No articles match
The post-estimation contract20 days ago
The promises | What decides a cell | The fits | The contract | Why each refusal | Decisions
estimatr 2.0: what changes, what does not, and why20 days ago
The short version | What does not change | What changes, and why | 1. horvitz_thompson() takes one probability argument | 2. Clustered fixed effects get a different default standard error | 3. Functions that are gone | Porting an existing script | What is new | Blocked designs whose blocks are not all the same shape | Speed | How this was checked | The estimatr issue list | Current status | References
Mathematical notes and guarantees for estimatr20 days ago
Written with AI, and checked accordingly | How the checking works | Notation | The data used throughout | lm_robust | Coefficients | Weights | Heteroskedasticity-robust variance | Leverage at and above one | Cluster-robust variance | Absorbed fixed effects | Rank | Weights and the hat matrix: where Stata differs | Confidence intervals and testing | Multiple outcomes | lm_lin | iv_robust | Variance | Which leverage | Correspondence with Stata | lh_robust | difference_in_means | Estimates | Variance for unblocked and clustered designs | Variance for blocked designs | What is refused | horvitz_thompson | What the declaration buys | Every promise in one table | What this does not cover | References
Getting started with estimatr20 days ago
The example data | lm_robust() | Clustered standard errors | Choosing a different variance estimator | Fixed effects | Weights | lm_lin() | iv_robust() | difference_in_means() | horvitz_thompson() | lh_robust() | Working with the output | In a simulation | Where to go next | References
Contributing a design22 days ago
The short version | 1. Fork | 2. Add your file | 3. Refresh | 4. Check | 5. Pull request | YAML fields | Checklist
DesignLibrary: a library of declared designs22 days ago
The idea | What a design file looks like | What designs are included? | Use DesignLibrary to make a design | Change parameters and simulate | Browse in Shiny | Maintainer tools
Upgrading from DesignLibrary 0.1: what changes, what does not, and why22 days ago
The short version | What does not change | What changes, and why | Removed functions | block_cluster_two_arm_designer() takes new arguments | multi_arm_designer() has a larger default sample | Two renamings that change no values | args_to_fix is ignored | Designs are now all DeclareDesign 2.0 designs | Sweeping a parameter | Porting an existing script | What is new
DeclareDesign 2.0: what changes, what does not, and why24 days ago
The short version | Getting it | What does not change | What changes, and why | In DeclareDesign 2.0 | redesign() reaches a name the design reads, not a number it writes down | inquiry = is the quoted label of the inquiry | make_groups = vars(...) is gone; group with group_by() | Dotted argument names: .design, .expand, .method, and .summary | draw_data() no longer runs part of a design on supplied data | reshape_diagnosis() reshapes; select columns afterwards | expand_design() is deprecated | set_diagnosands() is deprecated | Step-modification functions are deprecated | The code-comparison functions are removed; compare_diagnoses() is not | declare_potential_outcomes(), declare_reveal(), and declare_estimators() are gone | What comes back has one more column, in a different order | An estimator that fails on one draw is recorded, not fatal | In fabricatr 2.0 | Porting an existing script | What is new | declare_parameters() and declare_notes() | A design is a value | Step-level draws | Variance decomposition | Parallel execution | summary() says what each step did | Progress | Estimator arguments pass as written | Speed | Why rewrite at all | Current status | Bugs reported against 1.1.1 that 2.0 fixes
Getting started24 days ago
Declaration | Diagnosis | Redesign | A design is code
fabricatr: what changes, what does not, and why26 days ago
The short version | Getting it | What does not change | What changes, and why | nest = FALSE becomes declare_level() | join_using() becomes .by | by becomes .by in modify_level() | N must be named | The result is a tibble | modify_level() at a named level works at that level | reveal_outcomes() keeps the type of the outcome | ID columns are zero-padded character, as in 1.x | total_sd is a parameter, not a rescaling | The correlated draw does not depend on your machine | On reproducing an old seed | N is never a column | Porting an existing script | What is new | Existing data frames come in as levels | n() counts the rows in hand | Panels with persistent shocks | Speed | Why rewrite at all | Upstream issues | Other changes worth knowing
Getting started26 days ago
One level | Two levels | Potential outcomes | Draws that look like real variables | Resampling | Where to read next
Design and Analysis of Experiments with randomizr3 months ago
A hypothetical experiment | Simple random assignment | Complete random assignment | Simple and complete random assignment compared | Block random assignment | Balanced random assignment | Clustered assignment | Blocked and clustered assignment | Calculating probabilities of assignment | Best practices | Random assignment procedure = Random assignment function | Check probabilities of assignment directly | Save your random assignment
Introduction to balanced_ra3 months ago
Terms used in this article | balanced_ra guarantees | 1. Demonstration | Uneven blocks | Balancing a covariate | 2. Logic | Three C++ implementations | 2.1 Two-arm pivot | One draw: four units, target 2.3 | Leftover pairing across blocks | C++: cube_pivot_pass | 2.2 Multi-arm cube | One draw: four units, three arms | C++: the fair bet, the window, and per-block landing | 2.3 Cube-on-X | One draw: $x=(1,2,3,6)$, $p_i=1/2$ | C++: the main loop | C++: finding a direction | 3. Examples | 3.1 One winner from unequal chances | 3.2 Two-arm counts with blocks | 3.3 Three arms, no blocks | 3.4 Heterogeneous probabilities | 3.5 Clusters | Clusters with a covariate | 3.6 formula plus blocks | 3.7 Multi-arm totals with blocks are not tight | 3.8 An attainable covariate total that is still missed | 3.9 An example of an attainable covariate total that never appears at all | 4. Analysing an assignment | 4.1 Weight by the probability of the condition received | 4.2 Standard errors: fine without formula, conservative with it | 5. Declaring the design | 6. Caveats | References
What randomizr guarantees3 months ago
Written with AI, and checked accordingly | What a guarantee looks like here | The designs being checked | What randomizr promises | What this does not cover