A: Look at the last sentence you quoted.

Do you agree with the math on the hospitalization rates of 1.89% and 3.79%?
If so, how does that agree with the statement from the study?

Quote:

---

The incidence of hospitalization or death did not differ between groups (P = 0.29).

---

Numerically they are different. But there's a reason you do statistical hypothesis testing.
https://en.m.wikipedia.org/wiki/Statistical_hypothesis_testing

amercer said:

---

Skillet Shot said:

---

Do you agree with the math on the hospitalization rates of 1.89% and 3.79%?
If so, how does that agree with the statement from the study?

Quote:

---

The incidence of hospitalization or death did not differ between groups (P = 0.29).

---

---

(P = 0.29) means that the results are from random chance, not an actual effect.

---

Stuff like this is really hard for people to grasp if they don't deal with it regularly because it's so counter intuitive. Our brains have evolved to see all connections and correlations as causal.

We can now move on to Type 1 and Type 2 Errors. Quiz on Monday.

Infection_Ag11 said:

---

amercer said:

---

Skillet Shot said:

---

Do you agree with the math on the hospitalization rates of 1.89% and 3.79%?
If so, how does that agree with the statement from the study?

Quote:

---

The incidence of hospitalization or death did not differ between groups (P = 0.29).

---

---

(P = 0.29) means that the results are from random chance, not an actual effect.

---

Stuff like this is really hard for people to grasp if they don't deal with it regularly because it's so counter intuitive. Our brains have evolved to see all connections and correlations as causal.

---

I will admit this is tough for me to grasp, without any formal statistics training. I am looking into to try and understand better.

If anyone wants to walk through the P value calculation, I would very much appreciate it.

Quote:

---

We had originally designed the trial assuming an 8% incidence of hospitalization and 2% incidence of intensive care unit stay or death (10% in total for these adverse outcomes) (14, 15). Using a proportional odds model with an estimated 50% effect size to reduce these ordinal outcomes with a 2-sided level of 0.05 and 90% power, we had estimated 621 participants per group. With a novel internet-based trial, we had assumed that loss to follow-up might be higher than in a traditional trial; therefore, we had adjusted the sample size by 20% to 750 participants per group.
The primary analysis cohort included participants who completed at least 1 follow-up survey, so that change in symptom severity score could be assessed. The symptom severity score was self-assessed using a 10-point visual analogue scale (0 to 10, with 0.1-point increments). We assigned a severity score of 0 to those with no symptoms. Those who died of complications related to COVID-19 were assigned a severity score of 10 for any surveys missed up until the date of death. Both actual severity scores and changes in score from baseline were assessed for normality (Supplement Figure 4). We used a longitudinal mixed model, adjusted for baseline severity score, to analyze the primary end point of change in symptom severity through day 14. The absolute difference and 95% CI for change in severity score from baseline between groups are presented, along with the relative difference, calculated as [(hydroxychloroquine mean placebo mean) / placebo mean]. A priorispecified subgroups for the primary outcome included days of symptoms before enrollment, age, sex, and laboratory-confirmed infection versus probable COVID-19. The primary end point was additionally assessed by medication adherence, zinc use, or vitamin C use as post hoc analyses. The Supplement gives additional detail on statistical methods and sensitivity analyses.
Analysis of the ordinal secondary end point of no hospitalization, hospitalization, or admission to the intensive care unit or death was not done because of the low event rate. The overall incidence of hospitalization or death was compared between the groups with Fisher exact tests. The analysis cohort for the outcome of hospitalization or death included all randomly assigned participants with vital status known at any point during follow-up. The presence of symptoms at each time point was assessed with the Fisher exact test, and we analyzed change from baseline symptom severity score at each visit using linear regression, adjusted for baseline severity score. We did analyses with SAS software, version 9.4 (SAS Institute), according to the intention-to-treat principle (that is, all participants with data are included in the analyses regardless of their medication status) with a 2-sided type I error using an of 0.05. No adjustments for type I error were made to account for the number of secondary and subgroups analyses; therefore, subgroup analyses should be interpreted with caution.

---

Some of you need to go visit Ms. Julie Carrol. I somehow aced her tests when class averages were in the 50s. #notsohumblebrag

Too bad I forgot most of what I learned.

Glad you're taking the effort to learn.

Here's hopefully a good example. Suppose you have a hypothesis that Aggies are smarter than Longhorns. You randomly select 100 A&M and 100 UT students and give them a math test. The average scores were: A&M 92, UT 88. But does that mean you can reject the null hypothesis that there's no difference between these two populations and proclaim A&M students as smarter? No. You have to take into account sample size and variance within the groups' scores.

So what is .05? That means if you took other samples of the population (remember there's FAR more A&M and UT students than 100), 95% of the time you would observe A&M student sample's scores being greater than the UT sample.

Not sure if this explanation helps. Trying to discuss stats on a phone is challenging..

All I know about Stats is that I failed STAT in about the most epic fashion ever possible. That was the worst class I ever signed up for.

I understand the concept now. A p-value of 0.29 means there is a 29% probability of chance accounting for the results in the study.

Now I want to make sure I understand how it is calculated. I calculated a p-value of 0.24 using chi square distribution chart. Is this correct?

I'm following the example here: https://math.hws.edu/javamath/ryan/ChiSquare.html

Using the same methodology, but including deaths and hospitalizations I get a p-value 0.28, which is close to the study finding of 0.29

So the study shows HCQ yields a 50% reduction in hospitalization rate. HCQ = 1.89%, Placebo = 3.79%

BUT

Based on the P-value of 0.29. There is a 29% probability that the results are due by chance. Therefore, since p-value > 0.05, the 50% reduction is dismissed as statistically insignificant.

However, what if the hospitalization reduction rate of 50% is constant regardless of sample size? Let's say we triple the study size and keep the hospitalization rate constant.

Scenario 1: Triple the study size and hospitalization rate held constant



Observation: Same hospital rate for the larger study size yields a chi square value of 4.18 which would put the p-value <0.05. This hypothetical study would show that HCQ is an effective treatment and the study is statistically significant.

How do we know that HCQ treatment is not CLINICALLY significant even though the original sample size shows the conclusion is not STATISTICALLY significant.

Just because the statistics do not warrant enough confidence to confirm HCQ is effective with certainty, does not necessarily mean it is ineffective. I think the 50% reduction rate in hospitalizations found in the study to be encouraging and warrants additional studies with larger sample sizes.

The biostats for clinical trials is a field unto itself. Suffice it to say that you have to set all the numbers ahead of the trial to ensure that no one just squints at the data and sees what they want to see. Retrospective analysis of a failed trial is usually garbage.

This is my beef. A 50% reduction in hospitalization rate from HCQ treatment would be an obvious success.

The number of patients in the study and the number of hospitalizations in the placebo group are independent of HCQ effectiveness. So even if the hospital reduction rate is actually 50%, the result will be statistically insignificant by default.

Skillet Shot said:

---

This is my beef. A 50% reduction in hospitalization rate from HCQ treatment would be an obvious success.

The number of patients in the study and the number of hospitalizations in the placebo group are independent of HCQ effectiveness. So even if the hospital reduction rate is actually 50%, the result will be statistically insignificant by default.

---

One caution, a clinical trial only "proves" what it is originally designed to demonstrate so you cannot say that these outcomes are "significant" (statistically) either, especially when a few patients in either arm can swing dramatic differences. You would have to design another trial specifically to show that these number hold up. Off the cuff, that's about a 2400 person trial minimum.

I don't care at all if HCQ is a wonder drug or not, but is there any outpatient treatment right now that has been shown to reduce the chance of hospitalization or death at home from Covid?

I am just asking because I feel like it is a matter of time before I get it and want to know if there is something that I can go get that will give me a better chance or avoiding either of those two outcomes. Does aspirin help avoid clots? A blood thinner? Steroid shot? Anything?

Skillet Shot said:

---

This is my beef. A 50% reduction in hospitalization rate from HCQ treatment would be an obvious success.

The number of patients in the study and the number of hospitalizations in the placebo group are independent of HCQ effectiveness. So even if the hospital reduction rate is actually 50%, the result will be statistically insignificant by default.

---

Agree....isn't this how we are treating masks??? No one knows how effective they are. Everyone knows they aren't 100% effective...but we have been conditioned to think that they are effective at 'helping' lessen the viral load transmission. So masks along with other lifestyle changes slows spread (not going to debate if that is the best approach or not).

Yet masks are a must and HCQ is still not effective.