Can statistical tests be applied to analyze temperature anomalies in vaccine refrigerators across different shifts and days?

Question

Background: I'm analyzing the vaccine refrigerator's internal temperature data collected during COVID pandemic (2021-2022) from all over an country. I measured the internal temperature of the 1077 refrigerators every 30 seconds over a period of 1-2 years through IoT monitoring System.

However, some of the refrigerators have insufficient amount of data due to loss of connection and maintenance. Thus I filtered the refrigerators collected the data over 1 year at least. As a result, 448 vaccine refrigerators are remaining.

Aim of analysis: My aim is to analyze the frequency of temperature anomalies(under 2°C or over 8°C) and duration of anomalies based on two criteria:

1. Shift (Day/Evening/Night)
2. Day of the Week

I hypothesize that there will be differences in the occurrence of temperature anomalies and duration of anomalies in vaccine refrigerators by shift and day of the week.

For each refrigerator, I have calculated the total number of anomalies frequency and the duration of anomalies observed during each shift over 2 years. Here’s a sample of the data:

Data Example: The data examples below are samples of the data I collected.

1-1) Frequency by shift

1-2) Frequency by day of the week

2-1) Duration by shift

Although these duration data are presented as means, I have the total anomaly data. Since the total dataset is so large, it's difficult to review it all. Thus, I present the sample of duratino mean of anomalies.

(hh:mm:ss)

2-2) Duration by day of the week

I want to use this data to check if there are differences in the frequency and duration of temperature anomalies according to work hours and days of the week, using data from vaccine refrigerators measured over two years.

To do this, I tried to apply various methods such as Wilcoxon signed-rank test, Friedman test, and generalized linear mixed models, but I found it difficult to apply them in the way I wanted. Given that the data is not strictly repeated measures, Wilcoxon signed-rank test, Friedman test and GLMM are not directly applicable.

What would be a good way to test this? Or would it be sufficient to simply show it descriptively using statistical techniques without using hypothesis tests? Or should I try a different approach?

Plus, given the limitations of the original analysis, I explored the possibility of grouping the vaccine refrigerators by region and conducting a one-way ANOVA or Kruskal-Wallis test. Is this a viable approach?

Any help or suggestion would be greatly appreciated. Thank you!

Answer 1

I think you should be able to used mixed models for this. Refrigerator ID is a random effect, and day of the week or shift are fixed effects.

One thing to keep in mind is that my summarizing mean duration for each refrigerator, you're losing information. If you set up day and shift as predictors, then you have repeated measurements for each fridge, which is a good situation for using random effects. So you don't need to summarize means and regress means - you can just use the direct data.