Commom factor analysis is acceptable. Item Response Theory would be ideal.

Matt

Follow.

Hi Samrat,

it depends on your theory about the structure of the symptoms. As Matthew Kerry pointed out, the traditional/usual way of thinking is in terms of symptoms being the result of an underlying disease which renders the disease a unidimensional entity causing the symptoms. There were recently some challenges to this model. Especially the working group around Denny Borsboom published a bunch of paper in which they argued that a network structure (and model) would be more suitable. Perhaps this is worth considering in your point, too. Although the work by Borsboom is rather focused on psychopathology, the range of applications is broader. I included the main references below.

Best,

Holger

Borsboom, D., & Cramer, A. O. J. (2013). Network analysis: An integrative approach to the structure of psychopathology. Annual Review of Clinical Psychology, 9(1), 91-121. doi:10.1146/annurev-clinpsy-050212-185608

Borsboom, D. (2008). Psychometric perspectives on diagnostic systems. Journal of Clinical Psychology, 64(9), 1089-1108.

Epskamp, S., Borsboom, D., & Fried, E. I. (2018). Estimating psychological networks and their accuracy: A tutorial paper. Behavior Research Methods, 50(1), 195-212. doi:10.3758/s13428-017-0862-1

Epskamp, S., Cramer, A. O. J., Waldorp, L. J., Schmittmann, V. D., & Borsboom, D. (2012). Qgraph: Network visualizations of relationships in psychometric data. Journal of Statistical Software, 48(4), 1-18. doi:10.18637/jss.v048.i04

Thank you all for the prompt reply.

I have run the factor analysis using PCA (both oblimin and varimax). But am having difficulty in inferring the output and whether it is what I need. Can anyone help.

Hi Samrat,

a PCA is no factor analysis (see the papers below) but a theory-less data reduction technique that sorts items into groups based on their intercorrelations. The level of inference does not go behind the data. Hence the consequences are composites of measured data without any "surplus meaning". Applied to your field, you would sort symptoms into groups of co-occurring symptoms.

A factor analysis tries to identify latent variables that could have caused the data. Applied to diseases, this would resemble the examples by Matthew Kerry that reflect the notion of a singular latent disease variables that causes a set of symptoms. CFA and IRT reflect express such a conviction as EFA does.

If the factor model is not appropriate (but the network structure), then PCA may indeed the more appropriate way to go, as a "disease" would then rather be label for a set of symptoms rather than their underlying latent cause.

HTH

Holger

Fabrigar, L. R., Wegener, D. T., MacCallum, R. C., Strahan, E. J., MacCllum, R., & Strahan, E. J. (1999). Evaluating the use of exploratory factor analysis in psychological research. Psychological Methods, 4(3), 272-299. doi:10.1037/1082-989X.4.3.272

Bartholomew, D. J. (2004). Measuring intelligence - facts and fallacies. Camebridge: University Press.

Russel, D. W., & Russell, D. W. (2002). In search for underlying dimensions - the use (and abuse) of factor analysis in personality and social psychology bulletin. Personality and Social Psychology Bulletin, 28(2), 1629-1646. doi:10.1177/014616702237645

Preacher, K. J., & MacCallum, R. C. (2003). Repairing tom swift's electric factor analysis machine. Understanding Statistics, 2, 13-43.

Hello Samrat,

You have received some thoughtful advice above, and I suggest you think carefully about the applicability of that advice because it might feed in nicely to your research aims.

I would recommend approaching the problem from a completely different angle, however, if you are interested in effective screening for a disease. This approach involves determining metrics such as sensitivity and specificity, and perhaps predictive values associated with each of your 15 variables - or maybe even a combination of the "best" of those variables. About 14 months ago, I wrote an article about these things. Here is the reference:

Trevethan, R. (2017). Sensitivity, specificity, and predictive values: Foundations, pliabilities, and pitfalls in research and practice. Frontiers in Public Health, 5, 307. https://doi.org/10.3389/fpubh.2017.00307.

I mention it because it seems to be of interest and helpful to people, with nearly 13,000 views since it was published.

So:

It would seem that you know who, among your total sample, definitely have the disease, and who definitely do not have the disease. Presumably, you have a good basis for knowing that - which is referred to as a "reference standard". As I demonstrate in my article (and the basics of that article are by no means new), the mathematics for determining sensitivity and specificity are extremely simple (so your lack of knowledge about stats can be put aside), and I think you could quickly get an impression about which of your 15 variables were most and least associated with your target disease. It would take a bit of time to do the full set of calculations, and perhaps some combinations as you gained insights along the way - but the calculations can be done by hand and are by no means daunting. My hunch is that they could be enjoyable as you saw your results unfolding.

I am a fan of factor analysis (when used properly), but I have a hunch that it might not be the most appropriate strategy for your current aims.

Please feel free to get back to me if I might be able to offer further assistance. I am writing a subsequent article on the topic, but my original article might contain sufficient information for your current needs.

Robert.

If the your latent variable (disease X) is unidimensional, IRT models, especially Rasch analyses, may be the more appropriate than PCA or factor analyses. Rasch analyses share some conceptual ground with sensitivity/specificity discussed above, and may help you identify a smaller set of questions that will help you simplify your diagnostic tool.

Thank you all for all your inputs. I need to go through some of the articles that you all have posted. I freely admit that I am out of depth as this is the first time I am into something like this so I need to read a bit more. But I am throughly enjoying this new exploration. Thank you all for taking time out to reach out to me.

Holger Steinmetz Robert Trevethan Jeffrey M. DeVries

Hello Samrat,

Please feel free to get back if any of us might be able to help you more. Just quickly, I think you could try to get 15 different 2 x 2 cross tables (say, from Excel or a stats package). Your reference standard (whether the people definitely had the disease or not) would always be at the top of those crosstabs, and the symptoms (presumably coded as yes or no) would change for the rows of each crosstab. Then you could easily calculate the sensitivity and specificity for each of those crosstabs. That doesn't require any difficult maths. I think you don't need to worry about predictive values at this stage.

Then, I think you could see which of your 15 symptoms had the highest sensitivity and specificity. In my view, both of those values would ideally be at least 85%.

Robt.

Robert Trevethan Hello Robert,

I am doing the calculations. Thank you

Hello Samrat,

I hope you enjoy what you're doing. If you would like to share with me your findings, concerns, etc. in a less public arena, please feel free to contact me on my email address: [email protected].

I don't want to be a co-author on any publications, etc., etc.; I just would like to help you - and I find the kind of thing you are doing to be very interesting.

Robt.

Thats very generous of you. I would surely do so. Thanks Robert Trevethan

Hello all. I am working on this data set. I need some help with calculating Prevalence in age groups adjusted for sex and prevalence in sexes adjusted for age. I need either an online calculator/Excel format/ or any other. Robert Trevethan Jeffrey M. DeVries Holger Steinmetz Amit Choudhury

Hello Samrat,

If I understand your question correctly, all you need is a cross tabulation of age against sex. If you ask for row and column percentages in addition to numbers in the cells, you should be able to get the information you need.

Cross tabulation is easy to obtain, and the output is nice, if SPSS is used. I've not used Excel for this kind of task.

Robt.

Dr. Samrat,

You may try multi-criteria decision making methods (MCDMs) to identify the most relevant symptoms. or you can go with a decision tree approach, which will give you an idea of the factors associated with the disease.

You could analyze every symptom via a mesh field analysis in Pubmed and see which one has the highest prevalence with the disease. Here is a simple algorithm to identify the diagnosis field in MeSH.

( "Glaucoma/diagnosis"[Mesh] OR "Glaucoma/diagnostic imaging"[Mesh] )