RooStatsTutorialsJune2013 < RooStats

---+ *RooFit/RooStats tutorials  for INFN School of Statistics 2013*

This is a set of !RooFit/RooStats tutorials given at INFN school of statistics  (3-7 June 2013) , see [[http://agenda.infn.it/conferenceDisplay.py?ovw=True&confId=5719][the school agenda]].

The required materials used during the !RooFit/RooStats tutorials is described.  The materials consist of exercises that can be solved from what has been presented in the lecture. Two levels of help are provided: a hint and a full solution. 
%TWISTY{mode="div" showlink="Hint&nbsp;" hidelink="Hide&nbsp;" }%
Here the hint is shown.
%ENDTWISTY%
%TWISTY{mode="div" showlink="Solution&nbsp;" hidelink="Hide&nbsp;" }%
Here the solution is shown.
%ENDTWISTY%
The hint allows to create the ROOT macro required to complete the exercise step by step, while the solution provides the full running code. 
The hint consists of  code snippets, which can be combined to write a full ROOT macro to solve the exercise. If you are not able to write the macro yourself,  because you find too difficult or because the time is not sufficient, you can go  directly to the solution, which will contain the running code and (hopefully) you will be able to run the macro and get the result of the tutorial exercise. 

You can find the single code attached (links at the end of the page) or you can use the provided tar file. 
Some points linked to the exercises are optional and marked with a %ICON{hand}% icon: they are useful to scrutinise in more detail some aspects. Try to solve or look at them if you have the time.

Additional material can be found on: 
   * the ROOT [[http://root.cern.ch][Web page]];
   * the !RooFit [[http://root.cern.ch/drupal/content/roofit][Web page]] and [[ftp://root.cern.ch/root/doc/RooFit_Users_Manual_2.91-33.pdf][User Guide]],[[http://root.cern.ch/drupal/sites/default/files/roofit_quickstart_3.00.pdf][Quick Start Guide]];
   * the !RooStats Twiki WebHome where you find links also to past tutorials; 
   * the lecture slides (see [[https://agenda.infn.it/conferenceTimeTable.py?confId=5719#20130607][Timetable for June 7]]);
   * !RooFit and !RooStats tutorials available in =$ROOTSYS/tutorials/roofit= and =$ROOTSYS/tutorials/roostats=.

%TOC{title="Contents:"}%

---++ *Introduction* 

---+++ Getting started with the software

A few advice words before starting. The tutorials are based on the ROOT version 5.34, possibly the latest one, 5.34.07 which is also installed in the  Virtual Machine available for the school. 
   * See the [[https://agenda.infn.it/getFile.py/access?resId=0&materialId=7&confId=5719][computer setup instructions]] if you need to install the virtual machine.
   * If you are going to use your installed version of ROOT, make sure you have configured with !RooFit ( =./configure --enable-roofit=). You can check if roofit is installed in your version of ROOT by typing the command =root-config --has-roofit=. The answer must be =yes=. If not, you must re-configure ROOT and rebuild it or install a new binary version. See again the  [[https://agenda.infn.it/getFile.py/access?resId=0&materialId=7&confId=5719][computer setup instructions]] for a detailed description on how to install and configure ROOT. 
   * A basic knowledge of ROOT is assumed, in particular how to run a ROOT macro in interpreted mode (CINT) and in compiled mode (ACLIC)
<verbatim style="background-color: rgb(255,250,205);">
           root>  .x myMacro.C        // interpreted 
           root>  .x myMacro.C+      // compiled
</verbatim>
   * It is recommended to add the following lines at the beginning of  your ROOT macro, that will load automatically the Roofit and Roostats libraries (note that the !RooStats calculator are defined in the =RooStats= namespace): 
<verbatim style="background-color: rgb(255,250,205);">
           using namespace RooFit;
           using namespace RooStats;    
</verbatim>
   * Avoid running a macro multiple times in the same ROOT CINT sessions (it might unfortunately crash sometimes if you do it, because some objects might not be properly deleted in the macro)
   * Note that !Roostats methods start with capital letter (as in ROOT) while !Roofit ones start with lower letter
   * The !Roostats calculator are quite verbose, you can suppress the output by doing: 
<verbatim style="background-color: rgb(255,250,205);">
RooMsgService::instance().setGlobalKillBelow(RooFit::WARNING) ;
</verbatim>



Some links to documentation:
   * ROOT reference guide: http://root.cern.ch/drupal/content/reference-guide
   * !Roofit user's manual: http://root.cern.ch/drupal/content/users-guide
   * ROOT, !Roofit, !Roostats tutorials: http://root.cern.ch/root/html/tutorials/ or within the local ROOT installation under =$ROOTSYS/tutorials=
   

---+++ Terminology and Conventions
Here we give pragmatic definitions for a few basic concepts that we will use.
   * *observable* - something you measure in an experiment, for example a particle's momentum. Often, a function of measured quantities, for example an invariant mass of several particles
   * *global observable* or *auxiliary observable* - an observable from another measurement, for example, the integrated luminosity 
   * *model* - a set of probability density functions (PDFs), which describe distributions of the observables or functions of observables
   * *model parameter* - any variable in your PDF expression, which is not an observable
   * *parameter of interest (POI)* - a model parameter that you study, for example a cross section of your signal process
   * *nuisance parameter* - every other model parameter, which is not your parameter of interest
   * *data* or *dataset* - a set of values of observables, either measured in an experiment, or simulated
   * *likelihood* - a model computed for a particular dataset
   * *hypothesis* - a particular model, with specified observables, POI, nuisance parameters, and prior PDFs (in case of Bayesian inference)
   * *prior PDF* - a probability density for an observable or a model parameter, which is known _a priori_, i.e. before a measurement is considered. This is a Bayesian concept exclusively. Prior has no meaning or place in a frequentist type of inference
   * *Bayesian* - a type of statistical inference that usually produces probability of the hypothesis given the data. Requires a prior.
   * *frequentist* - a type of statistical inference that usually produces probability of the data given the hypothesis.


---+ RooFit Exercises: Create and fitting a model  

The aim of the !RooFit exercises is to learn how to build a !RooFit model, how to fit the model on a set of data that we will generate and how to plot the result. We will also learn how to save the full !RooFit workspace containing the model and data in a file which we can read back later, for example when we will do the !RooStats exercises. We will learn also how to generate the data from the model. The same generated data we will then use to fit the model 


Using the snippets of code, one can build a full ROOT macro which can be run interactively in ROOT or compiled using ACLIC. 

---++ Exercise 1: Gaussian model and fit it to random generated data

We will start with a very simple exercise. If you already know a bit of !RooFit you might skip this exercise and go directly to the next one. The aim of this exercise is to show and make you practise the basic functionality of !RooFit. 

We will create a Gaussian model from which we will generate a pseudo-data set and then we will fit this data set. After fitting, we will plot the result. You can find the instruction for each steps in the lecture slides. 
   *  Start creating the Gaussian model using the workspace factory, with the syntax introduced in the lecture slides. 
   *  Use the =generate(...)= method of the =RooAbsPdf= class to generate a data set with 1000 events. 
   *  Plot the data set using =RooPlot=  in a ROOT Canvas. 
   *  Fit the generate data using the =fitTo(...)= method of the =RooAbsPdf= class
   *  Plot the resulting fit function on top of the data 




%TWISTY{mode="div" showlink="Hint&nbsp;" hidelink="Hide&nbsp;" }%
Creating the model
   * Use the syntax of the =RooWorkspace= factory to create first the Gaussian p.d.f. function with the  variables (observables and parameters).
   * Every variable in !RooFit, needs to be created with a defined range [xmin,xmax]. Use very large value if you don't know the range. If you provided only a value, the variable is considered constant. If you provide only the minimum and the maximum, the initial value will be taken in the middle of the range. It is important, to avoid undesired side effect, to have  the initial value within the given range. 
   * Note one can create using the workspace factory syntax any existing !RooFit pdf class by using the Pdf class name stripped by the =Roo= suffix. The exact signature required (variables to be passed) is defined in the Pdf class constructor (You can browse the reference documentation [[http://root.cern.ch/drupal/content/reference-guide][here]]. 
   * You need to define the [value, min, max] of a variable only the first time you create in the factory. Afterwards you can reference the variable by its name.
%CODE{"cpp" style="background: yellow;"  }%
// create Gaussian Pdf
RooWorkspace w("w"); 
w.factory("Gaussian:pdf(x[-10,10],mu[1,-1000,1000],s[2,0,1000])");   
%ENDCODE%
Retrieving objects from the workspace.
   *  After you have created the variable and p.d.f in the workspace, you can access their pointers, by using =RooWorkspace::var= for variables or =RooWorkspace::pdf= for p.d.f. Example: 
%CODE{"cpp"}%
// retrieving model components
RooAbsPdf * pdf = w.pdf("pdf");   // access object from workspace
RooRealVar * x = w.var("x");   // access object from workspace
%ENDCODE%
Generating a dataset:
   * We need to pass the observable we want to generate the data on (=x=) and the number of events
   * Note that we can generate also binned data sets. In that case we have two options: 
      * use a =RooDataHist= and  =generateBinned=:   =RooDataHist * binData = pdf->generateBinned(*x, 1000);=. 
      * use the option =AllBinned()= which will generate a weighted data set: =RooDataSet * data = pdf->generate(*x, NumEvents(1000), AllBinned());=. 
%CODE{"cpp"}%
// generate an unbinned dataset of 1000 events
RooDataSet * data = pdf->generate( *x, 1000); 
%ENDCODE%
Plot the data:
%CODE{"cpp"}%
// create a RooPlot object and plot the data
RooPlot * plot = x->frame(); 
data->plotOn(plot); 
%ENDCODE%
Fit the data: you need to call the =RooAbsPdf::fitTo=.
%CODE{"cpp"}%
// fit the data and save its result. Use the optional =Minuit2= minimiser
RooFitResult * res = pdf->fitTo(*data, Minimizer("Minuit2","Migrad"), Save(true) );
%ENDCODE%
Plot the resulting fit function in the same plot
%CODE{"cpp"}%
pdf->plotOn(plot); 
plot->Draw();   // to show the RooPlot in the current ROOT Canvas
%ENDCODE%
   
%ENDTWISTY%

%TWISTY{mode="div" showlink="Solution&nbsp;" hidelink="Hide&nbsp;" }%
%CODE{"cpp"}%
%INCLUDE{"%ATTACHURL%/GaussianModel.C"}%
%ENDCODE%

   * Gaussian Fit Result: <br />

%IMAGE{"GaussianModelFit.png" type="frame" align="none" size="600"
caption="Fit of the Gaussian model using !RooFit"}%

%ENDTWISTY%

%ICON{hand}% The fit has only two shape parameters.  How can you also add the a normalisation fit parameter, the number of events ? 
%TWISTY{mode="div" showlink="Solution&nbsp;" hidelink="Hide&nbsp;" }%
You need to make an Extended likelihood fit. For doing this in !RooFit you need to create a =RooExtendPdf= class. This can be done as following: 
%CODE{"cpp"}%
w.factory("Gaussian:gaus(x[-10,10],mu[1,-1000,1000],sigma[2,0,1000])"); 
w.factory("ExtendPdf:pdf(gaus, nevt[100,0,100000])");
%ENDCODE%
%ENDTWISTY%

%ICON{hand}% The fit is an unbinned likelihood fit. How can you make a binned likelihood fit ? 
%TWISTY{mode="div" showlink="Solution&nbsp;" hidelink="Hide&nbsp;" }% 
Generate a binned data set using =generateBinned= which will return a =RooDataHist= class or by using =generate= with the option =AllBinned=. In this last case a =RooDataSet= class is returned. 
%ENDTWISTY%


---++ Exercise 2:  Fit of a Signal over an Exponential Background

The aim of tis exercise is to learn how to build a composite model in RooFit made of two p.d.f, one representing a signal and one a background distributions. We want to determine the number of signal events. For this we need to perform an extended maximum likelihood fit, where the signal events is one of the fit parameter. 
We will create the composite model formed by a Gaussian signal over a falling exponential background. We will have ~ 100 Gaussian-distributed signal event over  ~ 1000 exponential-distributed background. We assume that the location and width of the signal are known. The "mass" observable is distributed in range [0 - 10] and the signal has mean ~ 2 and sigma 0.3. The exponential decay of the background is 2. 

We will follow the same step as in the previous exercise: 

   *  Start creating the model using the workspace factory. Create the separate signal and background pdf and then combine together using the =RooAddPdf= class. 
   *  Use the =generate(...)= method of the =RooAbsPdf= class to generate a data set with 1000 events. 
   *  Plot the data set using =RooPlot=  in a ROOT Canvas. 
   *  Fit the generate data using the =fitTo(...)= method of the =RooAbsPdf= class
   *  Plot the resulting fit function on top of the data 
   *  Save the workspace in a file


%TWISTY{mode="div" showlink="Hint&nbsp;" hidelink="Hide&nbsp;" }%
Creating the model
   * Use the syntax of the =RooWorkspace= factory to create first the Gaussian p.d.f. function with the  variables (observables and parameters) and the Exponential
   * Create then a =RooAddPdf= using the Gaussian and the Exponential and the relative number of events. Note that we could instead of the number of events a relative fraction. In this last case we would fit only for the shape and not the normalisation. 
   * The =RooAddPdf= is created using the =SUM= operator of the factory syntax.  
%CODE{"cpp" style="background: yellow;"  }%
// create model
RooWorkspace w("w"); 
 w.factory("Exponential:bkg_pdf(x[0,10], a[-0.5,-2,-0.2])");
 w.factory("Gaussian:sig_pdf(x, mass[2], sigma[0.3])");

// create RooAddPdf
 w.factory("SUM:model(nsig[100,0,10000]*sig_pdf, nbkg[1000,0,10000]*bkg_pdf)");  // for extended model
%ENDCODE%
Retrieving objects from the workspace
   * After you have created the variable and p.d.f in the workspace, you can access their pointers, by using =RooWorkspace::var= for variables or =RooWorkspace::pdf= for p.d.f. Example: 
%CODE{"cpp"}%
// retrieving model components
RooAbsPdf * pdf = w.pdf("pdf");   // access object from workspace
RooRealVar * x = w.var("x");   // access object from workspace
%ENDCODE%
Generating a dataset
   * We generate the dataset using the =generate= function passing the observables we want to generate and the number of events. In case of an extended pdf (as the =RooAddPdf= we have created), if we don't pass the number of events, the number of events will be generated according to a Poisson distribution given the expected events (=nsig+nbkg=). 
   * Note that you can generate also binned data sets. In that case you have two options: 
         * use a =RooDataHist= and  =generateBinned=:   =RooDataHist * binData = pdf->generateBinned(*x, 1000);=. 
         * use the option =AllBinned()= which will generate a weighted data set: =RooDataSet * data = pdf->generate(*x, NumEvents(1000), AllBinned());=. 
%CODE{"cpp"}%
// generate an unbinned dataset of 1000 events
RooDataSet * data = pdf->generate( *x, 1000); 
%ENDCODE%
Plot the data:
%CODE{"cpp"}%
// create a RooPlot object and plot the data
RooPlot * plot = x->frame(); 
data->plotOn(plot); 
%ENDCODE%
Fit the data: you need to call the =RooAbsPdf::fitTo=.
%CODE{"cpp"}%
// fit the data and save its result. Use the optional =Minuit2= minimiser
RooFitResult * res = pdf->fitTo(*data, Minimizer("Minuit2","Migrad"), Save(true) );
%ENDCODE%
   * Plot the resulting fit function in the same plot. 
   * Plot also the signal and background fit components with different colour and style
%CODE{"cpp"}%
pdf->plotOn(plot); 
//draw the two separate pdf's
pdf->plotOn(plot, RooFit::Components("bkg_pdf"), RooFit::LineStyle(kDashed) );
pdf->plotOn(plot, RooFit::Components("sig_pdf"), RooFit::LineColor(kRed), RooFit::LineStyle(kDashed) );
plot->Draw();   // to show the RooPlot in the current ROOT Canvas
%ENDCODE%
Save the workspace in a file for re-using it without the need to generate the data
%CODE{"cpp"}%
w.writeToFile("GausExpModel.root",true);  // true means the file is re-created 
%ENDCODE%

%ENDTWISTY%

%TWISTY{mode="div" showlink="Solution&nbsp;" hidelink="Hide&nbsp;" }%
Note that in the solution we also create a =ModelConfig= object that we save in the workspace. This will be useful for the 
!RooStats exercises (see later the !RooStats exercise 1)
%CODE{"cpp"}%
%INCLUDE{"%ATTACHURL%/GausExpModel.C"}%
%ENDCODE%

   * Gaussian plus Exponential fit result: 

%IMAGE{"GausExpModelFit.png" type="frame" align="none" size="600"
caption="Fit of the Gaussian plus Exponential background model using !RooFit"}%

%ENDTWISTY%

---++ Exercise 3: Fit to a Combined model

In this exercise we will create a combined model based on two channels and we will perform a simultaneous fit on both channels. We will learn how to create a =RooSimultaneous= p.d.f. and a combined data set, which can be fit for one (or more) common parameters.  
   * The first channel will be similar to the one of the previous exercise. A Gaussian signal over an Exponential decay background. We assume as before that the location and width of the signal are known. The "mass" observable is distributed in range [0 - 10] and the signal has mean ~ 2 and sigma 0.3. The exponential decay of the background is 2. We assume we have 1000 background events.
   * The second channel will be similar to the first one, we will have the same signal model, but a different background model. We assume we have less and a flatter background. We have  an exponential decay of 4 and 100 background events in total. 
   * We assume the number of signal events in the first channel is 50 and in the second channel is 30. We introduce a  common rate parameter =mu= expressing the  signal strength.  

We will so the same steps as before:

   *  Create the models for the two channels
   *  Create a combined model using simultaneous pdf (class =RooSimultaneous=). 
   *  Generate a combined data set, fit and plot the result
  
%TWISTY{mode="div" showlink="Hint&nbsp;" hidelink="Hide&nbsp;" }%
   * Create first the two separate background models and a common Gaussian signal model
   * Express the number of signal events in term of the signal strength =mu=:  the number of signal events is  the product of =mu= with a fixed number (50 for the first channel and 30 for the second). 
%CODE{"cpp" style="background: yellow;"  }%
RooWorkspace w("w"); 
w.factory("Exponential:bkg1_pdf(x[0,10], a1[-0.5,-2,-0.2])");
w.factory("Exponential:bkg2_pdf(x, a2[-0.25,-2,-0.2])");
w.factory("Gaussian:sig_pdf(x, mass[2], sigma[0.3])");
// express number of events as the    
w.factory("prod:nsig1(mu[1,0,5],xsec1[50])");
w.factory("prod:nsig2(mu,xsec2[30])");

w.factory("SUM:model1(nsig1*sig_pdf, nbkg1[1000,0,10000]*bkg1_pdf)");  // for extended model
w.factory("SUM:model2(nsig2*sig_pdf, nbkg2[100,0,10000]*bkg2_pdf)");  // for extended model
%ENDCODE%
   * Make a simultaneous pdf
       * create first the category label to identify the channels (=RooCategory=)
       * use the factory operator =SIMUL= for creating the =RooSimultaneous= pdf. 
 
%CODE{"cpp" }%
// Create discrete observable to label channels
w.factory("index[channel1,channel2]");
// Create joint pdf (RooSimultaneous)
w.factory("SIMUL:jointModel(index,channel1=model1,channel2=model2)");
%ENDCODE%
   * Generate the events. We can generate directly a combined data sets from the model, with the number of events fluctuating around the number of expected events in each channel. 
   * Note that for creating a combined data set we need to add the Category variable in the data set as an additional observable
%CODE{"cpp" }%
// generate combined data set
RooDataSet * data = pdf->generate( RooArgSet(*x,*index)); 
%ENDCODE%
   * Plot the data. We make two plots representing the two channels. We need then to tell to the combined data set to select the channel we want
%CODE{"cpp" }%
RooPlot * plot1 = x->frame();
RooPlot * plot2 = x->frame();
data->plotOn(plot1,RooFit::Cut("index==index::channel1"));
data->plotOn(plot2,RooFit::Cut("index==index::channel2")); 
%ENDCODE%
   * Fit the data is as before using the =fitTo= method. 
   * Plot the resulting model on the two data sets. We need to project the simultaneous pdf on the corresponding channel. This is done as following: 
%CODE{"cpp" }%
 // fit 
 RooFitResult * r = pdf->fitTo(*data, RooFit::Save(true), RooFit::Minimizer("Minuit2","Migrad"));
 r->Print();

 // plotting (draw the two separate pdf's )
 pdf->plotOn(plot1,RooFit::ProjWData(*data),RooFit::Slice(*w.cat("index"),"channel1"));
 pdf->plotOn(plot2,RooFit::ProjWData(*data),RooFit::Slice(*w.cat("index"),"channel2"));

 // draw the two plots in two separate pads:
 c1 = new TCanvas();
 c1->Divide(1,2);
 c1->cd(1); plot1->Draw(); 
 c1->cd(2); plot2->Draw();
%ENDCODE%

%ENDTWISTY%

Code solution: 
%TWISTY{mode="div" showlink="Solution&nbsp;" hidelink="Hide&nbsp;" }%
%CODE{"cpp" }%
%INCLUDE{"%ATTACHURL%/CombinedModel.C"}%
%ENDCODE%

   * Combined Model fit result: 

%IMAGE{"CombinedModelFit.png" type="frame" align="none" size="600"
caption="Fit of the Combined model of two channels using !RooFit"}%


%ENDTWISTY%


---+ !RooStats Exercises  

In these exercises we will learn to use the !RooStats calculators to perform a statistical analysis on a model. We will use two types of models: 
   * Observable-based analysis. This is the model we have used in the !RooFit exercise 2. It is composed of 50 Gaussian-distributed signal event over ~ 1000 exponential-distributed background. The "mass" observable is distributed in range [0 - 10] and the signal has mean ~ 2 and sigma 0.5. We assume that the location and width of the signal are known. We can re-use the macro from the !RooFit exercise to create the model. 
   * Poisson model: let's assume you observe n events in data while expecting b background event. We  consider some systematic uncertainties in the background model value, b. We express this uncertainty as Gaussian. 
Optionally we can use a log-normal or a  a Poisson/Gamma distribution for expressing this uncertainty. We could also extend the model adding other systematic uncertainties. 
   
For these model we compute:

   * 68% CL 2-sided confidence interval  from the (profiled-) likelihood ratio. 
   * 68% credible interval using the !MCMCCalculator or the Bayesian Calculator (based on numerical integration). On the Poisson model we will compute the 95% Bayesian upper limit. 
 
For running the !RooStats calculators we read the workspace from the ROOT file. All the !Roostats calculators take as input a data set and a !ModelConfig object. It is therefore convenient to save the workspace (with model and data) in a ROOT file and then read it afterwards for running the calculators.

If you have problem creating the code for running the calculators,  you can run the provided solution or just the standard !Roostats tutorials (in $ROOTSYS/tutorials/roostats)

Tutorials you could use are:

   * =StandardProfileLikelihoodDemo.C=
   * =StandardBayesianNumericalDemo.C=
   * =StandardBayesianMCMCDemo.C=
 

They  can be run as following from the Root prompt by passing the workspace file name (e.g. _GausExpModel.root_ ), workspace name, model config name and data set name: 
 <verbatim style="background-color: rgb(255,250,205);"> 
root [0] .L $ROOTSYS/tutorials/roostats/StandardProfileLikelihoodDemo.C
root [1] StandardProfileLikelihoodDemo("SPlusBExpoModel.root", "w", "ModelConfig", "data" )
</verbatim>
 
 
Other tutorials go into more details to create various models and run the calculators. See them all at http://root.cern.ch/root/html/tutorials/roostats/index.html.

---++ Exercise 1: Create a =ModelConfig= from the Gaussian plus Exponential Model used in the !RooFit Exercise 2. 

We show first how to create the =ModelConfig= object which we need to import into the workspace in order to be able to run later the !RooStats calculators. 
We need to specify in the =ModelConfig=: 
   *  the p.d.f of the model 
   *  the parameters of interest (in this case the number of signal events)
   *  the nuisance parameters (the number of background events and the exponential shape parameter )
   *  the observables (in this case x, the reconstructed mass)

We will see later, that if we have a model where some external knowledge on a parameter is introduced, we would need to define the model global observables. 
 
Below is how we create a =ModelConfig= for the Gaussian signal plus Exponential background model:
%CODE{"cpp" }%
RooStats::ModelConfig mc("ModelConfig",&w);
mc.SetPdf(*pdf);
mc.SetParametersOfInterest(*w.var("nsig"));
mc.SetObservables(*w.var("x"));
// define set of nuisance parameters
w.defineSet("nuisParams","a,nbkg");

mc.SetNuisanceParameters(*w.set("nuisParams"));

// import the ModelConfig in the workspace
w.import(mc); 

// write workspace in the file 
w.writeToFile("GausExpModel.root"); 
%ENDCODE%
These lines must be included at the end of the macro we used for exercise 2 (they are already present in the solution). 


---++ Exercise 2:  Profile Likelihood Calculator

In this exercise we will see how to run the !RooStats Profile Likelihood calculator on the Gaussian signal plus Exponential background model. We will create a ROOT macro, doing the following: 
   * Read the workspace from the file and get a pointer to the =ModelConfig= and data set objects. 
   * Create  the Profile likelihood calculator class
   * Run the calculator by getting the interval at the corresponding desired confidence level
   * Plot the interval and the profile likelihood function


%TWISTY{mode="div" showlink="Hint&nbsp;" hidelink="Hide&nbsp;" }%
Here is the code to read the workspace from the file. Remember to add at the beginning of the macro  =using namespace RooStats;=.
%CODE{"cpp" }%
using namespace RooStats;

void SimplePL(  
   const char* infile =  "GausExpModel.root", 
   const char* workspaceName = "w",
   const char* modelConfigName = "ModelConfig",
   const char* dataName = "data" )
{

  // open input file 
  TFile *file = TFile::Open(infile);
  if (!file) return;

  // get the workspace out of the file
  RooWorkspace* w = (RooWorkspace*) file->Get(workspaceName);


  // get the modelConfig out of the file
  RooStats::ModelConfig* mc = (RooStats::ModelConfig*) w->obj(modelConfigName);

  // get the modelConfig out of the file
  RooAbsData* data = w->data(dataName);

%ENDCODE%
We create now the =ProfileLikelihoodCalculator class passing the =ModelConfig= object and a data set object (=RooAbsData=): 
%CODE{"cpp" }%
  ProfileLikelihoodCalculator pl(*data,*mc);
%ENDCODE%
Now the calculator has all information and can compute the interval. We need to set first the desired confidence level and then call =GetInterval(…)=:
%CODE{"cpp" }%
pl.SetConfidenceLevel(0.683); // 68% interval
LikelihoodInterval* interval = pl.GetInterval();
%ENDCODE%
We can then retrieve the actual interval lower/upper value for the desired parameter (e.g. the parameter of interest):
%CODE{"cpp" }%
// find the iterval on the first Parameter of Interest
RooRealVar* firstPOI = (RooRealVar*) mc->GetParametersOfInterest()->first();

double lowerLimit = interval->LowerLimit(*firstPOI);
double upperLimit = interval->UpperLimit(*firstPOI);


cout << "\n68% interval on " <<firstPOI->GetName()<<" is : ["<<
    lowerLimit << ", "<<
    upperLimit <<"] "<<endl;
%ENDCODE%
At the end we can plot the interval and the profiled likelihood function in a ROOT Canvas using the =LikelihoodIntervalPlot= class: 
%CODE{"cpp" }%
LikelihoodIntervalPlot * plot = new LikelihoodIntervalPlot(interval);
plot->SetRange(0,150);  // change range to have interval visible (default range is variable range)
plot->Draw(); 
%ENDCODE%
Note that for some complex model, which require time to evaluate, it can be convenient to use less points and also use the "tf1" option
%CODE{"cpp" }%
plot->SetPoints(50);
plot->Draw("tf1"); 
%ENDCODE%
%ENDTWISTY%

%TWISTY{mode="div" showlink="Solution&nbsp;" hidelink="Hide&nbsp;" }%
%CODE{"cpp" }%
%INCLUDE{"%ATTACHURL%/SimplePL.C"}%
%ENDCODE%

   * Profile Likelihood result on Gaussian signal plus Exponential background model:  

<verbatim>
68% interval on nsig is : [60.3382, 97.0782] 
</verbatim>
%IMAGE{"PLInterval.png" type="frame" align="none" size="600"
caption="Profile Likelihood scan for the number of signal events"}%


%ENDTWISTY%

---++ Exercise 3a:  Bayesian MCMC Calculator
  
In this exercise we will learn to use the !RooStats Bayesian calculators. We have the choice between using the MCMC calculator or the numerical one. We will start using the MCMC and optionally in the next exercise we will show how to use the numerical Bayesian calculator. We will use the same model used before for the Profile Likelihood calculator.
What we need to do is very similar to the previous exercise: 
   * Read the model from the file as in the previous exercise
   * Create the =MCMCCalculator= class 
   * Configure the calculator setting the proposal function, number of iterations, etc..
   * Call the =GetInterval= function to compute the desired interval. The returned =MCMCInterval= class will contained the resulting Markov-Chain. 
   * Plot the result using the =MCMCIntervalPlot= class

%TWISTY{mode="div" showlink="Hint&nbsp;" hidelink="Hide&nbsp;" }%
Assuming we have already the code for reading a model from the previous exercise, we create the =MCMCCalculator class passing the =ModelConfig= object and a data set object (=RooAbsData=) exactly as we did for the =ProfileLikelihoodCalculator=: 
%CODE{"cpp" }%
MCMCCalculator mcmc(*data,*mc);
mcmc.SetConfidenceLevel(0.683); // 683% interval
%ENDCODE%
Configure the calculator. 
   * We set the proposal function and the number of iterations. As proposal we use the =SequentialProposal= function. This function samples each parameter independently using a gaussian pdf with a sigma given by a fraction of the parameter range. We use in this case a value of 0.1 for the fraction. It is important to check that the acceptance of the proposal is not too low. In that case we are wasting iterations. 
%CODE{"cpp" }%
// Define proposal function. This proposal function seems fairly robust
SequentialProposal sp(0.1);
mcmc.SetProposalFunction(sp);
%ENDCODE%
Set then the number of iterations. We use a not too large number of iterations to avoid waiting too long in this tutorial example. 
%CODE{"cpp" }%
// set number of iterations and initial burning steps 
mcmc.SetNumIters(100000);         // Metropolis-Hastings algorithm iterations
mcmc.SetNumBurnInSteps(1000);       // first N steps to be ignored as burn-in
%ENDCODE%
Define the interval we want to compute. We will compute a central interval but we can compute 
   * a shortest interval (default and the only one defined in a multi-dimensional posterior) 
   * central interval (equal probability on both the left and right side)  
   * one sided (lower or upper limits) 
%CODE{"cpp" }%
 mcmc.SetLeftSideTailFraction(0.5); // 0.5 for central Bayesian interval  and 0 for upper limit
%ENDCODE%
Run the calculator (obtain the Markov chain and retrieve the actual interval lower/upper value for the desired parameter (e.g. the parameter of interest) by integrating the 
posterior function: 
%CODE{"cpp" }%
MCMCInterval* interval = mcmc.GetInterval();

// print out the iterval on the first Parameter of Interest
RooRealVar* firstPOI = (RooRealVar*) mc->GetParametersOfInterest()->first();
cout << "\n68% interval on " <<firstPOI->GetName()<<" is : ["<<
             interval->LowerLimit(*firstPOI) << ", "<<
             interval->UpperLimit(*firstPOI) <<"] "<< endl;
%ENDCODE%
At the end we plot the interval and the posterior function in a ROOT Canvas using the =MCMCIntervalPlot= class: 
%CODE{"cpp" }%
// make a plot of posterior function
new TCanvas("IntervalPlot");
MCMCIntervalPlot plot(*interval);
plot.Draw();
%ENDCODE%
%ENDTWISTY%


%TWISTY{mode="div" showlink="Solution&nbsp;" hidelink="Hide&nbsp;" }%
%CODE{"cpp" }%
%INCLUDE{"%ATTACHURL%/SimpleMCMC.C"}%
%ENDCODE%

   * MCMC Bayesian result on Gaussian signal plus Exponential background model:  

<verbatim>
Metropolis-Hastings progress: ....................................................................................................
[#1] INFO:Eval -- Proposal acceptance rate: 10.469%
[#1] INFO:Eval -- Number of steps in chain: 10469
68% interval on nsig is : [61.245, 98.2044] 
</verbatim>
%IMAGE{"MCMCInterval.png" type="frame" align="none" size="600"
caption="Posterior function for the number of signal events obtained using the =MCMCCalculator="}%
%ENDTWISTY%

---++ Exercise 3b:  Bayesian Numerical Calculator

We can try now to use also the numerical Bayesian calculator on the same model. As before we create the class and configure it setting for example the integration type or the number of iterations. 
Since the numerical integration can be tricky, it is better in this case to set before a reasonable range from the nuisance parameters, which will be integrated to obtain the marginalised posterior function.
A sensible choice is to use for example an interval using 10-sigma around the best fit values for the nuisance parameters. 

Note that we have not defined a prior distribution for the parameter of interest (the number of signal events).  If a prior is not defined in the model, it is assumed that the prior distribution for the parameter of interest is the uniform distribution in the parameter range. 

%TWISTY{mode="div" showlink="Hint&nbsp;" hidelink="Hide&nbsp;" }%
Again we create the =BayesianCalculator class passing the =ModelConfig= object and a data set object, =RooAbsData=:
%CODE{"cpp" }%
BayesianCalculator bayesianCalc(*data,*mc);
%ENDCODE%

%CODE{"cpp"}%
// define reasanable range for nuisance parameters ( +/- 10 sigma around fitted values) to facilitate the integral
RooArgList nuisPar(*mc->GetNuisanceParameters()); 
for (int i = 0; i < nuisPar.getSize(); ++i) { 
   RooRealVar & par = (RooRealVar&) nuisPar[i];
   par.setRange(par.getVal()-10*par.getError(), par.getVal()+10*par.getError());
}
%ENDCODE%

We can then configure the calculator:
%CODE{"cpp"}%
// set the type of interval (not really needed for central which is the default)
bayesianCalc.SetLeftSideTailFraction(0.5); // for central interval
  //bayesianCalc.SetShortestInterval(); // for shortest interval
bayesianCalc.SetConfidenceLevel(0.683); // 68% interval

// set the integration type (default is "ADAPTIVE")
// possible alternative values are  "VEGAS" , "MISER", or "PLAIN"  but they require  libMathMore 
// or  "TOYMC" (toy MC integration, work when nuisances exist and they have a constraints pdf)
bayesianCalc.SetIntegrationType("TOYMC"); 

// limit the number of iterations to speed-up the integral 
bayesianCalc.SetNumIters(1000);

// compute interval by scanning the posterior function
// it is done by default when computing shortest intervals
bayesianCalc.SetScanOfPosterior(50);
%ENDCODE%
We can then run the calculator and retrieve the result. The plot of the posterior can be retrieved directly from the class as a =RooPlot= object: 
%CODE{"cpp" }%
SimpleInterval* interval = bayesianCalc.GetInterval();
if (!interval) { 
     cout << "Error computing Bayesian interval - exit " << endl;
     return;
}

double lowerLimit = interval->LowerLimit();
double upperLimit = interval->UpperLimit();

RooRealVar* firstPOI = (RooRealVar*) mc->GetParametersOfInterest()->first();
cout << "\n68% interval on " <<firstPOI->GetName()<<" is : ["<<
             lowerLimit << ", "<<
             upperLimit <<"] "<<endl;

// draw plot of posterior function
RooPlot * plot = bayesianCalc.GetPosteriorPlot();
if (plot) plot->Draw();  
%ENDCODE%
%ENDTWISTY%

%TWISTY{mode="div" showlink="Solution&nbsp;" hidelink="Hide&nbsp;" }%
%CODE{"cpp" }%
%INCLUDE{"%ATTACHURL%/SimpleBayes.C"}%
%ENDCODE%

   * Numerical Bayesian result on Gaussian signal plus Exponential background model:  

<verbatim>
[#1] INFO:Eval -- BayesianCalculator::GetInterval - found a valid interval : [61.01 , 97.9247 ]

68% interval on nsig is : [61.01, 97.9247] 
</verbatim>
%IMAGE{"BayesInterval.png" type="frame" align="none" size="600"
caption="Posterior function for the number of signal events obtained using the =BayesianCalculator="}%
%ENDTWISTY%

---++ Exercise 4:  Create a Poisson Counting Model

We create now a counting model based on Poisson Statistics. Let's suppose we observe =nobs= events when we expect =nexp=, where =nexp = s+b= (=s= is the number of signal events and =b= is the number of background events. The expected distribution for =nexp= is a Poisson : =Poisson( nobs | s+b)=.
=s= is the parameter we want to estimate or set a limit (the parameter of interest), while =b= is the nuisance parameter. =b= is not known exactly, but  has uncertainty =sigmab=. The nominal value of =b= is = b0=. To express this uncertainty we add in the model a Gaussian constraint. We can interpret this term as having an additional measurement =b0= with an uncertainty =sigmab=: i.e. we have a likelihood for that measurement ==Gaussian( b0 | b, sigma)=. In case of Bayesian statistics we can interpret the constraint as a prior knowledge on the parameter b. 

   * Make first the !RooFit workspace using its factory syntax. Let's assume =nobs=3=, =b0=1=, =sigmab=0.2=. 
   * Create the =ModelConfig= object and import in the workspace. We need to add in the =ModelConfig= also =b0= as a "global observable=. The reason is that =b0= needs to be treated as an auxiliary observable in case of frequentist statistics and varied when tossing pseudo-experiments. 

%TWISTY{mode="div" showlink="Hint&nbsp;" hidelink="Hide&nbsp;" }%
We make first the !RooFit workspace using its factory syntax. We have =nobs=3=, =b0=1=, =sigmab=0.2=. 
%CODE{"cpp"}%
int nobs = 3;   // number of observed events
double b0 = 1; // number of background events
double sigmab = 0.2;   // relative uncertainty in b

RooWorkspace w("w");

// make Poisson model * Gaussian constraint
w.factory("sum:nexp(s[3,0,15],b[1,0,10])");
// Poisson of (n | s+b)
w.factory("Poisson:pdf(nobs[0,50],nexp)");
// Gaussian constraint to express uncertainty in b
w.factory("Gaussian:constraint(b0[1,0,10],b,sigmab[0.2])");
// the total model p.d.f will be the product of the two
w.factory("PROD:model(pdf,constraint)");
%ENDCODE%
We need also to define =b0=, the global observable,  as a constant variable (this is needed when we fit the model). However, =b0= often needs to have a range. 
w.var("b0")->setConstant(true); 

After creating the model p.d.f we create the =ModelConfig= object and we set =b0= as the "global observable=
%CODE{"cpp"}%
ModelConfig mc("ModelConfig",&w);
mc.SetPdf(*w.pdf("model"));
mc.SetParametersOfInterest(*w.var("s"));
mc.SetObservables(*w.var("nobs"));
mc.SetNuisanceParameters(*w.var("b"));
// need now to set the global observable
mc.SetGlobalObservables(*w.var("b0"));
// this is needed for the hypothesis tests
mc.SetSnapshot(*w.var("s"));

// import model in the workspace 
w.import(mc);
%ENDCODE%

We generate then an hypothetical observed data set. Since we have a counting model, the data set will contain only one event and the observable =nobs= will have our desired value (i.e. 3).
Remember to import the data set in the workspace and then save the workspace in a ROOT file. 
%CODE{"cpp"}%
// make data set with the namber of observed events
RooDataSet data("data","", *w.var("nobs"));
w.var("nobs")->setVal(3);
data.add(*w.var("nobs") );
// import data set in workspace and save it in a file
w.import(data);
w.writeToFile("ComtingModel.root", true);
%ENDCODE%
%ENDTWISTY%


%TWISTY{mode="div" showlink="Code Solution&nbsp;" hidelink="Hide&nbsp;" }%
%CODE{"cpp" }%
%INCLUDE{"%ATTACHURL%/CountingModel.C"}%
%ENDCODE%
%ENDTWISTY%

%ICON{hand}% Suppose you want to define the uncertainty in the background as log-normal. How do you do this ? 

%ICON{hand}% Suppose instead the background is estimated from another counting experiments. How do you model this ? (This model is often call on/off model)

You can look at this past !RooStats exercise (RooStatsTutorialsAugust2012) for these solutions. 

---++ Exercise 5:  Compute a 95% Upper Limit with the Counting Model

Using the previously defined Counting model we can compute a 95% upper limit on the signal parameter =s=. We could use : 
   * The Bayesian Calculator or the MCMC Calculator
   * The Profile Likelihood Calculator

Suppose we use the Bayesian calculator: 
%TWISTY{mode="div" showlink="Hint&nbsp;" hidelink="Hide&nbsp;" }% 
In this case we can re-use the macro =SimpleBayes.C=. We just need to change these lines, setting the tip elf interval and give as input the right ROOT file containing the workspace. 
%CODE{"cpp"}%
//bayesianCalc.SetConfidenceLevel(0.68); // 68% interval
bayesianCalc.SetConfidenceLevel(0.95); // 95% interval

// set the type of interval 
//bayesianCalc.SetLeftSideTailFraction(0.5); // for central interval
bayesianCalc.SetLeftSideTailFraction(0.); // for upper limit
%ENDCODE%
%ENDTWISTY%

This is the result we will obtain on the Bayesian calculator: 
%TWISTY{mode="div" showlink="Solution&nbsp;" hidelink="Hide&nbsp;" }%

   * Numerical Bayesian 95% Upper limit result on Counting Poisson model

<verbatim>
[#1] INFO:Eval -- BayesianCalculator::GetInterval - found a valid interval : [0 , 6.80149 ]

95% interval on s is : [0, 6.80149] 
</verbatim>
%IMAGE{"BayesUpperLimit.png" type="frame" align="none" size="600"
caption="Posterior function for the counting model (nobs = 3, b = 1)  obtained using the =BayesianCalculator="}%
%ENDTWISTY%

%ICON{hand}% Try to use the Profile Likelihood Calculator. Which result do you obtain ? 

%ICON{hand}% We can try to use a different prior on the signal instead of the default uniform in the case of the Bayesian calculator. For example we can try the Jeffrey reference prior, =P(s) = 1./sqrt(s)=. 
%TWISTY{mode="div" showlink="Hint&nbsp;" hidelink="Hide&nbsp;" }%
For using a different signal prior than the uniform, we need to define the prior pdf in the workspace and add it also in the =ModelConfig= so the Bayesian !RooStats are aware of its existence: 
%CODE{"cpp"}%
// define as a pdf based on an expression
w.factory("EXPR::prior_s('1./sqrt(s)',s)");
// define prior in ModelConfig
mc.SetPriorPdf(*w.pdf("prior_s"));
%ENDCODE%
%ENDTWISTY%

---++ Exercise 6:  Compute the significance (Hypothesis Test)

The aim of this exercise is to compute the significance of observing a signal in the Gaussian plus Exponential model. It is better to re-generate the model using a lower value for the number of signal events (e.g. 50), in order not to have a too high significance, which will then be difficult to estimate if we use pseudo-experiments. 
To estimate the significance, we need to perform an hypothesis test. We want to disprove the null model, i.e the background only model  against the alternate model, the background plus the signal. 
In !RooStats, we do this by defining two two =ModelConfig= objects, one for the null model (the background only model in this case) and one for the alternate model (the signal plus background). 

We have defined in the workspace saved in the ROOT file only the signal plus background model. We can create the background model from the signal plus background model by setting the value of =nsig=0=. We do this by cloning the S+B model and by defining a snapshot in the =ModelConfig= for =nsig= with a different value: 
%CODE{"cpp"}%
ModelConfig*  sbModel = (RooStats::ModelConfig*) w->obj(modelConfigName);
RooRealVar* poi = (RooRealVar*) sbModel->GetParametersOfInterest()->first();
// define the S+B snapshot (this is used for computing the expected significance)
poi->setVal(50);
sbModel->SetSnapshot(*poi);
// create B only model
ModelConfig * bModel = (ModelConfig*) sbModel->Clone();
bModel->SetName("B_only_model");      
poi->setVal(0);
bModel->SetSnapshot( *poi  );
%ENDCODE%

Using the given models plus the observed data set, we can create a !RooStats hypothesis test calculator to compute the significance. We have the choice between 
   * Frequentist calculator 
   * Hybrid Calculator
   * Asymptotic calculator 

The first two require generate pseudo-experiment, while the Asymptotic calculator, requires only a fit to the two models. For the Frequentist and the Hybrid calculators  we need to choose also the test statistics. We have various choices in !RooStats.  
Since it is faster, we will use for the exercise the Asymptotic calculator. The Asymptotic calculator it is based on the Profile Likelihood test statistics, the one used at LHC (see slides for its definition). For testing the significance, we must use the one-side test statistic (we need to configure this explicitly in the class). 
Summarising,  we will do: 
   * create the =AsymptoticCalculator= class using the two models and the data set; 
   * run the test of hypothesis using the =GetHypoTest= function. 
   * Look at the result obtained as a =HypoTestResult= object
%TWISTY{mode="div" showlink="Hint&nbsp;" hidelink="Hide&nbsp;" }% 
%CODE{"cpp"}%
AsymptoticCalculator   ac(*data, *sbModel, *bModel);
ac.SetOneSidedDiscovery(true);  // for one-side discovery test

// this run the hypothesis test and print the result
HypoTestResult * result = ac.GetHypoTest();
result->Print();
%ENDCODE%
%ENDTWISTY%

%TWISTY{mode="div" showlink="Asymptotic calculator Solution&nbsp;" hidelink="Hide&nbsp;" }%
%CODE{"cpp" }%
%INCLUDE{"%ATTACHURL%/SimpleHypoTest.C"}%
%ENDCODE%

   * Result of the significance test on the Gaussian Signal plus Background model (nsig=50, nbckg=1000): 

<verbatim>
Results HypoTestAsymptotic_result: 
 - Null p-value = 0.00118838
 - Significance = 3.0386
 - CL_b: 0.00118838
 - CL_s+b: 0.502432
 - CL_s: 422.787
</verbatim>
%ENDTWISTY%

%ICON{hand}% Run the Frequentist calculator
%TWISTY{mode="div" showlink="Hint&nbsp;" hidelink="Hide&nbsp;" }% 
For the Frequentist calculator we need to : 
   * set the test statistic we want to use 
   * set the number of toys for the null and alternate model
%CODE{"cpp"}%
FrequentistCalculator   fc(*data, *sbModel, *bModel);
fc.SetToys(2000,500);    // 2000 for null (B) and 500 for alt (S+B) 

// create the test statistics
ProfileLikelihoodTestStat profll(*sbModel->GetPdf());
// use one-sided profile likelihood
profll.SetOneSidedDiscovery(true);

// configure  ToyMCSampler and set the test statistics
ToyMCSampler *toymcs = (ToyMCSampler*)fc.GetTestStatSampler();
toymcs->SetTestStatistic(&profll);
  
if (!sbModel->GetPdf()->canBeExtended())
   toymcs->SetNEventsPerToy(1);
 
%ENDCODE%
We run now the hypothesis test
%CODE{"cpp"}%
HypoTestResult * fqResult = fc.GetHypoTest();
fqResult->Print();
%ENDCODE%
And we can also plot the test statistic distributions obtained from the pseudo-experiments for the two models. We use the =HypTestPlot= class for this. 
%CODE{"cpp"}%
// plot test statistic distributions
HypoTestPlot * plot = new HypoTestPlot(*fqResult);
plot->SetLogYaxis();
plot->Draw();
%ENDCODE%
%ENDTWISTY% 

%TWISTY{mode="div" showlink="Frequentist calculator Solution&nbsp;" hidelink="Hide&nbsp;" }%
Use the macro [[SimpleHypoTest.C]] remove the =return= statement in the middle and run it again

   * Result of the significance test on the Gaussian Signal plus Background model (nsig=50, nbckg=1000) using the Frequentist calculator: 

<verbatim>
Results HypoTestCalculator_result: 
 - Null p-value = 0.001 +/- 0.000706753
 - Significance = 3.09023 +/- 0.2099 sigma
 - Number of Alt toys: 500
 - Number of Null toys: 2000
 - Test statistic evaluated on data: 4.61654
 - CL_b: 0.001 +/- 0.000706753
 - CL_s+b: 0.496 +/- 0.02236
 - CL_s: 496 +/- 351.262
</verbatim>
%IMAGE{"FrequentistHTResult.png" type="frame" align="none" size="600"
caption="Test statistics distribution for the B model (red) and S+B model obtained with the Frequentist calculator"}%
%ENDTWISTY%

---++ Exercise 6b: Compute significance (p-value) as function of the signal mass  

As an optional exercise to lear more about computing significance in !RooStats,  we  will study the significance (or equivalent the p-value) we obtain in our Gaussian signal plus exponential background model as function of the signal mass hypothesis. 
For doing this we perform the test of hypothesis, using the !AsymptoticCalculator, for several mass points. We can then plot the obtained p-value for the null hypothesis (=p0=). 
We will also estimate the expected significance for our given S+B model as function of its mass. The expected significance can be obtained using a dedicated function in the =AsymptoticCalculator=: 
=AsymptoticCalculator::GetExpectedPValues= using as input the observed p-values for the null and the alternate models. See the Asymptotic formulae paper for the details. 

%TWISTY{mode="div" showlink="Solution&nbsp;" hidelink="Hide&nbsp;" }%
%CODE{"cpp" }%
%INCLUDE{"%ATTACHURL%/p0Plot.C"}%
%ENDCODE%
%IMAGE{"p0Plot.png" type="frame" align="none" size="600"
caption="p value obtained as function of the signal mass in the Gaussian plus exponential background model"}%
%ENDTWISTY%


---++  Exercise 7: Compute Limits using Hypothesis Test Inversion (=CLs= Limits). 

In our last exercise we will compute  an upper limit using the frequentist hypothesis test inversion method. We need to perform the hypothesis test for different parameter of interest points and compute the 
corresponding p-values. Since we are interesting in computing a limit, the test null hypothesis, that we want to disprove, is the in this case the S+B model, while the alternate hypothesis is the B only model.   
It is important to remember this, when we construct the hypothesis test calculator. 

For doing the inversion =RooStats= provides an helper class, =HypoTestInverter=, which is constructed using one of the !HypoTest calculator (Frequentist, Hybrid or Asymptotic).
To compute the limit we need to do: 
   *  Create/read a workspace with S+B and B only model. If the B model does not exist we can create as in the previous exercise (setting the poi to zero). 
   *  Create an !HypoTest calculator but using as null the S+B model and as alt the B only model
   *  Configure the calculator if needed (set test statistics, number of toys, etc..)
   *  Create the  =HypoTestInverter= class and configure it  (e.g. to set the number and range of points to scan, to use =CLs= for computing the limit, etc…)
   *  Run the Inverter using =GetInterval=  which returns an =HypoTestInverterResult= object. 
   *  Look at the result (e.g. limit, expected limits and bands)
   *  Plot the result of the scan and the test statistic distribution obtained at every point. 

%TWISTY{mode="div" showlink="Hint&nbsp;" hidelink="Hide&nbsp;" }% 
Suppose we use the =AsymptotiCalculator=. If we use the Frequentist is the same, just we need to set also the test statistics and number of toys
%CODE{"cpp"}%
 // asymptotic calculator
AsymptoticCalculator ac(*data, *bModel, *sbModel);
ac.SetOneSided(true);  // for one-side tests (limits)
AsymptoticCalculator::SetPrintLevel(-1);  // to avoid to much verbosity
%ENDCODE%
We create now the Inverter class and configure it for CLs 
%CODE{"cpp"}%
// create hypotest inverter 
// passing the desired calculator 
HypoTestInverter calc(ac);

// set confidence level (e.g. 95% upper limits)
calc.SetConfidenceLevel(0.95);
  
// for CLS
calc.UseCLs(true);
%ENDCODE%

We run now the Inverter: 
%CODE{"cpp"}%
int npoints = 10;  // number of points to scan
// min and max (better to choose smaller intervals)
double poimin = poi->getMin();
double poimax = poi->getMax();
//poimin = 0; poimax=10;

std::cout << "Doing a fixed scan  in interval : " << poimin << " , " << poimax << std::endl;
calc.SetFixedScan(npoints,poimin,poimax);

// run inverter  
HypoTestInverterResult * r = calc.GetInterval();
%ENDCODE%

We can then retrieve from the obtained result object the observed limit value and also the  expected limit value and bands  (median +/- 1 sigma bands):
%CODE{"cpp"}%
double upperLimit = r->UpperLimit();
// estimate of the error on the upper limit
double ulError = r->UpperLimitEstimatedError();

std::cout << "The computed upper limit is: " << upperLimit << " +/- " << ulError << std::endl;
  
// compute expected limit and bands 
std::cout << "Expected upper limits, using the B (alternate) model : " << std::endl;
std::cout << " expected limit (median) " << r->GetExpectedUpperLimit(0) << std::endl;
std::cout << " expected limit (-1 sig) " << r->GetExpectedUpperLimit(-1) << std::endl;
std::cout << " expected limit (+1 sig) " << r->GetExpectedUpperLimit(1) << std::endl;
%ENDCODE%  

Plot now the result of the scan using the =HypoTestInverterPlot= class: 

%CODE{"cpp"}%
 HypoTestInverterPlot *plot = new HypoTestInverterPlot("HTI_Result_Plot","Feldman-Cousins Interval",r);

// plot in a new canvas 
TCanvas * c1 = new TCanvas("HypoTestInverter Scan"); 
c1->SetLogy(false);

plot->Draw("CLb 2CL");  // plot also CLb and CLs+b 
 //plot->Draw("OBS");  // plot only observed p-value
%ENDCODE%

If we were using toys (e.g. FrequentistCalculator), we could plot in a different ROOT Canvas the test statistics distributions. 
%TWISTY{mode="div" showlink="Show Code&nbsp;" hidelink="Hide&nbsp;" }% 
%CODE{"cpp"}%
// plot test statistics distributions for the two hypothesis
// when distribution is generated (case of FrequentistCalculators)
const int n = r->ArraySize();
if (n> 0 &&  r->GetResult(0)->GetNullDistribution() ) {  
      TCanvas * c2 = new TCanvas("Test Statistic Distributions","",2);
      if (n > 1) {
         int ny = TMath::CeilNint( sqrt(n) );
         int nx = TMath::CeilNint(double(n)/ny);
         c2->Divide( nx,ny);
      }
      for (int i=0; i<n; i++) {
         if (n > 1) c2->cd(i+1);
         SamplingDistPlot * pl = plot->MakeTestStatPlot(i);
         pl->SetLogYaxis(true);
         pl->Draw();
      }
   }
%ENDCODE%
%ENDTWISTY%

%ENDTWISTY%
 
%TWISTY{mode="div" showlink="Solution&nbsp;" hidelink="Hide&nbsp;" }%
%CODE{"cpp" }%
%INCLUDE{"%ATTACHURL%/SimpleHypoTestInv.C"}%
%ENDCODE%
%IMAGE{"AsymptoticCls.png" type="frame" align="none" size="600"
caption="CLS limit for the Poisson model obtained using CLs"}%
%ENDTWISTY%


---++ Exercise 7b: Using the =StandardHypoTestInvDemo.C= tutorial

This is a macro in $ROOTSYS/tutorials/roostats which can run the hypothesis test inversion with several possible options. We can find the tutorial code also 
[[http://root.cern.ch/root/html534/tutorials/roostats/StandardHypoTestInvDemo.C.html][here]].
Suppose we want to compute the 95% =CLs= limit on the parametric unbinned model (Gaussian plus Exponential background), that we have used in the significance exercise: 

<verbatim> 
root [0] .L $ROOTSYS/tutorials/roostats/StandardHypoTestInvDemo.C+
root [1] StandardHypoTestInvDemo("GausExpModel.root", "w", "ModelConfig", "", "data", 2, 3, true, 20, 0, 150 )
</verbatim>
The macro will print out the observed limit, expected limit and +/1,2 sigma bands/ It will also produce a p-value (=CLs= in this case)  scan plot.
 
Next we can run the Frequentist calculation (it will take considerably more time, but we use only 500 toys):
 <verbatim> 
root [1] StandardHypoTestInvDemo("GausExpModel.root", "w", "ModelConfig", "", "data", 0, 3, true, 10, 0, 100, 500, false )
</verbatim>
Since we are using toys to estimate the test statistics distribution, the macro will produce the sampling distribution plots for each scanned point. 

Note that, since we are running on the number counting model, we need to set to true the last option
<verbatim> 
root [1] StandardHypoTestInvDemo("CountingModel.root", "w", "ModelConfig", "", "data", 0, 3, true, 10, 0, 10, 500, true )
</verbatim>
For reference, here is the list of input arguments to the macro: 
<verbatim>
StandardHypoTestInvDemo(const char * infile = 0,
                        const char * wsName = "combined",
                        const char * modelSBName = "ModelConfig",
                        const char * modelBName = "",
                        const char * dataName = "obsData",                 
                        int calculatorType = 0,
                        int testStatType = 0, 
                        bool useCLs = true ,  
                        int npoints = 6,   
                        double poimin = 0,  
                        double poimax = 5, 
                        int ntoys=1000,
                        bool useNumberCounting = false,
                        const char * nuisPriorName = 0){
</verbatim>
Where we have for the calculatorType: 
   * type = 0 Freq calculator 
   * type = 1 Hybrid calculator
   * type = 2 Asymptotic calculator  
 

   *  testStatType = 0 LEP
   *                       = 1 Tevatron 
   *                      = 2 Profile Likelihood
   *                       = 3 Profile Likelihood one sided (i.e. = 0 if mu < mu_hat)

   * useCLs  :       scan for CLs (otherwise for CLs+b)    

   * npoints:        number of points to scan , for autoscan set npoints = -1 

   * poimin,poimax:  min/max value to scan in case of fixed scans (if min >  max, try to find automatically)                           

   * ntoys:         number of toys to use 

   * useNumberCounting:  set to true when using number counting events 

   * nuisPriorName:   name of prior for the nnuisance. This is often expressed as constraint term in the global model
Attachments
Topic attachments
I	Attachment	History	Action	Size	Date	Who	Comment
png	AsymptoticCls.png	r1	manage	19.6 K	2013-06-04 - 23:16	LorenzoMoneta	Asymptotic CLs result on Counting model
png	BayesInterval.png	r1	manage	10.2 K	2013-06-03 - 23:05	LorenzoMoneta	Bayesian Interval result
png	BayesUpperLimit.png	r1	manage	9.4 K	2013-06-04 - 10:53	LorenzoMoneta	Bayesian 95% Upper limit result
c	CombinedModel.C	r2 r1	manage	3.3 K	2013-06-14 - 11:25	LorenzoMoneta	Combined Model ROOT macro
png	CombinedModelFit.png	r1	manage	22.6 K	2013-06-02 - 12:35	LorenzoMoneta	Combined model fit result
c	CountingModel.C	r1	manage	1.5 K	2013-06-04 - 10:52	LorenzoMoneta	ROOT macro for creating the counting model
png	FrequentistHTResult.png	r1	manage	11.3 K	2013-06-04 - 15:00	LorenzoMoneta	FrequentistCalculator result on a test of significance
c	GausExpModel.C	r2 r1	manage	2.4 K	2013-06-05 - 17:45	LorenzoMoneta	Gaussian plus Exponential ROOT macro
png	GausExpModelFit.png	r1	manage	17.0 K	2013-06-01 - 17:54	LorenzoMoneta	Gaussian plus Exponential fit result
c	GaussianModel.C	r1	manage	1.2 K	2013-06-02 - 23:20	LorenzoMoneta	Gaussian Model ROOT Macro
png	GaussianModelFit.png	r1	manage	16.4 K	2013-06-01 - 16:54	LorenzoMoneta	Gaussian Fit Result
png	MCMCInterval.png	r1	manage	7.1 K	2013-06-03 - 23:04	LorenzoMoneta	MCMC interval result
png	PLInterval.png	r1	manage	8.2 K	2013-06-02 - 23:08	LorenzoMoneta	Profile Likelihood interval result
c	SimpleBayes.C	r1	manage	3.0 K	2013-06-03 - 23:04	LorenzoMoneta	Bayesian Calculator macro
c	SimpleHypoTest.C	r1	manage	2.5 K	2013-06-04 - 15:01	LorenzoMoneta	Hypothesis Test (Significance) ROOT macro
c	SimpleHypoTestInv.C	r1	manage	4.5 K	2013-06-04 - 23:26	LorenzoMoneta	Hypothesis Test Inversion macro
c	SimpleMCMC.C	r1	manage	2.0 K	2013-06-03 - 23:03	LorenzoMoneta	MCMC Calculator macro
c	SimplePL.C	r1	manage	1.6 K	2013-06-02 - 23:07	LorenzoMoneta	Profile Likelihood Interval macro
c	p0Plot.C	r1	manage	2.9 K	2013-06-04 - 17:05	LorenzoMoneta	ROOT macro for scanning p-values
png	p0Plot.png	r1	manage	10.1 K	2013-06-04 - 17:05	LorenzoMoneta	Result of p value scan
css	tutorial.css	r1	manage	0.2 K	2013-06-02 - 11:58	LorenzoMoneta	style for code background colour
This topic: RooStats > WebHome > RooStatsTutorialsJune2013
Topic revision: r20 - 2013-06-14 - LorenzoMoneta
Copyright &© 2008-2023 by the contributing authors. All material on this collaboration platform is the property of the contributing authors.
or Ideas, requests, problems regarding TWiki? use Discourse or Send feedback