Big questions in this lecture
Stockholm Doctoral Course Program in Economics Development Economics II: Lecture 4 (1st half)
Firms Masayuki Kudamatsu
IIES, Stockholm University
23 November, 2011
Big questions in this lecture
1. Is return to capital high for firms in
LDCs?
2. Are capital and labor misallocated
across firms within each industry in LDCs?
Is return to capital / labor
- heterogeneous across firms within industry? (Banerjee & Duflo 2005)
- Use treatment as IV for estimating
RCT on firms
- production function Deal with attrition bias
- Estimate eterogeneous impacts >Monopolistic competi
1. de Mel, McKenzie, & Woodruff
(2008)One of the 1st RCTs on
- interventions to firms in LDCs
- few firms, lots of heterogeneity,
Which is usually difficult
non-persistent outcomes ⇒ low power to detect impact (McKenzie 2011)
Estimated parameters used in
- subsequent research
Besley, Burchardi, & Ghatak (2011),
1.1 Research question
What’s the return to capital for
- microenterprises in Sri Lanka?
- Small firms employ half or more labor
Important?
force in LDCs
- Microfinance • Is production function convex?
- No credible estimate in the literature
Original?
- Feasible?
- Observed level of capital stock: reflect (unobservable) entrepreneurial ability
- Cross-sectional estimation misleading
- Those who apply for credit: selected sample of firms
- Exploiting exogenous change in credit access (e.g. Banerjee-Duflo 2004): not the population average return to capital
1.2 Experimental design
Sample: 408 microenterprises
- (invested capital: ≤100,000 LKR) in Sri Lanka
- e.g. grocery stores
203 in retail sales
- e.g. sewing clothes, making bamboo
205 in manufacturing
products, food processing
- 1. 10,000 LKR in cash 2. 20,000 LKR in cash 3. 10,000 LKR equipment of their choice
Treatments: grant in cash or in kind
The treatments are a big capital
- injection
55% or 110% of median initial level of
- invested capital
⇐ Benefit of RCTs on firms in LDCs:
you can cheaply create a big shock
Survey of across-firm RCTs by
- Bandiera et al. (2011) mostly cite studies from LDCs
1.3 Data
9 waves of quarterly surveys
- (April 2005 to April 2007) Profits: elicited directly
- Better than asking revenues and
- expenditures in detail (de Mel et al. 2009)
Table I: balancing on covariates
1.4 Program evaluation
�
4 g
Y it = T + δ t + λ i + ε it
it g =1
Y Outcomes for firm i at time t it g
T indicator of treatment of type g
δ t Survey round FE λ i Enterprise FE1.4 Program evaluation (cont.)
Results (Table II) ↑
Capital stock
- ↑
Profits
- Owner hours worked ↑ for 10,000
- LKR treatments Family/hired labor hours: no change
- Robust to outliners (Table III)
1.5 Estimating return to capital
π = β K + δ + λ + ε
it i it t i it
π Profit itK Capital stock in 100 LKR at time t it
⇐ Instrumented by the amount of capital grant in 100 LKR (0 for control, 100 for treatment group 1 & 3, 200 for the rest)
1.5 Estimating return to capital
(cont.)Results (Table IV) Return to capital:
- per month
5.85% Issues with 2SLS
(i) Exclusion restriction
(ii) Weak instruments(iii) LATE
(i) Exclusion restriction
Instrument can affect labor inputs
- (in quantity or in quality)
Owner labor hours ↑ indeed (Table
- II(5))
Deduct from profits the value of
- owner’s labor hours
⇒ 4.59% to 5.29% per month Digression: control for outcomes
Can we avoid this issue by
- controlling for owner labor hour? No. Owner labor hour: correlated
- w./ K it In general, when a treatment affects
- two outcomes A and B, DO NOT regress A on treatment dummy and B
(A |B = x) − E(A |B = x) E 1i 1i 0i 0i = (A − A |B = x) E 1i 0i 1i
- {E(A |B = x) − E(A |B = x)} 0i 1i 0i 0i 1st term: Causal effect on A conditional on • B = x for those treated
E (A |B = x) − E(A |B = x) 1i 1i 0i 0i = E(A − A |B = x) 1i 0i 1i {E(A |B + = x) − E(A |B = x)} 0i 1i 0i 0i
- 2nd term: Selection bias = x by
Those induced to have B
- treatment: DIFFERENT from those
(ii) Weak instruments
1st stage F-stat on excluded
- instrument: 27.81
(iii) LATE
Some reasons to believe the IV
- estimate is ATE, not LATE it
- treatment & observables that would correlate with β IV
- per year Market interest rate: 12% to 20%
- per year Why don’t these firms borrow? >Survey responses: missing mar
- treatment status
- If attrition in control is higher for
- higher-profit making firms, we overestimate treatment effect Solution: Lee (2009)
- Another example of application in
- development RCT: Ashraf et al. (2010)
outcome and sample selection
equations(2) treatment induces attrition
- monotonically (ie. either more or less, not in both directions)
- Identify the excess # of observations
- missing due to treatment Trim the upper or lower tails of the
- non-missing due to treatment = 5.2% of non-missing treated observations
- of profit distribution in treated observations
- Program effect: 404 to 754 LKR
- Return to capital: 2.6% to 6.7%
- Lower profits in baseline increase
- attrition probability
- Program effect: 404 to 754 LKR
Return to capital: 2.6% to
- Lower profits in baseline increase
- attrition probability
- households (ie. hh that consumes and produces at the same time) to understand
- correlate with bigger return to capital How different theories (credit /
- insurance) predict heterogeneity differently
- markets, we have
- If credit market is imperfect so that
- ∗ > ¯ + A + nw...
- E(marginal utility of capital)
- < E(marginal return to capital) �� Difference is bigg
- more risk averse (u (c)) >higher variance in output (Var (ε))
- Π it =β i D it γ D it · X
- δ δ · X + λ + ε
- for household asset and household
- s
- δ δ · X + λ + ε
- estimate heterogeneous treatment effects by covariate x, never fail to allow time FEs to differ across x
- credit market imperfection, inconsistent w/ insurance market imperfection In addition, female owners have
- significantly lower return to capital
- At very low capital stock, return to capital is high
- But firms in the sample are those already in operation
- Entry may require a large fixed cost
- Then why don’t they reinvest profits to grow?
Lack of saving institutions &
- development: what explains huge differences in TFP across countries Only recently economists start >looking at inefficient allocations of capital & labor as the source of low TFP in poor countries This paper establishes this as
- across manufacturing firms in China and India? What industry characteristics
- associate with more misallocation? How much would TFP go up if no
- misallocation?
- price or access to credit, in a given industry s, revenue productivity
- reallocated to firms with higher TFPR si
- product and thus has monopoly power over price
- si
- si si (revenue), K , and L .
- s ,
- ln (TFPR /TFPR ) in US, China,
- TFPR: dispersed a lot more in India
- and China than US
- variance of ln (TFPR si ) on industry-level covariates Those robustly correlated positively
- are:
- India: size restriction
- Small firms believed to be
- if no distortion in India and China? To answer this question, Hsieh &
- Klenow set up a monopolistic competition model to derive expressions for industry TFP
- Each firm in an industry produces
- differentiated product and thus has monopoly power to set their output price With so many firms in each industry,
- each firm takes other firms’ behavior as given
- model.
- international trade Verhoogen (2008) & Bustos (2010) • use this model to analyze plant-level data in Mexico / Argentine (& exploit
natural experiments: currency
- (1 − τ )P Y − wL − (1 + τ )RK
- output subsidies
- si
- each monopolistic firm i in industry
- thus becomes: σ −1 σ σ −1 σ max �(1 − τ Y )θ s YY Y
- The demand curve equation P si
- Industry TFP is expressed as:
- (by model assumptions for the
- si
- P . si α s Treating A (1 − τ )/(1 + τ ) as si Y K si si
- one term will help
- Estimates in the literature ranges 3 to
- As gains
- (P ) si si
- Only P Y observed, not Y si si si
- recover Y si
- θ s as each industry share, we can compare aggregate TFP gains by
- China: 86.6 - 115.1%
- India: 100.4 - 127.5 %
- US: 30.7 - 42.9%
- Contract enforcement
McMillan & Woodruff (1999)
- Banerjee & Duflo (2000)
- Macchiavello & Morjaria (2010)
- Impact of export
- Verhoogen (2008a)
- Bustos (2011)
No correlation btw. ∆K after
i
ˆ cf. β : weighted average of β w./ weight i being the relative size of the change in the regressor due to the instrument (Card 2001: 939)1.5 Estimating return to capital
(cont.)
Implications 5.85% per month ⇒ more than 60%
1.6. Robustness to attrition
Attrition rate differs across
(control) vs (treated)
14.3% 9.6%
Assume:
(1) treatment is exogenous both in
Then the bound obtained by
⇐
( 0.143 - 0.096 ) / (1 - 0.096)
Trim 5.2% of the upper or lower tail
Bounds obtained:
Bounds obtained:
6.7 %
⇒ Relevant is the upper bound.
1.7 Heterogeneous return to
capital
Use a simple model of agricultural
What individual-level observables
We can test which theory is relevant ⇒ HH w/ asset A , n working-age members, and ability θ solves: � ��
� εf (K , θ) − rK max E u
K ,B,A ,I +r (A − A ) + (nw − I ) K K K K
s.t. K ≤ A + I + B
K K
B ≤ ¯ B (credit constraint) A K ≤ A
I K ≤ nw With perfect credit & insurance
�
f (K , θ) = r
∗
Denote this level of capital by K
K B Marginal return to capital: higher if lower A, lower n, higher θ
� � f (K, θ ) H f (K, θ ) L r If insurance market is imperfect...
s s
i �
s
s st i it t i s
D Amount of capital grant in 100 LKR
itat time t
s X i’s s-th characteristics s
If credit constraint matters, γ < 0
s
size; γ > 0 for proxies of talent If insurance constraint matters,
γ > 0 for risk aversion and self-reported uncertainty in profits
You must have these theoretical predictions before designing the questionnaire
�
s s
Π + it =β i D it γ D it · X
i � s s s
t i it t i s
In a panel regression, if you
Results in Table V: consistent w/
1-8 Taking stock
⇒ Production function is not
non-convex
2. Hsieh and Klenow (2009)
Long-standing question in macro
2.1 Research questions
Are capital and labor misallocated
2.2 Factor misallocation
With no distortion in the output
P Y
si si
TFPR si ≡ P si A si =
1 α −α s s
K L
si si
must be equal for every firm i Why? If not, capital and labor will be
⇒ Output of such firms ↑
⇒ Output price (P si ) for such firms ↓
Each firm produces differentiated⇒ Eventually TFPR ≡ P A si si si
P si Y si ≡ P =
TFPR si si A si
1 α −α s s
K L
si si
So let’s measure the actual TFPR
in US, China, and India to see if they are equal within each industry For each si, we observe P A
si si
For industry s’s capital share α
Figure II plots the distribution of
si s
and India
Normalized by industry average
⇒ Evidence for distortions in factor
allocation across firms in India and China
2.3 What industry characteristics
correlates w/ more variation in TFPR?
Run a regression of industry-level
China: share of state-owned firms
⇐
China (Table XII) India (Table XIII)
2.4 Monopolistic competition model
How much productivity would go up
3 features
ie. No strategic interactions as in oligopoly models
Free entry to each industry & # of • variety endogenously determined Krugman (1980) applies this model
to intl trade, to explain intra-industry trade Melitz (2003) also applies this
Now the workhorse model of
2.4.1 Model
A representative final good
producer combines the outputs of S manufacturing industries by technology �
θ s
Y = Y
s � s
where θ s = 1 and
s � � σ σ −1 � −1 σ Each firm i in industry s produces
differentiated products by
1 α −α s s
Y = A K L
si si si si
Their profits:
Y si si si K si si si τ output distortions Y si size restrictions, transportation costs,
τ capital price distortions K si
2.4.2 Analysis
Profit minimization w.r.t. Y by final
good producer implies: σ −1 σ θ YY
s s
P =
si σ 1 Y si
This is the demand curve faced by
Firm i’s profit maximization problem
s si si K si si ,L �
− wL − (1 + τ )RK
si K si si
si
si
Y
si
) P
K si
) (1 − τ
Y si
σ (1 − τ
R (σ − 1)
= α s
K
si
Y
si
)P
Y si
(1 − τ
(σ − 1) σ
= (1 − α s ) w
si
L
FOCs for profit maximization of firm i give us the optimal level of inputs
Y 1 σ si = θ s YY σ σ−1 s is used to obtain
2.5 Gains from reallocation
TFP
α s si
1 −α s
L si �
α s � � i
K si �
i
) � σ σ −1 � �
1 −α s si
L
K
s
(A
i
= � �
1
−α ss
L
α s s
K
≡ Y s
si
Plug in the optimal level of K and
L si . Then, lots of tweaks in algebra
Use the demand curve equation for
obtains � � σ � � σ−1 � 1 TFPR −1
s
TFP = A
s si
TFPR si
i Now we need σ and A
si Use σ = 3
10
↑ more w/ higher σ, take lower bound si σ−1 σ A is obtained by
Y κ s α s s 1 −α K L si si
By using the demand curve equation,
we have TFPR s = TFPR si . Which means � � � 1
efficient σ−1 σ −1
TFP = A
s si i
Aggregating by Cobb-Douglas with
3. Other topics on firms in LDCs
cf. See Eric Verhoogen’s lecture notes
(2008b, 2008c) for these topics and
Ashraf, Nava, James Berry, and Jesse M Shapiro. 2010.
“Can Higher Prices Stimulate Product Use? Evidence from
a Field Experiment in Zambia.” American Economic Review 100(5): 2383-2413.Bandiera, Oriana, Iwan Barankay, and Imran Rasul. 2011. “Field Experiments with Firms.” Journal of Economic
we ask how the sausage is made?” Journal of Development
Economics 88(1): 19-31. de Mel, Suresh, David McKenzie, and Christopher Woodruff. 2010. “Returns to Capital in Microenterprises: Evidence from a Field Experiment.” Quarterly Journal of Economics 123(4): 1329-1372.McMillan, John, and Christopher Woodruff. 1999.
“Interfirm Relationships and Informal Credit in Vietnam.”
Quarterly Journal of Economics 114: 1285-1320.Melitz, Marc J. 2003. “The Impact of Trade on Intra- Industry Reallocations and Aggregate Industry