ECME algorithm, 0 (default) implement multivariate normal EM algorithm.

_AMfully

vector indicating column number of variables included in imputation dataset that are fully observed. If specified, conditional algorithm is implemented. Default is "." indicating no variables selected.

_AMidvar

indicates the position of identification variables (country names, identification numbers) which will not be used in the imputation model but will remain in the imputed datasets.

_AMkknp

indicates number of parties in imputation model of effective vote for multiparty vote share data.

_AMkkpfo

indicates the position of the vote share data for the party that is fully observed (contesting all districts).

_AMmthd

0 (default) to implement EMis; 1 to implement EM; 2 to implement EMs; 3 to implement IP; 4 to implement SIP (a vectorized version of IP); 5 to implement VIP.

_AMnoms

column vector with column numbers of each nominal variable in the dataset. For each variable, ${\mathfrak{A}melia}$ will create $k-1$ dichotomous variables to represent the $k$ unique values of this nominal variable and then reconstruct the nominal variable after the procedure. Do not use this for dichotomous variables.

_AMords

column vector with column numbers of each ordinal variable in the dataset. Use this only if you need imputations that correspond to the same scale values (so a missing value in an input variable taking on values 1, 2, and 3 will not get imputed as 2.43 for example). Leaving the exact imputation is better if it won't disturb your analysis model.

_AMordmd

method to use to ordinalize desired variables. 1 (default) will scale the continuous imputation to a probability and use this probability in a binomial distribution with N equal to the range between the upper and lower observed ordinal values. 0 will simply round the continuous imputation and bound it by the upper and lower observed ordinal values.

_AMprt

1 (default) to print output and intermediate results to screen; 0 for no printing

_AMpatch

Set to 1 to use the dynamic library (which must be installed for Gauss version, see the README file; this option cannot be used with the conditional model in the current version) or 0 (default) to use internal default procedures. Setting this to 1 can increase the speed of the program by a factor of ten or more; see section 9.

_AMsave

0 (default) not to save theta matrix from each iteration of the estimation procedures (means and covariances); 1 to save them in data buffer

_AMstart

starting values for theta matrix: 1 (default) for listwise deleted matrix of data means and covariances; 2 for identity matrix for $[2:p+1,2:p+1]$ and -1 in [1,1] and zeros elsewhere; or user provided starting value matrix

_AMtype

string name of file extension for imputed datasets; options are WKS for Lotus v1.0, XLS for Excel v2.1, WQ1 for Quattro v1.0, WRK for Symphony v1.0, DB2 for dBase II, DBF for dBase III, DB for Paradox v3.0, CSV, TXT, or ASC for ASCII character delimited, PRN for ASCII formatted, DAT for GAUSS data set (default=DAT). If running Gauss in operating system other than Windows, must set _AMtype to default setting.

_AMvarnm

A vector of variable names. If the input to amelia is a string, the default is the variable names stored in the Gauss data set. Otherwise it is var1, var2, etc.

_AMmupr

hyperparameter for prior on means, set externally only if _AMempri=-3

_AMtaupr

hyperparameter for weighting of prior (i.e., how many imaginary observations is prior on means based on), set externally only if _AMempri=-3

_AMmpr

hyperparameter of prior, equals df for inverse wishart distribution, set externally only if _AMempri=-3; if _Amempri $>$0 then _AMmpr=_AMempri and ridge prior is implemented

_AMsigpr

hyperparameter of prior, equals prior on the variance covariance matrix, set externally only if _AMempri=-3.

_AMimpte

set to 1 indicates that imputed data sets are created (default), set to 0 indicates that no imputations are made

_AMgap

number of iterations for IP, SIP, and VIP between any two concurrent final imputations (default=100) (this global lets the user control how many iterations of each method are run between creating an imputed data set).

_AMimps

1 to save thetas when making imputations; 0 not to save thetas (default).

_AMnew

string name for the new completed datasets. Appended to this name is a number for each dataset, consequently _AMnew should be 6 characters or less. Default is file if the program was called as amelia("file") and newdat if a matrix was fed to amelia.

_AMnds

number of datasets created by imputation procedures (5 is default).

_AMsn

number of draws from the approximating distribution for importance sampling is equal to the number of datasets to be imputed _AMnds $\times$ _AMsn (default for _AMsn is 10). Consider increasing this if the number of resamplings is about the same as _AMnds.

_AMsfac

factor by which to multiply the variance matrix used in the importance sampling, 1.1 is the default

_AMst

Defines the approximation distribution for importance sampling. 0 (default) for multivariate normal distribution; greater than 2 for multivariate t distribution, with _AMst degrees of freedom.

_AMvcm

Calculate variance matrix using 1 the VIP simulation method (default); 2 the inverse of the negative Hessian; 3 outer product gradient method.

_AMts

Scalar indicating column of variable which indicates time.

_AMcs

Scalar indicating column of variable which indicates each unique cross-section for dataset.

_AMusets

Scalar equal to 1 if the time index variable defined by _AMts should be included in the imputation model, 0 if not (default).

_AMusecs

Scalar equal to 1 if the cross sectional index variable defined by _AMcs should be included in the imputation model, 0 if not (default).

_AMlagvs

Column vector whose elements indicate which variables to include lags for in the imputation model.

_AMtstep

Distance between 2 successive observations by time (i.e. if you have yearly data this should equal 1).