MOST from a Methodological Perspective: Module 6 study notes

Source anomalies found in Module 6 (flagged, not silently corrected). (1) The Lesson 2 and Lesson 4 knowledge checks are true/false items for which the course marks both “True” and “False” as correct; both checkboxes are reproduced exactly as in course.json and each item is flagged, not endorsed. (2) The caption transcripts contain transcription artefacts, for example “Corsera” for Coursera, “Guestero” for Guastaferro, “pure” for peer, “dieice” for dice, “premuted” for permuted, “boneroni”/“bonferoni” for Bonferroni, “Gallas at all” for Gallis et al., “Guads at all” for Gwadz et al., “Catio”/“kaddio.org” for CADIO/cadio.org, “basian” for Bayesian, “longestablished” for long-established, and the spoken approximations “equals33” for 0.0033 and “0.16” for 0.0016; these are preserved inside narration quotes and noted where they occur.

The module's argument in one paragraph. This module is the course's account of rigorous and responsible conduct of optimization trials, built around four requirements a factorial optimization trial must satisfy. First, all factors must be manipulable independently: every combination of factor levels must make sense, and each level of each factor must be implemented consistently wherever it occurs; conditions that cannot be implemented cannot simply be deleted (that destroys the balance property), and inconsistent implementation can manufacture interactions that would otherwise not exist (Lesson 1). Second, participants must receive at least the standard of care, whether through factor-specific control levels coded as standard of care or through a constant component offered to everyone; a constant component cannot be evaluated (no estimable main effect or interaction) and results generalise only to settings in which it is provided (Lesson 2). Third, assignment to conditions must be correct: unrestricted randomisation yields only approximately balanced conditions, permuted block randomisation is recommended (in the running example, P = 16, n = 22, N = 352), stratification balances important characteristics, larger blocks mitigate the predictability criticism, and cluster structures change the arithmetic of both within-cluster (replicate the whole experiment per cluster) and between-cluster assignment (Lesson 3). Fourth, the decision whether to apply a Type I error rate correction such as Bonferroni must be reasoned through: corrections preserve the experiment-wise alpha at a steep cost to power and sample size (in a 2⁴ experiment, α/15 = .05/15 = .0033; in a 2⁵, α/31 = .05/31 = .0016), and the lesson recommends against corrections under the decision-priority perspective, favouring decisions based on observed effect sizes instead (Lesson 4).

Lesson 1: Ensuring That All Factors Can Be Manipulated Independently

The lesson sets out the necessary assumption behind every factorial experiment, that all factors can be manipulated independently, and works through its two implications using the course's running 2⁴ example (four candidate components, each no/yes). The first implication is that all combinations of factor levels make sense; the second is that each level of each factor is implemented consistently across all experimental conditions. Violations matter: unimplementable conditions cannot be deleted without destroying balance, and inconsistent implementation can create spurious interactions.

The running example: “We are developing an intervention aimed at reducing viral load among HIV positive individuals who drink heavily.” “Suppose there are four candidate components all with levels no and yes.” “Motivational interviewing, peer mentoring, text message support and mindfulness meditation.” “This would be a 2 to the 4th factorial experiment with 16 experimental conditions.” “This table shows the abbreviated names of the four factors, namely MI, pure, text, and mind and the 16 experimental conditions.” (Caption artefact: “pure” is peer.)Narration, Lesson 1; deck slides 4 to 6

The necessary assumption: “Now, I'd like to discuss a necessary assumption in all factorial experiments.” “This is the assumption that all the factors can be manipulated independently.” It has two implications: “The first one is that all combinations of factor levels make sense.” “The second one is that each level of each factor is implemented consistently across all experimental conditions.” “In other words, each level of each factor is implemented in the same manner irrespective of which other levels of other factors it is combined with.”Narration, Lesson 1; deck slides 7 to 8

The booster-session example: “The investigators would like to test the effectiveness of a booster session to be delivered 6 months after the intervention has been completed.” “they are considering adding a fifth factor to the experiment.” “This factor would have two levels, booster set to yes and booster set to no.” “there would be a condition in which MI, peer, text, and mind are all set to no and booster is set to yes.” “Would it make sense for someone who has not been assigned any of these other components to receive a booster?” “Under some circumstances, it might. Under other circumstances, this will not make sense.”Narration, Lesson 1; deck slide 10

Implication 2: consistent implementation: “how a level of a factor is implemented should be exactly the same across experimental conditions and should not vary depending on which other components and component levels are combined with it in an experimental condition.” The MI-and-mindfulness staff example: “Suppose the same staff member is to deliver motivational interviewing and the training and mindfulness meditation.” Some conditions call for MI only, “these are conditions 9, 11, 13, and 15.” Some call for mindfulness only, “These would be conditions 2, 4, 6, and 8.” Some call for both, “These would be conditions 10, 12, 14, and 16.” “This staff member must deliver motivational interviewing and the training and mindfulness meditation in the same manner in all of these experimental conditions.” “This means for example not combining them into one briefer treatment to save time in conditions 10, 12, 14 and 16.”Narration, Lesson 1; deck slides 13 to 20

Figure 1: The necessary assumption and its two implications (redrawn)

Element	Content
The assumption	All the factors can be manipulated independently.
What it does not mean	That the factors' effects are expected to be independent; the factors may interact, and studying interactions is often a primary motivation for a factorial experiment.
Implication 1	All combinations of factor levels make sense; review all experimental conditions when planning.
Implication 2	Each level of each factor implemented in the same manner irrespective of which other levels of other factors it is combined with.
If conditions cannot be implemented	Usually you cannot conduct the experiment; leaving conditions out destroys the balance property of the factorial experiment.
If consistency is violated	An interaction can appear between factors where otherwise there would be none; some redundancy may be inevitable (e.g. streamlining when components are presented together), but keep it to a minimum.

Redrawn from Lesson 1 deck slides 7 to 12 and 21 to 22, and the narration.

Figure 2: Worked examples on the running 2⁴ design (redrawn)

Example	What the lesson shows	Conditions involved
Text messages sent by the peer mentor	Conditions in which TEXT is set to yes but PEER is set to no cannot be implemented as specified, so the factorial design cannot proceed as planned (deck slide 9 highlights these rows)	3, 4, 11, 12
Booster session as a candidate fifth factor	Adding a second level (booster no/yes) creates a condition with every other factor at no and booster at yes; whether it makes sense must be evaluated case by case	All 16 conditions re-examined for the new factor
Consistent implementation of MI and mindfulness by one staff member	Deliver both in the same manner everywhere; do not merge them into one briefer treatment to save time where both appear together	MI only: 9, 11, 13, 15; MIND only: 2, 4, 6, 8; both: 10, 12, 14, 16

Redrawn from Lesson 1 deck slides 9 to 10 and 13 to 20 (the 16-condition table appears on deck slide 6). Condition numbers follow those printed on the slides and spoken in the narration.

Knowledge check

MULTIPLE RESPONSE: Which of the following statements about factorial optimization trials are correct? There may be more than one.

☐ It is necessary to assume that the components are completely independent and do not interact.
☑ It is necessary to assume that the components can be experimentally manipulated independently.
☐ If an experimental condition turns out not to be implementable, it should be omitted from the experiment.
☑ Implementation of a level of a factor should be exactly the same across experimental conditions.

Feedback when incorrect (source text, retained as provided): Answer A. is incorrect. There is no assumption that the components do not interact; in fact, the factorial experiment enables estimation of interactions. Answer C. is incorrect. Omitting an experimental condition will destroy balance. If an experimental condition is not implementable the entire experiment must be redesigned.

Reading for All of Module 6

This lesson has no video or slide deck; it lists the set reading for the module (spacing normalised; names as listed by the course).

Lesson 2: Ensuring That All Participants Are Provided with at Least the Standard of Care

The lesson states the standard-of-care requirement for optimization trials and shows two ways to honour it inside a factorial design: factor-specific control levels coded as the relevant standard of care, or a constant component offered to all participants. The constant-component route has consequences for what can be estimated and how far results generalise, worked through with the ART and behavioural-skills-training examples.

How the control levels read in the design table: “This table shows the experimental conditions in a factorial experiment with four factors.” “In this table, the acronym S represents standard of care.” “In this table, each factor uses its respective standard of care as a control.” “You can see that for factor A, the experimental treatment is provided for conditions 9 through 16 and standard of care is provided for conditions 1 through 8.” “The main effect of factor A then estimates the average difference between the experimental treatment and the standard of care.” “The same reasoning applies to the remaining factors.”Narration, Lesson 2; deck slide 6

Reporting the constant component: “When you describe or report on an optimization trial that includes a constant component, it is of course necessary to describe the constant component.” “In general, I recommend including the constant component in any visual representation of an optimization trial.” The lesson shows a published example: “This is one example from a published article by Guads at all.” (Caption artefact: “Guads at all” is Gwadz et al.; the deck slide 9 reprints Figure 2 from Gwadz et al. 2017, showing the conditions in the fractional factorial design, as a way of including a constant component.)Narration, Lesson 2; deck slide 9

The behavioural-skills-training example: “Suppose in our example, the investigators decide to include a behavioral skills training component as a constant component.” “Suppose this experiment is conducted and the results indicate that the peer factor has a main effect in the desired direction.” Then “this result could mean that peer mentoring is effective or it could mean that peer mentoring is effective only when behavioral skills training is provided.” “Suppose in reality peer mentoring is highly effective when behavioral skills training is provided and is not effective at all when behavioral skills training is not provided.” “If behavioral skills training is a constant component, there is no way to detect this.” “If behavioral skills training is represented by an experimental factor, this would be reflected in a strong synergistic interaction.” The lesson's conclusion: “The point here is that any results about the performance of the experimental factors cannot be generalized to a setting in which the constant component is not provided.”Narration, Lesson 2; deck slides 12 to 15

Figure 3: Standard of care in the factorial design (redrawn)

Approach	How it works	Consequence
Factor-specific control levels (S = standard of care)	Each factor's control level is that factor's standard of care; the other level is an experimental treatment that exceeds standard of care (or participants receive standard of care later). In the deck example, factor A is at standard of care in conditions 1 to 8 and at the experimental treatment in conditions 9 to 16	The main effect of each factor estimates the average difference between the experimental treatment and the standard of care
Constant component (e.g. ART prescription + basic education)	A component that is not experimentally manipulated; all participants are offered it; in the deck tables the constant component column is set to yes for all 16 conditions	The constant component's main effect and interactions cannot be determined; results cannot be generalised to settings where it is not provided
Worked warning (behavioural skills training as constant)	A significant peer main effect could mean peer mentoring works, or that it works only when behavioural skills training is also provided; with BST as a constant component there is no way to detect the difference, whereas as an experimental factor it would show as a strong synergistic interaction	Be thoughtful about whether and what to include as a constant component; ask whether it is really standard of care rather than a treatment with open questions worth testing

Redrawn from Lesson 2 deck slides 5 to 15. Deck slide 9 reprints Figure 2, “Conditions in the fractional factorial design”, from Gwadz et al. (2017) as an example of including a constant component in a published design.

Knowledge check

MULTIPLE CHOICE: Is the following statement true or false? When a constant component is included in a factorial optimization trial, any results about the performance of the experimental factors can be generalized only to settings in which the constant component is provided.

☑ True
☑ False

Flagged source anomaly: for this true/false item the course marks both “True” and “False” as correct. Both marks are reproduced exactly as they appear in the course source; the item is preserved and flagged rather than silently corrected, and is not endorsed.

Lesson 3: Random Assignment

The lesson covers how to assign participants to conditions in a factorial optimization trial: unrestricted random assignment, the recommended permuted block approach and its variations (larger blocks, stratification, minimization), the adaptations needed for cluster structures, and the logistics of doing it correctly.

Unrestricted random assignment: “One approach to random assignment is unrestricted random assignment.” “This is the simplest approach.” The mechanics: “So the total sample size represented by an uppercase N equals the number of participants per condition times the number of conditions that is lowercase N * P.” “you can fashion a virtual die by using a random number generator.” “Find a random number generator that is based on a uniform distribution of integers between 1 and P and use it to generate a list of N random numbers.” “If you use this approach, balance will be only approximate.” “the larger the sample size, the closer you are expected to come to perfect balance.” (The deck's example: a 2⁴ factorial has P = 16 conditions; 10 participants per condition gives N = 160, and each of the 160 participants is assigned by a random integer between 1 and 16.)Narration, Lesson 3; deck slides 5 to 6

The worked example: “You want 22 participants per experimental condition.” “How can you achieve balance?” “This example has P equals 16 conditions lowercase N= 22 participants per condition for a total sample size of 16 * 22 = 352.” “To conduct permuted block randomization in its simplest form, there are variations we will talk about soon.” “You would take the integers from 1 to 16 and randomly permute them.” “Do this 22 times.” “Now you have 22 permuted blocks.” “The slide shows an example of what one of the permuted blocks might look like.” “It's just the integers 1 through 16 in a random order.” “Suppose this random order which as the slide shows begins with the number two then the number 13 and so on is the first permuted block.” “Then the first participant would be assigned to condition two. The second would be assigned to condition 13 and so on.” “I hope you can see that this will spread the first 16 participants evenly across the 16 conditions, one per condition.” “You can then use the next permuted block to assign the next 16 participants and so on until all 352 participants have been assigned.” “At the conclusion of this process, there will be 22 participants in each condition.” (The deck's example block, printed on slides 10 to 11, is: 2 13 12 10 7 14 4 11 3 9 8 6 1 15 5 16.)Narration, Lesson 3; deck slides 9 to 12

The blindness criticism and larger blocks: “A criticism of permuted blocks randomization has been that some people feel it is not quote unquote blind enough because once all but one of the conditions have been assigned within a block, it is known what the assigned condition of the last participant in that block is going to be.” “This may not be a problem if random assignment is kept separate from implementation and program staff are kept blind to the content of each condition.” “One approach that some people might find more palatable is to use larger blocks.” “you could use a block size equal to twice the number of experimental conditions.” “In these larger blocks, each random number appears twice.” “This mitigates the problem but does not completely eliminate it.”Narration, Lesson 3; deck slide 13

Stratification: “Often investigators wish to use stratification in random assignment.” “The purpose of stratification is to distribute certain characteristics of participants evenly across experimental conditions.” The health-literacy example: you measure literacy at intake and “you don't know what the proportions are”, so “You just create separate lists of random numbers.” “if the individual to be assigned is above the threshold, you would use list A for random assignment.” “If the individual is below the threshold, you would use list B.” “This will evenly distribute each subgroup.” “if it turns out that 30% of your sample are below the threshold, each condition will include about 30% below the threshold.” Practical advice: “it is practical to stratify on one variable only.” “More than that tends to get messy quickly.”Narration, Lesson 3; deck slides 14 to 16

Cluster structures: “If you can reasonably conduct within cluster assignment, the best way to preserve power is to replicate the entire experiment within each cluster.” “This can be done however you are planning to conduct random assignment by having a separate list of random numbers for each cluster.” The clinic example: “suppose the optimization trial is to be conducted in five different HIV clinics and you are responsible for the random assignment.” “you would create five different lists of permuted blocks of the integers 1 through 16 and provide one of the lists to each clinic.” With stratification: “you would provide each clinic with two lists of random numbers, one for each category.” “You would need a total of 10 lists.” Between-cluster assignment: “If you are using between cluster randomization, you will have a much smaller number of quote unquote participants to assign because each cluster counts as a participant.” “if you were conducting a school-based study with 32 schools and you were using permuted blocks random assignment, you would need just two permutations of the integers 1 through 16.” “When you assign a particular school to a condition, you assign all the students within it.”Narration, Lesson 3; deck slides 17 to 19

Logistics and documentation: “Random assignment is something you don't want to mess up.” “Technology can be extremely helpful.” “At this time, I recommend the use of the REDCap software.” “There is a tutorial on this topic on the Catio website, kaddio.org.” (Caption artefacts: “Catio” and “kaddio.org” are CADIO and cadio.org; the deck slide 20 prints the address.) “it is important that random assignment be done correctly.” “Keep track of any deviations from random assignment and also keep track of any protocol deviations of any kind.”Narration, Lesson 3; deck slide 20

Figure 4: Random assignment approaches in the running example (redrawn)

Approach	Mechanics	Notes and example numbers
Unrestricted random assignment	For each of N participants, draw a uniform random integer from 1 to P (a virtual P-sided die); the number is that participant's condition	Balance only approximate; approaches perfect balance as the sample grows. Deck example: 2⁴, P = 16 conditions, n = 10, N = 160
Permuted block randomization (recommended)	Permute the integers 1 to P once per block, n times, giving n blocks; assign the first P participants from the first block, then the next P from the second, and so on	Example: P = 16, n = 22, N = 16 × 22 = 352. First block shown on deck slides 10 to 11: 2 13 12 10 7 14 4 11 3 9 8 6 1 15 5 16; first participant to condition 2, second to condition 13, and so on; after all blocks, 22 participants per condition
Larger blocks	Block size twice the number of conditions (2p); each random number appears twice per block	Mitigates the criticism that the last assignment in a block is predictable; does not eliminate it
Stratification	Separate randomisation lists per stratum (e.g. list A for participants above a health-literacy threshold, list B for those below); assign from the matching list	Balances the stratum across conditions (if 30% fall below the threshold, each condition holds about 30%); practical for one variable only
Minimization	Named as an alternative approach	Not covered in the lesson; reviewed by Gallis et al. (2018)
Cluster, within-cluster assignment	Replicate the whole experiment inside each cluster; separate randomisation list(s) per cluster	Five HIV clinics → five lists of permuted blocks of 1 to 16; with two stratification categories, ten lists in total
Cluster, between-cluster assignment	Each cluster counts as one “participant” to be assigned; assigning a cluster assigns everyone in it	32 schools need just two permutations of the integers 1 to 16

Redrawn from Lesson 3 deck slides 4 to 20 and the narration. Logistics: the lesson recommends REDCap for random assignment and keeping track of any deviations from random assignment or protocol; a tutorial is on the CADIO website (cadio.org).

Lesson 4: Whether or Not to Use a Type I Error Rate Correction

The lesson works through the question of applying a correction (such as Bonferroni) to maintain the experiment-wise Type I error rate in a factorial optimization trial. It reviews the power logic of the 2ᵏ design, shows what adding a factor does to the number of hypotheses (15 → 31), prices the correction in sample size, and recommends against corrections under the decision-priority perspective.

The setup: efficiency of the 2ᵏ design: “In this course, we have emphasized the efficiency and economy of the 2 to the K factorial experiment.” “We've explained how it is possible to add one or more factors to a factorial experiment without necessarily needing to increase the sample size to maintain power.” The reference design: “Suppose this experiment is adequately powered with an overall sample size of 352, which would be a per condition sample size of 22.” And the power logic: “in factorial experiments, the per condition sample size does not determine power. What matters for power is the sample size per level of each factor.” “To convince yourself of this, recall how main effects are calculated, that is by comparing means, each of which is computed based on all the experimental conditions that make up a factor level.”Narration, Lesson 4; deck slides 4 to 8

Adding a fifth factor: “Now suppose you want to add a candidate component behavioral skills training with levels low intensity and high intensity” and “suppose the size of the anticipated main effect is no smaller than the effect size used to do the original power analysis.” “This shows the new 2 to the 5th experimental design. It has 32 experimental conditions twice the number in the previous design.” “If the overall sample size is kept the same, the new design has 11 participants per condition.” “Yet, it has the same statistical power as the previous design for detection of all five main effects with one important caveat which we will discuss now.”Narration, Lesson 4; deck slides 8 to 10

The Bonferroni correction defined: “An example of such a correction is the familiar boneroni correction.” (Caption artefact: “boneroni” is Bonferroni.) “The boneroni correction calls for dividing the alpha among the number of hypotheses being tested.” “Usually the alpha is divided evenly.” “So if the chosen alpha is 0.05 and there are five hypotheses to be tested each one would be tested using alpha equals 0.01.” “However the alpha can be allocated unevenly as long as this is done a priori.” Whether to correct at all: “Whether to apply the bonferoni correction or a similar correction or not in an optimization trial is controversial.”Narration, Lesson 4; deck slides 11 to 12

Adding a factor makes it worse: “adding a factor makes things worse because it increases the number of hypotheses.” “In a 2 to the 5th experiment, there are up to 31 hypotheses to test as compared to 15 in the 2 to the 4th.” “you would use alpha / 31.” The deck prints α/31 = .05/31 = .0016 (the narration renders it “0.16”). “it would be necessary to increase the sample size considerably, even beyond what was needed for the 2 to the 4th experiment.” “the need for additional participants when we added the fifth factor is due entirely to the imposition of the boneroni correction.”Narration, Lesson 4; deck slides 14 to 15

Conclusion-priority versus decision-priority: “A major consideration is are you working from a conclusion priority or decision priority perspective.” “From the conclusion priority perspective, the purpose of conducting the experiment is to enable you to conclude whether or not particular effects exist.” “From the decision priority perspective, the purpose of conducting the experiment is to enable you to make a decision about which components and component levels are to make up the optimized intervention.” “From the conclusion priority perspective, it may be necessary to maintain the experiment-wise type 1 error rate because of longestablished scientific conventions.” “From the decision priority perspective, maintaining power is critical.” (Caption artefact: “longestablished” is long-established.)Narration, Lesson 4; deck slides 17 to 18

Figure 5: The arithmetic of the Bonferroni correction and the priority question (redrawn)

Design	Up to how many hypotheses	Even split at experiment-wise α = .05	Sample-size implication
2⁴ (four factors: MI, PEER, TEXT, MIND)	15	α/15 = .05/15 = .0033	To maintain power, sample size would need to increase dramatically over the no-correction requirement; alternatively allocate unevenly (more to primary main effects, less to interactions) as decided a priori
2⁵ (add behavioural skills training as a fifth factor)	31	α/31 = .05/31 = .0016	Increase considerably beyond the 2⁴ requirement; the extra participants when adding the fifth factor are due entirely to the correction
Power context	-	-	With N = 352 held constant, the 2⁵ gives 11 per condition but the same power for all five main effects as the 2⁴ with 22 per condition, and the power of a comparable t-test on N = 352, with the caveat that this assumes no correction
Variations	-	Subset Bonferroni (others treated as exploratory); approaches other than Bonferroni	All preserve the experiment-wise Type I error rate at a cost to the Type II error rate
Conclusion-priority perspective	-	-	Purpose: to conclude whether or not particular effects exist; maintaining the experiment-wise Type I error rate may be necessary because of long-established scientific conventions
Decision-priority perspective	-	-	Purpose: to decide which components and levels make up the optimised intervention; maintaining power is critical; the lesson does not recommend a Bonferroni-type correction and favours decisions based on observed effect sizes

Redrawn from Lesson 4 deck slides 7 to 20. Arithmetic as printed on the slides (slides 13 and 15); the narration's approximations “equals33” for 0.0033 and “0.16” for 0.0016 are caption artefacts, preserved in the quotes and noted. In a balanced 2ᵏ factorial analysed with effect coding, all effect estimates are uncorrelated (deck slide 19; narration).

Knowledge check

MULTIPLE CHOICE: Is the following statement true or false? Whether you are working from the conclusion-priority or the decision-priority perspective is an important consideration when deciding whether to use a Bonferroni or similar correction.

☑ True
☑ False

Synthesis (my own reading: not part of the course)

This module is the course's quality-control chapter, and its four requirements are best read as four failure modes being defused in advance. The independent-manipulability lesson is about the design's internal coherence: factorial logic presumes every cell of the design can exist and that a level means the same thing in every combination. The two examples are instructive because they pull in opposite directions, the text-from-peer example fails because a combination cannot be built at all, while the booster example fails only conditionally, which is why the lesson's advice is a review process rather than a checklist. The consistency requirement is subtler: streamlining that looks like kindness (one briefer combined session) silently turns a zero interaction into a positive one, so the rule protects estimability, not aesthetics.

The standard-of-care lesson prices an ethical requirement in inferential currency. Factor-specific control levels keep the standard of care inside the design at the cost of making every treatment comparison a comparison against care-as-usual; a constant component is cheaper to operate but removes the component from the estimable space entirely. The behavioural-skills-training hypothetical makes the trade concrete: a constant component can mask exactly the conditional effect that a factorial is built to find, so the choice is between protection and knowledge, and the lesson's closing question, is this really standard of care, or an open question dressed as one, is the right filter.

The randomisation lesson is mostly engineering, and its recommendations scale sensibly: permuted blocks for balance, larger blocks against guessability, stratification for one key covariate, separate lists per cluster to preserve power under within-cluster assignment, and REDCap plus a deviations log for the audit trail. The Type I error lesson is the most opinionated in the module: it lays out the correction arithmetic honestly (α/15 =.0033; α/31 =.0016, with the sample-size consequences) and then, on the decision-priority reading that matches an optimization trial's purpose, declines to recommend it. That position is coherent with the course's repeated claim about effect coding, uncorrelated estimates in a balanced 2ᵏ design, but it is a judgement about the purpose of the experiment, not a statistical theorem; the lesson says as much by naming the Bayesian direction as the more natural fit. As with earlier modules, the two flagged true/false items (Lessons 2 and 4) mark both “True” and “False” correct; the notes preserve the source marks and flag the incoherence rather than guessing an intended key, and caption artefacts are preserved inside quotes.

Glossary

Sources

Video list

Prepared 13 Sep 2026. Notes paraphrase the narration unless marked as verbatim quotes; figures are redrawn from the course slides; quotes follow the YouTube caption transcripts. Transcripts were retrieved on 13 Sep 2026 through the browser transcript panel (Show transcript) on each video page, which is the supported route on this host; each transcript carries a video ID and timestamp coverage (L1 0:00 to 8:26, 64 segments; L2 0:01 to 5:48, 42; L3 0:01 to 12:51, 97; L4 0:01 to 9:20, 67). Caption transcription artefacts are preserved inside quotes and flagged in the anomaly note. Slide images were extracted from the four deck PDFs (90 slides) into 17 montages; all montages were processed with OCR, and eight montages were additionally inspected visually (Lesson 1 slides 1 to 18; Lesson 2 slides 1 to 12; Lesson 3 slides 7 to 18; Lesson 4 slides 13 to 18). Some deck titles in the PDF text layer are template text and do not match the slide images; slide numbers cited here follow the numbers printed on the slides themselves. The course's four knowledge checks are reproduced with the course answer key; the Lesson 2 and Lesson 4 items marking both True and False correct are flagged source anomalies, not endorsed. No separate quiz distinct from these items appears in the course.