Difference between revisions of "Data Analysis"

Jump to: navigation, search
 
(32 intermediate revisions by 6 users not shown)
Line 1: Line 1:
Data analysis is the process of exploring and describing trends and results from data. Data analysis typically occurs in two stages: exploratory analysis and final analysis. This page provides guidance on how to organize analysis files and output results in an orderly and [[Reproducible Research | reproducible]] manner.


==Read First==
* Always [[Data Cleaning | clean]] data before conducting data analysis.
* Place exploratory analysis files in a separate, well-organized "exploratory" folder; place the final analysis file in the "final"  analysis folder.
* Create tables and figures via [[Reproducible Research|replicable]] '''do-files''' in such a way that the result files are ordered, named, placed, and formatted appropriately.


==Exploratory vs. Final Analysis==


Exploratory analysis focuses on producing easily understood summaries of the trends in the data so that researchers can begin to outline reports, [[Publishing Data|publications]], presentations, and summaries. Final analysis is the fine-tuned culmination of exploratory analysis and requires re-written code appropriate for public release with the results.


==Introduction==
== Organizing Analysis Files ==
Data analysis refers to the full process of exploring and describing trends and results from data. Data analysis typically has two stages:
Place exploratory analysis programs in an "exploratory" folder, separated according to topic. When folder syncing over [https://www.dropbox.com Dropbox] or [[Getting Started with GitHub|Github]], separating these files by function (rather than combining them into a single "analysis" file) allows multiple researchers to work simultaneously and modularly.
 
# Exploratory Analysis
# Final Analysis
 


In exploratory analysis, emphasis will be on producing easily understood summaries of the trends in the data so that the reports, publications, presentations, and summaries that need to be produced can begin to be outlined. Once those stories begin to come together, the code is re-written in a "final" form which would be appropriate for public release with the results.
When the final analysis workflow is agreed upon for a given publication or other output, a final analysis file should be collated for that output only in the "final analysis" folder. This allows selective reuse of the code from the exploratory analyses, in preparation for the final release of the code if required. This allows any collaborator, referee, or replicator to access only the code used to prepare the final outputs and [[Reproducible Research | reproduce]] them exactly.


== Preparing the Dataset for Analysis ==
== Implementing Analysis ==


Once data is collected, it must be recombined into a final format for analysis, including the construction of derived variables not present in the initial collection. See [[Data Cleaning]].
Below follows a sampling of specific analytical methods.  


== Organizing Analysis Files ==
*[[Spatial Analysis |Spatial/GIS Analysis]] uses geospatial data to explore relationships mediated by proximity or connectiveness. This can be descriptive (i.e. map illustrations) or informative (i.e. distance to and quality of the nearest road).
*[[Randomization Inference | Randomization inference]] techniques replace the "normal" p-values from regression analyses with values based on the '''treatment assignment''' methodology. They are generally recommended for reporting in experiments whose estimates are of [[Randomization|randomly]] assigned '''treatment''' controlled by the implementer and researcher.
*[[Cost-effectiveness Analysis | Cost-effectiveness analysis]] compares the cost and effectiveness per unit of a given program to determine whether the value of an intervention justifies its cost.
*[[Regression Discontinuity|Regression discontinuity]] analysis is a [[Quasi-Experimental Methods | quasi-experimental]] '''impact evaluation''' design which attempts to find the causal effects of interventions by assigning a threshold (cutoff point) above and below which the '''treatment''' is assigned.
*[[Propensity Score Matching | Propensity score matching]] is another '''quasi-experimental impact evaluation''' technique to estimate the effects of a '''treatment''' by matching control group participants to '''treatment group''' participants based on '''propensity score'''.
*Heterogeneous Effects Analysis uses an [https://web.stanford.edu/~jgrimmer/het.pdf ensemble of methods] to understand how effects vary across sub-populations
*[[Data visualization]] is a critical step in effectively communicating your research results.


Analysis programs that is exploratory in nature should be held in an "exploratory" folder and separated according to topic. Particularly when folder syncing over [www.dropbox.com Dropbox] or [www.github.com Github] is being used, separating these files by function (rather than combining them into a single "analysis" file) allows multiple researchers to work simultaneously and modularly.
The [[Stata Coding Practices | Stata]] cheat sheet on [http://geocenter.github.io/StataTraining/pdf/StataCheatSheet_analysis_201615_June-REV.pdf data analysis] gives guidelines on relevant '''Stata''' code for analysis. The cheat sheet on [http://geocenter.github.io/StataTraining/pdf/StataCheatSheet_programming15_2016_June_TE-REV.pdf Stata programming] is a good resource for more advanced analytical tasks in '''Stata'''.
 
When the final analysis workflow is agreed upon for a given publication or other output, a final analysis file should be collated for that output only in the "final" analysis folder. This allows selective reuse of the code from the exploratory analyses, in preparation for the final release of the code if required. This allows any collaborator, referee, or replicator to access only the code used to prepare the final outputs and reproduce them exactly.


== Outputting Analytical Results ==
== Outputting Analytical Results ==


Since the final analysis do-files are intended to be fully replicable, and the code itself is considered a vital, shareable output, all tables and figures should be created in such a way that the files are ordered, named, placed, and formatted appropriately. Running the analysis dofile should result in ''only'' necessary files in the "outputs" folder, with names like "figure_1.png", "table_1.xlsx", and so on.
Since the final analysis '''do-files''' are intended to be fully [[Reproducible Research|replicable]] and the code itself is considered a vital, shareable output, all tables and figures should be created in such a way that the files are ordered, [[Naming Conventions | named]], placed, and formatted appropriately. Running the analysis '''do-file''' should result in only necessary files in the "outputs" folder, with names like "figure_1.png", "table_1.xlsx", and so on.
 
For some applications (such as creating internal presentations or simple Word reports, file types like PNG and XLSX are sufficiently functional. For larger projects with multiple collaborators, particularly when syncing over a [www.github.com GitHub] service, plaintext file types such as EPS, CSV, and TEX will be the preferred formats. Tables and figures should at minimum be produced by this file such that no further mathematical calculations are required; they should furthermore be organized and formatted as nearly to the published versions as possible. Figures are typically easy to do this in by using an appropriate <code>graph export</code> command in Stata or the equivalent. [www.latex-project.org LaTeX] is a particularly powerful tool for doing this with tables. DIME provides several guides on both processes.
 
== Resources for Specific Analytical Tasks ==
 
===Heterogeneous Effects Analysis===
 
===Randomization Inference===
 
[[Randomization Inference]] techniques replace the "normal" p-values from regression analyses with values based on the treatment assignment methodology, and are generally recommended for reporting in experiments whose estimates are of randomly assigned treatment controlled by the implementer and researcher.


===Principal Components Analysis===
For some applications (i.e. creating internal presentations or simple Word reports), file types like PNG and XLSX are sufficiently functional. For larger projects with multiple collaborators, particularly when syncing over a [[Getting Started with GitHub|GitHub]] service, plaintext file types such as EPS, CSV, and TEX will be the preferred formats. Tables and figures should at minimum be produced by these files such that no further mathematical calculations are required. They should furthermore be organized and formatted as nearly to the [[Publishing Data|published]] versions as possible. It is typically easy to do this with figures by using an appropriate <code>graph export</code> command in [[Stata Coding Practices|Stata]] or the equivalent. [https://www.latex-project.org LaTeX] is a particularly powerful tool for doing this with tables. DIME provides several guides on both processes. See [[Exporting Analysis |exporting analysis results]] for more details and more resources.
 
[[Principal Component Analysis (PCA)]] is an analytical tool looks to explain the maximum amount of variance with the fewest number of principal components.  
 
=== Cost Effectiveness Analysis ===
 
[[Cost-effectiveness Analysis]]


== Additional Resources ==
== Additional Resources ==
 
*Angrist and Pischke's [https://www.researchgate.net/publication/51992844_Mostly_Harmless_Econometrics_An_Empiricist's_Companion Mostly Harmless Econometrics: An Empiricist's Companion]


[[Category: Data Analysis ]]
[[Category: Data Analysis ]]

Latest revision as of 18:27, 9 August 2023

Data analysis is the process of exploring and describing trends and results from data. Data analysis typically occurs in two stages: exploratory analysis and final analysis. This page provides guidance on how to organize analysis files and output results in an orderly and reproducible manner.

Read First

  • Always clean data before conducting data analysis.
  • Place exploratory analysis files in a separate, well-organized "exploratory" folder; place the final analysis file in the "final" analysis folder.
  • Create tables and figures via replicable do-files in such a way that the result files are ordered, named, placed, and formatted appropriately.

Exploratory vs. Final Analysis

Exploratory analysis focuses on producing easily understood summaries of the trends in the data so that researchers can begin to outline reports, publications, presentations, and summaries. Final analysis is the fine-tuned culmination of exploratory analysis and requires re-written code appropriate for public release with the results.

Organizing Analysis Files

Place exploratory analysis programs in an "exploratory" folder, separated according to topic. When folder syncing over Dropbox or Github, separating these files by function (rather than combining them into a single "analysis" file) allows multiple researchers to work simultaneously and modularly.

When the final analysis workflow is agreed upon for a given publication or other output, a final analysis file should be collated for that output only in the "final analysis" folder. This allows selective reuse of the code from the exploratory analyses, in preparation for the final release of the code if required. This allows any collaborator, referee, or replicator to access only the code used to prepare the final outputs and reproduce them exactly.

Implementing Analysis

Below follows a sampling of specific analytical methods.

  • Spatial/GIS Analysis uses geospatial data to explore relationships mediated by proximity or connectiveness. This can be descriptive (i.e. map illustrations) or informative (i.e. distance to and quality of the nearest road).
  • Randomization inference techniques replace the "normal" p-values from regression analyses with values based on the treatment assignment methodology. They are generally recommended for reporting in experiments whose estimates are of randomly assigned treatment controlled by the implementer and researcher.
  • Cost-effectiveness analysis compares the cost and effectiveness per unit of a given program to determine whether the value of an intervention justifies its cost.
  • Regression discontinuity analysis is a quasi-experimental impact evaluation design which attempts to find the causal effects of interventions by assigning a threshold (cutoff point) above and below which the treatment is assigned.
  • Propensity score matching is another quasi-experimental impact evaluation technique to estimate the effects of a treatment by matching control group participants to treatment group participants based on propensity score.
  • Heterogeneous Effects Analysis uses an ensemble of methods to understand how effects vary across sub-populations
  • Data visualization is a critical step in effectively communicating your research results.

The Stata cheat sheet on data analysis gives guidelines on relevant Stata code for analysis. The cheat sheet on Stata programming is a good resource for more advanced analytical tasks in Stata.

Outputting Analytical Results

Since the final analysis do-files are intended to be fully replicable and the code itself is considered a vital, shareable output, all tables and figures should be created in such a way that the files are ordered, named, placed, and formatted appropriately. Running the analysis do-file should result in only necessary files in the "outputs" folder, with names like "figure_1.png", "table_1.xlsx", and so on.

For some applications (i.e. creating internal presentations or simple Word reports), file types like PNG and XLSX are sufficiently functional. For larger projects with multiple collaborators, particularly when syncing over a GitHub service, plaintext file types such as EPS, CSV, and TEX will be the preferred formats. Tables and figures should at minimum be produced by these files such that no further mathematical calculations are required. They should furthermore be organized and formatted as nearly to the published versions as possible. It is typically easy to do this with figures by using an appropriate graph export command in Stata or the equivalent. LaTeX is a particularly powerful tool for doing this with tables. DIME provides several guides on both processes. See exporting analysis results for more details and more resources.

Additional Resources