aimror Change in attitude and encourage a culture of fact

making, Thatwill based decisiol

make Analytics all pervasive across the
organization.
1 STEPS OF ANALYTICS PROCESS
Many analysts would not agree that Analytics is a structured process. The
Business Analytics
is a structured procesS
argument they put forth is that the business problems are quite varied and
comparable to a typical that require innovative approach every time it is undertaken.Even while
business research pro this argument has some merit, it cannot be considered as not having any
cess from a higher level. structure atall. Hence,we tried to abstract the steps of a typical Analytics
initiative.
The difference is in the
details. Figure2.2 below gives the steps of atypical Analyticsinitiative. It starts
with understanding the business need driving the requirement of an
Analytics initiative.
Business Need or Exigency
solution or a
At this stage, the business considers a problem that need a covered
so far
suboptimal process that need improvement. The discussion
required such analytics approaches. Few more exam
a range of issues that
be as follows.
plesunder this category could
telecom service provider noticed that eftort spend on approval deci
"A Standard. It would like to reduce
sion is much higher than industry
as possible.
manual effort as much
 Analytics Proes3
Fig. 2.2 Steps of Analytics Process
Business need or exigencies

Objective of analytics

Analytics approach and

data collection

ProcesS inputs

Exploratory analysis Intermediate output

Build a relationship
(Model) Utilize model result

Implement the model/track

the performance

" Ahospital is being prodded bv an investment firnm to improve the duty

allocation of itsstaff and docors. The firm foundthat thepractice being
ollowed by the hospital is subopimal compared to modern hospitals
andthere is scopc for improvement.
" A consumerconnpany allocates its marketing budget based on a man
Business need to
ager survey andtop management opinion. Itwould like to improve this
processby optimallv allocating the budget to maximize the profit while improve elficiency of a
process or reduCe man
meeting other requirements. ual intervention can be
" Aretail chain likes to identify stores that nced more support consider considercd as the start
ing the current and future potential. Currently. it got a process which ing point of an Analytics
ihe fhrm feels not good enough as it is not doing justice to the future initiative.
potential.
Abusinessoperating horse race across country is worried by decreasing
collection. It is wondering what are the levers it can apply to improve
the revenue.
" Acity administration is interested in improving the security by deploy
ing its police force more optimally.
As discussed above, most business problems are statement of improve
ment required by an organization. This needs to be converted into a set of
objectives that can beaddressed through Analytics. This requires deeper
knowledge of strengths and weaknesses of Analytics. data and state of art
applications in various domains. Below we willdiscuss how we can oet io
objectives.
 Analytics Objective
The business problem is converted into Analytics problem at this
and the expertise of Analytics professionals is critically stap
needed.
They
be able to make an evaluation of the existing approaches towards a wil
Business problem is hle solutionfor this problem. An assessment about the suitability orposs-
val:
converted to Analyt
ics objectives. This is ability of the approaches also can be made. It is important at this
done by evaluating the avoid force fitting an unsuitable solution to a problem... kind of stage to
informationgap that is
required for developing
searching for a problem.
The Analytics solutions suitable for business problems could wo
'solution
the solution. widely. These could vary in terms of accuracy, cost and time
tion. This will be decided by analyst in consultationwith business. of comple
The term that will help us to convert business problem to
Analvics
objective is information gap. Think critically what information will belk
you to improve the situation or solve the problem. Don't worry too
much
about the feasibility of obtaining this.
Let us take each of the business problenm
which information can help you to address mentioned
it.
above and evaluate
Atelecom service provider noticed that
efforts spent on approval deci
sion is much higher than industry standard. It would like to reduce
manual effort as much as possible.
Inthis case, the firm must have been
using a set of rules to take approval
decisions manually. The efficiency of this process can be improved if
we can automate the process by
applying an algorithm that mimic the
manualprocess. This algorithm must consider
the application and decide the details available in
approval by mimicking the manual process.
Hence, the requirement is an algorithm
BM-|3s approval process. that mimics the manual
Specifically, the information gap is the chance of
approval given the application information. The objective of Analytics
effort will be to develop a solution that will provide this information
MAT based on the details in the
Ahospital is being application.
Information gap can allocation of its
prodded by an investment firm to improve the duy
be identified by staff and doctors.The firm found that
thinking
through the expected followed by the hospital is the practice beiny
solution to address the hospitals and there is scope for suboptimal when compared to mode
busincss problem. This Hospitals are
its service operating in a complex
improvement.
rcquire knowledge of taking multiple environment with the demandfor
domain and Analytics. emergency, planned and dimensions. It can be regular consultations.
staff can be made unplanned admission, etc. The aallocation of
ious sources. Thisoptimally if we know the expected demand from var-
by day for the needs to be
next few months.estimated as detailed ass possible; by hhour
Hence, the
hour and byinformation
day for
gap in
this case is expected hospitall demand by
the Analytics next few months. this information is
objective of Estimating
willbe used to build an first stage. In the next thisinformation
algorithm to schedule staffstage,deployme
 Analytic oe
"
Aconsumerbusiness allocates its
sIVey and top management opinion. marketingbudget basedona mal.e
cess by optimally allocating the budgetIt would like to impove his pro
to maxinnize the prolit while
mcctingother requirement.
In this case, we require an
optimizing
sensitivity of sales to various marketingalgorithm which considers the
elements (like advertisenent,
trade promotion, etc.), budget and other
Constraints are readily constraints. Few of these Marketing budget can
estimation.
available but sensitivity requires stat istical be allocated efficiently
if sensitivity of market
Hence, sensitivity of various marketing elements is the information ing clennents to sales
gaps and it will be the Analytics objective. is known. liencc, here
" A retail chain likes to identify stores that information yap is the
need more support consider
ing the current and future potential. Currently, it got a sensitivity.
the firm feels not good enough as it is not doing justiceprocess which
to the future
potential.
The current process adopted by the retail chain must be based on cer
tain performance parameters of stores. The drawback of this approach
is that high performers may be near their peak and further investment
may not provide proportionate return. Hence, it makes sense to con
sider the potential. Potential for growth depends on the economic pros
pects of catchment area and extended area surrounding it. Additionally,
it also depends on the equity created by the store among its customers
as reflected by loyalty, share of wallet, etc. To a great extent, these can be
established with athorough analysis of past growth of stores and vari
ous drivers.
Hence, informationgap is the sensitivity of store growth to economic
environment of catchment area, customer equity and various other
influencing factors. We also need forecasted values of these to evalu
ate future sales. Connected to these information gaps will be many
Analytics objectives like sensitivity estimation, customer equity assess
iment,forecasting and simulation. A retailchain can allo
cate the investment opti
worried about
" Abusiness operating horse race across country is mally if the potential
decreasingcollection. It is wondering what are the levers it can apply to of the store is known.
Improve the revenue. Hence, the information
by numerous methods by gap in this case is store
As we know, the revenue can be increased business.
course in this potential.
changing price, promotion,service mix,etc. Of horses,tracks, jock
to
there willbe many additional variables related
applied by a manager should be
eys, etc. Ideally, themarketing leversreliable information may be miss
relatedto influencers of revenue; but
stakeholders got preconceived
ing. In most of the organizations, themany of them are very confident
notions about these influencers and
about it. ThisAnalytics initiative is to test ifthe data is supportingthese
level of influence.
notions and to estimate the
 26 Chapter 2

Hencc, information gap in this case is the sensitivity of

marketing levers to revenue. Knowing this sensitivity. m various
decide which lever to apply more accurately. Analytics managerSin Can
:will beto analysc historical data and estimate these objective this
Casc

" Acity administration is interested inimproving the sensitivities.

security by
ing itspolice force more optimally. deploy-
Thecity will have various localities with varying crime profiles.
can be improved if police can be deployed proactively rather than rear
tively. This can be achicved if we can predict the OCcurrence of a
.Security
for various localities. crime
A city administration Hence, information gap in this case is the chance of occurrence of .
can optimally allocate crime at a particular locality by day and time. This information shoula
police forcc if they can be availablc carly so that police force can plan
predict the chance of deployment. The Analyt.
ics objective is toevaluate historical crime information
rime at various localj by locality and
tics. Hee, the informa
builda predictive algorithm. After deployment, this needs to be tracked
lo evaluate the impact.
tion gap is this hancc.
(onversion of a business problem to Analytics problem relies a lot on
industry practices to0. Industries will have standardapproaches to address
usual issues. These approaches are developed and refined over a
timeand there is some merit in giving priority to these while period of
aproblenm. Hence, it makessense to evaluate the addressing
choosing a ncw approach. problem carefully before

Analytics Approach and Data Collection

At thisstage the analyst will connect
approach. By 'Analytics Approach', weAnalytics mean
objectives with Analytics
lication, optimization, ctc., that will help us tostatistical
achieve the
modelling,classi
quantity and quality of data that is required for analysis isobjectives.to The
chosen approach. Some of the linked the
data going back many periods of time. multivariate approach will require detaie
I is always a
It requires domainchallenge toandconnect Analytics objectives with da.
knowledge Analytics knowledge in equal measult
What helps this process is a
model). This diagram connets conceptual the model (also called theoretical
and the influencing successively phenomenon we exanu
Analytic pproah two stages. At the
variables. is a multistage approach and usually got
first
It
refers Io the
lors at the highest level.stage,
In the
the phenomenon isSconnectedto various
tehuiquc that will be down to next stages, cach of thcse factors aare broken
applied los develop individual
ability of data or thevariables. At this stagge, we will not consider theavail
objective is to create anpracticality
ing the solution, Ana of collecting it or the COSt involved: the
lyst wll opc data
(onsider the exhaustive list.
Tequirenent with Ana iny and thus case of abusiness organization irinterestcd in understand
Io identifythereducing employee attrition. The
lytics aPproh.
the study was
of attrition of fators influencing attrition and objective
to predic the propensily
cadh of the
below links a number of employees. The conceplual diagram (Figure2.)
fctors thal influence
attrition.
 Analytics Process 7
Fig. 2.3 Coneptual Diagram of
Employee Attntion

Employee Age. gender, experience.

education, location of residence

Manager Manager personality/att1tude,

performance sCore, experience
Employee
attrit1on

Business Size, growth, employee

friendliness

Macro
economic Industry growth, inflation,
unemployment rate

It shows that attrition can be

ness and Macrofctors. (We call these influenced by Employee, Manager, Busi
ables). Ennplovee related factor consistsfactors' as it is a collection of vari
and other personal variables. It is also of variables like age, experience
of her innmediate manager influenced by the characteristics
Of course, resignation is also (people leave managers, not companies').
influenced
This looks like a wish list as it is quite by business and macro factors.
large consisting of qualitative and
quantitative variables. Conceptual diagram
Allorganizations willhave variables like is helpful for
profle, etc., available at one place. However, salary, increment, employee assessing
data rcquircment for
have variables like performance appraisal some organization may achicving the Analytics
ratings
department. business units etc. They may even adopt kept in silos divided by objectives.
of assessments. Variables like different standards
attitude, personality, etC., may not be even
available and would require extensive study design and data
buildit.For analvsts, it is critical to decide if he collection to
that are readily available or would spend will proceed with variables
much effort in collecting and
collating that are not available.
Consider the case of a
to estimate the sales due tobusiness evaluating itS sales. The objective is
marketing
and promotions. This will help them tointerventions like advertisements
keting operations next year. Apossible calculate Rol and plan for mar
provided below. conceptual diagram (Figure 2.4) is
 Anotherissuefrequently faced by analyst isthe CxISlence of ouli,.
Ireferstovalues of variables that are unlike what is expccted..Such valu,
parameters andshould be
canlead to biased estimation of the outliers using standard
corrected. Thc
initial cffort will be to identify
instead of deleting the methods. (referThis
(usually capped)
too is corrected for a revievw ofmethods). observation
Tiwariet al. 2007

Mark43- Catching Criminals with Analytics

QBox2.4
fightino
operating in an unlikely fheld for Analytics - crime The
Mark43 is and
objective of the organization isto
make policing smarter
for
efficient.
enter
Its plat-
it easier and faster Cops to arrest and
form is a web app that makes hours on legacy sv.
several
incident reports, reducing a process that can take
tems to about 30 minutes!,
fhling a police report.
The system walk the officers through the maze of
automatically filling in helds
It expands and contracts the decision tree, even
on clicking
when information is available. They can scarch for a suspect and
shots of associates
the suspect's page pulls up a mug shot along with mug records are made
(same way as Faccbook displays friends of friends)2. Police
into single scarchable body in the cloud; which is definitely a better system.
Currently, it is apatchwork of isolated paper and digital fles.
The company has added a layer of algorithmic intelligence connecting data
such as social media activity and phone records (obtained via warrant) to police
reports. This give cops a more complete picture of a criminal or a syndicate.
It can display relevant information tht is as readable as the feeds on Linkedn
or Twitter. This way of visualizing data by sketching out aweb of phone cals or
following movements across amap help them to identify key players in acrinie
ring or exonerate the usual suspects much faster.

http:/www.fastcompany.com, Innovative Companics

2015?Mark 43), as accessed on 20 (Navigation>Home>Most
November 2017 at 8.15pm 1ST
http:timc.com, 2015?Mark 43). as
accessed on 20 (Navigation>Home>Most Innovative
November 2017 at 8.1 5pm IST Companies

soured
Correct Mismatch in Levels of
from various processes that areMeasurement: The data could be anal1sis
to procced, all mcasured at different levels. For singke

data set. For variables should be at same level and mergedintocouldhe' 7

example,
at regional level while inadvertisement
a
marketing Analytics
data at the national
saleslevel.
data Ourpr
erence would beto model at the level of sales study.
and hence, we will
tryt0
disaggregate advertthanisement data into
approach rather This will be
the
usual

regional level.
tryingto summarize sales nationallevel.
into
 Analytics Process 31
Summarization: Sometimes the data available will be from operational
systems that are too granular for analysis. Such data will be summarized at
weekly or monthly level before it is used. For example, operational
of a credit card business will be at system
transactional level containing many
details about each transaction like amount, retailer, time, location, etc.
But
for mostof the analytics application,what is required will be a
the transactions at weekly or monthly level for each summary of
customer. Usually data generated
by operational systems
Exploratory Analysis are too granular for
Analytics applications
The objective of exploratory data analysis is tounderstand the nature of and would require sum
the data that was collected for analysis. It usually results in an marization.
presentation tothe decision makers or the users of the resultsintermediate
(refer Myatt
2006 for detailed coverage). Asthe name indicates, it is for
insight intothe data and relationships before building final developing an
solution, usu
ally a model. Input data was collected with certain
and at this stage it is tested to examine if it holds.
assumptions in mind
Exploratory analysis includes univariate and bi-variate analyses. In the
case of univariate analysis, each variable is evaluated on its own. It will
cover measures of central tendency, variation and distribution. In the case
ofbivariate analysis, each of the variable is analyzed against dependent vari
able (also pairs of independent variables chosen judiciously). This analysis
willhelp us to have an early indication of important variables and the type
of relationships (linear, different types of non-linearity, etc.) that exists.
Most importantly, it results in a valuable dialogue with decision makers.
It will ensure that all variables are being considered and no variable is
omitted inadvertently. It helps to get the feedback from them on certain
behaviour of the data which is not explained otherwise. For example,
whileanalysing sales for a snack item, analysts were surprised to noticea
dip insales during holiday season. The brand manager clarified that asize The objective of
able proportion of the consumption of this brand was as snacks for school descriptive Analytics is
students and hence,sales was impacted during holidays. Such insights will to understand the nature
be useful as the study moves to next stage. of data and interrela
In general, the findings will make intuitive sense to decision makers as tionships.
they are living inthis environment with intimate knowledge of the details.
Still, some of the fndings will surprise them too. Thisstage scts the back
ground for the final solution and they are prepared on what to expect. It
also will convince them that theanalysis is progressing well and thefinal
solution willmeet the expectations.
Build a Relationship
The most critical step in the whole process is to build arelationship
between dependent and independent variables. This relationship could be
in the nature of amathematical model built using statistical or operations
research methods. In most of the Analytics Cxercise, this stage will be there
 may not move to thËs
but therc can be initiativcs wherC il stage (00.
to Bor
such studies,the insight gencratedl: at cxploratory stagc could be
answer the questionsand take
decisions.
non-lincar
enoughto
In statistical approach, it could be alincar or model.
based on the nature of the dependent variable (continuous or This is
Most of the Analyt and the nature of underlying relationship. If the approach is catoperati
egorical)
ics initiatives involve research, it will be alinear or nonlinear optimization model. This
building a mathemat critical asthe quality of final solution is based on the outcome of thisstage is
ical model connecting stage.
Depending on the approach, this stage itself will gothrough anumbe
depending and indepen well-defined steps.
dent variables.
In the statistical modelling approach it usually passes through the fol.
lowing steps. We will consider each of the variables and make it
for modelling, If it is aquantitative variable, usually it doesn't ready
require any
treatment. However, if the relationship is not meeting the fundamental
assumption of the techrnique, it may require transformation. For exam
ple, when we apply linear regression to build models, the assumption is
linear relationship between dependent and independent variable. In case
the relationship is not so, the independent variable will require a
to make it linear (for example, take log). Similarly, if the nature treatment
of therela
tionship is highly iregular, it doesn't make sense to use it inthe original
form. In such cases, we may convert the variable into
provides an overview of these categories. Figure 2.5
steps.
In case of qualitative variables, a
conversion to quantitative form is
required before it can be used. Categorical
binary variables if the number of categoriesvariable will be converted to
is not too large. If
large numbers of categories like are there
to reduce the number. For zipcodes, it will require some processing
example,
region and then these can be made into zipcodes can be mapped to states or
It is quite usual to have large binary.
stage and hence, the effort requirednumber of variables at the
at this stage is modelling
Models are validated to appropriate modelling approach, this can be reduced. significant. choosin8
By
ensure that performance the umbrella of selection processes Usually it falls under
js
not compromised on the modelling objective, (forward, backward, etc.). Dependiny
when it is implemented maybe chosen. forward, backward or stepwise approa
in a It rarely happens that a
different environ
ment.
Usually, anumber of modelsfinal are
model is developed in the first attemy
For models that are
focused on
developed and it is tested before finaliz1ng
can be made more accuracy, the decision about best model
casily as it involves
Model Validation: The picking the model with lowest error.
oped model retains its objective of validation is to ensure that thedevel-
it is predictive ability
in adeveloped. This concern is raised as outside the environment
where
phenomenon
model fits the data called
on 'overfitting' . This
modelling
refers to a
approach can he
Situation when the
the'quirks' of the
data. which is modelling 'too well' and accommodates
it
while fails
it
miserably
following are someof theoutside
Performance of the model
this data:; when it is will turn out to be good
practice. The
specifics that could lead toapplied this
in
situation.
 Fig. 2.5 Data
Processing Steps
Consider variables
One by one

Is variable
quantitative?
Quantitative variable
Qualitative variable
Check and
transform Create binary
variables

Pool variables to
make master set

Choose modelling
approach

Finalize model

" It is
common to have large
Asmore variables are added,number
of variables under
consideration.
the chance of Type Ierror (error
a variable is erroneously included in the
model) increases.
whereby
This also reduces the number of degrees of
freedom and chance of over
fhtting increases.
" It is especially critical when the data is
lying distribution of the data is changing non-stationary. That is the under
over time. For example, con
sider the sales of aproduct in the growth mode. Sales whenthe model is
g0ing to be applied could be very different when itwas modelled.
Although, thhe ultimate test of validity is the performance of the model
practice, we could trv and control factors like overfitting. The usual To validate a model,
the dataset is divided
approach is to keepa part of the modelling dataset (titled hold out sample) into Training and Valida
Coto
cent asidetestwhile
the model. Typical industry practice is to keep about 25 per tion samples. The mnodel
ensuring that the sample sizeis enough to provide areli- isdeveloped using Train
able estimate.. Note that the size of the hold out sample will reduce the data ing sampleand tested on
available for development. Figure 2.6 below shows the steps involved in validation sample.
the process.
The process starts with splitting the modelling sample into develop
nent and hold out samples randomly. These datasetsare also called 'Train
Ing' and Test'. 7The model is developed using the development dataset and
 Validation Process Steps
Fig. 2.6 Hold Out Sample
Total modelling
sample (100%)

development
Split data into random
and holdout at

Hold out sample

Development
sample (70-80%) (20-30%)

Develop model Test the model

Modify model Pass the test?

inputs No

Yes

Proceed to next
step

the developed model is run on the hold out

compare the models between these two datasets. sample. The next step is to
we will conclude that the If the difference is minol;
model is validated well. This
most commonly used method but there are
many
approach S
could be applied as well in other approaches tie
suitable situations.
Cross Validation: The above
the validity of a model. This method depends on asingle dataset toodecide
In cross could be misleading if we happen to geta
In the case of time
split of bad
series data, models are datasets.
ments and validation, the idea is to create multiple develop-
validated using out of hold out samples
The models are then instead of each as we discussed above.
time validation. Here, a
time window is left for dation sets. evaluated based onone the results of each of theseevali-
validation.
Out of Time
impact of timeValidation: This validation is understandthe
varying applied to
for data. Model is characteristics and is relevant only for
series time
develo2.7ped on a dataset leaving a window oftime
validation alone. Figure
validation'.
The figure shows the data for'Out ofTime
ysis. Data of shows that
l data with
splitting
vintage to 12
monthsvintage of
was used15for months
foranal-
was used

development
 Analytics Process 35
Fig. 2.7 Out of Time Validation

Development period Validation

period Application period
1 2 3 4 5 7 8 10 1 12 13 14 15 16 17 18 19 20 21
Months

of months 13 to 15 was kept aside for

validation. The advantage of this
method is that validation data is truly 'new'
ability of model to adjust for any change in and will be able to capture the
characteristics due to time.
Training, Validation and Test Process: It is not always a
that finalizes the two-step process
model. Since the two step process improve the
using the result of validation; use the model
information to influence the model Model
implemen
development, it doesn't make sense to the validation sample to finalize
the model.We needa third database forusefinal tation is the practical
cess we will divide the dataset into three parts.decision. Hence, in this pro applicationof the devel
oped model. At a funda
Training sample for developing the model. mental level, the focus is
on the coefficients.
Validation sample for validation. Note that we will go back and forth
between these two samples to develop a validated model.
" Test sample to test the model for a go or
no-go decision.
Model Implementation
The final step in the process is to leverage the model results and
could take many a different forms. At the most fundamental level, this step
the
attention could be only on the coefficients and the interpretations. It could
also form part of a decision support system for regular
Hence, we can divide implementation into two steps namely decision making.
model results and leverage the leverage the
model.
Leverage Model Results: The focus in this case is the detailed result of the
model, not exactly the model as an entity. Take the case of a simple model
built to explain the salary of managers in an organization.
The coefficients of the
S=a+bE + cG (2.1) developed model can
be used to derive the
Here insights and that can be
S= annual salary objcctive of the eKercise
E= expericncc (in years) in sonne cases.
G= Gender (Male =0, Female = 1)
Here the model gives an estimate of average starting salary (a for male,
a+cfor female), estimate of average increment (value of b) and uncover
any evidence of gender discrimination (valuc ofc). The objective of the
nodelling exercise could have becn to get these estimates and in such
Cases, the model served its purposc by cstimating these values. It can be
 further extended by splitting the salary into components like
salary and increment received. Depending on the situation, such applie starting
tionmay bemeaningful.
Leverage the Model: In this case the interest is in the model as a whola
rather than the values of the coefficients. Model is applied to a decision
situation to automate or improve the process.
In the casc of predic For example, consider the earlier model. Ifit is an elaborate one with edu.
tive applications, the cation,department, etc., it can be used for estimating a salary for making
focus is on the model as an offer to anew employee. Of course the organization will not adopt the
awhole and it is applicd estimate without change as the estimate is abase to kickstart discussions.
to automatc or improvc
Similarly, a model predicting probability of delinquency can be used to
a process.
automate loan approval decision. The model will be implemented in the
system and as the financial institution receives application, the informa
tion will beentered into the system and themodel will estimate the prob
ability of delinquency for the applicant. This probability is invaluable and
willbe used along with other factors toapprove the loan.
Marketing mixmodel connccts sales of abrand to influencers like price,
promotion, advertisement, etc. It can be used as decision aid to evaluate
various choices available with managers. In this case, the model is
with auser-friendly front nd that allows a meshed
to cvaluate the impact (a decision manager to vary the factors
reduce theprice by 2.5 per cent andsuPport
increase
system). For example, she can
the
55 per cent to evaluate the cumulative impact. advertisenment support by
It is expected that most of the model
results will be rclevant for about one
vear. Beyond this, the strength of relationship bct
weaken or new variables may come into play. The ween the variables may
environment, behaviour and attitude ol people, etc., business and cconomic
resulting in a weak relationship with variables in the would change thus
formance is deteriorated beyond specified limit, the model. When the per
In situations where models are used regularly for taking model may be rebuil.
decisions like loan approval, the pertormance is for taking financially critical
dels als0 undereo
tracked more freuuentlv.