factorial designs (between person)

Please leave suggestions for studies using a factorial design for between-person studies.

Consider Marlatt et al (1975; https://doi.org/10.1037/0021-843X.84.6.652) Provocation to anger and opportunity for retaliation as determinants of alcohol consumption in social drinkers. Here's the abstract:

Assigned 30 male and 30 female college students identified as heavy social drinkers to 1 of 6 groups in a 3 * 2 factorial design. In addition to the S sex factor, the 3 main treatment groups were provocation to anger with no opportunity to retaliate, provocation with opportunity for retaliation, and a no-provocation, no-retaliation control group. Provoked Ss were angered by an insulting confederate, whereas controls experienced a neutral interaction with the confederate. In the retaliation condition, Ss were given the opportunity to deliver a fixed number of shocks to the confederate who had provoked them. Drinking rates in all Ss were then determined by their participation in a standardized taste-rating task, which permitted an unobtrusive measure of alcohol consumption. Results show that group members who were provoked and expressed their anger by retaliating against the confederate consumed significantly less alcohol than provoked Ss in the no-retaliation condition. Controls drank an intermediate amount of alcohol but did not differ significantly from the other 2 groups. Sex was not a significant determinant of alcohol consumption

This is a nice quasi-experimental 2 X 3 design where 30 men and 30 women who identified heavy social drinkers were randomized into 1 of 3 conditions. The outcome is ounces of alcohol drank, which would make for a nice example with distributions like the gamma or lognormal.

Consider Clark et al (2021; https://doi.org/10.1111/add.15072) Impact of health warning labels communicating the risk of cancer on alcohol selection: An online experimental study. Here's the abstract:

Aims Evidence from tobacco research suggests that health warning labels (HWLs) depicting the adverse consequences of consumption change smoking behaviours, with image-and-text (also known as 'pictorial' or 'graphic') HWLs most effective. There is an absence of evidence concerning the potential impact of HWLs placed on alcohol products on selection of those products. This study aimed to obtain a preliminary assessment of the possible impact of (i) image-and-text, (ii) text-only, and (iii) image-only HWLs on selection of alcoholic versus non-alcoholic drinks. Design A between-subjects randomised experiment with a 2 (image: present versus absent) × 2 (text: present versus absent) factorial design. Setting The study was conducted on the online survey platform Qualtrics. Participants Participants ($n = 6024$) were adults over the age of 18 who consumed beer or wine regularly (i.e. at least once a week), recruited through a market research agency. Interventions Participants were randomised to one of four groups varying in the HWL displayed on the packaging of alcoholic drinks: (i) image-and-text HWL; (ii) text-only HWL; (iii) image-only HWL; and (iv) no HWL. HWLs depicted bowel cancer, breast cancer and liver cancer, which were each displayed twice across six alcoholic drinks. Each group viewed six alcoholic and six non-alcoholic drinks and selected one drink that they would like to consume. Measurements The primary outcome was the proportion of participants selecting an alcoholic versus a non-alcoholic drink. Findings Alcoholic drink selection was lower for all HWL types compared with no HWL (image-and-text: 56%; image-only: 49%; text-only: 61%; no HWL: 77%), with selection lowest for HWLs that included an image. Image-and-text HWLs reduced the odds of selecting an alcoholic drink compared with text-only HWLs (OR = 0.80, 95% CI = 0.69, 0.92), but increased the odds of selecting an alcoholic drink compared with image-only HWLs (OR = 1.34, 95% CI = 1.16, 1.55). Conclusions Health warning labels communicating the increased risk of cancers associated with alcohol consumption reduced selection of alcoholic versus non-alcoholic drinks in a hypothetical choice task in an online setting; labels displaying images had the largest effect. Their impact in laboratory and real-world field settings using physical products awaits investigation.

This is a posttest-only 2 X 2 factorial experimental design where $n = 6024$ adult drinkers were randomized into 1 of 4 conditions. The primary outcome is the probability of selecting an alcoholic beverage in an online selection task, with a control condition being compared to conditions using text-only, image-only, or text-and-image health warning labels. This would make for an easy example with binomial regression and probability contrasts. You could compare the conventional y ~ 1 + dummy1 + dummy2 + dummy1 : dummy2 approach with a y ~ 1 + factor(group) approach (where control is the reference category). As the authors concluded the previous research on this topic for alcohol was scant, of small samples sizes, and somewhat contradictory, this would be an easy example to use weakly-regularizing priors.

Also, the authors made the data openly available on the OSF at https://osf.io/pr8zu/ (see the Alcohol study 2 full dataset.xlsx file in the Study 2 folder.

You can also use this as an example of using baseline covariates with logistic regression.

Consider Berg et al (2020; https://doi.org/10.3389/fpsyt.2020.00503) The role of learning support and chat-sessions in guided internet-based cognitive behavioral therapy for adolescents with anxiety: A factorial design study. Here's the abstract:

Background: Increased awareness of anxiety in adolescents emphasises the need for effective interventions. Internet-based cognitive behavioural therapy (ICBT) could be a resource-effective and evidence-based treatment option, but little is known about how to optimize ICBT or which factors boost outcomes. Recently, the role of knowledge in psychotherapy has received increased focus. Further, chat-sessions are of interest when trying to optimize ICBT for youths. This study aimed to evaluate the role of learning support and chat-sessions during ICBT for adolescent anxiety, using a factorial design.

Method: A total of 120 adolescents were randomised to one of four treatment groups, in a 2x2 design with two factors: with or without learning support and/or chat-sessions.

Results: Anxiety and depressive symptoms were reduced (Beck Anxiety Inventory-BAI; Cohen’s $d = 0.72$; Beck Depression Inventory-BDI; $d = 0.97$). There was a main effect of learning support on BAI ($d = 0.38$), and learning support increased knowledge gain ($d = 0.42$). There were no main effects or interactions related to the chat-sessions. Treatment effects were maintained at 6-months, but the added effect of learning support had by then vanished.

Conclusion: ICBT can be an effective alternative when treating adolescents with anxiety. Learning support could be of importance to enhance short-term treatment effects, and should be investigated further.

This is a pretest-posttest randomized controlled trial with a 2 X 2 design where 120 adolescents in Sweden were treated with a 9-week protocol of internet-based cognitive behavioural therapy (ICBT) for anxiety and comorbid depression. The primary outcomes were the BAI and BDI-II sum scores, which make for a nice comparison of the conventional Gaussian and beta-binomial likelihoods.

Consider Aminpour et al (2022; https://doi.org/10.1186/s12889-022-14801-6), The choice of message and messenger to drive behavior change that averts the health impacts of wildfires: An online randomized controlled experiment. Here's the abstract:

Background To reduce the negative health effects from wildfire smoke exposure, effective risk and health communication strategies are vital. We estimated the behavioral effects from changes in message framing and messenger in public health messages about wildfire smoke on Facebook.

Methods During September and October 2021, we conducted a preregistered online randomized controlled experiment in Facebook. Adult Facebook users ($n$ = 1,838,100), living in nine wildfire-prone Western U.S. states, were randomly assigned to see one of two ad versions (narrative frame vs. informational frame) from one of two messengers (government vs. academic). We estimated the effects of narrative framing, the messenger, and their interactions on ad click-through rates, a measure of recipient information-seeking behavior.

Results Narrative frame increased click-through rates by 25.3% (95% CI = 22.2, 28.4%), with larger estimated effects among males, recipients in areas with less frequent exposure to heavy wildfire smoke, and in areas where predominant political party affiliation of registered voters was Republican (although not statistically different from predominantly-Democrat areas). The estimated effect from an academic messenger compared to a government messenger was small and statistically nonsignificant (2.2%; 95% CI = − 0.3, 4.7%). The estimated interaction effect between the narrative framing and the academic messenger was also small and statistically nonsignificant (3.9%; 95% CI = − 1.1, 9.1%).

Conclusions Traditional public service announcements rely heavily on communicating facts (informational framing). Shifting from a fact-focused, informational framing to a story-focused, narrative framing could lead to more effective health communication in areas at risk of wildfires and in public health contexts more broadly.

The research team "used a 2×2 factorial design where Facebook users were randomly assigned to see one of two ad versions (Informational vs. Narrative messages) from one of two messenger types (Government vs. Academic sources). Randomization was achieved by employing the A/B split testing functionality embedded in Facebook ads" (p. 3). The dependent variable is click-through rates (CTR), which are binary. The authors have high-quality prior research on typical CTR's, which makes for nice prior information. Importantly, the data are available in aggregated form on the OSF at https://osf.io/9ax6n/?view_only=55488a95825444aa830eadbd38fe3565. The data include several baseline covariates, making this a nice way to showcase aggregated binomial ANCOVAs'. Given the large sample size, this also makes for nice discussions on the strength of default weakly-regularizing priors, and the importance of aggregated data.

ASKurz / Experimental-design-and-the-GLMM

factorial designs (between person) #6