Too many zeros and/or highly skewed? A tutorial on modelling health behaviour as count data with Poisson and negative binomial regression

Green, James

doi:10.31219/osf.io/ux9et

Cited by 5 publications

(5 citation statements)

References 29 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Next, the distribution of the outcome variable was checked, and found to be best described by a negative binomial distribution (Supplementary material Figure 1), which is a discrete probability distribution with lower bound at 0, and variance much larger than mean, suggesting the presence of overdispersion. 38 To avoid losing information by dichotomizing the data or using non-parametric tests, a generalized linear model (GLM) for a negative binomial distribution with a log link function was chosen for data analysis. Due to the data structure (providers nested in PHCCs and PHCCs within country), generalized linear mixed models were initially used to test for the inclusion of random effects.…”

Section: Discussionmentioning

confidence: 99%

Factors Associated with Primary Health Care Providers’ Alcohol Screening Behavior in Colombia, Mexico and Peru

et al. 2021

View full text Add to dashboard Cite

Background: Screening for unhealthy alcohol use in routine consultations can aid primary health care (PHC) providers in detecting patients with hazardous or harmful consumption and providing them with appropriate care. As part of larger trial testing strategies to improve implementation of alcohol screening in PHC, this study investigated the motivational (role security, therapeutic commitment, self-efficacy) and organizational context (leadership, work culture, resources, monitoring, community engagement) factors that were associated with the proportion of adult patients screened with AUDIT-C by PHC providers in Colombia, Mexico and Peru. Additionally, the study investigated whether the effect of the factors interacted with implementation strategies and the country. Methods: Pen-and-paper questionnaires were completed by 386 providers at the start of their study participation (79% female, M age ¼ 39.5, 37.6% doctors, 15.0% nurses, 9.6% psychologists, 37.8% other professional roles). They were allocated to one of four intervention arms: control group; short training only; short training in presence of municipal support; and standard (long) training in presence of municipal support. Providers documented their screening practice during the five-month implementation period. Data were collected between April 2019 and March 2020. Results: Negative binomial regression analysis found an inverse relationship of role security with the proportion of screened patients. Self-efficacy was associated with an increase in the proportion of screened patients only amongst Mexican providers. Support from leadership (formal leader in organization) was the only significant organizational context factor, but only in non-control arms. Conclusion: Higher self-efficacy is a relevant factor in settings where screening practice is already ongoing. Leadership support can enhance effects of implementation strategies.

show abstract

Section: Discussionmentioning

confidence: 99%

Factors Associated with Primary Health Care Providers’ Alcohol Screening Behavior in Colombia, Mexico and Peru

et al. 2021

View full text Add to dashboard Cite

show abstract

“…To account for this, we used negative binomial regression models. These models are commonly used for count data, but as the sMFQ scores are discrete, independent, and have no negative values, models using count distributions are still applicable (Green, 2020;Kandola et al, 2020). The outcome for these models is interpretable as a percentage change in sMFQ scores.…”

Section: Main Analysismentioning

confidence: 99%

Prospective relationships of adolescents’ screen-based sedentary behaviour with depressive symptoms: the Millennium Cohort Study

et al. 2021

View full text Add to dashboard Cite

Background Frequent use of screen-based devices could be a modifiable risk factor for adolescent depression, but findings have been inconsistent and mostly from cross-sectional studies. We examined prospective associations of video gaming, social media, and internet use with depressive symptoms in adolescents. Methods A total of 11 341 adolescents from the Millennium Cohort Study, a representative, UK population-based. The main outcome was depressive symptoms from a Moods and Feelings Questionnaire (age 14). Exposures were frequency of video game, social media, and internet use (age 11). Physical activity (effect modifier) was measured by self-report. Results The fully adjusted models indicated that boys playing video games most days, at least once a week, and at least once a month at age 11 had lower depression scores at age 14 by 24.2% (IRR = 0.77, 95% CI 0.66–0.91), 25.1% (IRR = 0.75, 95% CI 0.62–0.90), and 31.2% (IRR = 0.69, 95% CI 0.57–0.83), compared with playing less than once a month/never. In girls, compared with less than once a month/never, using social media most days at age 11 was associated with 13% higher depression scores at age 14 (IRR = 1.13, 95% CI 1.05–1.22). We found some evidence of associations between using the internet most days and depressive symptoms compared with less than once a month/never in boys (IRR = 0.86, 95% CI 0.75–1.00). More frequent video game use was consistently associated with fewer depressive symptoms in boys with low physical activity, but not in those with high physical activity. Conclusions Different types of screen-time may have contrasting associations with depressive symptoms during adolescence. Initiatives to address adolescents’ screen-time may require targeted approaches.

show abstract

“…Continuous outcomes were modeled using negative binomial regression. Negative binomial regression is a generalization of the Poisson distribution and is suitable to model non-negative count data with overdispersion [42]. To improve interpretation, regression-adjusted coefficients were exponentiated and interpreted as percentages.…”

Section: Discussionmentioning

confidence: 99%

Changes in physical activity after joining a bikeshare program: a cohort of new bikeshare users

Auchincloss

Michael

Niamatullah

et al. 2022

Int J Behav Nutr Phys Act

View full text Add to dashboard Cite

Background There are hundreds of bikeshare programs worldwide, yet few health-related evaluations have been conducted. We enrolled a cohort of new bikeshare members in Philadelphia (Pennsylvania, USA) to assess whether within-person moderate and vigorous physical activity (MVPA) increased with higher use of the program and whether effects differed for vulnerable sub-groups. Methods During 2015–2018, 1031 new members completed baseline and one-year follow-up online surveys regarding their personal characteristics and past 7-day MVPA minutes per week (minutes per week with- and without walking). Participants were linked to their bikeshare trips to objectively assess program use. Negative binomial (for continuous outcomes) and multinomial (for categorical outcomes) regression adjusted for person characteristics (socio-demographics, health), weather, biking-infrastructure, and baseline biking. Results Participant median age was 30, 25% were of Black or Latino race/ethnicity, and 30% were socioeconomically disadvantaged. By follow-up, personal bike ownership increased and 75% used bikeshare, although most used it infrequently. Per 10 day change in past year (PY) bikeshare use, non-walking MVPA min/wk increased 3% (roughly 6 min/wk, P < 0.014). More days of bikeshare was also associated with change from inactive to more active (odds ratio for ≥ 15 days in PY vs. no bikeshare use 1.80, CI 1.05–3.09, P < 0.03). Results were consistent across vulnerable sub-groups. In general, impacts on MVPA were similar when exposure was personal bike or bikeshare. Conclusions Bikeshare facilitated increases in cycling, slightly increased non-walking MVPA, and showed potential for activating inactive adults; however, for larger program impact, members will need to use it more frequently.

show abstract

Too many zeros and/or highly skewed? A tutorial on modelling health behaviour as count data with Poisson and negative binomial regression

Cited by 5 publications

References 29 publications

Factors Associated with Primary Health Care Providers’ Alcohol Screening Behavior in Colombia, Mexico and Peru

Factors Associated with Primary Health Care Providers’ Alcohol Screening Behavior in Colombia, Mexico and Peru

Prospective relationships of adolescents’ screen-based sedentary behaviour with depressive symptoms: the Millennium Cohort Study

Changes in physical activity after joining a bikeshare program: a cohort of new bikeshare users

Contact Info

Product

Resources

About