chi2

So chi square can be used here?

It sounds like the way you're describing hypertension --- as 1x, 2x, 3x, and so on --- that it would be treated as an ordinal variable, not as a nominal categorical variable.

So if i treat it as ordinal data, chi square works?

I hate when people give you a one word answer to a complex question. It's always wrong.

A chi-squared test does not take into account the ordered nature of the data. And event counts are interval, not ordinal, because they are based on a defined unit.

What you have is the incidence of events, which inevitably involves the concept of duration of exposure. In fact, you have incidence density data. Incidence density is the rate of occurrence of events per unit of exposure time. (The other kind of incidence is cumulative incidence, which is the rate at which members of the population will experience one or more events in a given unit of time, so it counts people, not events.)

You do not mention the period during which the patient was exposed to risk of hypotension, but assuming that it's something like the duration for which they were receiving a prescribed drug, then you can model the number of events in that period using Poisson or negative binomial regression.

I would also have a look at this graphically, using, say, a lowess smoother to see what shape the relationship between number of episodes and age has.

If using Poisson or NBregression, use robust standard errors.

Nibras Husein Mhawesh

Your outcome variable (i.e., the incidence of hypotension) seems ordinal with five levels. Therefore, you could go for Spearman's rank-order correlation, which examines the strength of association between two ordinal or interval variables. The following could provide good guidance.

Khamis, H. (2008). Measures of association: How to choose? Journal of Diagnostic Medical Sonography, 24(3), 155-162. https://doi.org/10.1177/8756479308317006

Laerd Statistics. (2018). Spearman's rank-order correlation. SPSS Statistics Tutorials and Statistical Guides | Laerd Statistics. https://statistics.laerd.com/statistical-guides/spearmans-rank-order-correlation-statistical-guide.php

Good luck,

Mohialdeen Alotumi I disagree. The incidence is a counted variable, so it is not ordinal.

Ronán Michael Conroy, when you say counts are not ordinal, I suspect what you mean is that counts are not merely ordinal. Is that fair?

Maheen Wahid, you described the count variable as "once, twice, thrice, four times, five times". Does that represent the actual range of observed counts? I.e., there were no 0 counts? And there were no counts > 5 that were lumped into a 5+ category? If there are no 0 counts, you may need to use a zero-truncated model. E.g., here is a page that discusses zero-truncated Poisson regression with Stata examples:

https://stats.idre.ucla.edu/stata/dae/zero-truncated-poisson-regression/

HTH.

Bruce Weaver Sir there is a zero count as well, and the maximum value is 7 times. How about I label 0 as "no hypotension' and 1 as 'yes hypotension did occur (regardless of number of times it occured'

Then it is a simple yes no variable and i can straightaway go for chi square

According to my knowledge, You can simply use an bar chart and see subjectively. To get a conclusion; you have to do a test and see.

The other variable I mean the occurrence isn't just a categorical variable. It's an ordinal categorical one. Actually, you can use chi-square for categorical one.But for ordinal categorical, there is a separate test unless you may neglect some data. It's a waste of data. Can't you categorize the age group. It's also a data reduction but I think it may useful.

Dear Ronán Michael Conroy

Dear Sal Mangiafico

Dear Bruce Weaver

Dear Mohialdeen Alotumi

one question. I want to learn and know myself. Thank you for your reply

Is it possible for us to consider the study as a linear model below?

DV (age)ij = Intercept + Alphai (incidence of hypotension) + Epsilonij

where in i=1,2,3,4,5 j=1,2,3,...,n

And in the next step, we will use Pairwise comparisons (Estimated Marginals Means) to compare different groups of incidence of hypotension.

Respectfully

Chi-square would have been suitable if the data for both the response and independent variables were frequency data with two or more categories or levels (in the case of ordinal data). Since the independent variable of your study is age (measured continuously) and the dependent variable is categorical (ordinal), the most appropriate statistical method for use in such situations is the "Discriminant Analysis." Discriminant analysis works in the same way as ANOVA but in the opposite direction (read more here https://www.statisticssolutions.com/discriminant-analysis/). However, should you decide to use Chi-square, then you will need to transform the continuous age data from continuous to categorical data, then count the frequency of membership. One way to do this is using the Latent Cluster Analysis. You can also do the transformation manually by taking a range of ordered age groups (for example 10 -19 years, 20 -29 years, 30 - 39 years and so on) then assign respondents to the groups based on the continuous age data.

I hope this helps.

OVJ

Valentine Joseph Owan –absolutely no. Chi-squared does not take account of the ordered nature of the data, and ordinal methods do not take account of the interval nature of the data (counted events).

Furthermore, changing a continuous variable into a series of ordered categories reduces your statistical power, considerably if the number of categories is small.

The question is about counted events and a continuous predictor. There are well-established methods for modelling these.

Useful paper : Collins GS, Ogundimu EO, Cook JA, Manach YL, Altman DG. Quantifying the impact of different approaches for handling continuous predictors on the performance of a prognostic model. Statistics in medicine. 2016 Oct 15;35(23):4124-35.

Assuming it is a count variable—albeit not clear from the rendered description, you could go for Poisson regression model. The book by Cameron and Trivedi (2013), referenced below, could be helpful.

Cameron, A. C., & Trivedi, P. K. (2013). Regression analysis of count data (2nd ed.). Cambridge University Press. https://www.cambridge.org/core/books/regression-analysis-of-count-data/2AB83B406C5798030F7C91ECC99B1BE4

Good luck,

Abolfazl Ghoodjani , To me it would make more sense if Incidence of Hypotension were the dependent variable. If so, a good approach would be to follow Ronán Michael Conroy 's suggestion of Poisson or negative binomial regression. .... I guess you _could_ view the problem as you have proposed (treat Age as the DV, and treat Hypotension as individual groups), if that's really the way you want to view the data.

Sal Mangiafico Thank you for your answer

To add to Sal Mangiafico 's point, I don't see why age would be the dependent variable, since hypotensive episodes don't cause one to age, but age may change risk of hypotensive episodes.

And as to what causes one to age, the major risk factors are teenagers, committees and curriculum development.

Given Seán R. Millar's enjoyment of Ronán Michael Conroy's last post, I'll risk knocking this thread right off the rails by giving a cheeky response to one of Ronan's earlier posts--the one where he wrote this:

I hate when people give you a one word answer to a complex question. It's always wrong.

Forty-two. :-|

Just to establish a relationship an ordinal measure of association would be sufficient. I'm fond of Kendall's tau in this context but rho, gamma and several others are available. Indeed Pearson's r between age and the hypotension measure would likely be serviceable.

If you want a more fine grained model (e.g., for prediction) then an ordinal logistic regression with age as a predictor would be a reasonable choice. You can potentially capture more subtle relationships by including quadratic effects or splines.

I agree, Thom Baguley , that an ordinal measure of association would be sufficient, but it doesn't necessarily give a measure of effect size that can be interpreted.1 An incidence rate ratio from a counted model would be preferable. While age is not a modifiable risk factor for hypotensive episodes, number of prescribed medications is, and it would be useful if prescribers knew how each additional medication they prescribe increases the risk of a hypotensive episode.

1 Kendall's tau does have an interpretation but not one that would impress a clinician!

I think that the data are parametric assuming that the hypotension data could be converted to numbers. In that case Pearson correlation would be applied fulling its other assumptions. If assumptions (numbers of samples) of parametric tests cannot be fulfilled then non-parametric approached should be adopted. Best wishes

Yadunath Bajgai - in fact, the data are numbers. But data cannot be parametric or nonparametric. A parameter is a property of a population, not a property of a measurement. Some statistical procedures estimate parameters (including the Wilcoxon Mann-Whitney test!) while others do not. And there are even tests that are both parametric and nonparametric (Cox proportional hazards regression is a notable example).

@ Ronán Michael Conroy, Thank you much very for enriching my understanding of the statistical procedures. Your invaluable observation is very useful for me and others following the questions. Best regards.

Thank you everyone for your input. I have indeed learnt a lot from this discussion!

Chi-square test of association

Nibras Husein Mhawesh

Most of the responses suggested above are helpful. As the incidence of hypotension variable in this particular scenario is a count variable (1, 2, 3, 4 and 5) and the age variable is continuous, it would be helpful to set incidence of hypertension as dependent and age as independent in the potential count regression models(Poisson, NB, Zero Inflated models, Hurdle models). The choice may depend on the result of model assumptions checking.

The other potential solution would be to convert the variable into categorical variable and use the chi-square analysis as usual.

Categorize age and do chisquare test for association

Ravishankar Suryanarayana , why would you want categorize a continuous variable (age) ? And why would you want to treat an ordinal variable (hypotension) as a nominal variable ? In both cases you are loosing information from the original data. ... Some of these points have been brought up in previous comments.