PDA

View Full Version : How to calculate trend

Purple Princess
06-12-2003, 01:47 PM
Can anyone give me advice on where I could find some really basic information on how to calculate claim trends in group health insurance? I'm mainly looking for stuff on which data should be used, i.e. should I take only employees who had coverage for the full period I'm looking at, etc....

Dr T Non-Fan
06-12-2003, 05:16 PM
It does depend on what you plan to use the trend for.

I suggest:
1. Using all exposure of risk.
2. Normalizing for change in age distribution.
3. Normalizing for change in benefit design distribution.
4. Realizing that historical trend might have no bearing on future trend.

Toll Free
06-16-2003, 12:00 AM
The simplest calculation of trend is twelve months of incurred claims, per member per month (paid in 15 months) divided by the prior twelve months of incurred claims PMPM, also paid in 15 months. For example, claims incurred January-December 2002, paid January 2002-March 2003, divided by member months (or, more clasically, employee months) in 2002, produces Claims PMPM (or PEPM). Repeat for 2001, and 2002 PMPM divided by 2001 PMPM, minus 1, "equals" trend. This method, although readily attainable by most pricing departments, is chock full of errors, such as:

- Does not account for changes in demographics (age, gender, area, industry, etc.) [DTNF said to correct for this]
- Does not account for changes in plan design (groups migrating to higher deductibles) [See DTNF]
- If individual or small group, does not account for changes in average duration.
- Is actually an approximation of trend about 15 months ago (assuming trend is the "slope" of some imaginary claim cost PMPM curve, the definition above is an estimation of the slope of the curve 15 months ago)

etc.

That's why more sophisticated trend models will account for more than simple "total exposure" risk - the theory being that smaller parts may be easier to predict. For example:

- Most folks try to estimate severity (the cost of a service) and utilization (how often services are used) independently. (Really fancy models will also use intensity - the substitution of one high cost service, such as an MRI, for another, such as an X-ray.)
- You may want to examine trend separately for inpatient hospital, outpatient hospital, physician, and drug charges separately.
- You may want to fit some sort of regression line to the claim costs, as opposed to period-over-period analysis.
- You may want to examine "billed", or some other charge level pre cost-sharing, to remove the effect of plan design changes
- You may want to account for the fact that October claims will be higher than September claims, by virtue of the fact that there is one more day in October. This may not sound like a big deal, but one day is 3%.
- If you trust your reserve factors, you may want to use completed charges, rather than a "12/15" definition"

To answer your question simply - use all the data, use a PMPM approach, so that it does not matter whether or not an individual employee was covered for the entire period, and normalize your information as much as possible.

Good luck.

Toll Free
06-19-2003, 11:39 PM
Naturally, I forgot the most common method of calculating trend:

Use whatever your boss read in National Underwriter, Medical Benefits, or the Wall Street Journal this morning, or perhaps some time in the last three years.

Purple Princess
06-20-2003, 10:34 AM
Thanks for the ideas guys :bighug:

Ponderer
06-24-2003, 12:57 PM
DTNF or TF, care to elaborate 'normalizing ...' - what does that mean and how to do that?

Dr T Non-Fan
06-24-2003, 02:09 PM
Love to.

Suppose there is a change in age/sex distribution of a block of business that you're trying to determine trend for. Well, that change will affect the claims PMPM, and will be part of the trend. Since age/sex distrbution is priced for in most companies, it should not be considered part of the trend that will be used to increase prices. It would be a double-count.

For each exposure period: determine the age/sex factor; divide the claims by the age/sex factor.

This "normalizes" the claims. It adjusts the claims, so that each exposure period now can compared without concern for the change in distribution.

In many cases, the age/sex factor should not change too much. But 1% here or there means a lot to some actuaries.

JMO
11-21-2011, 08:08 AM
Love to.

Suppose there is a change in age/sex distribution of a block of business that you're trying to determine trend for. Well, that change will affect the claims PMPM, and will be part of the trend. Since age/sex distrbution is priced for in most companies, it should not be considered part of the trend that will be used to increase prices. It would be a double-count.

For each exposure period: determine the age/sex factor; divide the claims by the age/sex factor.

This "normalizes" the claims. It adjusts the claims, so that each exposure period now can compared without concern for the change in distribution.

In many cases, the age/sex factor should not change too much. But 1% here or there means a lot to some actuaries.
Someone in another thread referred to this post. Let's keep the discussion here.

Suppose there is a change in age/sex distribution of a block of business that you're trying to determine trend for. Well, that change will affect the claims PMPM, and will be part of the trend. Since age/sex distribution is priced for in most companies, it should not be considered part of the trend that will be used to increase prices. It would be a double-count.

For each exposure period: determine the age/sex factor; divide the claims by the age/sex factor.

This "normalizes" the claims. It adjusts the claims, so that each exposure period now can compared without concern for the change in distribution.

In many cases, the age/sex factor should not change too much. But 1% here or there means a lot to some actuaries.

Could you provide an insight on how to calculate age/sex factor and benefit factor for normalisation?

Jolaolu
11-21-2011, 08:12 AM
Ok Great. Thanks Carol. SO DR TN-F,

I await your response, or from anyone with valuable insights...

T-roy
11-21-2011, 11:09 AM
Can use Milliman for A/G.

I am doing this now.

Dr T Non-Fan
11-21-2011, 11:51 AM
Could anyone provide an idea as to exact determination of age/sex factor for calculating PMPM?
This question doesn't make any sense.
I don't use an age/sex factor for calculating PMPM. I just divide claims by member-months.
Did you forget the woord "trend"?
As T-Roy said in the other thread, you can use a disinterested third party's age/sex factor matrix, which is a set of adjustment factors for each age-band-by-sex.
To determine the age-sex factor of a pool of members for a period of time, distribute the members by age bands and gender, exactly as your age-sex factor matrix from your disinterested party is in. 5-year age bands, for example. There is usually an "age x and above" section.
Multiply the number of members in each cell by the age/sex factor for the cell. Add up these products, and divide by the total number of members. Voila: age sex factor of the pool of members.

Now, what to do with it? By itself, it doesn't tell you anything, except that the claims for that pool should be higher (if Age/sex > 1) than the disinterested party's definition of a 1.0 person, or that the claims should be lower (if Age/sex < 1) than the disinterested party's definition of a 1.0 person. You'll want to relate this to some other pool of members, which requires the same exercise.

Dr T Non-Fan
11-21-2011, 12:14 PM
Please rephrase your question so that I, or a similarly intellected three-year old, can understand it.

Jolaolu
11-21-2011, 12:17 PM
Dr T Non-Fan:

the question should have been "age-sex factor for normalizing PMPM". I agree with you that PMPM is just diving claims by Member-months, however, normalizing such PMPMs with age-sex factor and other factors such as benefit factor was actually the point i was driving at.

I was going through a material from SOA's resource pool, and the age-sex and benefit factors have already been computed. I just need to get around such computations.

It was also mentioned that Rates are typically set for the richest plans option and are then adjusted down for other plan design.

I hope you understand the question better now

Dr T Non-Fan
11-21-2011, 12:47 PM
Ah, you want to know how to MAKE the age/sex adjustment factor matrix?

There is a reason why Milliman makes the matrix: There are so many other factors that affect claim costs that you have to isolate them in order to make a matrix of your own. And you might not have enough credible data. Milliman has, or claims to have, a lot of data.

Area, Benefit, and Trend are the biggest other factors. So isolate them by using one benefit design and one area. You'll need enough data in one benefit and one area in order to determine the age/sex factor adjustment matrix.
For trend, I usually trend all claims to one point in time. Now, what trend to use, since this is seemingly a circular argument? Well, you hope there will be some kind of Newton's Method convergence. Test a few trends, see what the sensitivity of the choice of trend is. Maybe it won't matter for the exercise, but if you're using two or three years of data (due to a small benefit/area choice), the trend should be close to reality.

So, you'll need to get benefit-area-trend-neutral claims into age/sex buckets, and then divide by member-months to get PMPMs.

I prefer every single age, because I smooth them using Excel's linear (adding some constant amount for each age increase) or exponential regression (multiplying by a constant for each age increase) tools. I prefer exponential, saying that each age increases claims, say, 2%-4%.

You should use covered charges, and not paid claims, since the one benefit design you choose to isolate might affect the age-sex factor. Plan design affects the slope of the age-sex factor. Steeper for leaner benefits, flatter for richer benefits.
(Well, you should do both, to see the difference.)

Jolaolu
11-21-2011, 12:50 PM
Thank you Dr T Non-Fan

Thanks Guys

Dr T Non-Fan
11-21-2011, 12:57 PM
No problem. I hope you get a good result. Sometimes, the regression curve is a mess (graph it to see it better), and sometimes you can't get the data the way you want it, but you never know until you try.

daphnag
11-29-2011, 06:34 AM
Also consider neutralizing high cost claimants - depending on the size of the group these may throw off your comparison.

Dr T Non-Fan
11-29-2011, 11:01 AM
Also consider neutralizing high cost claimants - depending on the size of the group these may throw off your comparison.
Good point.

Now, to define "high cost claimant"....

FormLetter
11-29-2011, 11:21 AM
Good point.

Now, to define "high cost claimant"....

The answer is to model the definition of "high cost claimant" such that your trend model produces the trend you want!

I jest.

In seriousness, do you think it wise to continue to toggle the level that defines "high cost claimant" until you reach a level at which the trend answer isn't particularly sensitive to it?

Dr T Non-Fan
11-29-2011, 12:37 PM
The answer is to model the definition of "high cost claimant" such that your trend model produces the trend you want!

I jest.

In seriousness, do you think it wise to continue to toggle the level that defines "high cost claimant" until you reach a level at which the trend answer isn't particularly sensitive to it?

I wrote that dilemma partly serious (it needs to be done), and partly in jest (it's not easy).
I wouldn't do that. I go in hoping that my subgroup is large to absorb the large claims without affecting the trend calculation.

One can create claim probablility distributions for each year of the trend calculation to see if there is anything unusual about the distributions, comparing:
1. Each year between the subgroup and the larger population;
2. The larger population between year 1 and year 2;
3. The subgroup between year 1 and year 2;

This exercise assumes high-cost claimants are more truly random, and are not somehow attracted or repelled by the subgroup.