In Greek mythology the goddess Ariadne handed a clew of thread to Theseus to guide his way back out of the Minotaur's labyrinth. In the labyrinth of multimorbidity we also depend on a guide to find our way. As we have no other clue (or: clew), we must rely on statistical methods to explore the different pathways, i.e. the chains of associations in the pool of chronic conditions.
We made the following assumptions on the associations between the diseases: We supposed that there is a limited number of multimorbidity patterns (i.e. clusters of diagnosis groups that are significantly associated with each other). We expected that some diseases are associated with other diseases, while some diseases are rather independent of other diseases. We also assumed that some diseases are part of more than one pattern.
Previous research on multimorbidity patterns used cluster analyses to identify morbidity patterns 
. At a given hierarchical level however, this cluster analysis would assign each disease to only one cluster, while in reality some diseases may be frequently associated with different patterns. In our data set this applies to seven diagnosis groups, i.e. depression, dizziness, renal insufficiency, atherosclerosis/peripheral arterial occlusive disease, anemias, chronic stroke and cardiac insufficiency. For this reason a cluster analysis of diseases might have produced an artificial pattern structure. Instead, a factor analysis fitted better with our assumptions. For each disease we know the degree of association with each pattern, so that diseases may become part of one or even several patterns. If the diseases are rather independent, there is little factor loading to each pattern.
Usually, the analysis of multimorbidity patterns is based upon a multimorbid population. This did not seem to be appropriate for our approach. In the factor analyses based on tetrachoric correlations, the dichotomous diagnoses were recalculated into continuous variables (see methods chapter), i.e. the illness data used for the exploratory analyses also reflect predispositions for illnesses as assumed by correlations between diagnosis groups. If we had excluded persons not being multimorbid or having no diseases at all we might have overestimated correlations between diagnosis groups and therefore biased the correlation matrix. For this reason we decided to base the analysis upon all insurants in the age group 65+ independently of the persons' number of diseases.
Persons are usually assigned to factors by performing a cluster analysis with the persons' individual factor scores 
. We deviated from this method, because we presumed that the complexity of overlapping patterns (cf. and ) may not adequately be expressed by a limited number of clusters (normally 2 to 4). Instead, we assigned patients to a pattern if they had a diagnosis in at least three pattern-specific groups (cf. section statistical analyses). In doing so, we were also able to calculate prevalence rates for multimorbidity patterns and show how often these patterns overlap in individual patients.
Strengths and weaknesses of the study
This study is the first to apply a factor analysis for identifying multimorbidity patterns. The factor analysis did perform well with the given data set. We had a limited number of three factors for both genders, a good model fit expressed by a high rate of cumulative percent (75% and 78%, respectively) and a sufficient sampling adequacy (Kaiser-Meyer-Olkin measure of 0.84 and 0.85, respectively).
A definite strength of our approach relates to a comprehensive picture of chronic diseases in the individual patients. We included all highly prevalent chronic conditions (≥1% in the age group 65+) into our diagnosis groups. For that reason we are quite sure that our statistical model is adjusted for noticeable influences of confounding diagnoses that may bias our results.
Although accidental and transitory diagnoses were excluded, in some cases diagnoses may be imprecise, ambiguous or incomplete because they were not clinically verified by trained professionals. This is a general problem in insurance claims data, but in our view, the benefits of claims data outweigh their disadvantages: We are provided with a large unselected population, representing real-world conditions and including persons living in protected institutions/nursing homes as well as frail individuals and the oldest olds, all frequently not included in survey and field studies. In choosing insurance claims data, we also avoided selection bias concerning service providers and as a matter of course there is no recall bias concerning diagnosis data.
We found three matchable multimorbidity patterns in both genders. Altogether 50.3% of female and 48.2% of male persons in our sample at least belong to one multimorbidity pattern. As the patterns differed only in singular diseases between men and women (see below) they are presented together.
The first pattern includes cardiovascular and metabolic disorders, an association that has long been known and became more clearly defined in the 1980s, as the term “metabolic syndrome” was established to designate the cluster of risk factors and diseases that come together in a single individual. The main features include insulin resistance (precursor of diabetes), hypertension and obesity 
, all of these conditions are found in this pattern. Gout was recently shown to be associated with the metabolic syndrome as well 
. This coherence can be interpreted as an example for causal comorbidity, as these diseases probably tend to co-occur because they share the same risk factors (e.g. diabetes and gout 
The second pattern found in our study covers anxiety, depression, somatoform disorders and pain. Groups of symptoms described in this pattern have appeared under different labels like “food intolerance,” 
“chronic pain syndrome” 
or "medically unexplained symptoms," 
meaning that a definite medical diagnosis explaining the symptoms is often not established and a reasonable organic explanation is lacking 
. While some of the diseases are clearly psychogenic (e.g. anxiety or depression) or clearly organic (e.g. arthrosis or osteoporosis), some can be assigned to mental and/or somatic causes, such as chronic back pain, gastritis or migraine. The association of anxiety, depression and somatic symptoms displayed in this pattern is well described 
. A depressive or anxiety disorder is reported in about 30% of patients presenting physical complaints 
. Others also report a close connection between anxiety, depression and gastrointestinal symptoms like gastritis or intestinal diverticulosis 
As a third pattern we found a group of diseases mainly consisting of neuropsychiatric disorders. Most combinations can be explained by causal comorbidity, such as dementia and Parkinson's disease probably being the causes for urinary incontinence. Other disease connections are more complex, for example cardiac insufficiency is a risk factor for stroke 
which increases the risk for vascular dementia 
. Another intricate association may result in the presence of anemia in the female pattern of neuropsychiatric disorders. There is an increased risk of carotid atherosclerosis and stroke (and therefore vascular dementia) in patients with renal insufficiency 
, which may also lead to anemia as a typical result of reduced kidney function 
The large overlapping of neuropsychiatric and cardiovascular/metabolic disorders (cf. and ) may also be explained by the relation between cardiovascular disorders, stroke and dementia. In addition, the pattern of neuropsychiatric disorders shows interference with the ADS and pain pattern. This correlation is well described, e.g. depression is a frequently described comorbidity in people after stroke 
as well as patients with Alzheimer's disease or other forms of dementia 
Age- and gender-specific differences in patterns
The prevalence of all patterns rises with the age of the patients. As age per se is the major risk factor for cardiovascular diseases 
and metabolic syndrome 
, it is not surprising that the prevalence of this pattern strongly increases with age. Also, the prevalence of the pattern with neuropsychiatric disorders gains with increasing age as expected and already reported elsewhere 
Interestingly, there seems to be much less age dependency in the pattern of anxiety, depression, somatoform disorders and pain. This might in part be explained by underdiagnosis in older age, e.g. because older patients with clinically significant mental disorders tend to underreport their symptoms 
or because the focus of doctors might move to manifest somatic diagnoses with increasing age of patients 
There are considerable differences in the composition of the patterns between the genders. The biggest differences can be found in the female pattern of neuropsychiatric disorders, which includes pre-terminal conditions such as chronic ischemic heart disease, renal insufficiency and anemia, suggesting an association of neuropsychiatric disorders and frailty in female patients. This might also explain the larger growth of this pattern with age in females than in males.
Gender differences are not always easy to explain. On one hand, a part of the differences between male and female patterns belongs to gender-specific morbidity as prostatic hyperplasia and sexual dysfunction in the male and non-inflammatory gynecological problems in the female ADS and pain pattern. On the other hand, gender differences in prevalence rates might to some extent account for the different composition of the patterns, e.g. rheumatoid arthritis exclusively belongs to female pattern of ADS and pain and has a prevalence in our sample of 3.9% in female, but only 1.7% in male patients and tobacco abuse exclusively belongs to the male cardiovascular/metabolic pattern and has 1.6% in male patients, but only 0.9% in female.
Comparison with other studies
Three other research groups also made efforts to find clusters of multimorbidity. The results are quite diverse due to differences in study design and inclusion criteria: The studies differ in data sources (i.e. administrative data 
, survey data 
and data from clinical examinations 
), populations (e.g. US Veterans 
or American Indian elders 
) and number and type of diagnosis groups (i.e. 10 conditions including tuberculosis 
, 15 condition including hip fracture 
or 23 conditions including HIV, post traumatic stress disorder and schizophrenia 
). As mentioned above, all three studies conducted a cluster analysis.
Despite these differences in approach, there are still some common results in these studies. All studies report a similar cluster of cardiovascular and related diseases, in one case combined with metabolic disorders 
like in our study, in another case combined with stroke 
which was also found for male patients in our study. One other group also found an anxiety/depression cluster 
, but without the associated somatic disorders we found in our pattern. The other two studies did not include psychiatric disorders at all 
or only in the form of depression 
. Two studies included a neuropsychiatric cluster, one combined with peripheral vascular disease and seizures 
, another combined dementia, depression and hip fracture 
In an overall view, our results fit well with previous evidence. There are some minor discrepancies to our study, but they can be explained well by a new approach of identifying multimorbidity patterns, an unselected patient group and a more comprehensive list of included diagnosis groups in our study.
The underlying structure of the labyrinth of multimorbidity sharpens if we allow for a little more complexity. As stated in our predetermined hypotheses the single multimorbidity patterns seem to share some diagnosis groups, to influence each other and to overlap in a large part of the population. We accommodated with these hypotheses by choosing a factor analysis based on tetrachoric correlations as our method. Also, it was important to base the exploratory analysis on an unselected population and a comprehensive selection of diagnosis groups to avoid blurring the subtle ramifications of the labyrinth.
Our clew of thread leads us through three prevalent pathways of multimorbidity, i.e. clusters of statistically significant co-occurrence of chronic diseases. About 50% of all persons of 65 years and older belong to at least one multimorbidity pattern. The patterns of cardiovascular/metabolic, neuropsychiatric and anxiety/depression/somatoform disorders and pain fit well with existing evidence. Research is still needed concerning the impact of the different patterns. Future studies should especially focus on interactions between the patterns and (negative) synergy effects of multiple patterns in individual patients. In recognizing the full complexity of multimorbidity we might improve our ability to predict needs and achieve possible benefits for elderly patients who suffer from multiple chronic conditions.