Frantic - need advice on building a report POC using SSAS 2012 [closed]

StackOverflow https://stackoverflow.com/questions/21210129

Question

I apologize in advance for the wordiness of this post. I never know what will be important so I usually write more than needed.

A little background

I have a week and a half to create a POC for using OLAP behind client-facing reports. We aren't committed to a reporting tool and we have a lot of data, so OLAP seems to make sense(?). Right now we use SSRS 2005 against a somewhat flattened reporting database but we have a new client who's asking us to become sophisticated quickly.

I worked with SSAS 10 years ago - built cubes and simple pivot tables off of them - no MDX. I'm savvy with SSRS, but against relational sources. We have no dimensional model so I have to mock one up. I was going to do the mocking in SSAS 2012 (against a SQL 2005 DB).

Requirements for the POC

Report information about physician performance across multiple attributes

Dimensions:

  • Time (Date/Month/Year)
  • Physician
  • Physician Specialty (1:M with Physician)
  • Affiliation 1 > Affiliation 2 > Affiliation 3 (hierarchy, 1:M with Physician)
  • Registry > HEDIS Measure (hierarchy, M:M with Physician, HEDIS hereinafter because "Measure" gets confusing)

Measures (lowest granularity is Physician/Date/HEDIS - drillthrough to patient data will be necessary at some point):

  • Patient Population
  • Patients Seen (a portion of Patient Population)
  • Score (this is our KPI - quotient % of Patients Seen / Patient Population)
  • Quartile (based on Score; measured for Physician, Affiliations 1, 2 & 3, HEDIS and Specialty across Day/Month/Year)

I have created a functional cube with all of the dimensions and the two additive measures (Patient Population & Patients Seen). I went to add Score and Quartile and froze. Now I'm in analysis paralysis mode and panicking. I didn't know MDX was so unintuitive (or maybe I'm just dense), and that calculating quartiles would be such a problem!

So now I'm trying to throw something together in the database with views and static tables. I'm pretty junior when it comes to dimensional modeling. I need to design the tables for the quickest, easiest cube development and report turnaround possible. It doesn't need to be perfect, but I haven't seen a project like this all the way through and I would love some advice on how to avoid running into obvious "gotchas" during cube and report development due to poor DB design choices. Can someone can give me a generalized "what would you do in my shoes" run down?

Here are some of my questions/concerns in the form of mental vomit

So I have fully additive and non-additive measures, right? (I'm not even sure what Score is - I don't think it qualifies as semi-additive?). Anyway, I'm torn as to whether to store these non-fully-additive metrics in a Measure Dimension or multiple fact tables at different granularities.

It seems like going the fact table route might be less confusing, but then does each fact table get its own cube and drill down/across would be accomplished through some sort of linking in Excel or SSRS? E.g., you're looking at the YTD Physician Scores by Annual HEDIS Quartile...how do you drill down to the MTD Physician Scores by Monthly HEDIS Quartile if they are in different cubes? Or would they be in the same cube in different Measure Groups maybe...? Or if I go with the Measure Dimension and use a single cube, how do I protect the user from the scenario above...they're looking at the YTD Physician Scores by Annual HEDIS Quartile, then replace Annual HEDIS Quartile with Monthly HEDIS Quartile - how is such a thing that prevented? Or is this scenario even legit?

I'd love to throw this in SSRS where I can control things through parameters but how hard is it to such against an OLAP source? Not to mention interactive graphs?

I am so confused right now I don't even know if these questions are making sense. Any help (or even links to succinct documentation that you found helpful) would be great!

Was it helpful?

Solution

Some thoughts:

  • I would put physican, specialty, and the affiliations into one dimension, say "physican" as separate attributes, and the build a user hierarchy from the affiliations.
  • This leaves you with three dimensions: time, physican, and registry (I do not understand the details of this from your description), and one fact table.
  • The fact table should contain the two measures "patients population" and "patients seen". "patients population" is the number of patients per physican, time, and registry. I would assume it is not aggregatable across time. You would have to decide which aggregation you use across time, e. g. LastChild, i. e. each month shows the value of its last day, each year the value of its last month, etc. For the "patients seen", I am not sure what you want to do: Should this be distinct patients seen? In this case, you cannot aggregate across time. You would have to have the data on patient level, and link the patient dimension (possibly just consisting of the patient id and invisible to end users) via a many-to-many relationship to the main fact table. Then the measure could be a DistinctCount.
  • "Score" would be a simple calculated measure - simply [Measures].[patients seen] / [Measures].[patients population].
  • YTD and MAT is normally implemented as calculated members in MDX as well. The best solution would be using a utility dimension (see e. g. http://sqlblog.com/blogs/marco_russo/archive/2007/09/02/datetool-dimension-an-alternative-time-intelligence-implementation.aspx). But for the POC, you can also just implement it on a per query base in the WITH clause: Define it as a member of your time hierarchy using the Ytd and ParallelPeriod functions.
  • For the Quartile, I think the best way - at least for the POC - woudl be implementing this in a query specific WITH clause in the MDX that you write in SSRS. I would rank the set on which you want to calculate the quartiles, and then use a CASE expression with four WHEN clauses for the four quartiles, something like `WHEN rank >= count / 4 and rank < count / 2 THEN "Q2", where you would have to decide how to exactly handle the borderline cases.

OTHER TIPS

OK, I found an article that pointedly and succinctly answered most of my nagging questions about non-aggregatable measures and then a YouTube video that showed how easy it actually is to create an SSRS report from an OLAP cube.

Here they are for posterity:

Article - http://www.packtpub.com/article/measures-and-measure-groups-microsoft-analysis-services-part2 Video - http://www.youtube.com/watch?v=o2N3ZCinpPs

I think I will create multiple fact tables (views for now) at different granularities and introduce each one into the cube as its own Measure Group, linking it to the conforming dimensions at different levels of detail. Then, for the reporting layer, I'll use the trick in the link above under the title "Non-aggregatable measures: a different approach".

It looks like Score should just be a simple matter of creating a calculated measure against Population and Seen.

I'll handle the quartiles by mocking up a dimension, then calculating the quartile in the DB at each relevant granularity and storing that number in the facts. Then I'll just use that number as the key by which to join to the quartile dimension in SSAS. Seems quick and straight forward.

I'd like to leave this open for another day or so, just in case someone out there disagrees with this approach?

Licensed under: CC-BY-SA with attribution
Not affiliated with StackOverflow
scroll top