Pastor Andrew

Pharmacist-epidemiologist

[Pastor Andrew @Andrew_Simwanza]

Significance without effect size is an incomplete result. In a large dataset, it is almost meaningless.

[Pastor Andrew @Andrew_Simwanza]

The chi-square test told you the association was significant.
It did not tell you how strong it was.
A chi-square statistic is a function of both the effect size and the sample size.

[Pastor Andrew @Andrew_Simwanza]

In a large enough dataset, even a trivially weak association will produce a significant chi-square. After the chi-square, report Cramér's V or the phi coefficient. These measure the strength of the association independently of sample size.

[Pastor Andrew @Andrew_Simwanza]

Adding variables to a regression model without understanding their causal role does not improve the analysis. It can destroy it.
Draw your directed acyclic graph before you build your model.

[Pastor Andrew @Andrew_Simwanza]

A confounder and a mediator look identical in a regression model.
Both are associated with the exposure and the outcome. Both change the coefficient when you add them to the model.
But they require opposite analytical responses.

[Pastor Andrew @Andrew_Simwanza]

You adjust for a confounder to remove bias. You do not adjust for a mediator, because the mediator is the pathway through which the exposure causes the outcome. Adjusting for it blocks the very effect you are trying to estimate.

[Pastor Andrew @Andrew_Simwanza]

It will give you a correlation coefficient that misrepresents the actual association.
The choice of correlation method is not a formatting decision. It is an analytical one. Make it deliberately.

[Pastor Andrew @Andrew_Simwanza]

Pearson or Spearman?
Most researchers pick Pearson by default and never think about it again.
Pearson measures the strength of a linear relationship between two continuous, normally distributed variables.

[Pastor Andrew @Andrew_Simwanza]

Spearman measures the strength of a monotonic relationship between two variables and makes no assumption about normality or linearity.
If your data is skewed, if your variables are ordinal, or if the relationship between them is not linear, Pearson is the wrong choice.

[Pastor Andrew @Andrew_Simwanza]

You need both answers. The p-value only gives you one.

[Pastor Andrew @Andrew_Simwanza]

The t-test came back significant.
p = 0.04.
The difference between groups was 0.3 mmHg in systolic blood pressure.
Clinically meaningless. Statistically significant.
This happens when sample sizes are large enough that even trivial differences produce small p-values.

[Pastor Andrew @Andrew_Simwanza]

The test detected a real difference. The difference just does not matter to any patient alive.
Statistical significance answers one question: is this difference likely to be real? Clinical significance answers a different one: does this difference matter?

[Pastor Andrew @Andrew_Simwanza]

A before-and-after comparison without a pre-intervention trend is not an evaluation. It is a description of two time points with a story connecting them.
Use interrupted time series. Use the data you already have. Use it properly.

[Pastor Andrew @Andrew_Simwanza]

The data said the intervention worked.
The stockout rate dropped from 34% to 11% in six months.
Nobody asked whether the improvement was already underway before the intervention started.
Nobody plotted the pre-intervention trend.

[Pastor Andrew @Andrew_Simwanza]

Nobody tested whether the change in level and slope after the intervention was larger than what the existing trajectory would have predicted.
The intervention may have worked. The analysis did not establish that it did.

[Pastor Andrew @Andrew_Simwanza]

The hazard ratio tells you how much faster or slower the event occurs in one group relative to another, at any given point in follow-up.
Time is not a covariate in survival analysis. It is the outcome. Treat it accordingly.

[Pastor Andrew @Andrew_Simwanza]

Survival analysis is the only method that respects the full structure of time-to-event data.
A patient who dies at three months and a patient who dies at six years are both events.

[Pastor Andrew @Andrew_Simwanza]

Treating them as equivalent in a binary outcome model discards the most clinically meaningful dimension of the data.
Cox regression models the rate at which the event occurs over time.

[Pastor Andrew @Andrew_Simwanza]

Antimicrobial stewardship without drug utilisation surveillance is clinical advice without epidemiological grounding. Both are necessary. Neither works without the other.

[Pastor Andrew @Andrew_Simwanza]

A new antibiotic enters the market.
Prescriptions rise. Consumption doubles within two years.
Resistance rates begin climbing eighteen months later.

[Pastor Andrew @Andrew_Simwanza]

Nobody connects the two datasets because antimicrobial consumption data and resistance surveillance data sit in separate systems, managed by separate departments, reported to separate audiences.
The signal was there. The infrastructure to detect it was not.