Introduction to Accounting Research
Lecture 1
Accounting research
Accounting research is interesting because it can provide useful information and insights for
regulators, auditors, tax consultants and other practicing accountants.
Areas of accounting research:
- Financial accounting
Deals with how managers produce financial information for economic agents
(investors, financial analysts) outside the organization and how these agents
respond to variations in accounting methods and estimates.
- Auditing
Deals with the general question how auditors audit financial statements and
examines the antecedents and consequences of variations in auditor characteristics
and work methods.
- Managerial accounting
Examines how economic agents within the organization (managers and employees)
produce and use accounting information.
Advanced data analytics play an important role everywhere → trading, risk analysis, loan
origination (can someone take a loan?) marketing and consumer behaviour, HR.
Research process
Steps:
1. Formulate a research question
- Yes/no questions are recommended
- Positive, not normative, questions
- Often questions about causal effects of X on Y
Research questions are most interesting when there is some tension or uncertainty
in the answers.
2. Literature review
- Work backwards
- Identify one or two key papers related to the research topic (top journal
publications or heavily cited)
- Read papers that are cited by the key papers
- Also read the papers which cite the key papers
3. Develop a theory → critical reasoning about the why
4. Formulate (testable) hypotheses
Hypotheses should be:
- Unambiguous (clearly stated)
- Simple
- Specify a specific (directional) relation between X and Y
- Testable
5. Design empirical tests of the hypothesis
* Qualitative/case study
* Experimental
1
, * Survey
* Archival (using data)
Main challenge = develop generalizable, credible inferences about causal
effects in the overall population from a sample of imperfect data.
6. Interpret results
Lecture 2
DeAngelo (1981)
Purpose of positive theoretical studies:
- Provide deeper insights into how two constructs (X & Y) are related
- Develop causal predictions that can then be tested empirically
- Some phenomena are hard or impossible to measure (such as auditor independence)
Research question = Does “low balling” (= requiring an audit fee less than the cost price)
on initial audit engagement impair auditor independence?
Motivation behind the question:
- Accepting an audit engagement for a price below your actual cost of auditing, is basically
the same as an unpaid audit fee →will only be collected if the client doesn’t terminate the
audit engagement → the client is less likely to terminate if he is happy with the auditor →
auditor does what client wants (i.e. is less independent).
Why does “low balling” exist?:
* Auditors incur higher costs in initial period to
acquire client-specific knowledge (A + K)
* Costs are lower and constant in future periods
(A2 … At = A)
* Future audit fees are increased to above
auditing cost (F2 … Ft = F)
* Value from obtaining the engagement:
π = (F1 – A1) + (F – A)/r
* Expectations of the future quasi rents lead to
competition for the initial audit engagement,
driving down the initial audit fee F1
How far can you increase the future audit fees F? → the client won’t switch auditors after
the fee increase in t = 2 as long as the PV of all future fees to the current auditor (F + F/r) is
lower than the PV of fees to another auditor plus the switching costs (SC).
The paper implies that competition will ensure zero profits for any ‘new’ auditor. Therefore,
the PV of the ‘new’ auditor’s fees equals the present value of his total costs (A 2 + A/r), where
A2 is A+K.
2
,So the PV of the fees of the current auditor is defined by:
F + F/r < A2 + A/r + SC → replace A2 by A+K: F + F/r < A + K + A/r + SC
multiply by r: rF + F < rA + rK + A + rSC
simplify: F(1 + r) < r(K + SC) + A(1 + r)
you get: F < r(K + SC)/(1 + r) + A
optimal fee (F*) current auditor = A + r(K + SC)/(1 + r) – ε
In t = 1, all competing auditors have rational expectations about the future advantages of
incumbency (quasi rents). Again, competition will drive profits down to zero. This means
that auditors will “low ball” until: π* = (F1* – A1) + (F* – A)/r = 0
From the last equation, we know that F* > A. So we know that
(F* – A)/r is positive. Therefore, F1* – A1 is negative; market
equilibrium requires that initial fees should be less than costs
(i.e. that there is “low balling”).
But, since “low balling” is sunk in future periods, it neither causes future rents nor impacts
future auditor independence. The expectation of future quasi rents does reduce
independence because these rents are only collected if the client doesn’t terminate the
engagement.
Conclusions:
- Regulation that decreases “low balling” doesn’t improve auditor independence
- The expected quasi rents increases “low balling” and it reduces auditor independence
Statistical refresher
Construct = abstract idea which is not directly observable or measurable and should be
operationalized for empirical testing → from the theory domain.
Variable = observable item which can assume different values and is used to measure a
theoretical construct → used in empirical analysis.
Example = intelligence (IQ), performance (ROA), earnings management (abnormal accruals),
information asymmetry (bid-ask spread).
Important determinants of the validity of the empirical analysis:
- Construct validity → the degree to which a variable (operationalization of a construct)
captures the underlying theoretical construct it is supposed to measure.
To what extent does IQ capture the construct of intelligence?
- Reliability → the degree to which a variable provides consistent estimates.
If you measure the same person’s IQ 10 times, do you get 10 similar results?
Categorical variable = takes category values and places a unit of observation into one or
several groups (e.g. nationality, ethnics etc.)
Quantitative variables = * Discrete random variable → a random variable that has a
specific set of possible values (throwing a dice)
* Continuous random variable → can take on a continuous
range of possible values
We use statistics generated from a sample to make inferences about the parameters (= how
many would vote for Biden? How many people are blond?) of the population.
3
, Population = all individuals/companies that are of interest to a researcher
Sample = a subset of the population
Sample selection bias = bias resulting from the non-random sampling of a set of
observations from the population
Illustrative example question = What is the mean value of a certain variable in the
population? E.g. how tall are students in the lecture hall on average?
Expected value/population mean = weighted average of all possible values of a random
variable (= height) using the probability of each outcome (p) as its weight:
𝜇X = E[X] = p1 * x1 + p2 * x2 + … + pn * xn
But when the entire population, and the probability of different outcomes is unknown, we
need to estimate the population mean using a sample → law of large numbers = the sample
average converges to the population mean 𝜇 as the sample size increases (= the sample
estimates get more accurate with a larger sample).
Question = how accurate is a statistic estimated from a sample?
Distributions
How to read a probability density function (PDF):
* On the y-axis → likelihood of observing a certain value
* On the x-axis → possible values of your random variable
* The total area under the PDF is 1
* The area between two values gives you the probability that an
observed value will be between these thresholds
Mathematical statistics:
- If repeated random samples of a given size 𝑛 are taken from a population where the
population mean is 𝜇 and the population standard deviation is 𝜎, then the mean of all
sample means 𝑥̅ is population mean 𝜇. This means that in expectation, the sample means
are going to give the true value.
σ
- The standard deviation of all sample means 𝑥̅ is , where 𝜎 is the population standard
√n
deviation and 𝑛 is the sample size.
- Since the square root of sample size 𝑛 appears in the denominator, the standard deviation
decreases as the sample size increases (= gets more accurate).
- Even if the population distribution of the underlying variable is not normal, the sampling
distribution of sample means will be approximately normal as long as the sample size is
large enough (Central Limit Theorem).
Implications = 𝑥̅ is a point estimate of 𝜇 (= the best guess of what the population mean is).
We can also say something about how accurate this point estimate is by giving a confidence
interval about how sure we are that the true value 𝜇 will be within a certain range of
σ
values around 𝑥̅ → 𝑥̅ ± z * , where the second part is the margin of error, and z depends
√n
of the desired level of confidence.
4
Lecture 1
Accounting research
Accounting research is interesting because it can provide useful information and insights for
regulators, auditors, tax consultants and other practicing accountants.
Areas of accounting research:
- Financial accounting
Deals with how managers produce financial information for economic agents
(investors, financial analysts) outside the organization and how these agents
respond to variations in accounting methods and estimates.
- Auditing
Deals with the general question how auditors audit financial statements and
examines the antecedents and consequences of variations in auditor characteristics
and work methods.
- Managerial accounting
Examines how economic agents within the organization (managers and employees)
produce and use accounting information.
Advanced data analytics play an important role everywhere → trading, risk analysis, loan
origination (can someone take a loan?) marketing and consumer behaviour, HR.
Research process
Steps:
1. Formulate a research question
- Yes/no questions are recommended
- Positive, not normative, questions
- Often questions about causal effects of X on Y
Research questions are most interesting when there is some tension or uncertainty
in the answers.
2. Literature review
- Work backwards
- Identify one or two key papers related to the research topic (top journal
publications or heavily cited)
- Read papers that are cited by the key papers
- Also read the papers which cite the key papers
3. Develop a theory → critical reasoning about the why
4. Formulate (testable) hypotheses
Hypotheses should be:
- Unambiguous (clearly stated)
- Simple
- Specify a specific (directional) relation between X and Y
- Testable
5. Design empirical tests of the hypothesis
* Qualitative/case study
* Experimental
1
, * Survey
* Archival (using data)
Main challenge = develop generalizable, credible inferences about causal
effects in the overall population from a sample of imperfect data.
6. Interpret results
Lecture 2
DeAngelo (1981)
Purpose of positive theoretical studies:
- Provide deeper insights into how two constructs (X & Y) are related
- Develop causal predictions that can then be tested empirically
- Some phenomena are hard or impossible to measure (such as auditor independence)
Research question = Does “low balling” (= requiring an audit fee less than the cost price)
on initial audit engagement impair auditor independence?
Motivation behind the question:
- Accepting an audit engagement for a price below your actual cost of auditing, is basically
the same as an unpaid audit fee →will only be collected if the client doesn’t terminate the
audit engagement → the client is less likely to terminate if he is happy with the auditor →
auditor does what client wants (i.e. is less independent).
Why does “low balling” exist?:
* Auditors incur higher costs in initial period to
acquire client-specific knowledge (A + K)
* Costs are lower and constant in future periods
(A2 … At = A)
* Future audit fees are increased to above
auditing cost (F2 … Ft = F)
* Value from obtaining the engagement:
π = (F1 – A1) + (F – A)/r
* Expectations of the future quasi rents lead to
competition for the initial audit engagement,
driving down the initial audit fee F1
How far can you increase the future audit fees F? → the client won’t switch auditors after
the fee increase in t = 2 as long as the PV of all future fees to the current auditor (F + F/r) is
lower than the PV of fees to another auditor plus the switching costs (SC).
The paper implies that competition will ensure zero profits for any ‘new’ auditor. Therefore,
the PV of the ‘new’ auditor’s fees equals the present value of his total costs (A 2 + A/r), where
A2 is A+K.
2
,So the PV of the fees of the current auditor is defined by:
F + F/r < A2 + A/r + SC → replace A2 by A+K: F + F/r < A + K + A/r + SC
multiply by r: rF + F < rA + rK + A + rSC
simplify: F(1 + r) < r(K + SC) + A(1 + r)
you get: F < r(K + SC)/(1 + r) + A
optimal fee (F*) current auditor = A + r(K + SC)/(1 + r) – ε
In t = 1, all competing auditors have rational expectations about the future advantages of
incumbency (quasi rents). Again, competition will drive profits down to zero. This means
that auditors will “low ball” until: π* = (F1* – A1) + (F* – A)/r = 0
From the last equation, we know that F* > A. So we know that
(F* – A)/r is positive. Therefore, F1* – A1 is negative; market
equilibrium requires that initial fees should be less than costs
(i.e. that there is “low balling”).
But, since “low balling” is sunk in future periods, it neither causes future rents nor impacts
future auditor independence. The expectation of future quasi rents does reduce
independence because these rents are only collected if the client doesn’t terminate the
engagement.
Conclusions:
- Regulation that decreases “low balling” doesn’t improve auditor independence
- The expected quasi rents increases “low balling” and it reduces auditor independence
Statistical refresher
Construct = abstract idea which is not directly observable or measurable and should be
operationalized for empirical testing → from the theory domain.
Variable = observable item which can assume different values and is used to measure a
theoretical construct → used in empirical analysis.
Example = intelligence (IQ), performance (ROA), earnings management (abnormal accruals),
information asymmetry (bid-ask spread).
Important determinants of the validity of the empirical analysis:
- Construct validity → the degree to which a variable (operationalization of a construct)
captures the underlying theoretical construct it is supposed to measure.
To what extent does IQ capture the construct of intelligence?
- Reliability → the degree to which a variable provides consistent estimates.
If you measure the same person’s IQ 10 times, do you get 10 similar results?
Categorical variable = takes category values and places a unit of observation into one or
several groups (e.g. nationality, ethnics etc.)
Quantitative variables = * Discrete random variable → a random variable that has a
specific set of possible values (throwing a dice)
* Continuous random variable → can take on a continuous
range of possible values
We use statistics generated from a sample to make inferences about the parameters (= how
many would vote for Biden? How many people are blond?) of the population.
3
, Population = all individuals/companies that are of interest to a researcher
Sample = a subset of the population
Sample selection bias = bias resulting from the non-random sampling of a set of
observations from the population
Illustrative example question = What is the mean value of a certain variable in the
population? E.g. how tall are students in the lecture hall on average?
Expected value/population mean = weighted average of all possible values of a random
variable (= height) using the probability of each outcome (p) as its weight:
𝜇X = E[X] = p1 * x1 + p2 * x2 + … + pn * xn
But when the entire population, and the probability of different outcomes is unknown, we
need to estimate the population mean using a sample → law of large numbers = the sample
average converges to the population mean 𝜇 as the sample size increases (= the sample
estimates get more accurate with a larger sample).
Question = how accurate is a statistic estimated from a sample?
Distributions
How to read a probability density function (PDF):
* On the y-axis → likelihood of observing a certain value
* On the x-axis → possible values of your random variable
* The total area under the PDF is 1
* The area between two values gives you the probability that an
observed value will be between these thresholds
Mathematical statistics:
- If repeated random samples of a given size 𝑛 are taken from a population where the
population mean is 𝜇 and the population standard deviation is 𝜎, then the mean of all
sample means 𝑥̅ is population mean 𝜇. This means that in expectation, the sample means
are going to give the true value.
σ
- The standard deviation of all sample means 𝑥̅ is , where 𝜎 is the population standard
√n
deviation and 𝑛 is the sample size.
- Since the square root of sample size 𝑛 appears in the denominator, the standard deviation
decreases as the sample size increases (= gets more accurate).
- Even if the population distribution of the underlying variable is not normal, the sampling
distribution of sample means will be approximately normal as long as the sample size is
large enough (Central Limit Theorem).
Implications = 𝑥̅ is a point estimate of 𝜇 (= the best guess of what the population mean is).
We can also say something about how accurate this point estimate is by giving a confidence
interval about how sure we are that the true value 𝜇 will be within a certain range of
σ
values around 𝑥̅ → 𝑥̅ ± z * , where the second part is the margin of error, and z depends
√n
of the desired level of confidence.
4
