ols vs logistic regression

[Saulius Ringaila]

Good evening ladies and gents

May I ask for your expertise in choosing the above. I want to test the hypothesis of: adolescents time spent in MVPA on weekends vs weekdays (moderate to vigorous physical activity) association with: age, gender, ethnicity, household income. After running multiple regression analysis with the dependant variable MVPAweekend for example, the R2 is very small 0.2 and everything seems to be insignificant apart from gender (male p>0.001 ), income level (p>0.05), and bmi (obese p>0.05). Should I run logistic regression analysis instead? dependant variable converted to 1-if 60min mvpa per day is achieved and 0- if not.
Thanks if anyone can be bothered

[Roman Mostazir]

I don't think you will get anything different with a logit model. You also lose validity of the variable mvpa by dichotomising it as it loses valueable information when you change it to categorical from continuous. The logit model will also have very different interpretations (Odds ratios) about the predictors compared to what you are seeking from your current model.

[Jeff Wooldridge]

Saulius: Since your Y >= 0, and I suspect you have some zeros, you can try Poisson regression with robust standard errors. I’d probably use log of income to estimate an elasticity in that case.

[Duong Le]

Saulius: Since your Y >= 0, and I suspect you have some zeros, you can try Poisson regression with robust standard errors. I’d probably use log of income to estimate an elasticity in that case.

Dear Professor Jeff Wooldridge,

Is it possible to use Poisson regression in my case? For example, my Y is the number of caregiving hours per day so its range is from 0-24 (around 30% of zeros) but not continuous because there are jumps (e.g., from 10h to 13h). My Y also have some fraction values, i.e. 0.3 or 0.5 hours per day. I would be grateful if you can suggest me which regressions are appropriate for my case.

Thank you.

[Jeff Wooldridge]

Jason: Do you also have values at 24? Or close to 24?

[Saulius Ringaila]

Thank you Jeff, very appreciated for the advice and your time, will look into this further. Also do you think that, perhaps it would be better to go with multiple regression models: each model would show if adolescents with particular characteristics are less likely to engage in physical activity on weekends
1 model - weekmvpa (continuous var) :age, gender, ethnicity, income, bmi categories. (eststo weekmvpa)
2 model weekendmvpa + same vars. (eststo weekendmvpa)
3 model mvpa in 10 minute bouts weekday +same vars (eststo weekbouts)
4. weekend mvpa10min bouts + same vars (eststo weekendbouts)
(esttab, r2 ar2 se scalar(rmse)

[Duong Le]

Jason: Do you also have values at 24? Or close to 24?

Dear Professor Jeff Wooldridge,

My Y also has values at 24.

Thank you.

[Jeff Wooldridge]

Jason: I’d go with the binomial quasi-MLE:

glm y x1 ... xk, fam(bin 24) link(logit) vce(robust)

and then use the margins command.

[Duong Le]

Dear Professor Jeff Wooldridge,

Thanks a lot for your useful suggestions.

[Saulius Ringaila]

excuse me gents (or ladies) could anyone voice their opinions on income, am not too sure am interpreting it correctly. Picture attached.
Dependant var. is physical activity in minutes per day, mean 142.33 SD74; min. 1 - max. 504.
Income: total household income per week. Mean 451£/week ; SD175 min. 0 - max. 986.85 £/week.
So, am thinking: for every extra 1£ in total income/ per week, you more likely to spent extra 15 seconds? coef. 0.015, p=0.128
Please correct me if am wrong, kindly appreciate any input.

[Saulius Ringaila]

Table attached.