Estimate a mixture model with a Poisson or negative binomial count model and a point mass at zero. The predictors can be different for the two models. The estimated model can be saved and used for predictions on new data.
- IBM SPSS Statistics 18 or later and the corresponding IBM SPSS Statistics-Integration Plug-in for R.
Note: For users with IBM SPSS Statistics version 23 or higher, the STATS ZEROINFL extension is installed as part of IBM SPSS Statistics-Essentials for R.
- Open IBM SPSS Statistics
- Navigate to Utilities -> Extension Bundles -> Download and Install Extension Bundles
- Search for the name of the extension and click Ok. Your extension will be available.
STATS ZEROINFL
MODELSOURCE = ESTIMATE^** or FILE or WORKSPACE
MODELFILE = "file"
DEPENDENT = variable
COUNTMODEL = variables
ZEROMODEL = variables
SAMEREGRESSORS = YES or NO^**
COUNTOFFSET = variable
ZEROOFFSET = variable
COUNTDIST = POISSON^** or NEGBIN or GEOMETRIC
ZEROLINK = LOGIT^** or PROBIT or CLOGLOG or CAUCHIT or LOG
/OPTIONS
STARTVALUES = GENLIN^** or EM
OPTMETHOD = BFGS^** or NELDERMEAD or CG or LBFGSB or SANN or BRENT
MAXITER = number
TOL = number
MISSING = EXCLUDE^** or FAIL
/SAVE
DATASET=dataset
ID = variable
WORKSPACEACTION = CLEAR^** or RETAIN
WORKSPACEOUTFILE = "file"
/HELP
^* Required
^** Default
Example:
STATS ZEROINFL DEPENDENT=y COUNTMODEL = x1 x2 x3
ZEROMODEL= w1 COUNTDIST= NEGBIN
/SAVE DATASET=results ID=id WORKSPACEOUTFILE="C:/myproject/zero.Rdata".
STATS ZEROINFL /HELP displays this help and does nothing else.
This command operates in two modes. It can be used to estimate a model, or it can be used with a model previously estimated and saved in memory or a file to make predictions for new data.
MODELSOURCE specifies the model source. The choices are
- ESTIMATE: estimate a model using data in the current active dataset
- WORKSPACE: use the model in the currently loaded workspace. The model must come from a previously run command in the current session with MODELSOURCE=ESTIMATE and retained in memory with WORKSPACEACTION=RETAIN.
- FILE: use a previously created model that was saved to a file with WORKSPACEOUTFILE.
MODELFILE specifies the file containing the model and is required if MODELSOURCE=FILE is used.
DEPENDENT specifies the dependent variable, which is expected to contain counts. It is required in estimation mode. It must have a scale measurement level.
COUNTMODEL specifies one or more independent variables for the count and is required in estimation mode.
ZEROMODEL can specify a different set of predictors for the zero/nonzero portion of the model. If omitted, the model depends on the setting for SAMEREGRESSORS.
SAMEREGRESSORS specifies whether the predictors for the zero model are the same as for the count model. If it is NO, and no ZEROMODEL variables are supplied, the zero model is the same for all cases, i.e., the zero model predictors consist just of a constant term.
COUNTOFFSET can specify an offset for the count model.
ZEROOFFSET can specify an offset for the zero model. The SAMEREGRESSORS setting does not apply to the zerooffset specification.
COUNTDIST specifies the distribution for the count model. The choices are
- POISSON: Poisson
- NEGBIN: Negative binomial
- GEOMETRIC: Negative binomial with a size (number of failures) parameter of 1.
ZEROLINK specifies the link function for the binary zero-inflation model. The choices are
- LOGIT: logistic
- PROBIT: probit
- CLOGLOG: Complementary log log
- CAUCHIT: Cauchy quantiles
- LOG: log
STARTVALUES specifies whether the intial values for the iterative estimation algorithm come from a generalized linear model or from an EM (estimation maximization) model. The EM algorithm uses a generalized linear model first step, so it is possible to get a failure message referring to glm when EM is the setting.
OPTMETHOD specifies the iterative solving algorithm. The choices are
- BFGS: Quasi Newton
- Nelder-Mead: A robust but slow method that does not use gradients
- CG: Conjugate gradient
- SANN: variant of simulated annealing
MAXITER specifies the maximum number of iterations and defaults to 1000.
TOL specifies the convergence tolerance and typically defaults to 1.64 * 10^-10.
MISSING specifies whether cases with missing values are omitted or cause the procedure to stop.
DATASET specifies the output dataset name. The name must not already be in use. For estimation, the dataset will contain the fitted values and residuals. For prediction, it will contain the predicted values and is required.
ID optionally specifies an id variable to be included in the dataset. If one is not specified, the cases are sequentially numbered.
WORKSPACEACTION specifies whether to clear the in-memory workspace after the procedure (CLEAR) or to retain it for following prediction calculations.
WORKSPACEOUTFILE specifies whether to save the estimated model to a file. It only applies to estimation model.
- Apache 2.0
- JKP, IBM SPSS