Changelogs » Zepid



Labeling fix for `RiskDifference` summary

Adding option to extract standard errors from `TMLE` and `AIPTW`


Warning for upcoming change for `IPTW` in v0.8.0. To better align with other causal estimators, `IPTW` will no longer
only return a vector of weights. Behind the scenes, `IPTW` will be able to estimate the marginal structural model
and provide the results directly in v0.8.0. `IPTW` will still allow access to the `Weight` column. Other tweaks are
coming, such as `IPTW` estimators built for different data types. For example, `SurvivalIPTW` for survival data (like

Stochastic treatments can be estimated with the new `StochasticIPTW` class. This class is different from `IPTW` in that
it provides the estimated mean of the outcome given the treatment plan. For comparisons, multiple versions of treatment
plans need to be specified, calculated, then compared. For confidence intervals, a bootstrap procedure should be used


G-estimation of structural nested models (for a single time point) are now available through `GEstimationSNM`. Psi
parameters can be calculated using a closed form solution or via a `scipy` optimization procedure

Survival analysis g-formula is now implemented with `SurvivalGFormula`. This g-formula implementation is for
time-to-event data, where the treatment/exposure is determined at baseline. This does not allow for time-varying
exposures. For time-varying exposures, `MonteCarloGFormula` or `IterativeCondGFormula` should be used instead

`summary()` functions have been updated to provide more information regarding the model

Added a calculator function for Rubin's Rule to merge multiple imputation results. Input is a list of point estimates
and a list of variance estimates for `rubins_rules()`. This function returns a summary point estimate and summary

Weighted models are switched from `GEE` to `GLM` when possible. GEE takes extra computation time. GLM provides the
correct point estimates, but wrong variance. Since I don't need the variance to be correct from most models, I switched
to GLM. This improves the speed of fitting weighted models. Especially important for bootstrapping procedures

Aligned `exposure` and `outcome` references with the causal functions. All classes now use the same labels for the
exposure and the outcome column labels.

Updated ReadTheDocs website


AIPTW`` now supports continuous outcomes (normal or Poisson). Format is the same as `TMLE`.

`AIPTW` and `IPTW` now include the optional argument `weights`

Fixed `TMLE` attribute for average treatment effect confidence intervals, from `average_treatment_effect_ic` to

Fixed issue in `IPTW` assumption calculations. Depending on when `positivity()` was called, it changed the results of


MonteCarloGFormula` now includes a separate `censoring_model()` function for informative censoring.
Additionally, I added a low memory option to reduce the memory burden during the Monte-Carlo procedure

``IterativeCondGFormula`` has been refactored to accept only data in a wide format. This allows for me to handle more
complex treatment assignments and specify models correctly. Additional tests have been added comparing to R's `ltmle`

There is a new branch in `zepid.causal`. This is the `generalize` branch. It contains various tools for generalizing
or transporting estimates from a biased sample to the target population of interest. Options available are
inverse probability of sampling weights for generalizability (`IPSW`), inverse odds of sampling weights for
transportability (`IPSW`), the g-transport formula (`GTransportFormula`), and doubly-robust augmented inverse
probability of sampling weights (`AIPSW`)

`RiskDifference` now calculates the Frechet probability bounds

``TMLE`` now allows for specified bounds on the Q-model predictions. Additionally, avoids error when predicted
continuous values are outside the bounded values.

``AIPTW`` now has confidence intervals for the risk difference based on influence curves

``spline`` now uses `numpy.percentile` to allow for older versions of NumPy. Additionally, new function
`create_spline_transform` returns a general function for splines, which can be used within other functions

Lots of documentation updates for all functions. Additionally, `summary()` functions are starting to be updated.
Currently, only stylistic changes


While conducting further testing, I found an error in `AIPTW`. I have since corrected it and added additional tests
to `tests/`. Please rerun any analyses ran that used `AIPTW`


Added a fix to ``TMLE`` for machine learning libraries and missing outcome data


Support for Python 3.7 has been added

``AIPW`` has been removed. It has been replaced with ``AIPTW``

``TMLE`` now supports continuous outcomes (normal or Poisson) and allows for missing outcome data to be missing at
random. This matches more closely to the functionality to R's `tmle`

``IPMW`` allows for monotone missing data.

``MonteCarloRR`` for probabilistic bias analysis allows for random error to be incorporated


``TimeVaryGFormula`` is separated into ``MonteCarloGFormula`` and ``IterativeCondGFormula``. This change is for
maintenance of the estimators and to avoid confusion since they are sufficiently distinct. Originally, I was unaware of
the iterative conditional estimator, which is why the original name was based on time-varying g-formula. While they are
related, it is more confusing to wrap them both in the same class. ``TimeVaryGFormula`` will stick around to v0.6.0.
Going forward it will be cut. It will not be updated going forward

L'Abbe plots are now supported. These plots are useful for visualizing additive and multiplicative interactions for
risk estimates. These are valid for either associations or causal effects.

``IPTW.plot_love`` now displays the legend. I have previously not included this in the function (I should have)

``TMLE`` refactored to estimate machine learners via an outside function. Also converts all pd.Series to np.array to
avoid some unhappiness with sklearn / supylearner models



``TMLE`` defaults to calculate all possible measures (risk difference, risk ratio, odds ratio) rather than individual

``TimeFixedGFormula`` allows stochastic interventions for binary exposures. For a stochastic intervention, *p* percent
of the population is randomly treated. This process is repeated *n* times and mean is the marginal outcome. Stochastic
interventions may better align with real-world interventions (often you intervention will **not** be able to treat
*everyone*). Additionally, conditional probabilities are implemented for stochastic interventions. For example, those
with *C=1* might be treated randomly at *p*, while those with *C=0* are treated at *q*.

``IPTW.standard_mean_difference`` and ``IPTW.plot_love`` both support categorical variables. Categorical variables must
be modeled with ``patsy``'s ``C(.)`` keyword. Otherwise, the dummy variables will be treated as binary variables


Added case-control example data set. ``load_case_control_data()``

Changed rounding in Table 1 generator



``TimeFixedGFormula`` supports Poisson and normal distributed continuous outcomes

``IPTW``'s ``plot_kde`` and ``plot_boxplot`` can plot either the probabilities of treatment or the log-odds

``IPTW`` allows for sklearn or supylearner to generate predicted probabilities. Similar to ``TMLE``

``IPTW`` now allows for Love plot to be generated. These plots are valuable for assessing covariate balance via absolute
standardized mean differences. See Austin & Stuart 2015 for an example. In its current state ``IPTW.plot_love`` is
"dumb", in the sense that it plots all variables in the model. If you have a quadratic term in the model for a
continuous variable, it plots both the linear and quadratic terms. However, it is my understanding that you only need
to look at the linear term. These plots are not quite for publication, rather they are useful for quick diagnostics

``IPTW.standardized_mean_differences`` now calculates for all variables automatically. This is used in the background
for the ``plot_love``. For making publication-quality Love plots, I would recommend using the returned DataFrame from
this function and creating a plot manually. *Note* it only returns standardized differeneces, not absolute standardized
differences. Love plots use the standardized differences.
*WARNING:* standardized differences **only** supports binary or continuous variables. Categorical variables are NOT
supported. This will be fixed in v0.4.2 update


Website updated to reflect above changes and correcting errors I had missed on last check



``TMLE`` has been modified to estimate the custom user models now, rather than take the input. This better corresponds
to R's tmle (however, R does the entire process in the background. You must specify for this implementation). The reason
for this major change is that ``LTMLE`` requires an iterative process. The iterative process requires required fitting
based on predicted values. Therefore, for ``LTMLE`` an unfitted model must be input and repeatedly fit. ``TMLE`` matches
this process.

``TimeVaryGFormula`` supports both Monte Carlo estimation and Sequential Regression (interative conditionals) this
added approach reduces some concern over model misspecification. It is also the process used by LTMLE to estimate
effects of interventions. Online documentation has been updated to show how the sequential regression is estimated and
demonstrates how to calculated cumulative probabilities for multiple time points

All calculator functions now return named tuples. The returned tuples can be index via ``returned[0]`` or

Documentation has been overhauled for all functions and at ReadTheDocs

Tests have been added for all currently available functions.

Travis CI has been integrated for continuous testing


``AIPW`` drops missing data. Similar to ``TMLE``

``IPTW`` calculation of standardized differences is now the ``stabilized_difference`` function instead of the previously
used ``StandardDifference``. This change is to follow PEP guidelines

The ``psi`` argument has been replaced with ``measure`` in ``TMLE``. The print out still refers to psi. This update is
to help new users better understand what the argument is for

Better errors for ``IPTW`` and ``IPMW``, when a unstabilized weight is requested but a numerator for the model is



``TMLE`` now allows estimation of risk ratios and odds ratios. Estimation procedure is based on ``tmle.R``

``TMLE`` variance formula has been modified to match ``tmle.R`` rather than other resources. This is beneficial for future
implementation of missing data adjustment. Also would allow for mediation analysis with TMLE (not a priority for me at
this time).

``TMLE`` now includes an option to place bounds on predicted probabilities using the ``bound`` option. Default is to use
all predicted probabilities. Either symmetrical or asymmetrical truncation can be specified.

``TimeFixedGFormula`` now allows weighted data as an input. For example, IPMW can be integrated into the time-fixed
g-formula estimation. Estimation for weighted data uses statsmodels GEE. As a result of the difference between GLM
and GEE, the check of the number of dropped data was removed.

``TimeVaryGFormula`` now allows weighted data as an input. For example, Sampling weights can be integrated into the
time-fixed g-formula estimation. Estimation for weighted data uses statsmodels GEE.


Added Sciatica Trial data set. Mertens, BJA, Jacobs, WCH, Brand, R, and Peul, WC. Assessment of patient-specific
surgery effect based on weighted estimation and propensity scoring in the re-analysis of the Sciatica Trial. PLOS
One 2014. Future plan is to replicate this analysis if possible.

Added data from Freireich EJ et al., "The Effect of 6-Mercaptopurine on the Duration of Steriod-induced
Remissions in Acute Leukemia: A Model for Evaluation of Other Potentially Useful Therapy" *Blood* 1963

``TMLE`` now allows general sklearn algorithms. Fixed issue where ``predict_proba()`` is used to generate probabilities
within ``sklearn`` rather than ``predict``. Looking at this, I am probably going to clean up the logic behind this and
the rest of ``custom_model`` functionality in the future

``AIPW`` object now contains ``risk_difference`` and ``risk_ratio`` to match ``RiskRatio`` and ``RiskDifference``



TMLE now allows user-specified prediction models (like machine learning models). This is done by setting the option
argument `custom_model` to a fitted model with the `predict()` function. For a full tutorial (with SuPyLearner), see
the website.

Updated API for printing model results to the console. All branches have been updated to
use print_results now. (Thanks Cameron Davidson-Pilon)

Semi-Bayesian function now calculates a check on the compatibility between the prior and data. It generates a warning
if a small p-value is detected (p < 0.05). The full information on this check can be read in *Modern Epidemiology* 3rd
edition pg340.


Addition of Targeted Maximum Likelihood Estimation (TMLE). No current timeline developed

Addition of IPW for Interference settings. No current timeline but hopefully before 2018 ends

Further conforming to PEP guidelines (my bad)


TimeVaryGFormula speed-up: some background optimization to speed up TimeVaryGFormula. Changes include: pd.concat()
rather than pd.append() each loop . Shuffled around some statements to execute only once rather than multiple times. In
some testing, I went from 22 seconds to run to 3.4 seconds



IPW all moved to zepid.causal.ipw. zepid.ipw is no longer supported

IPTW, IPCW, IPMW are now their own classes rather than functions. This was done since diagnostics are easier for IPTW
and the user can access items directly from the models this way.

Addition of TimeVaryGFormula to fit the g-formula for time-varying exposures/confounders

effect_measure_plot() is now EffectMeasurePlot() to conform to PEP

ROC_curve() is now roc(). Also 'probability' was changed to 'threshold', since it now allows any continuous variable for
threshold determinations


Added sensitivity analysis as proposed by Fox et al. 2005 (MonteCarloRR)

Updated Sensitivity and Specificity functionality. Added Diagnostics, which calculates
both sensitivity and specificity.

Updated dynamic risk plots to avoid merging warning. Input timeline is converted to a integer (x100000), merged, then
back converted

Updated spline to use np.where rather than list comprehension

Summary data calculators are now within zepid.calc.utils


All pandas effect/association measure calculations will be migrating from functions to classes in a future version.
This will better meet PEP syntax guidelines and allow users to extract elements/print results. Still deciding on the
setup for this... No changes are coming to summary measure calculators (aside from possibly name changes). Intended as


Removed histogram option from IPTW in favor of kernel density. Since histograms are easy to generate with matplotlib, just dropped the entire option.

Created causal branch. IPW functions moved inside this branch

Added depreciation warning to the IPW branch, since this will be removed in 0.2 in favor of the causal branch for organization of future implemented methods

Added time-fixed g-formula

Added simple double-robust estimator (based on Funk et al 2011)


Fix to 0.1.4 and since PyPI does not allow reuse of library versions, I had to create new one. Fixes issue with ipcw_prep() that was a pandas error (tried to drop NoneType from columns)


Updates: Added dynamic risk plot

Fixes: Added user option to allow late entries for ipcw_prep()


Updates: added ROC curve generator to graphics, allows user-specification of censoring indicator to ipcw,


Original release. Previous versions (0.1.0, 0.1.1) had errors I found when trying to install via PyPI. I forgot to include the `package` statement in `setup`