Risk Factors for Multiple Falls in Men Over the Age of 65.

Martin Arrigotti; Tyler Bennett; Anna Booman; Collin Hawkinson; Matt Hoctor

BSTA 513, Spring 2021

Study Background

• Osteoporosis is the most common bone disease in the U.S.
• Prevalence is almost 20% in women over age 50, and 4.4% in men over age 50.¹
• People with osteoporosis are at significantly higher risk of bone fractures
• Medical expenditures associated with osteoporosis are estimated to reach $25.3B by 2025.²
• Most studies have focused on post-menopausal women, as they make up the majority of cases of osteoporosis
• Men account for 30% of hip fractures
• Mortality after bone fractures is higher in men than in women.³

The MrOS Study⁴

• Prospective cohort study of men over age 65
• Average age = 73.7, sample size = 5994
• Assesses the significance of bone mass, bone geometry, lifestyle, anthropometric and neuromuscular measures, and fall propensity as risk factors for bone fracture
• Effect of bone fractures on quality of life
• association between osteoporosis and prostate disease
• Participants fill out a baseline survey with potential risk factors, confounders, and demographic data
• Follow up surveys completed every 4 months
• Collects information on falls, fractures, prostate cancer, and death

Study Objective

Our study aims to assess potential risk factors for experiencing more than one fall in a 12-month period for men over age 65.

Exploratory Analysis

GGally::ggpairs, skimr::skim, dplyr::glimpse

​

​

Methods

After exploratory analysis…

We used Hosmer-Lemeshow’s seven-step purposeful variable selection protocol:

1. Univariate analysis
2. First Multivariable Model
3. Screen for Confounders
4. Checking Removed Covariates & Regrouping of Categorical Covariates
5. Check Linearity Assumption for Continuous Variables
6. Screening Interactions
7. Checking the Fit of the Preliminary Final Model

Initial Univariate Analysis

• Subjective health rating collapsed into binary variable:
• 0: “Excellent”-”Good”
• 1: “Fair” - “Very Poor”
• PDX was referent for site

​

Univariate Analysis

• Removed total lean body mass & corrected hip BMD variables

Missingness:

step2_narm <- MrOs %>%

dplyr::select(-b1tblkg,-b1thd,-mhfallv2) %>%

drop_na()

​

Initial Multivariable Model

• Excluded site, DM2 hx, stroke hx, cancer hx, and PACE score variables to create the reduced model
• However, likelihood ratio test showed a significant result
• PACE score was added back in to create the reduced model (tested re-addition of each exclusion individually).

Reduced Model

• None of the excluded variables resulted in a change in coefficients > 0.20; therefore the reduced model passed on to step 4
• Removed models were checked for significance; only lean body mass was added back to the model

Linearity Assumption for Continuous Variables

• Re-parameterization of walking speed and grip strength was considered
• Re- parameterized grip strength did not significantly improve the model in the presence of inverse-square walking speed
• Inverse square walking speed was retained

Exploring Interactions

Conclusion:

• None of the models with interaction terms proved better than the reduced model based on likelihood ratio tests.

Procedure:

• Generate all potential EMM models (78 total).
• Test significance of each interaction term.
• Nine interactions were found to be statistically significant (p < 0.1 = α)
• Only consider clinically relevant and statistically significant terms for inclusion.

​

Goodness of Fit of the Preliminary Final model

• Hosmer-Lemeshow test p-value: 0.83 (g=378)
• AUROC 66.3% (64.1, 68.4)
• Pseudo-R²: 0.057
• AIC: 4057.6
• BIC: 4149.1

Assessment of Outliers in the Preliminary Final model

• Other diagnostic plots unremarkable
• Four high leverage observations
• Corresponded to four slowest participants
• Considered re-parameterized walking speed: no improvements
• Considered refitting the model with four most leveraged observations removed

Goodness of Fit of the Alternative Model

• Hosmer-Lemeshow test p-value: 0.14 (g=378)
• AUROC 66.2% (vs 66.3%)
• Pseudo-R²: 0.055 (vs 0.57)

​

• Preliminary final model used as final model

Results: Demographics

• Overall N: 5,994; analytic N: 5,092
• In the analytic sample:
• Mean age: 73.1 years
• 5% had a history of stroke, 0.6% had a history of Parkinson’s disease, 9.8% had a history of COPD, 46.2% had a history of arthritis
• 88% rated their baseline health as “excellent” or “good”
• Mean BMI: 27.4 kg/m² (overweight category)
• Included and excluded individuals statistically significantly differed for every covariate in the model

Results: Odds Ratios

• OR among those with a history of Parkinson’s disease compared to no Parkinson’s disease: 5.00 (2.38, 10.4)
• Every ten-year increase in age: OR 1.30 (1.11, 1.52)
• Fair/Poor/Very Poor health rating: OR 1.36 (1.08, 1.71) compared to Excellent/Good health rating
• Every 5 kg/m² increase in BMI: 32% decrease in predicted odds (13.9%, 46.6%)

Discussion

• Missing data
• Generalizability
• Bias
• Less precise estimates
• Model summary
• Age, history of Parkinson’s disease, COPD, arthritis, subjective health rating, BMI, average grip strength, walking speed, neck bone mineral density, total body fat, and lean body mass are significantly associated with the risk of falling more than once within a year of baseline
• History of strokes and PASE score were not significant (Wald p-values 0.053 and 0.059, respectively) but are included because of clinical relevance

Conclusions

• These results reaffirm existing practices aimed mitigating fall risk and fracture risk for elderly men with Parkinson’s and low bone mineral density.
• Efforts to increase BMI in underweight adults may decrease risk of falls in elderly men

Future Research

• Include medication history to see if medications, especially opioids, increase risk of falls
• Longer than one year follow up
• BPH (prostate enlargement) and number of bathroom trips

Role Descriptions

Background/Objectives

• Marty Arrigotti

Methods

• Matt Hoctor
• Colin Hawkinson

Results

• Anna Booman

Discussion/Conclusion

• Tyler Bennett
• Marty Arrigotti
• Anna Booman