MA Task 4 (Syndicate 2)

Jolin Guan
Nick Gao
Virginia Li
Zhengwei Zhong
Jason Yang
Q1
The assumption is that the demand curve is linear, fixed and variable costs don’t affect the
revenue, the price elasticity of demand for the ice cream is 0.895.

Q ! − Q" 180 − 200

[(Q ! + Q" )/2] (180 + 200)/2
Price elasticity of demand = = = −0.8947
P! − P" 4.5 − 4
[(P! + P" )/2] (4.5 + 4)/2
Q2
We assume that there is a linear demand curve and fixed and variable costs are not
affected, let’s assume the cups of coffee he will buy per week when the price is $5.00 is x
and apply the midpoint formula.
x−8
(x + 8)/2
−1.3 =
5−6
(5 + 6)/2

x=10.144, Peter will buy 10 cups of coffee when the price is $5 per cup
Q3
a)
An instrument in this context should only affect sales through the price of the banana. A
promotion might attract customers and raise their interest about banana even price is not
changed, hence, it might affect the dependent variable. Whether it is a stormy period is
unlikely to affect sales directly but only through its effect on price. Hence, the Stormy
dummy variable could be an instrumental variable to study price elasticities. Apple as a
substitute of banana can also affect sales of banana, but unlikely to affect banana price.
Hence, can be included in the regression to reduce error.

b)
We first calculate the correlation between Banana price and Stormy, it is shown that they
are negatively correlated with a correlation coefficient of -0.67, indicating a strong IV.
Then, we run a regression with Stormy as the instrumental variable that explains price. The
result is shown as below:

It is shown that each variable has parameter estimate significantly different from 0 (at 5%
level). The price elasticity is shown as -1.12, indicating that customers are price elastic.
Promotion is positively related to sales controlling for banana price. The cross price elasticity
of Apple is negative, indicating the apple is more of a complement instead of a substitute for
banana, which is different from what we think.
There are several test diagnostics to see whether the instrumental variable regression is
valid, and their results are shown below:
We reject the null hypothesis in the Weak Instrument test, with a statistic>10, which
indicates that the relationship between the instrumental variable and the dependent
variable is strong. The Wu-Hausman test is also rejected indicating that PriceBanana does
suffer from endogeneity problem and instrumental variable is needed. As we have only one
instrumental variable, the Sargan test does not return any value.

Hence, we can conclude that consumers are price elastic towards bananas, a reduction in
price would attract more customers and maximise profit.

c)
We would ideally need further instruments. Transportation and packaging cost could
be both exogenous and relevant. If we could obtain a time series, we may be able to
use lagged price values as instruments. If we cannot obtain any instruments but need
to know the exact price elasticity, then we may need to run a randomised controlled
trial (or regional experiment/ quasi-experiment) where price is the treatment.

Q4

See figure above for the causal diagram. SizeProportion is the proportion of avocados sold, as
the proportion of small Hass, Large Hass, XLarge Hass, small bag, large bag, XLarge bag
summing up to 1. SizeProportion, Region and Type can both affect the change in AveragePrice
and TotalVolume. PLag, Gas, Wage, and OtherP have direct causal relationships with
AveragePrice. Region also causally affects Gas and Wage. Here, we assume that the lag of
price only directly affects the current price, similarly, the previous lag price (lag2) only directly
causally affects the last lag. Also, different regions would have causal effects on the average
price and total volume sold, as well as the gas and wage.
The relationship of most variables seems to lead to our standard IV situation with price. Type,
Region and SizeProportion (the effect of size proportion is however insignificant) are known
confounders that confound the causal relationship between AveragePrice and TotalVolume.
They need to be included to obtain accurate causal effects, otherwise, we have omitted
variable biases. Furthermore, there might still be unobserved confounders, such as income
per person which would affect both the AveragePrice and TotalVolume. Hence, to deal with
such situation, we recommend using 2SLS regression that utilises one or more instrumental
variables for the endogenous variable AveragePrice.

b)
We need to apply our general framework:
Contextual:
For the association of avocado retailers, we are just told to ‘make more money’. We do not
have information on the cost structure, such as fixed and variable costs. So, we can only
provide recommendations based on increasing revenue based on our analysis of the price
elasticity of avocados. Importantly, profitability of a company is a result of comprehensive
success; price is an important factor but not the only factor that contributes to profitability.
Considering the information on hand, we can only make suggestions on whether to increase
or decrease price.
Theoretical:
Based on information on hand, we generated a causal diagram and assumed that we are
dealing with unobserved confounders and need to use 2SLS regression. Now, the problem
becomes that “which possible variable could be a theoretically valid instrumental variable
(IV)”.
a) PLag: the price of the product of the previous period may be used as IV to address
endogeneity concerns. However, the key assumption is that Price_lag has no direct
causal effect on the dependent variable or on the unobserved confounders. Else, it is
not a valid IV.
b) Wage or Gas: wage or gas cannot be chosen as IVs based on our causal diagram due
to the precence of confounder Region.
c) OtherP: the price of the other type of avocadoes could be an instrument as it may be
subject to similar cost shifters as this type. Similarly, the price lag or the previous lag
of the other type can be IVs, the problem is whether they are strong instruments. The
key assumption here is that these variables have no direct causal effect on the
dependent variable or on the unobserved confounders. Else, it is not a valid IV.
Analytical:
Following the causal diagram above, we run an OLS and a 2SLS regression for the price
elasticity. The OLS model1 gives an elasticity of -1.64254 (p= 1.41e-06). Best IV regression
seems to be using PLag and PLag2 as instruments2. The price elasticity coefficient = -1.52243
(p = 0.000828).
Our diagnostics:
Statistics p-value
Weak Instrument 18.971 <0.0001
F(2,41)

1
lm(log(TotalVolume)~ log(AveragePrice)+Type+log(LabourWage)+Region,data=data)
2
ivreg(log(TotalVolume) ~ Type+Region | log(AveragePrice) | log(OtherP)+log(Plag))
 Wu-Hausman 0.153 0.698
F(1,41)
Sargan 0.006 0.941
Chi-square (1)
The weak instrument test shows that instrumental variables have strong correlations with our
independent variable (as the test statistics being 18.9>10), as we reject the null of weak
instrumental variables. The Wu-Hausman test shows a p-value significantly larger than 0.05,
suggesting that we cannot reject the null hypothesis that 2SLS regression and OLS regression
and are equally consistent. The Sargan test gives a p-value much greater than 0.05, showing
that we have no problem of overidentification.
Conclusion:
The 2SLS regression gave us an estimate of price elasticity being -1.522, with a p-value less
than 0.001, suggesting that the estimate is significant. As it is over -1, it suggests that
customers are price elastic, indicating that an increase in price will lead to a decrease in
demand such that revenue decreases. To increase revenue, we suggest the company to
decrease the price for avocados.
However, there are several concerns. First concern is regarding using PLag or PLag2 as
instruments, as it is still plausible that these variables have direct causal effect on the
dependent variable or on the unobserved confounders. Plus, the 2SLS regression fails to pass
the Wu-Hausman test, suggesting that it may be equally consistent with OLS regression and
that based on the data available we cannot identify an endogeneity problem. Lastly, due to
the dataset being extremely small, we don’t have enough evidence to support our conclusion,
further testing need to be conducted on larger dataset.