ACADEMIC WRITING GUIDE

How to Conduct and Report Regression Discontinuity Design in a Dissertation (With Worked Examples)

By Acadelyra Editorial TeamPublished

A university offers academic support to students whose eligibility score falls below a fixed threshold. Students just below and just above that threshold may have similar underlying prospects, yet face different access to support. Regression discontinuity design (RDD, also called RD) uses that change in assignment to investigate a local treatment effect.

This guide explains how to conduct and report regression discontinuity in a dissertation, from checking the assignment rule to estimating a discontinuity, examining threats to validity and writing a careful results paragraph. A reproducible educational example connects descriptive means, local regression, uncertainty and bandwidth sensitivity.

All students, scores and outcomes in the worked example are synthetic teaching data. Numerical results were independently verified in Python and JavaScript. R was unavailable, so the R syntax was reviewed but not executed. The calculated confidence intervals are conventional HC3 Wald illustrations; robust bias-corrected inference was not performed. Adapt the reporting templates to your own executed research and institutional authorship requirements.

What is regression discontinuity design?

RDD studies an intervention whose assignment changes abruptly at a known value of an observed score. That score is the running variable, also called the assignment or forcing variable. The cutoff is the threshold in the assignment rule. The outcome is measured after assignment, while the running variable must describe the information actually used to allocate treatment.

An eligibility score is one possible running variable. An income measure, examination mark or population count can also determine access. Merely dividing a dataset into low and high scores does not create an RDD: the threshold must change an intervention, offer or receipt probability in the real allocation process. Imbens and Lemieux's guide provides the classic practical introduction.

In a sharp RDD, treatment receipt follows the cutoff rule exactly. In a fuzzy RDD, crossing the cutoff changes the probability of receiving treatment, but some eligible people do not participate or some ineligible people do. Record eligibility, the offer and actual receipt separately; they answer different questions.

When is RDD appropriate?

Start with administrative evidence: the written allocation policy, original score records, treatment dates, exceptions and implementation history. Establish which score version was used and whether staff could override it. Specify the outcome, estimand, cutoff, analysis window and sensitivity strategy in your dissertation data analysis plan.

Decisions that determine whether a dissertation can use RDD
Design questionEvidence to seekConsequence for analysis
Did a threshold actually change assignment?Allocation records and treatment receipt on both sidesWithout a jump in assignment or receipt, a cutoff comparison does not identify the intended treatment effect
Is the original running variable available?Unrounded administrative scores recorded before treatmentRounded or reconstructed scores can misclassify eligibility and weaken local comparisons
Are there enough observations and distinct scores near the cutoff?Counts, score support and missingness by sideA large overall sample cannot compensate for sparse local information
Did something else change at the same threshold?Policies on scholarships, teaching and assessmentThe discontinuity may identify a package rather than the support programme alone
Could students or staff precisely sort across the threshold?Rescoring procedures, appeals and density evidenceSelective placement can undermine the design
Can outcomes be measured comparably on both sides?Common assessment procedures and follow-upDifferential measurement or outcome availability can create an artificial jump

RDD is unsuitable when the threshold was chosen after inspecting outcomes, treatment has no discontinuity there, the relevant score is missing, or a substantial gap leaves no observations near the threshold. If eligibility depends jointly on several scores or thresholds, a simple single-score analysis needs further justification and often a different design.

Why compare observations near the cutoff?

The intuition is that a small score difference should normally be associated with a small difference in expected outcomes. If treatment access changes abruptly while other determinants evolve smoothly, an outcome jump at that location can isolate an intervention effect.

The continuity-based approach assumes that the conditional mean of each potential outcome, with and without treatment, is continuous at the cutoff. We cannot observe both potential outcomes for any student, so this is an identifying assumption supported by knowledge of allocation and falsification evidence. Lee and Lemieux explain the design's interpretation and limits.

Students near the cutoff were not literally randomized simply because their scores are close. A separate local-randomization framework makes additional assumptions within an explicit window. Do not treat that framework as an automatic justification for a continuity-based analysis.

Continuity also does not mean that baseline achievement must be constant across scores. Outcomes may trend with the score, and slopes can differ across the threshold. The question is whether untreated prospects would have jumped at that exact point. Interference, such as supported students sharing programme materials with nearby controls, can further complicate the treatment contrast.

Manipulation, sorting and the meaning of local effects

Suppose staff selectively lower borderline scores for students they expect to benefit most. The students on each side may then differ in unobserved motivation or need. An institutional promise of objective grading is insufficient; inspect how decisions were actually made. Conversely, awareness of a cutoff alone does not establish precise manipulation.

In a sharp continuity-based RDD, the target is the average treatment effect at the cutoff. It is not automatically the average effect for every eligible student or the whole university. The students far below the threshold may face different barriers, receive different treatment intensity or respond differently. Moving the cutoff is a policy question requiring additional assumptions about effects away from the original threshold.

Worked example: academic support below an eligibility score

Consider 400 synthetic students with a cutoff of 50. A lower eligibility score indicates greater support need. Define x = score - 50 and Z = 1 when the score is below 50; otherwise Z = 0. Every eligible student receives the support and every ineligible student does not, so actual treatment D = Z in this main example. A score exactly equal to 50 is ineligible, although no generated record equals the cutoff.

The later outcome is an attainment score measured in points. It is a different assessment from the eligibility score. The deterministic construction is:

Student index i = 1, ..., 400
x_i = (i - 200.5) / 20
score_i = 50 + x_i
Z_i = 1 if x_i < 0; otherwise 0
e_i = ((37 * i modulo 101) - 50) / 20
Y_i = 60 + 0.4*x_i + 6*Z_i + 0.15*x_i*Z_i + 0.04*x_i^2 + e_i

The modular sequence supplies reproducible variation without a random seed. It is not evidence about a real population or an independent random sample. The construction embeds a treatment component of 6 + 0.15*x; its cutoff value is 6 points. We know that value because we wrote the teaching model, not because a fitted regression reveals a real causal truth.

Example A: descriptive means within one score point

First restrict attention to students with abs(x) < 1 and compute unweighted means. The rule selects 20 students on each side, giving 40 students in this descriptive window.

Descriptive outcome means within one eligibility-score point of the cutoff
GroupNumber of studentsMean outcome, points
Below 50, receiving support2065.698325
Above 50, not receiving support2060.055825
Below minus above40 total5.642500

The descriptive difference is 65.698325 - 60.055825 = 5.642500 points. It compares the observed means of two finite score intervals. It does not adjust for within-window slopes or estimate the two limiting regression functions at the cutoff. Report it as a descriptive comparison, rather than calling it the formal RDD estimate.

Example B: fit separate local lines at the cutoff

Fit a local linear model using observations with abs(x) < h:

Y = alpha + beta*x + tau*Z + gamma*x*Z + residual
weight = 1 - abs(x)/h, within the chosen window

The interaction allows separate slopes. The fitted limit just above the threshold is alpha; just below it is alpha + tau. Because treatment is on the lower-score side, tau is the below-minus-above discontinuity. This sign convention must be stated explicitly: many RD packages default to above minus below.

For the teaching bandwidth h = 6, there are 120 observations below and 120 above the cutoff, giving 240 observations with positive triangular weights. The fitted below-cutoff limit is 65.881054 points, and the above-cutoff limit is 59.787307 points. Their difference, using unrounded coefficients, is 6.093747 points. The model therefore estimates a positive discontinuity of about 6.09 outcome points in the support direction.

This estimate differs from the descriptive gap because it fits score trends and extrapolates each local line to the boundary. It also differs slightly from the constructed cutoff effect because finite-sample variation and local approximation affect the fit. A real causal interpretation would require the allocation and continuity arguments; the synthetic calculation supplies neither institutional evidence nor a diagnostic test of those arguments.

Calculate an explicitly conventional confidence interval

The following calculation uses a heteroskedasticity-consistent HC3 sandwich covariance for the weighted local regression. With design matrix X, diagonal weight matrix W, fitted residuals u_i and weighted leverages l_i, the covariance calculation is:

A = inverse(X' W X)
l_i = w_i * X_i' A X_i
V_HC3 = A [sum_i w_i^2 * u_i^2 / (1-l_i)^2 * X_i X_i'] A
SE(tau) = sqrt(the tau diagonal element of V_HC3)
Conventional approximate 95% Wald CI = tau +/- 1.96*SE(tau)

For this fit, the HC3 standard error is 0.424609 points. Using unrounded values, the margin is 0.832234 points, and the conventional approximate 95% interval is 5.261513 to 6.925981 points. Rounded for prose: estimate 6.09 points, SE 0.42, conventional 95% CI 5.26 to 6.93.

HC3 is not robust bias correction for RDD. It addresses a variance calculation, not the smoothing bias of a local polynomial. This normal-approximation interval demonstrates computation; it is not a coverage guarantee for these deterministic data. In an actual study, dependence within classes or schools also requires a defensible clustering strategy, and few independent clusters can make standard approximations unreliable.

For a real dissertation, a confidence interval communicates uncertainty under the design and inference assumptions. It does not assign a probability that a fixed effect lies inside the realized interval. Our guide to calculating and interpreting confidence intervals develops that distinction.

Bandwidths, kernels and polynomial choices

The cutoff is a boundary for each side's regression. Each fitted line uses observations from its own side and is evaluated at that boundary. A triangular kernel gives more fitting weight to observations closer to the cutoff. These weights express proximity, not the probability of being treated or the reliability of an individual student's test score.

Bandwidth controls the local sample. A narrower window generally reduces exposure to curvature but leaves less information; a wider window often improves precision while risking more approximation bias. The best balance depends on the data and objective, rather than a universal number of score points. Report the chosen selector, both side-specific bandwidths if they differ, and local sample sizes.

The verified example uses a prespecified teaching bandwidth, not an estimated optimal bandwidth. To show sensitivity, the same model and covariance calculation were fitted to the same data at narrower and wider windows.

Bandwidth sensitivity using the same synthetic students and triangular local linear fits
Bandwidth on each side, score pointsStudents below / aboveBelow-minus-above estimateHC3 SEConventional approximate 95% CI
480 / 806.1702570.5275605.136239 to 7.204275
6, primary teaching choice120 / 1206.0937470.4246095.261513 to 6.925981
8160 / 1606.0772990.3657885.360356 to 6.794243

The positive estimates change modestly across these particular windows, and the displayed SE decreases as the window widens. This is a description of these computed fits, not proof of unbiasedness, correct coverage or stability under every reasonable specification. Do not choose the window because it gives the most attractive significance result.

Local linear estimation is a sensible starting point. A local quadratic specification can be a justified sensitivity analysis, and a higher-order local fit can estimate bias in an inference procedure. Those uses differ from fitting one high-order polynomial across the entire score range. Gelman and Imbens explain why global high-order polynomials can produce unstable estimates and unreliable intervals.

Why robust bias-corrected inference matters

An MSE-oriented bandwidth balances squared bias and variance for estimating the discontinuity. It does not ensure that the remaining bias is negligible for an ordinary confidence interval. Calonico, Cattaneo and Titiunik develop inference that estimates the leading bias and accounts for the uncertainty introduced by that correction.

Robust bias-corrected (RBC) inference therefore involves both a corrected centering and an appropriate variance calculation. Replacing an ordinary SE with an HC3 SE does not accomplish this. Calonico, Cattaneo and Farrell also distinguish bandwidth choices aimed at point estimation from choices aimed at confidence-interval coverage.

When using an established RD implementation, report which point estimate is displayed and which robust interval accompanies it. Do not construct an interval by casually combining a conventional point estimate with an SE belonging to a different bias-corrected statistic. Retain the software output and exact options.

Reproduce the example in R

The first block creates every student and reproduces the conventional weighted-fit procedure without additional packages. Its covariance formula matches the independently executed Python and JavaScript calculations. The R block itself was not run in preparing this article.

i <- seq_len(400)
d <- data.frame(id = i, x = (i - 200.5) / 20)
d$score <- 50 + d$x
d$Z <- as.integer(d$x < 0)
d$Y <- 60 + 0.4*d$x + 6*d$Z + 0.15*d$x*d$Z +
  0.04*d$x^2 + ((37*i %% 101) - 50)/20

near <- subset(d, abs(x) < 1)
aggregate(Y ~ Z, data = near, FUN = mean)

local_hc3 <- function(h) {
  a <- subset(d, abs(x) < h)
  a$w <- 1 - abs(a$x)/h
  fit <- lm(Y ~ x * Z, data = a, weights = w)
  X <- model.matrix(fit)
  w <- a$w
  A <- solve(crossprod(X, w * X))
  lev <- w * rowSums((X %*% A) * X)
  adjusted <- w * residuals(fit)/(1 - lev)
  V <- A %*% crossprod(X, adjusted^2 * X) %*% A
  tau <- unname(coef(fit)["Z"])
  se <- sqrt(V["Z", "Z"])
  c(h = h, n = nrow(a), estimate = tau, SE = se,
    lower = tau - 1.96*se, upper = tau + 1.96*se)
}
t(vapply(c(4, 6, 8), local_hc3, numeric(6)))

The official R documentation for weighted linear models explains the fitting interface. The manually constructed covariance above is used instead of the default model-based SE returned by summary(lm(...)).

Use established RD tools for a fuller analysis

The next block is an unexecuted extension, not the source of any numerical result in this article. It requires the rdrobust package. The maintainers' rdrobust documentation describes the polynomial orders, bandwidths, kernels and variance options.

# Requires rdrobust to be installed in your own R environment.
library(rdrobust)
rdplot(y = d$Y, x = d$x, c = 0, p = 1,
       x.label = "Eligibility score minus 50",
       y.label = "Later attainment score")

# Data-driven bandwidths and robust bias-corrected inference:
rbc <- rdrobust(y = d$Y, x = d$x, c = 0,
                p = 1, q = 2, kernel = "triangular",
                bwselect = "mserd", vce = "hc3")
summary(rbc)
rbc$bws
rbc$coef
rbc$se
rbc$ci
packageVersion("rdrobust")
sessionInfo()

Here p specifies the local estimation order and q the bias-estimation order. Letting the package select bandwidths will generally change the fitted result from the fixed-window teaching calculation. Inspect the returned estimates, standard errors and confidence intervals and identify the robust output using your installed version's documentation.

Sign warning: this call uses the original centered score and reports above minus below. Because support is below the cutoff, reverse the sign to describe the support-direction contrast. For any reported interval [L, U], the sign-reversed interval is [-U, -L]. Apply that transformation consistently to the chosen estimate and its matching interval; never change only the estimate's sign.

The binned plot helps reveal the score-outcome pattern. It is not a replacement for the local estimator, and the number or width of plotting bins is not the estimation bandwidth. Show the cutoff, units and separate fits clearly; avoid letting a visually dramatic axis determine your interpretation.

Diagnostics and sensitivity checks to report

The Foundations guide by Cattaneo, Idrobo and Titiunik discusses validation and falsification alongside estimation. Choose checks for the actual allocation process and report unfavorable evidence as carefully as favorable evidence.

RDD diagnostic evidence and what it can establish
CheckPractical implementationInterpretation limit
Treatment discontinuityPlot and estimate actual receipt against the running variableA sharp eligibility rule does not guarantee sharp receipt
Predetermined covariate continuityExamine baseline attainment and other genuinely pre-treatment attributes using local RD methodsNonsignificant estimates do not prove that all relevant attributes are continuous
Running-variable densityInspect counts, heaping and an appropriate density discontinuity testA rejection may flag sorting or recording problems; non-rejection does not certify no manipulation
Placebo outcomesTest outcomes measured before support could affect themUnexpected jumps challenge the design but require institutional interpretation
Placebo cutoffsFit comparisons where no assignment changes, with windows entirely away from the true thresholdSearching many cutoffs can generate chance findings; a contaminated window is not a clean placebo
Bandwidth and local orderCompare justified windows and low-order specificationsStability strengthens description but cannot establish identification
Donut sensitivityIf justified by heaping or special handling, exclude a small region nearest the cutoffRemoving the most local observations increases extrapolation and can change what is learned

Covariate adjustment can improve precision when appropriately specified using predetermined covariates continuous at the cutoff. It does not repair a design whose relevant untreated characteristics jump there. Avoid controlling for attendance, engagement or other variables changed by support when estimating its total effect.

For density checks, Cattaneo, Jansson and Ma provide the underlying local-polynomial density method. The official rddensity manual documents its implementation:

# Separate unexecuted diagnostic template for real assignment data:
# library(rddensity)
# density_check <- rddensity(X = study$running_score, c = study_cutoff)
# summary(density_check)
# Covariates and pre-treatment outcomes need their own RD analyses.

No density, covariate-balance, placebo or donut results were computed for the teaching example. Its evenly spaced synthetic scores were imposed by construction, and it has no generated baseline covariates. Reporting that these data "passed manipulation tests" would misrepresent what was done. For coarse, heaped or discrete real scores, examine the number of distinct values and the assumptions behind inference; mass-point software adjustments do not automatically solve identification with little local support.

Fuzzy RDD: what changes with imperfect compliance?

Suppose real students can decline support and some students above the cutoff obtain it through exceptions. Retain eligibility Z and record actual receipt D. The outcome jump is then an effect of cutoff-induced eligibility or access, rather than automatically an effect of receiving support.

Fuzzy RDD divides the outcome discontinuity by the discontinuity in treatment receipt, using the same orientation for both. Under appropriate continuity, exclusion and monotonicity conditions, and a nonzero first stage, the ratio can identify an average treatment effect for compliers at the cutoff: those whose receipt responds to eligibility. The Extensions guide develops these distinctions.

Exclusion requires that crossing the eligibility threshold affect the outcome only through the defined treatment. It is questionable if the threshold also assigns a scholarship or changes academic labeling independently of support. Monotonicity rules out students whose treatment response to eligibility runs in the opposite direction under the chosen treatment coding. Neither condition follows from a large first-stage estimate.

Use fuzzy = D in rdrobust with the actual receipt vector when that design is justified. Report both jumps, their uncertainty, the ratio and the inferential method. A weak receipt jump can make ratio estimates unstable; a simple division of two reported numbers with an ordinary CI is inadequate. The main synthetic data have perfect compliance, so no separate fuzzy numerical result is claimed here.

Missing data and other threats to validity

Investigate whether follow-up outcome availability changes at the cutoff. If supported students are more likely to sit the later assessment, the observed score jump may partly reflect who was measured. Report exclusions and outcome availability by side and local window, with a strategy addressing the reasons for missingness.

Do not automatically drop incomplete records or impute an eligibility score and then treat the reconstructed threshold classification as observed allocation. Recover original administrative information where possible; distinguish missing running variables, missing receipt and missing outcomes. An imputation model or weighting analysis needs its own assumptions and sensitivity assessment, particularly around the cutoff. Our missing-data guide explains the broader distinction between missingness patterns and assumptions; it does not make survey imputation an automatic solution for RDD.

Check for score rounding, outcome scale changes, staff overrides, simultaneous policies and spillovers. A support offer that also changes grading rules might create a discontinuity in recorded marks without the intended change in learning. Document the timing of measurement and define the treatment package precisely.

Local sample size and distinct score support determine precision. Plan feasibility around those quantities, rather than counting everyone in a university database. Account for clustered allocation or correlated outcomes where relevant, protect student information, and obtain the required ethical and administrative permissions before accessing real records.

RDD versus difference-in-differences

RDD identifies a discontinuity at a score threshold under continuity or separately justified local-randomization conditions. Difference-in-differences compares outcome changes over time between treatment and comparison groups under an appropriate parallel-trends assumption. A single post-treatment outcome can support a valid RDD; repeated before-and-after outcomes are central to the usual DiD setup.

The main example compares students around an eligibility threshold; it does not subtract a pre-period change. Use the difference-in-differences dissertation guide when intervention timing and comparative trends supply identification. Data containing both a cutoff and repeated periods do not automatically validate either design or a combined method.

Dissertation-ready methodology and results paragraphs

Methodology paragraph for the verified teaching analysis

This synthetic analysis illustrates a sharp regression discontinuity design for academic support assigned below an eligibility score of 50. The running variable was centered at the threshold, and treatment receipt equaled eligibility. Separate local linear trends were estimated through a treatment-by-score interaction with triangular weights and a prespecified bandwidth of six score points on each side. The primary fit used 240 observations, with 120 on each side. HC3 heteroskedasticity-consistent standard errors and conventional normal-approximation intervals were calculated. Bandwidths of four and eight score points were examined as sensitivity checks. The calculation does not implement robust bias-corrected inference or provide empirical evidence for continuity, absence of sorting or generalizability.

In a real methodology chapter, add the policy provenance, population and recruitment, score precision, outcome timing, missing-data decisions, clustering, selected bandwidth method and planned diagnostics. State the causal assumptions separately from the regression formula. Make clear which decisions preceded outcome inspection.

Results paragraph for the verified teaching analysis

At the primary bandwidth, the below-minus-above discontinuity was 6.09 outcome points (HC3 SE 0.42; conventional approximate 95% CI 5.26 to 6.93). The corresponding descriptive within-one-point mean difference was 5.642500 points and was not the local regression estimate. Estimates at narrower and wider bandwidths remained positive in this synthetic dataset. These results demonstrate a local estimation procedure; they are not findings from students or evidence that the identifying assumptions hold in an institution. Any real causal interpretation would concern the effect at the eligibility cutoff, rather than the effect for the university population.

Keep the interval label, sign convention and local target in the text, not solely in table notes. Describe the practical meaning of outcome points in your setting. The statistical-results reporting guide helps connect estimates and uncertainty to research questions.

Reporting checklist

  • Identify the original score, exact cutoff, equality rule, allocation date and treatment direction.

  • Separate eligibility, offer and receipt, and justify sharp or fuzzy classification.

  • Define the local target effect, outcome units, measurement timing and population limits.

  • Explain continuity, possible sorting, concurrent interventions and interference.

  • Provide local counts, distinct-score information, exclusions and missingness by side.

  • Report the kernel, polynomial orders, bandwidth selector and actual bandwidths.

  • Name the variance method, clustering level, bias correction and interval type.

  • Include a readable RD plot and diagnostic evidence, together with limitations.

  • Show justified sensitivity analyses and disclose deviations from the analysis plan.

  • Retain reproducible code, software versions and output without exposing student records.

Common mistakes to avoid

Inventing a threshold after seeing the outcome. A statistically convenient split does not establish an allocation discontinuity. Anchor the design in the actual policy.

Calling a mean difference an RDD estimate. A finite-window mean gap and a fitted difference at the boundary are different quantities. Label both.

Using one slope without justification. An interaction or separate fits allows the score relationship to differ across the threshold. Do not force equal trends merely for convenience.

Treating robust SEs as bias correction. Heteroskedasticity robustness and RD smoothing-bias correction solve different inferential problems.

Interpreting non-rejection as validation. A weak density or covariate test can miss meaningful problems. Combine uncertainty with institutional evidence.

Generalizing across the entire score range. A local effect does not establish what support would achieve for students far from the threshold.

Frequently asked questions

Is regression discontinuity an experiment?

It is commonly treated as a quasi-experimental design. In a continuity-based analysis, assignment is determined by the score rule, rather than literally randomized. A local-randomization analysis requires additional, explicit assumptions about a chosen window.

Do I need pre-treatment and post-treatment outcomes?

Not necessarily. A post-treatment outcome and the original pre-treatment assignment score can support the basic design. Earlier outcomes are useful for falsification and understanding the setting, but RDD does not depend on the ordinary DiD subtraction.

How many observations do I need?

There is no universal minimum. Local density, distinct score values, outcome variability, expected effect, bandwidth and dependence matter. A large total sample with few useful observations on either side of the cutoff may be inadequate.

Does a density test prove that the score was not manipulated?

No. A non-rejection can reflect limited power, and some allocation problems may not create a detectable marginal density jump. Investigate score production, appeals, overrides and relevant baseline characteristics as well.

Can I use a discrete examination mark?

Possibly, but coarse support changes the inferential problem. Inspect score frequencies, gaps and heaping and consider methods designed for the actual support. Do not describe repeated integer marks as an arbitrarily fine continuous running variable.

Should I control for baseline covariates?

Appropriately chosen predetermined covariates may improve precision. They cannot generally repair a discontinuity in untreated characteristics, and post-treatment controls can change the target effect. Explain the specification and whether bandwidth selection accounts for covariates.

What if some eligible students refuse support?

The eligibility rule may still define a sharp design for the offer, while actual receipt is fuzzy. Define the treatment of interest and assess the additional assumptions before interpreting a fuzzy ratio as an effect of receipt.

Can I call the worked interval robust bias-corrected?

No. The executed example calculates an HC3 covariance and a conventional approximate interval. The separate rdrobust block describes an unexecuted route to RBC inference; its outputs must be run, checked and reported before making that claim.

Support with your dissertation analysis

If you need help evaluating a cutoff rule, understanding local estimates or improving the clarity of your reporting, explore Acadelyra's academic support services. Support should strengthen your understanding and your own analysis. Retain responsibility for research decisions, verify executed outputs, follow your institution's external-assistance rules and acknowledge help where required. No result or grade can be guaranteed.

Sources and further reading