Quantitative Research Methods

Introduction and context

Quantitative research converts observations into numerical evidence so that patterns, relationships and differences can be examined systematically. In accounting and finance, it is used to investigate questions such as whether audit quality is associated with reporting quality, whether leverage predicts financial distress, whether investors respond to earnings announcements, or whether a budgeting intervention improves organisational performance.

Quantitative methods do not replace professional judgement or theoretical reasoning. They provide structured evidence for evaluating propositions derived from theory. A sound quantitative dissertation therefore links five elements: a research question, a conceptual framework, measurable variables, an appropriate design, and a transparent analytical procedure.

A useful distinction is between population and sample. The population is the complete set of entities relevant to the study, such as all listed manufacturing companies in a country. A sample is the subset actually observed. The researcher uses sample statistics to estimate population parameters, while recognising sampling and measurement uncertainty.

Learning objectives

By the end of this chapter, students should be able to:

1. Distinguish descriptive statistics from inferential statistics.
2. Define and operationalise accounting and finance variables.
3. Select appropriate survey, archival and observational designs.
4. Apply correlation, regression and hypothesis testing correctly.
5. Evaluate validity, reliability and statistical assumptions.
6. Interpret results without confusing association with causation.
7. Conduct robustness checks and report findings ethically and transparently.

From research question to measurable variable

A quantitative investigation begins with a question that identifies a relationship, difference or effect. For example, “Does audit committee independence affect financial reporting quality?” is more useful than the broad question “What affects reporting quality?” because it identifies a potential explanatory variable and an outcome.

A conceptual model translates the question into constructs and expected relationships. A construct is an abstract idea, such as audit quality, investor confidence or financial distress. An operational definition specifies how the construct will be measured. Audit committee independence might be measured as the proportion of committee members classified as independent under a relevant governance code. Financial distress might be measured using a distress score, an interest-coverage ratio or an indicator variable.

Variables may be classified as follows:

Measurement scales also matter. Nominal data identify categories without order. Ordinal data have an order but unequal distances may not be meaningful. Interval data have equal distances but no meaningful zero, while ratio data have equal distances and a meaningful zero. Accounting amounts and ratios are usually treated as ratio-scale variables, although their distributions may require transformation.

A hypothesis is a testable statement. A null hypothesis, H₀, commonly states that there is no relationship or difference. An alternative hypothesis, H₁, states that a relationship or difference exists. A directional hypothesis predicts the sign or direction, whereas a non-directional hypothesis predicts only that an effect exists.

Research designs and data sources

Archival and secondary data

Archival research uses existing records, including annual reports, regulatory filings, stock-exchange databases and commercial financial databases. Its advantages include relatively large samples, reduced respondent burden and the ability to study historical trends. Limitations include missing observations, inconsistent definitions, survivorship bias and differences in accounting policies.

Researchers must document the source, extraction date, inclusion criteria, treatment of missing values and any adjustments. For example, if firm size is measured as the natural logarithm of total assets, the study should state whether assets are measured at year-end and whether values are adjusted for inflation.

Surveys

A survey is appropriate when the variables concern perceptions, attitudes, practices or self-reported behaviour. A questionnaire should use clear items, avoid double-barrelled questions, define the reference period and provide balanced response options. A five-point agreement scale may range from “strongly disagree” to “strongly agree,” but the choice of scale should reflect the construct and respondent population.

Sampling may be probability-based, such as simple random, stratified or systematic sampling, or non-probability-based, such as convenience or purposive sampling. Probability sampling supports stronger generalisation when the sampling frame is adequate. Convenience sampling may be practical but introduces greater risk of selection bias.

Experimental and quasi-experimental designs

Experiments manipulate a treatment and compare outcomes with a control condition. In accounting research, an experiment might test whether different assurance labels affect users’ investment judgements. A quasi-experiment studies a policy or event without full random assignment. Difference-in-differences designs, for example, compare changes over time between an affected group and a comparison group. Such designs require careful consideration of whether the groups would have followed parallel trends in the absence of the intervention.

Descriptive statistics

Descriptive statistics summarise the observed sample; they do not by themselves establish causality or generalise beyond the sample.

For observations x₁, x₂, ..., x_n, the arithmetic mean is:

Here, x̄ is the sample mean, n is the number of observations and x_i is observation i. The mean is informative when values are reasonably balanced, but it is sensitive to extreme observations.

The median is the middle ordered observation, or the average of the two middle observations when n is even. It is often more representative than the mean for highly skewed variables such as firm size or executive compensation.

The sample variance and sample standard deviation are:

The denominator n−1 provides an unbiased estimator of population variance under the usual independent random-sampling framework. Standard deviation is expressed in the same units as the variable. The coefficient of variation is:

It is meaningful mainly for ratio-scale variables with a positive, non-zero mean and permits relative dispersion comparisons.

The range is the maximum minus the minimum. Quartiles divide ordered observations into four parts. The interquartile range, IQR = Q₃ − Q₁, is resistant to extreme values. A box plot can display the median, quartiles and potential outliers.

For proportions, percentages should be reported with the denominator. “Thirty companies, or 60% of the sample, reported losses” is more informative than “losses increased by 60%” when the latter could be confused with a percentage-point change or a relative increase.

Worked example: descriptive statistics

Suppose five firms have return on assets of 2%, 4%, 4%, 6% and 9%. The mean is:

x̄ = (2 + 4 + 4 + 6 + 9) / 5 = 5%

The ordered middle value is 4%, so the median is 4%. The deviations from the mean are −3, −1, −1, 1 and 4 percentage points. The sum of squared deviations is 28, and the sample variance is 28/(5−1) = 7 percentage-points-squared. Thus, the sample standard deviation is approximately √7 = 2.65 percentage points. The difference between the mean and median indicates some right-skew caused by the 9% observation.

Sampling, estimation and hypothesis testing

A sample statistic is subject to sampling variability. A confidence interval communicates plausible values for a population parameter under the model and sampling assumptions. For a mean with known population standard deviation, a 95% interval is:

where σ is the population standard deviation and z_0.975 ≈ 1.96 for a two-sided 95% normal-theory interval. In practice, σ is usually unknown, so a t critical value and sample standard deviation are used:

This interval assumes independent observations and a suitable sampling distribution; for small samples, approximate normality is particularly important.

Hypothesis testing normally follows these steps:

1. State H₀ and H₁.
2. Choose a significance level, commonly α = 0.05, before examining results.
3. Select a test consistent with the design, measurement scale and assumptions.
4. Calculate a test statistic and p-value or compare the statistic with a critical value.
5. Report the estimate, uncertainty and substantive meaning.

A p-value is the probability, assuming H₀ is true and the test assumptions hold, of observing a result at least as extreme as the one obtained. It is not the probability that H₀ is true, and statistical significance does not establish practical importance. A small effect can be statistically significant in a large sample, while an important effect may lack significance in a small or noisy sample.

A Type I error rejects a true null hypothesis; its probability is controlled by α. A Type II error fails to reject a false null hypothesis; its probability is commonly denoted β. Statistical power is 1−β. Researchers should consider effect sizes, confidence intervals and power rather than relying only on a binary significant/not-significant conclusion.

For two independent groups, a difference in means may be examined with a t-test when observations are independent and the outcome is approximately continuous. Welch’s t-test is generally preferable when group variances differ. Paired tests are suitable when the same units are measured before and after an intervention. Chi-square tests can examine association between categorical variables, but expected cell counts should be checked.

Correlation analysis

Pearson’s correlation measures linear association between two quantitative variables:

The coefficient lies between −1 and +1. A positive value indicates that larger values of one variable tend to accompany larger values of the other; a negative value indicates the opposite. A value near zero indicates little linear association, not necessarily no relationship. Pearson correlation is sensitive to outliers and can be misleading when relationships are nonlinear.

Spearman’s rank correlation is based on ranks and is useful for ordinal variables or monotonic relationships that do not meet Pearson assumptions. Neither correlation proves causation. A third variable, reverse causality, common trends or measurement error may explain an observed association.

Regression analysis

Simple linear regression

A simple regression models the conditional mean of an outcome:

Here, y_i is the dependent variable, x_i is the predictor, β₀ is the intercept, β₁ is the expected change in y associated with a one-unit increase in x, and ε_i is the error term. The coefficient is an association unless the research design supports a causal interpretation.

The fitted value is ŷ_i, and the residual is e_i = y_i − ŷ_i. Ordinary least squares chooses coefficients that minimise the sum of squared residuals:

The coefficient of determination is:

where SST = ∑(y_i − ȳ)² is total variation around the sample mean. R² is the proportion of sample variation explained by the fitted model under the chosen specification; it is not a percentage of causation and does not show that the model is correctly specified.

Multiple regression

Accounting and finance outcomes usually depend on several factors:

Each slope coefficient estimates the expected difference in y associated with a one-unit increase in its predictor, holding the other included predictors constant. This “holding constant” interpretation depends on the model and does not automatically remove omitted-variable bias.

Categorical variables can be represented by dummy variables. If industry has four categories, three indicators are normally included with one category omitted as the reference group. Including all four indicators alongside an intercept creates perfect multicollinearity, known as the dummy-variable trap.

An interaction can test moderation:

The marginal association of x is β₁ + β₃z. Interaction effects should be interpreted at meaningful values of z, rather than by interpreting β₃ alone.

Regression assumptions and diagnostics

Important assumptions include a correctly specified conditional mean, independent observations or appropriately modelled dependence, no perfect multicollinearity, and an error structure suitable for the reported standard errors. Normality of errors is not essential for unbiased ordinary least-squares coefficients in large samples, but it can matter for small-sample inference. Homoscedasticity means constant error variance; heteroscedasticity affects conventional standard errors and can make significance tests unreliable.

Diagnostics should include residual plots, leverage and influence measures, distributional checks, and a review of functional form. Variance inflation factors can indicate potentially problematic collinearity, although no universal cut-off should substitute for theoretical judgement. Clustered observations, such as repeated firm-year records, may require firm-clustered standard errors or a panel-data model.

Robust standard errors can address certain forms of heteroscedasticity, but they do not correct omitted variables, reverse causality or poor measurement. Transformations such as ln(total assets) can reduce skewness and make coefficients more interpretable, but the transformation and treatment of zero or negative values must be justified.

Survey analysis: validity and reliability

Reliability concerns consistency of measurement. Internal consistency is often assessed with Cronbach’s alpha:

Here, k is the number of items, σ_j² is the variance of item j, and σ_T² is the variance of the total score formed from the items. Alpha tends to increase with the number of items and their intercorrelation. A high alpha does not prove that the scale measures one construct, and very high values may suggest redundant items. Item wording, reverse-coded items and missing responses must be checked.

Validity concerns whether the instrument measures what it is intended to measure. Content validity asks whether items adequately cover the construct. Construct validity concerns convergent and discriminant evidence. Criterion-related validity examines association with a relevant external measure. Face validity, although weaker, concerns whether items appear appropriate to respondents.

Exploratory factor analysis may help identify dimensions when theory is incomplete. Confirmatory factor analysis is more suitable when a measurement structure is specified in advance. Factor analysis requires an adequate sample, meaningful correlations and defensible decisions about extraction and rotation; it should not be used simply to produce attractive results.

Common-method bias may arise when predictor and outcome data are collected from the same respondent using the same instrument at the same time. Procedural remedies include separating measurements, protecting anonymity and varying item formats appropriately. Statistical remedies can support, but cannot replace, careful design.

Robustness, reproducibility and ethical practice

A robustness check asks whether the conclusion survives a reasonable alternative analysis. Examples include:

• replacing a variable with a theoretically justified alternative measure;
• using median or winsorised treatment for extreme observations, with full disclosure;
• estimating models with and without influential observations;
• using heteroscedasticity-robust or clustered standard errors;
• testing alternative functional forms or lagged predictors;
• analysing relevant subgroups;
• checking whether conclusions change under different missing-data rules.

Robustness is not a licence to search until a preferred p-value appears. The researcher should distinguish confirmatory analyses specified in advance from exploratory analyses conducted after viewing the data. Selective reporting, p-hacking, undisclosed exclusion and HARKing—hypothesising after results are known—undermine credibility.

Missing data should be diagnosed rather than automatically coded as zero. Listwise deletion may reduce sample size and bias results if missingness is systematic. Imputation can be useful under defensible assumptions, but the method and sensitivity of results should be reported.

Ethical quantitative research requires informed consent for surveys, confidentiality, secure data storage, lawful use of personal data and honest reporting. Researchers must not identify respondents or firms unnecessarily, manipulate data, fabricate observations or present a model as causal without appropriate design. Data and code should be archived where legally and ethically possible, with a data dictionary explaining variable construction.

Software packages can calculate statistics efficiently, but they do not choose a valid design. AI tools may assist with coding, documentation or plain-language explanations, yet researchers remain responsible for checking generated code, protecting confidential data, identifying fabricated references and disclosing material assistance according to institutional policy.

Applied case: governance and reporting quality

Background

The following is an illustrative case. A student investigates whether audit committee independence is associated with financial reporting quality among 180 listed firms observed over three years. Reporting quality is measured using an accrual-based index, while independence is the percentage of independent audit committee members. Firm size, leverage, profitability and industry are included as controls.

Problem and analysis

The student first defines inclusion rules and constructs a firm-year dataset. Descriptive results show that independence averages 62%, but the median is 67%, suggesting some lower-independence observations. A scatter plot indicates a weak positive relationship, but several extreme firms are visible. Pearson correlation is positive, while Spearman correlation is slightly smaller, signalling sensitivity to skewness or influential observations.

The student estimates:

Firm-clustered standard errors are used because observations from the same firm may not be independent. The coefficient on independence is positive and statistically significant in the primary model. However, the coefficient becomes smaller when a different reporting-quality measure is used and is not significant in the smallest industry subgroup.

Interpretation and outcome

The appropriate conclusion is that the data provide evidence of a positive association in the primary specification, but the strength and generality of the evidence are limited. The study cannot claim that independence causes higher reporting quality because firms with stronger governance may differ in unobserved ways, and governance reforms may respond to reporting problems. The robustness results should be presented as qualification, not concealed as failure.

Learning points

The case demonstrates the importance of operational definitions, control variables, clustered inference, sensitivity analysis and cautious causal language. It also shows why a statistically significant coefficient is only one part of a research argument. Theory, measurement quality, design and transparent reporting are equally important.

Professional interpretation and reporting

A strong results chapter should report sample selection, descriptive statistics, a correlation matrix where appropriate, model specifications, coefficient estimates, standard errors, confidence intervals, sample sizes and key diagnostics. Tables should define every variable and state whether standard errors are conventional, robust or clustered. The narrative should explain magnitude and practical meaning. For example, a coefficient should be translated into the expected change in the outcome for a realistic change in the predictor, while preserving the predictor’s units.

Researchers should distinguish statistical significance, economic significance and policy relevance. They should report unexpected findings, limitations and alternative explanations. Reproducible workflows—using saved scripts, versioned data-processing steps and clearly named files—reduce transcription errors and permit review.

Chapter Summary

Quantitative research methods provide a framework for converting accounting and finance questions into measurable, analysable evidence. Descriptive statistics summarise distributions; confidence intervals express estimation uncertainty; hypothesis tests evaluate evidence against specified null hypotheses; and correlation describes association without proving causation. Regression models estimate conditional relationships, but interpretation depends on measurement, design, specification and assumptions. Survey research requires attention to sampling, validity and reliability. Diagnostics and robustness checks help assess whether conclusions depend on influential observations, alternative measures or fragile assumptions. Ethical practice requires accurate data handling, transparent reporting, protection of participants and responsible use of software and AI. The strongest dissertation integrates theory, design, evidence and cautious interpretation.

Key Terms

Construct: An abstract concept that must be represented through observable measures.

Operationalisation: The procedure used to convert a construct into a measurable variable.

Population: The complete set of units to which a study intends to relate.

Sample: The observed subset drawn from a population.

Descriptive statistic: A numerical summary of observed data.

Parameter: A numerical characteristic of a population.

Confidence interval: An interval procedure expressing uncertainty about a population parameter.

Null hypothesis: A statement usually representing no effect or no difference.

P-value: The probability of results at least as extreme as those observed, conditional on the null and model assumptions.

Type I error: Rejecting a true null hypothesis.

Statistical power: The probability of rejecting a false null hypothesis under specified conditions.

Correlation: A numerical description of association between variables.

Regression: A model for the conditional mean of an outcome given explanatory variables.

Residual: The observed outcome minus its fitted or predicted value.

Multicollinearity: Strong linear dependence among predictors.

Heteroscedasticity: Non-constant variance of regression errors.

Reliability: Consistency of a measurement instrument.

Validity: The extent to which a measure supports the intended interpretation.

Robustness check: An alternative reasonable analysis used to assess stability of conclusions.

Common-method bias: Distortion arising when variables are measured using the same method or source.

Review Questions

Short Answer Questions

1. What is the difference between a population parameter and a sample statistic?
2. Why can the median be preferable to the mean for a skewed financial variable?
3. What does a p-value represent, and what does it not represent?
4. Distinguish reliability from validity in survey measurement.
5. Why does correlation not, by itself, establish causation?

Long Answer Questions

1. Explain how a broad accounting research question can be developed into a conceptual model, operational definitions and testable hypotheses.
2. Compare archival, survey and experimental research designs, including their principal strengths and limitations.
3. Explain the assumptions of ordinary least-squares regression and describe suitable diagnostic procedures.
4. Discuss how statistical significance, confidence intervals and economic significance should be integrated in dissertation reporting.
5. Explain how missing data, influential observations and alternative variable definitions can affect quantitative conclusions.

Analytical / Application Questions

1. A researcher finds a positive correlation between leverage and financial distress but obtains a non-significant leverage coefficient after adding firm size and profitability. Provide at least three theoretically plausible explanations and propose diagnostic or robustness checks.
2. Design a survey study examining whether perceived audit transparency affects investor confidence. Specify the population, sampling approach, constructs, example measures, reliability and validity procedures, and ethical safeguards.
3. A regression coefficient indicates that a one-unit increase in an independence ratio is associated with a 0.04-unit increase in reporting quality, with a 95% confidence interval from 0.01 to 0.07. Explain the statistical and practical interpretation, identify what cannot be concluded, and suggest one additional analysis.

Multiple Choice Questions with Answers

Question 1

Which statement best describes a p-value?

A. The probability that the null hypothesis is true

B. The probability of results at least as extreme as those observed, assuming the null hypothesis and model assumptions hold

C. The probability that the research hypothesis is false

D. The percentage of variation caused by the independent variable

Correct Answer: B. The probability of results at least as extreme as those observed, assuming the null hypothesis and model assumptions hold

Question 2

For the five observations 2%, 4%, 4%, 6% and 9%, what is the median?

A. 2%

B. 4%

C. 5%

D. 9%

Correct Answer: B. 4%

Question 3

Which statistic is generally most resistant to extreme observations?

A. Mean

B. Variance

C. Median

D. Range

Correct Answer: C. Median

Question 4

In a multiple regression, the interpretation of a slope coefficient usually refers to the expected change in the dependent variable when the predictor changes by one unit while:

A. All observations are deleted

B. The other included predictors are held constant

C. The p-value is set to zero

D. The residuals are forced to be normally distributed

Correct Answer: B. The other included predictors are held constant

Question 5

What is a principal consequence of heteroscedasticity when conventional standard errors are used?

A. The dependent variable becomes categorical

B. The sample size automatically doubles

C. Inference about statistical significance may be unreliable

D. Correlation becomes proof of causation

Correct Answer: C. Inference about statistical significance may be unreliable

Question 6

Which measure is most directly concerned with the internal consistency of a multi-item survey scale?

A. Cronbach’s alpha

B. R-squared

C. The interquartile range

D. The variance inflation factor

Correct Answer: A. Cronbach’s alpha

Question 7

If a categorical variable has four groups and a regression includes an intercept, how many dummy variables are normally included?

A. One

B. Two

C. Three

D. Four

Correct Answer: C. Three

Question 8

Which statement about R-squared is correct?

A. It proves that the model is causal

B. It is the proportion of sample variation explained by the fitted model under its specification

C. It is always larger when the model is theoretically better

D. It is the probability that all coefficients are correct

Correct Answer: B. It is the proportion of sample variation explained by the fitted model under its specification

Question 9

Which practice is an appropriate robustness check?

A. Changing the analysis repeatedly until a preferred p-value appears without disclosure

B. Replacing a measure with a theoretically justified alternative and reporting the result

C. Deleting all observations with inconvenient values without explanation

D. Treating every missing value as zero

Correct Answer: B. Replacing a measure with a theoretically justified alternative and reporting the result

Question 10

What does a positive correlation coefficient close to zero most directly indicate?

A. A strong positive causal effect

B. Little linear association, although nonlinear association may still exist

C. That the null hypothesis is certainly false

D. That the variables have identical units

Correct Answer: B. Little linear association, although nonlinear association may still exist