Title
ptvtools -- Various tools for PTV analysis
Introduction
The ptvtools package is a small collection of tools for the analysis of PTVs. In electoral survey research, this acronym refers to Propensities To Vote, batteries of items concerning the self-reported probability that the respondent will ever vote for a specific party, and intended as indicators of what Downs (1957) referred to as "electoral utilities" (see van der Eijk and Franklin, Choosing Europe, University of Michigan Press 1996; van der Eijk et al., Electoral Studies 2006; van der Eijk and Franklin, Elections and Voters, Palgrave Macmillan 2009).
PTV analysis makes some specific demands both on the data and on the methods of analysis.
Firstly, data stacking is usually involved, implying that - in Stata terminology - data are better analyzed when in long rather than in wide form (see reshape). However, the tools facilitate reshaping of a dataset with multiple contexts, (eg. countries) each with possibly different numbers of items (eg. parties) in the battery(ies) to be stacked (different numbers of items having missing data for all cases in the context). After stacking, each respondent is represented by multiple rows in the data matrix, one for each of the items that was not entirely missing.
Secondly, the dependent variable is usually not the propensity to vote for a particular party, but for a generic party (yielding comparability across contexts if these involve different party systems, as is common). This also affects independent variables, which have to be specially treated - reformulated it terms of distances (proximities) or in terms of some other type of affinity, for instance so-called y-hats - before they can be used in a stacked analysis. In this, ptv analysis resembles analysis of discrete choice models for which it serves as a substitute, dealing with directly-measured preferences (utilities) rather than deriving these from the analysis of choices made.
Description
The ptvtools package includes the following commands:
gendist (Context-wise) generation of distances for a battery of spatial items, after optionally plugging missing data on the spatial items iimpute (Context-wise) incremental simple or multiple imputation of a set of variables genstacks (Context-wise) reshaping of a dataset for PTV analysis genyhats (Context-wise) generation of y-hat affinity measures linking indeps to ptvs gendummies generation of a set of dummy variables, with specific options
These tools largely duplicate existing commands in Stata but operate on data with multiple contexts (eg. countries) each of which requires separate treatment. Moreover, all of the commands in ptvtools have additional features not readily duplicated with existing Stata commands even for data that relate to a single context. If the contextvars option is not specified, each of the individual tools treats the data as belonging to a single context.
The commands take a variety of options, as documented in individual help files, some with quite cumbersome names. However, ALL options can be abbreviated to their first three characters and many can be omitted (as documented).
The commands also save a variety of indicator variables (as documented). Most of these start with an underscore character and can be deleted by "drop _*" if you don't want them to clutter your dataset. Three variables created by the command genstacks are needed by other tools and do not start with the underscore character. These are genstacks_stack, genstacks_item and genstacks_nstacks.
Workflow
The genstacks command always operates on an unstacked dataset, reshaping it into a stacked format. Other commands may operate either before or after stacking. No means is provided for unstacking a previously stacked dataset.
The commands, gendist, iimpute and genyhats by default assume the data are stacked and treat each stack as a separate context to be taken into account along with any higher-level contexts. These commands can, however, be used on unstacked data or they can be forced to ignore the stacked structure of the data by specifying the nostack option. This option has no effect on gendummies or on a command that is being used on unstacked data (since unstacked data have only one stack per case). With stacked data the nostack option has the effect of making the command ignore the separate contexts represented by each stack. This might be considered desirable if prior analysis has established that there is no stack-specific heterogeneity relevant to the estimation model for which these operations are being conducted, or in order to impute a variable that is completely missing in one stack (for example a question that was not asked for a particular choice option).
For logical reasons some restrictions apply to the order in which commands can be issued. In particular: (1) iimpute (when used for its primary purpose of imputing missing values for a battery of items) requires that data are not stacked, since members of that battery (eg. PTVs) are used for the imputation of other members of the same battery;
(2) gendist can be useful for removing missing data from items that can then be named in the addvars option to help iimpute missing data on other variables;
(3) if genyhats is used before stacking, it will have to be used once for each of the individual depvars that will, after stacking, become a single (generic) depvar;
(4) if gendist is employed after stacking, the items to which distances are computed have themselves to have been reshaped into long format by stacking them; and finally
(5) after stacking and the generation of y-hat affinity variables, the number of indeps required for a final (set of) iimpute command(s) will generally be greatly reduced, making it easier to adhere to Stata's 30 variable limit on indeps used for imputation.
Consequently, a typical workflow would involve iimpute to fill out the battery of PTVs by imputing any missing data, followed by genstacks to stack the data. This would probably be followed by genyhats, used to transform indeps (those that will not be transformed into distance measures) into y-hat affinity measures linking these indeps to the stacked depvar. The gendist command would then be used to plug missing values on item location variables and generate distances to be used in a final (set of) iimpute command(s) that will eliminate remaining missing data. This entire workflow sequence would need to be repeated as many times as datasets are required for multiple imputation.
Considerable flexibility is available, however, to transform a dataset in any sequence thought appropriate, since any of the other commands can be employed either before or after genstacks. Moreover, the researcher can use the nostack option to force the production of y-hats that "average out" contextual differences by regarding all stacks (and perhaps also all higher-level contexts) as a single context. This might be thought desirable after having established that there were no significant differences between these contexts in terms of the behavior of variables to be included in an estimation model.
Authors
Lorenzo De Sio - European University Institute; Mark Franklin - European University Institute; includes previous code and other contributions by Elias Dinas
Referencing
If you use ptvtools in published work, please use the following citation: De Sio, L. and Franklin, M. (2011), PTVTOOLS: A Stata package for PTV analysis (version 0.9), Statistical Software Components, Boston College