Applied Microeconometrics

Assignment

Time Use Survey Analysis 2019-20

Team Number: 10
Team Members: Amrita (EE/2023-25/003)
Kirthika (AE/2023-25/010)
Shivani V
B (EE/2023-25/023)
Suraj R (EE/2023-25/024)

Reference :
Economics 217 -Multinomial Choice Models. (n.d.). Retrieved
November 4, 2024, from
[Link]
Chandrashekhar 2007 Willingness to Contribute Sanitation
Scott Long - Regression Models for Categorical and Limited
Dependent Variables

1
Question 0: PREPARATORY
The summary Statistics of the variables used in the regression models segregated as male and
female:

TABLE 1: SUMMARY STATISTICS

Variable Category Obs Mean Std. Dev. Min Max

Male 4,403 0.3311 0.4707 0 1

in_educ Female 4,131 0.2692 0.4436 0 1
Total 8,534 0.3011 0.4591 0 1
Male 4,403 1.7543 0.5968 1 3
activity Female 4,131 2.2886 0.8625 1 3
Total 8,534 2.0169 0.7665 1 3
Male 4,403 21.6296 4.2149 15 29
age Female 4,131 22.0133 4.2529 15 29
Total 8,534 21.8187 4.234 15 29
Male 4,403 7.4231 0.5533 5.6021 9.69
lnmpce Female 4,131 7.3492 0.5532 5.7038 10.5321
Total 8,534 7.3865 0.5532 5.6021 10.5321
Male 4,403 4.8369 2.0007 1 20
hhsize Female 4,131 5.0482 1.9951 1 20
Total 8,534 4.9405 1.9971 1 20
Male 4,403 1.4522 0.5068 1 3
cook_fuel Female 4,131 1.4563 0.5011 1 3
Total 8,534 1.4543 0.504 1 3
Male 4,403 1.0182 0.1336 1 2
lighting Female 4,131 1.0182 0.1335 1 2
Total 8,534 1.0182 0.1335 1 2
Male 4,403 1.4733 0.4993 1 2
dwelling Female 4,131 1.4735 0.4994 1 2
Total 8,534 1.4734 0.4993 1 2
Male 4,403 3.4708 2.6243 1 9
social_group Female 4,131 3.3014 2.5327 1 9
Total 8,534 3.3861 2.5807 1 9
Male 4,403 1.136 0.7142 1 9
religion Female 4,131 1.1465 0.7431 1 9
Total 8,534 1.1412 0.7286 1 9

2
 Male 4,403 1.3732 0.4837 1 2
Sector
Female 4,131 1.3663 0.4818 1 2
Total 8,534 1.3697 0.4827 1 2
Male 4,403 116.9884 208.9555 0 930
time_learn Female 4,131 94.0208 192.7052 0 990
Total 8,534 105.8535 201.2791 0 990
Male 1,217 5.9225 0.5877 2.708 6.8352
ln_time_learn Female 932 5.8883 0.6386 2.708 6.8977
Total 2,149 5.9066 0.6123 2.708 6.8977
Male 4,403 1.0952 0.2935 1 2
type_of_the_day Female 4,131 1.0794 0.2704 1 2
Total 8,534 1.0875 0.2821 1 2

Test of hypothesis for balancing the variables across male and female
TABLE 2

Variable Chi² T-statistics P-value

in_educ 38.8704 - 0
activity 2500 - 0
age 40.7262 - 0
lnmpce - 6.1669 0
hhsize - -4.881 0
cook_fuel 7.3848 - 0.025
lighting 0 - 0.996
dwelling 0.0003 - 0.987
social_group 13.1696 - 0.004
religion 4.6193 - 0.202
Sector 0.4353 - 0.509
time_learn 5.3706 - 0

ln_time_learn - 1.2894 0.1974

type_of_the_day 6.6302 - 0.01

sh_learn_time - 5.3706 0

3
In the above table 2 t-test was conducted for continuous variables (lnmpce, hhsize,
ln_time_learn and sh_time_learn) and Chi-squared test was conducted for remaining
categorical variables. The hypothesis tested regarding the balance of variables between males
and females for various continuous and categorical variables revealed significant associations
for educational level (in_educ), activity type, age, household size (hhsize), cooking fuel
(cook_fuel), social group, and learning time (time_learn), with p-values falling below 0.05.
This suggests that these distributions differ significantly between genders.

On the other hand, variables such as lighting, dwelling type, religion, sector, and logged
learning time (ln_time_learn) did not show significant differences, with p-values equal to or
exceeding 0.05, indicating a balance between males and females for these aspects. Overall,
while many variables demonstrate significant gender differences, others remain balanced,
emphasizing the need to consider gender when analysing these characteristics.

Question 1: Binary Choice Model:

The objective of using a binary choice model is to analyse the factors influencing the probability
of attending an educational institution (in_educ = 1). The dependent Variable used here is
in_educ, representing the probability of being in education and the independent Variable
includes factors such as age, gender, education level (edu_level), marital status (married or
unmarried), type_of_the_day, socioeconomic indicators (lnmpce, hhsize, and so on). The two
models used here are,

Probit Model: This model assumes that the probability of attending an educational institution
is a cumulative normal distribution function. The probit model is appropriate when the
dependent variable (whether a youth is in education or not) is binary, and we want to estimate
the probability that this variable equal 1. The functional form for probit model is:

F (Xiβ) = Φ(X′β) =

Where, Φ is the cumulative distribution function (cdf) of the standard normal distribution. ϕ is
the probability density function (pdf) of it, interpreting the Coefficients of Gender.

Logit Model: This model assumes a logistic distribution function for the probability of in_educ
= 1. It is often used interchangeably with the probit model for binary outcomes, but it has
slightly different properties in terms of the distribution tails. Both models provide coefficients
that help us understand how each explanatory variable (like age, gender, edu_level, mpce)
influences the probability of attending an educational institution. The functional form of logit
model is given by:

F (Xi

Where, Ω is the cdf of logistic distribution. This function could also be written in terms of log
linear model.

Ln_Ω(x) = xβ where

4
 , is the odds of the event given x and the Ln_Ω(x) is
The log of the odds, known as the logit.

When we take the differentiation of this function we get, k

When the model is linear, βk can be interpreted like this: for a unit change in X k, we expect the logit
model to change by βk, holding all other variables constant.
The output difference in probit and logit model: While both logit and probit models produce similar
directional effects (in terms of sign and relative magnitude), the coefficients differ due to a scaling factor
(otherwise the functional form). Logit coefficients are approximately 1.81 times larger than probit
coefficients, a difference arising from the distinct shapes of the logistic and normal distribution
functions underlying each model.

(a) Estimating probit and logit models:

TABLE 3:

LOGIT PROBIT RATIO

Variable IN_EDUC Z IN_EDUC Z IN_EDUC Z
Age -0.48322*** 131.06 -0.26528*** -142.08 1.821532 -0.92244
[Link] 0.183296*** 9.34 0.102935*** 9.33 1.780698 1.001072
Edu_Level
Primary 2 0.367171*** 3.36 0.127741** 2.26 2.874332 1.486726
Middle & Sec. 3 1.118787*** 13.95 0.483106*** 11.6 2.315821 1.202586
Higher
2.157534*** 29.77 1.078889*** 29.5 1.999774 1.009153
Secondary 4
Grad And
3.592201*** 49.82 1.83979*** 51.39 1.952506 0.969449
Above 5
Marital_Status
Currently
-2.73606*** -65.06 -1.34313*** -70.44 2.037088 0.923623
Married 2
2 .Type_Of_The
0.960084*** 32.25 0.540817*** 33.22 1.775248 0.970801
_Day
Lnmpce 0.334822*** 16.81 0.179776*** 16.17 1.862442 1.039579
Hhsize 0.043287*** 7.48 0.024705*** 7.77 1.752152 0.962677
Cook_Fuel
Conventional 2 -0.30473*** -13.01 -0.17194*** -13.07 1.772255 0.995409
None 3 0.822692*** 8.29 0.474805*** 8.46 1.732693 0.979905
Lighting
Others 2 -0.00141 -0.03 -0.00099 -0.04 1.423268 0.75
Dwelling
Others 2 -0.05433** -2.44 -0.03516*** -2.82 1.545417 0.865248
Social_Group
2 -0.08761** -2.29 -0.04841** -2.27 1.809644 1.008811

5
 3 -0.05905 -1.66 -0.03417* -1.73 1.728221 0.959538
9 0.1062*** 2.86 0.053929*** 2.6 1.969263 1.1

Religion -0.33854*** -0.19258***

2 0.577876*** -11.76 0.312767*** -11.92 1.757882 0.986577
3 -0.23226*** 12.46 -0.11741*** 12.25 1.847624 1.017143
4 0.051258 -3.43 0.021937 -3.07 1.978206 1.117264
5 0.309*** 0.29 0.161851*** 0.22 2.336606 1.318182
6 0.085385 3.25 0.056213 3.04 1.909159 1.069079
7 0.678885*** 0.19 0.370631*** 0.22 1.518957 0.863636
9 7.45 7.31 1.8317 1.019152
Sector 0.123497*** 0.072234***
Urban 2 3.922595*** 5.53 2.326268*** 5.8 1.709692 0.953448
_Cons -0.48322*** 21.28 -0.26528*** 22.95 1.686218 0.927233
Observations 8529 8529
Standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

Marginal Effects:

This measures the effect at the mean values of all variables. For probit:

Where ϕ represent the density functions for the probit model.

Average Marginal Effect:

This is the average of marginal effects across all observations, calculated as:

Mean

) (probit)

The Average Marginal Effect (AME) calculates the marginal effect for each observation and
then averages these effects across all individuals, providing a comprehensive estimate of how
a predictor influences the probability across the entire sample. By capturing variation in
characteristics like ln_mpce and gender, the AME reflects the predictor's impact on the
likelihood of education across diverse profiles within the sample. In contrast, the Marginal

6
Effect at the Mean (MEM) computes the marginal effect by holding all variables at their mean
values, giving an estimated impact for a "typical" individual based on average characteristics.
This approach highlights the predictor’s effect for an average profile rather than across the
sample’s full variation.

The probit and logit models provide estimated coefficients indicating the likelihood of
attending educational institutions based on various factors. The pseudo-R-squared values for
both models suggest a strong model fit, with 0.5138 for the logit model and 0.5045 for the
probit model. These values indicate that a substantial portion of the variation in the probability
of attending an educational institution (in_educ) is explained by the predictors included in the
models. Some key findings are: Age has a negative and significant effect on the likelihood of
attending an institution for both the models with -0.25455 for probit model and -0.4746781 for
logit model, implying that older individuals are less likely to attend.
Higher Education Levels have strong positive effects, with "Grad and above" showing the
highest likelihood (1.490177 for probit and 2.982061 for logit).
Marital Status (married individuals) is associated with a lower probability of attending
educational institutions as the coefficient values for probit and logit models are -1.078965 and
-2.164969 respectively.

(b) Interpretation of the coefficient of gender separately using: (i)

Probit model for discrete difference:

Pr(y = 1| x, xj + δ) − Pr(y = 1| x, xj)

The coefficient for [Link] (female) in the Probit model is 0.0341, but it is not statistically
significant (p=0.408). This implies that gender does not have a statistically significant impact
on the probability of attending an educational institution after controlling for other factors in
the Probit model.

(ii) Logit model using odds ratio:

Odds Ratio=

In the Logit model, the coefficient for females is 0.0323, which also lacks statistical
significance (p=0.658). When interpreting this through an odds ratio, we find that the odds of
females attending an educational institution are about 3.2% higher than males. However, due
to the lack of significance, this effect is not considered impactful.

(c) Interpretation the coefficient of ln_mpce separately using:

(i) Probit model for continuous variable:

_mpce

7
The coefficient for ln_mpce in the Probit model is 0.250 (p < 0.001), indicating a positive
relationship with the probability of attending an educational institution. For each unit increase
in ln_mpce, the likelihood of being in education increases significantly, suggesting that higher
per capita expenditure, as a proxy for socioeconomic status, is strongly linked to educational
attendance.

(ii)Logit model using odds ratio:

In the logit model, the odds ratio for ln (mpce) is calculated as 𝑒βln (⁡𝑚𝑝𝑐𝑒). The logit model
coefficient for ln_mpce is 0.480 (p < 0.001), with an odds ratio interpretation. This means that
with each unit increase in ln_mpce, the odds of attending an educational institution increase by
approximately 48%. This substantial effect aligns with the socioeconomic findings in education
research, where increased expenditure often facilitates educational opportunities.

(d) Graph the predicted probability across age separately for males and females:

This graph illustrates the predicted probability of attending educational institutions i.e. Pr
(ln_Educ) across different ages for males and females. The vertical axis shows the predicted
probability, while the horizontal axis represents age, ranging from 15 to 29.

There is a clear downward trend in the probability of attending educational institutions as age
increases. Younger individuals, especially those around age 15, have a higher likelihood of
attending educational institutions, with a probability close to 0.6. However, this probability
declines steadily with age, approaching zero by age 29. The lines for males and females are
nearly overlapping throughout the age range, indicating minimal difference in predicted

8
probabilities between genders. Both male and female probabilities follow an almost identical
trajectory, suggesting that age has a similar effect on educational attendance likelihood for both
genders. The steep decline suggests that individuals are progressively less likely to attend
educational institutions as they grow older, possibly due to transitions into the workforce or
other commitments. The similar patterns for both genders may imply that age-related
educational engagement factors affect males and females similarly within this dataset.

Question 2: Polychotomous Choice Model

TABLE NO 4: MULTINOMIAL CHOICE MODEL

ACTIVITY EDUCATION NEET

Variables RRR Std. Err. RRR [Link].

Age .6000173*** .0095285 .9269559*** .0104088
Lnmpce 1.459772*** .1277605 .79396*** .0615737
Hhsize 1.061789*** .0235849 1.007658 .0180422
[Link] 5.310773*** .4979306 26.71528*** 2.117126
edu_level
Primary 2 .692124 .2652211 .6710867** .1101253
Middle & Sec. 3 1.256495 .3463827 .7715171** .0976628
Higher secondary
4 2.930268*** .736262 .9107154 .1048697
Grad and above 5 23.7354*** 6.04189 1.402612 .1668785
marital_status
Currently married
2 .1213569*** .0190015 1.110871 .1033149
2.type_of_the_day 2.128903*** .2801042 .324511*** .0509082
cook_fuel
Conventional 2 .7520816*** .0705186 .9229647 .0746931
None 3 2.594949 1.754999 4.057102** 2.886404
lighting
Others 2 2.496372*** .7507802 .81677 .2065094
dwelling
Others 2 .6416389*** .0571886 .7906892*** .0607598
social_group
2 2.021994*** .2661412 2.263498*** .2490014
3 1.8042*** .1956824 2.129236*** .1973474
9 2.574223*** .3603433 2.26334*** .2809068
religion1
Muslim 2 .6179067*** .0973273 1.114356 .145049
Christian 3 1.286572 .79841 1.343981 .8036525

9
 Others 9 2.254131* 1.055033 2.621636** 1.133323
Sector
Urban 2 1.084525 .1091835 1.353144*** .1197619
_cons 91.93179*** 71.16134 2.235604 1.478067

N = 8534; Reference category - No Contribution, RRR – Relative Risk Ratio

Standard errors in parentheses

*** p<0.01, ** p<0.05, * p<0.1

Additive Random Utility Model

The multinomial choice model, which is theoretically based on the Additive Random Utility
Model, assumes that an individual chooses between three Alternatives: Education, NEET, and
Employment, based on the utility they derive from each choice. Here, the Individual evaluates
each choice and picks the one that maximizes his utility. The utility an individual gains from
each option depends on observed factors (like age, gender, and household characteristics) and
an unobserved, random component.

In ARUM, the utility Uij that individual i derives from choosing alternative j (eg: education,
employment, NEET) is represented as:

Uij = Vij + εij where

• Vij is the deterministic part of utility (a linear combination of explanatory variables

such as age, education level, etc.).

• εij is the random component that captures unobserved factors affecting the choice.

This connects to the latent variable framework because the observed choice (activity status) is
a manifestation of the underlying latent utility

We observe the choice j if and only if Uij > Uik for all k ≠ j

This theoretical Framework in ARUM and latent variable framework helps explain why we
observe systematic patterns in activity choices based on observable characteristics while
accounting for unobserved factors through the error term.

In summary, ARUM justifies using the multinomial Choice model by providing a way to model
choices as a utility-maximizing decision process, where the relative utility across alternatives
determines the choice probabilities. The RRRs from the multinomial Choice model allow us to
interpret the probabilities in relation to a reference category (Employment), which aligns with
how ARUM conceptualizes utility in discrete choice scenarios.

(b) The likelihood ratio chi-square statistic (8365.24, with a p-value of 0.000) indicates that the
model is significant overall. Since Employment is a reference category, interpretations are
made relative to it. When looking at each variable's significance, existence of statistically

10
insignificant variables are found to be highly particularly NEET has relatively high number of
insignificant variables, indicating to study deeper on the relevance of explanatory variables in
this analysis.

1. Gender (Female):
• For the category Education, the RRR for gender (Female) is 5.31 with a high statistical
significance (p < 0.01) indicating that females are 5.31 times more likely to be in
Education rather than Employment compared to males, holding other variables
constant. Similarly for NEET, the RRR for gender (Female) is 25.47 (p < 0.01). This
suggests that females are 26.71 times more likely to be in the NEET group rather than
Employment compared to males. This high RRR reflects a strong gender disparity,
indicating a significant likelihood for females in this population to be outside
employment in either educational pursuits or the NEET category.

• These findings align with broader socio-economic patterns where women may face
barriers to employment, either due to educational, cultural factors, societal pressure or
family needs. The Higher proportion of NEET indicates potential constraints or
sociocultural expectations that influence the decision to not participate in employment
among females in Madhya Pradesh which should be further analysed to find the key
contributory factors that prevent women from education

2. Education Level (edu_level):

• Primary Education: For individuals with primary education, the RRR of 0.69 (not
significant) indicates no statistically significant effect on the likelihood of being in
Education rather than Employment. The RRR of 0.67 (p < 0.05) for NEET indicates
that individuals with primary education are 33% less likely to be in NEET compared
to Employment. This might suggest that primary education provides enough
qualification for basic employment (vocational jobs), reducing NEET status

• Middle & Secondary Education: The RRR of Middle & Secondary education in the
Education category is not statistically significant. This means that having a middle or
secondary education level does not significantly impact the likelihood of being in
Education over Employment. Thus, we cannot draw any meaningful conclusions from
this RRR for the Education category. For NEET, the RRR is 0.77, though statistically
significant (p < 0.05), this value suggests a 23% less likely to be in NEET compared
to Employment, suggests that these individuals are relatively less likely to be inactive.
This suggests that secondary education may provide sufficient qualification for
employment, thereby reducing the likelihood of NEET status.

• Higher Secondary Education: Individuals with higher secondary education have an

RRR of 2.93 for Education (p < 0.01), showing that they are 2.93 times more likely to
be in Education rather than Employment. For NEET, the RRR of 0.91 (not significant)
suggests that, for individuals with Higher Secondary Education, there is no clear
evidence that they are more or less likely to be in NEET compared to Employment.

• Graduate and Above: The RRR for individuals with a graduate or higher education
level is 23.74 for Education (p < 0.01), showing they are 23.74 times more likely to be

11
 in Education than Employment, reflecting a strong inclination towards further
educational engagement over employment. For NEET (RRR = 1.40, not significant):
There is no statistically significant effect of having a graduation level of education or
higher on the likelihood of being in NEET compared to Employment

• From the above education levels, it's interesting to note that RRR rises multiple folds
from HSC to graduate showing the importance of education in having better livelihood
opportunities. RRRs with higher educational levels underscore the importance of
education in influencing career choices. Higher education is strongly associated with
continued engagement in educational activities, potentially indicating aspirations for
advanced qualifications or specialized roles not readily available in current employment
sectors.

• For NEET, the statistical insignificance indicates that having a higher secondary or
graduate education level does not reliably predict being in NEET compared to
Employment, likely due to alternative paths (employment or further study) being more
prevalent for individuals at these education levels.

3. lnmpce (Log of Monthly Per Capita Expenditure):

The RRR for lnmpce is 1.46 (p < 0.01) showing a 1-unit increase in the natural logarithm
of monthly per capita expenditure increases the relative risk of being in the Education
category (compared to Employment) by 46%. This indicates that individuals with
significantly higher income households are more likely to engage in education rather
than employment possibly due to better financial support and fewer immediate
pressures to join the workforce, when analysing the same for NEET, the RRR for
lnmpce is 0.79 (p < 0.01), indicating a lower likelihood of being in NEET i.e., a 1%,
increase in lnmpce decreases the likelihood of being in NEET (as opposed to
Employment) by 21%. This implies that individuals from higher-income households
are less likely to be NEET, potentially due to greater access to educational and
employment opportunities. Further higher lnmpce likely correlates with better
socioeconomic conditions, enabling individuals to pursue education over immediate
employment. The lower likelihood of NEET in higher-income brackets suggests
economic resources play a strong role against economic inactivity.

4. Age:
• The RRR for education is 0.60 (p < 0.01) shows that, for each additional year of age,
the relative risk of participating in Education compared to Employment decreases by
40%. This significant reduction suggests that older individuals are less likely to engage
in Education, possibly due to increased workforce participation or other life
responsibilities that make further education less feasible or desirable as they age.
Similarly RRR for NEET is 0.93 (p < 0.01) indicates that, for each additional year of
age, the relative risk of being in the NEET category compared to Employment
decreases by 7%. This significant finding indicates that older individuals are also less
likely to be in NEET relative to Employment. This may be because, as people age, they
are more likely to find employment or establish more stable career paths, reducing the
chances of being in NEET.

12
 • The negative relationship between age and both Education and NEET relative to
Employment suggests that as individuals get older, they are more likely to be employed
than to remain in or return to education, or to be in NEET. This likely reflects a trend
where employment becomes a more common and stable status with age as well as
increased family responsibilities to provide better support and livelihood to the family

5. Dwelling (Others 2)
• The RRR for education is 0.64% ( P < 0.01) suggests that, individuals living in the
“Others 2” category of dwelling (possibly indicating a non-standard or less common
housing arrangement), the relative risk of being in Education compared to Employment
is 36% lower. This statistically significant finding may suggest that individuals in these
dwelling types have fewer resources, opportunities, or motivations to pursue education,
potentially due to economic or geographic constraints associated with their housing
situation. Similarly, RRR for NEET is 0.79% (P < 0.01) indicates that individuals in
the "Others 2" dwelling type are 21% less likely to be in NEET relative to Employment.
This significant effect could imply that despite limited educational engagement, people
in these types of dwellings are still more inclined towards employment over NEET.
This may reflect socioeconomic conditions that necessitate seeking work rather than
remaining inactive, even if they face barriers to accessing education.

• Overall, Dwelling (Others 2) suggests that certain living conditions may reduce
educational engagement but still motivate employment over inactivity, reflecting
possible socioeconomic pressures particularly poverty.

Overall Summary
The model illustrates strong associations between gender, education level, lnmpce, age, and
dwelling unit with individual activity choices in Madhya Pradesh. Females are significantly
more likely to be in Education or NEET rather than Employment, reflecting gendered
socioeconomic dynamics and stereotypes. Higher education levels and lnmpce are positively
correlated with educational engagement, while higher-income individuals are less likely to be
NEET. Age and Dwelling unit also impact activity choices, suggesting that when individual
gets older the responsibility to provide and support family increases, thus decreases the
likelihood of being in education and NEET, similarly poor living conditions typically reduce
educational attainment but still motivate employment over inactivity, reflecting possible
socioeconomic pressures. These insights underscore the complex interplay of socio-economic
factors in influencing career and educational paths in Madhya Pradesh

c) The Independence of Irrelevant Alternatives (IIA) assumption is a key characteristic of

the multinomial logit model used in discrete choice modelling. This assumption states that the
relative odds of choosing between any two alternatives are unaffected by the presence or
characteristics of other alternatives.

In the context of our study on activity choices (education, employment, NEET), the IIA
assumption implies that the relative probability of a youth choosing, for example, education

13
over employment should not be influenced by the presence or characteristics of the NEET
option.

i.e., w.r.t our study

1. The ratio of probabilities of choosing Education over Employment should be
independent of the NEET option

2. The ratio of probabilities of choosing NEET over Employment should be independent

of the Education option

This assumption simplifies the model but might be unrealistic if the alternatives are similar or
substitutes for one another, where changes in one option might affect the relative attractiveness
of the others.

To check IIA assumptions, Hausman test is done.

Stating Hypotheses
Null Hypothesis (H0): The IIA assumption holds. This implies that the difference in
coefficients between the restricted and full model is not systematic, and the relative
probabilities of choosing between two options are unaffected by the presence of other
alternatives.

Alternative Hypothesis (H1): The IIA assumption does not hold. This implies that the
difference in coefficients between the restricted and full model is systematic, indicating that
the relative probabilities of choosing between two options depend on the presence or
characteristics of other alternatives.

Test Statistic : chi2( 21) Prob > chi2

883.89 0.0000

Here P < 0.05

The Hausman-McFadden test produced a chi-square statistic (883.89) with a p-value of 0.0000,
which leads to the rejection of the null hypothesis. This result suggests that the IIA assumption
does not hold.

The violation of IIA suggests that the odds of choosing between any two activities (Education,
NEET, and Employment) are not independent of the third option. This indicates that these three
activities are likely to be perceived as substitutes or complements rather than independent
alternatives

Possible Reasons for Violations:

1. Substitutability or Similarity of Choices: Individuals may view certain options as
substitutes or more related to each other than to other options. For example, if the option
"education" becomes less attractive, some youths might switch to "employment" or
"NEET" based on their similar factors, violating the independence assumption.

14
 2. Shared Characteristics Across Choices: The alternatives (employment, education,
NEET) may share common characteristics or influencing factors, such as
socioeconomic status, family background, or regional conditions. These shared
characteristics can create correlations in the unobserved factors affecting each choice,
violating the IIA assumption.

3. Correlated Error Terms: The multinomial logit model assumes that the error terms
for each choice are independent. However, in real-world scenarios, these error terms
might be correlated (e.g., individuals considering education and employment as related
options), leading to an IIA violation.

NOTE: The multinomial logit has a pretty sharp property that is usually not good in practice:
Independence of Irrelevant Alternatives (IIA). Precisely, when choosing between two goods,
substitution with other goods does not matter (Economics 217 -Multinomial Choice Models,
n.d.).

Suggestion: Alternative models, like the nested logit or multinomial probit, which can
accommodate correlated choices or group similar options can be considered to account for the
complexities associated d)

Stating Hypotheses
Null Hypothesis (H0): There is no difference between the coefficients of the "Educ" and
"NEET" categories i.e., the coefficients of the variables are equal across the two categories.
This means that all predictors (e.g., age, education level, gender) have the same effect on both
"Educ" and "NEET" activity choices.

Alternative Hypothesis (H1): There are differences in the coefficients between the "Educ"
and "NEET" categories for at least one predictor, indicating that "Educ" and "NEET" cannot
be combined.

Wald Test

Test Statistic : chi2( 21) Prob > chi2

1717.38 0.0000

Here P-value<0.05

Since the p-value is highly significant (p < 0.05), we reject the null hypothesis that the
coefficients for the "Educ" and "NEET" categories are identical. This outcome indicates
substantial differences in the effect of predictors, such as age, education level, and gender, on
the likelihood of choosing 'Educ' versus 'NEET' activities. Therefore, combining the 'Educ' and
'NEET' categories is statistically inappropriate, as it would mask the distinct influence of these
predictors on each activity type.

15
e) The objective is to test if there are significant differences in the religion categories across
different activity choices (Education and NEET).

Stating Hypotheses
Null Hypothesis (H0): There is no difference in the effect of religion categories on the
likelihood of choosing a given activity

For Education: H0, 1: RRR (Muslim) 2=RRR (Christian) 3=RRR (others) 9

For NEET: H0, 2: RRR (Muslim) 2=RRRReligion (Christian) 3=RRR (others) 9

Alternative Hypothesis (H1): At least one religion category differs significantly in its effect
on the likelihood of choosing the activity.

Wald Test
For Education

Test Statistic : chi2( 2) Prob > chi2

7.78 0.0205

Here P-value<0.05

The chi-squared statistic 7.78 with a p-value of 0.0205 is significant. This result suggests that
we can reject the null hypothesis for the Education activity. Therefore, there are significant
differences in the effect of religion categories on the likelihood of choosing Education as an
activity. Different religious affiliations influence the likelihood of selecting this activity in
distinct ways.

For NEET

Test Statistic : chi2( 2) Prob > chi2

3.66 0.1602

Here P-value>0.05

NEET (Not in Education, Employment, or Training): Similarly, the chi-squared statistic of 3.66
with a p-value of 0.1602 also indicates no significant difference in religion effects on activity
choice for the "NEET_3" category. We fail reject the null hypothesis for the NEET activity
choice suggesting religious affiliation might not influence activity choices.

16
Overall, the contradicting results of both Hypothesis test can be due to various reasons such as
sociocultural or economic factors, for eg: if the education group has more youth actively
engaged in religious or community activities, it might drive a stronger association compared to
the NEET group. Particularly in the context of India, where religion and caste based reservation
tends to have strong correlation, plays a significant role in getting merit based admission in
various institutions.

Question 3: TOBIT Model

(a) The graphs provided illustrate various aspects of learning time in a dataset. Let's break down
each one:

Graph 3.1: Time Spent on Learning

The histogram shows an extreme concentration of observations near zero, with the frequency
dropping sharply as values increase. The majority of individuals spend little to no time on
learning, while a small subset allocates significantly more time.

Interpretation: This high frequency at the lower end could reflect that, in this sample, a large
portion of individuals do not engage heavily in learning activities. This might indicate barriers
to dedicating time to learning, such as lack of access, motivation, or competing time demands.
It suggests a right-skewed distribution where the mean is pulled by a small number of high time
learners.

17
 (PTO)
Graph 3.2: Natural Logarithm of Learning Time

Although the logarithmic transformation is often used to reduce skewness, the distribution of
ln_time_learn remains heavily skewed with a high concentration of values near the minimum.
This is partly because the transformation still reflects the original high frequency of low values
but compresses the range for larger values, making variations at higher levels of learning time
appear less distinct.

Interpretation: The log transformation provides a more compact view of time allocation across
individuals who do spend time learning, but it’s still clear that most observations fall at the
lower end. This indicates that the majority of people in the sample allocate minimal time to
learning, even after the log transformation.

Graph 3.3: Learning Time Share Out of Total Time

18
The histogram for sh_learn_time shows a similar trend with values clustered around zero,
indicating that the share of time individuals devotes to learning activities is very low for most
people. There are few observations where learning occupies a significant share of time.

Interpretation: This share variable, which is bounded between 0 and 1, shows that the
proportion of time allocated to learning remains minimal for most of the sample. This may
indicate that other activities dominate individuals' schedules, or that learning is not prioritized.

b)
TABLE 5: TOBIT MODEL (3 DEPENDENT VARIABLES)

Model 1: time_learn Model 2: ln_time_learn Model 3: sh_learn_time

Coefficient SE Coefficient SE Coefficient SE

Variables
Age -14.41*** -7.39 -0.0825*** -6.11 -0.0191*** -7.32

lnmpce 41.65*** -3.84 0.304*** -4.01 0.0517*** -3.54

hhsize -4.479 -1.59 -0.0149 -0.76 -0.00533 -1.41

female -11.86 -0.84 -0.0826 -0.84 -0.0178 -0.95

urban 29.99 -1.88 0.251* -2.25 0.0306 -1.43

femurb 5.893 -0.29 0.00149 -0.01 0.00856 -0.32

emp -727.7*** -37.4 -5.157*** -38.22 -0.983*** -37.66

neet -564.8*** -33.58 -3.944*** -33.73 -0.760*** -33.71

primary -9.069 -0.16 -0.0778 -0.2 -0.0171 -0.22

middle 17.34 -0.43 0.121 -0.44 0.0249 -0.47

high 51.83 -1.44 0.357 -1.45 0.0692 -1.44

grad 168.5*** -4.74 1.200*** -4.97 0.220*** -4.67

married -196.9*** -9.78 -1.469*** -10.64 -0.268*** -9.93

other_day -318.3*** -19.28 -1.875*** -16.37 -0.415*** -18.76

conven -21.87 -1.74 -0.171 -1.96 -0.0365* -2.17

nofuel 293.6*** -4.87 1.697*** -3.98 0.353*** -4.34

19
lighting -28.29 -0.72 -0.148 -0.54 -0.0263 -0.5

dwelling -36.21** -3.06 -0.196* -2.38 -0.0430** -2.7

scgroup 45.79* -2.52 0.352** -2.78 0.0569* -2.34

obcgroup 39.73** -2.59 0.292** -2.73 0.0482* -2.35

other_group 30.43 -1.67 0.233 -1.84 0.0333 -1.37

islam -60.15** -2.87 -0.474** -3.25 -0.0896** -3.18

Christi 155.6** -2.64 0.591 -1.4 0.186* -2.33

other_relig 121.5* -2.4 0.757* -2.13 0.155* -2.28

_cons 165.3 -1.7 3.416*** -5.05 0.267* -2.05

Variance 83192.4*** -29.92 4.151*** -29.2 0.150*** -29.81

N 8534 8534 8533

Standard errors in parentheses

*** p<0.01, ** p<0.05, * p<0.1

Interpretation
(i) Gender-Sector Interactions
The results show a consistent trend: females in rural areas spend less time on learning activities
compared to both rural males and urban females. The interaction variable for rural females is
statistically significant at the 5% level in Models 1 and 2, and at the 10% level in Model 3,
confirming that this effect is robust across different model specifications. Urban females also
spend less time on learning compared to urban males; however, this effect is minor, with the
urban female variable being statistically insignificant in Models 1 and 2 and reaching marginal
significance at the 10% level in Model 3. This suggests that differences in learning time for
urban females are limited. Additionally, males in urban areas allocate more time to learning,
with significance at the 10% level in Model 2, indicating a possible trend of higher engagement
among urban males.

(ii) Age

Age consistently shows a significant negative impact on time spent learning across all models,
suggesting that as individuals grow older, they allocate less time to learning. (Keeping other
factors constant)

20
 • Model 1: Each additional year of age reduces learning time by 14.4 minutes, indicating
that younger individuals may have more opportunities or incentives to engage in
learning.
• Model 2: Each additional year of age decreases learning time by 0.083% on a logged
scale, reflecting a smaller proportional decline with age.

• Model 3: The share of time dedicated to learning drops by 0.019 minutes with each
additional year of age.

Age remains statistically significant across all three models at the 5% level, reinforcing its
consistent negative relationship with time allocated to learning.

(iii) Activity
The type of primary activity (education, NEET, or employment) strongly influences learning
time across all models, with education and NEET individuals dedicating significantly more
time to learning than those who are employed.

• Model 1: Those involved in education activities spend 727.7 more minutes on learning
compared to employed individuals, emphasizing a major time commitment.

• Model 2: Those in education spend approximately 5.16% more time on learning,

showcasing a similar trend in proportional terms.

• Model 3: Education activity participants allocate about 0.983 more minutes of their
shared time to learning.

NEET individuals also spend more time learning than employed individuals, although less
than those in education, suggesting that while NEET individuals have more free time, they
do not prioritize learning to the same extent as those actively pursuing education.

(iv) Log of Monthly Per Capita Expenditure (lnmpce)

Higher household income, represented by lnmpce, positively impacts learning time across all
models. This suggests that individuals from higher-income households likely have greater
access to resources, enabling them to dedicate more time to learning.
(Keeping other factors constant)

• Model 1: Each unit increase in lnmpce is linked to a 41.65-minute rise in learning time.

• Model 2: lnmpce positively affects learning time by approximately 0.304% in

proportional terms.

• Model 3: The share of time dedicated to learning increases by about 0.052 minutes with
each unit rise in lnmpce.

This positive and statistically significant effect across all models underscores the role of income
in facilitating learning.

21
Most Appropriate Dependent Variable
Model 1, with time_learn as the dependent variable, captures the total time individuals spend
on learning. However, this variable’s distribution is positively skewed due to many zero values,
which may distort the interpretation. Model 3, using sh_learn_time as the dependent variable,
also suffers from skewness for the same reason, reflecting the proportion of time rather than
absolute time, which could limit interpretability.

Model 2, with ln_time_learn as the dependent variable, provides a more balanced distribution
due to its logarithmic transformation, which reduces skewness and minimizes the influence of
extreme values. This transformation allows coefficients to be interpreted as elasticities, offering
insights into proportional changes in learning time with respect to age, income, and other
variables. Therefore, Model 2 (ln_time_learn) emerges as the most suitable dependent
variable, providing clearer and more meaningful interpretations while delivering a better model
fit.

c) Based on the results from table 6, the findings as follows:

The analysis models time allocated to learning activities using three different dependent
variables—time spent in learning (in minutes), its logarithmic transformation, and the share of
time spent on learning. The coefficients indicate how various factors impact learning time. For
instance, the interaction between gender and sector variables suggests that females in rural
areas might have different learning time allocations compared to their urban counterparts,
highlighting potential gender and locational disparities. Age appears significant, with older
youth potentially allocating less time to learning, likely reflecting the shift towards employment
or other responsibilities. The activity categories (from the previous multinomial choice model)
serve as endogenous variables here, as the time allocated to learning could influence or be
influenced by one's primary activity (education, employment, or NEET status). Instrumental
Variable (IV) techniques were applied to address this endogeneity, ensuring unbiased and
consistent estimates for the activity coefficients, thereby capturing a clearer causal relationship.
Higher-income positively correlates with increased learning time, reflecting how greater
resources enable more learning opportunities. Urban females spend less time on learning
compared to urban males, while rural females allocate the least time, underscoring socio-
economic and possibly cultural barriers. The IV Tobit model’s adjustment provides clearer
insights into these relationships by controlling for endogeneity in activity status, delivering
more precise interpretations of learning behaviours across different demographic groups.

TABLE 6

(1) (2) (3)

Variable
time_learn sh_learn_time ln_time_learn
[Link]#1.r 0 0 0
SE (.) (.) (.)
[Link]#2.r 8.815 0.00353 -0.058
SE -19.14 -0.0259 -0.0383
[Link]#1.r 0.239 -0.000229 -0.0336
SE -18.12 -0.0246 -0.0323
[Link]#2.r 44.20* 0.0506 -0.0527

22
 SE -20.05 -0.0271 -0.0399
age -68.93*** -0.0931*** -0.0577***
SE -2.334 -0.00316 -0.00541
lnmpce 86.82*** 0.113*** 0.045
SE -14.16 -0.0191 -0.0262
2.edu_level -54.26 -0.079 -0.185
SE -87.68 -0.119 -0.263
3.edu_level 44.11 0.0606 0.0594
SE -62.27 -0.0848 -0.155
4.edu_level 180.0** 0.243** 0.102
SE -56.32 -0.0766 -0.145
5.edu_level 501.3*** 0.673*** 0.246
SE -56.15 -0.0763 -0.143
2.marital_status -418.6*** -0.567*** -0.561***
SE -25.28 -0.0342 -0.0989
2.type_of_the_day -224.2*** -0.289*** -0.553***
SE -20.27 -0.0277 -0.0439
hhsize -0.18 0.000393 -0.0194**
SE -3.376 -0.00455 -0.00654
2.cook_fuel -46.04** -0.0686** 0.0176
SE -16.01 -0.0216 -0.0287
3.cook_fuel 337.8*** 0.414*** 0.276*
SE -76.14 -0.0958 -0.115
[Link] -75.42*** -0.0970*** -0.0895**
SE -14.93 -0.0202 -0.0287
2.social_group 80.38*** 0.106*** -0.00214
SE -23.66 -0.0321 -0.04
3.social_group 58.31** 0.0757** -0.0404
SE -20.22 -0.0274 -0.0343
9.social_group 86.12*** 0.111*** -0.0546
SE -23.39 -0.0317 -0.0415
[Link] -116.9*** -0.166*** -0.138*
SE -26.91 -0.036 -0.0652
[Link] 169.5* 0.206* 0.434***
SE -68.88 -0.0917 -0.129
[Link] 53 0.105 -0.255
SE -126.6 -0.179 -0.305
[Link] 140.0* 0.171 0.278*
SE -68.77 -0.0922 -0.133
[Link] 367.9** 0.506** -0.0631
SE -118.8 -0.167 -0.0461
[Link] -891.6*** -1.211***

23
 SE -85.31 -0.115
[Link] -1594.5*** -2.164***
SE -89.76 -0.121
_cons 320.4* 0.470** 6.700***
SE -131.6 -0.179 -0.246
var(e.time_learn) 151768.8***
SE -4213.2

var(e.sh_learn_time) 0.278***

SE -0.0077

var(e.ln_time_learn) 0.292***

SE -0.0154
N 8534 8533 2149
Standard errors given below each variable
*** p<0.01, ** p<0.05, * p<0.1
---------END OF THE ASSIGNMENT-------

24